Elon Musk’s Terafab project is best understood as a proposal to reorganize the physical production of AI compute around a much tighter relationship between design, fabrication, packaging, deployment, and energy supply. The public discussion around the project has often reduced it to a single image: another ambitious Musk factory. That framing is too narrow. Terafab appears instead as a manufacturing architecture meant to support the next stage of Tesla, xAI, and SpaceX simultaneously. It links semiconductor production to robotics, autonomy, launch capability, and the possibility of orbit-based compute infrastructure. The project’s core proposition is that the present chip supply chain, while sophisticated, is too slow and too segmented for the scale and cadence of compute Musk believes his companies will require.

Recent reporting suggests that this proposition is now beginning to translate into early industrial activity. Musk has outlined an initial “advanced technology fab” in Austin, expected to cost approximately $3 billion and designed primarily as a research facility capable of producing a few thousand wafers per month. The stated purpose is to test new approaches to chip design, fabrication, and integration within a tightly controlled environment. In parallel, teams have begun coordinating directly with semiconductor equipment suppliers, in some cases offering premium pricing to accelerate access to constrained tools. This suggests that Terafab is no longer only a systems argument, but an attempt to reorganize real industrial inputs around speed and integration.

Terafab sits inside a much larger theory of technological development. Musk frames the project in civilizational terms, invoking the Kardashev scale and arguing that human progress will eventually be bounded by the rate at which civilization can capture and use energy. He notes that humanity is not yet a true Type I civilization and emphasizes how little of the sun’s available energy is actually being converted into useful power on Earth. The underlying implication is that future abundance depends on scaling both power and compute by orders of magnitude, not through marginal improvements to present systems but through a different industrial configuration. Terafab is presented as a first practical step in that direction: large by current standards, small by the standard of the long-term system Musk has in mind.

At the same time, Musk has begun to attach more explicit quantitative framing to this vision. The long-term target is approximately one terawatt of annual compute production, which is roughly double current U.S. compute capacity, and achieving that scale could require between $5 trillion and $13 trillion in capital expenditure. These numbers define the scale at which Musk believes future energy and compute systems must operate if they are to support robotics, autonomy, and large-scale AI deployment.

Terafab is described at once as a fab, an AI hardware program, a robotics enabler, and a bridge to orbital infrastructure. In ordinary semiconductor discourse, those categories would be separated into different sectors, different timelines, and often different firms. Here they are collapsed into a single argument. SpaceX provides mass to orbit, Tesla provides robotics and large-scale industrial execution, xAI provides growing internal compute demand, and Terafab is meant to supply the silicon and packaging substrate that binds them together. Intel’s entry into the project reinforces this interpretation.

More concretely, the emerging structure suggests a division of roles: Tesla is expected to build and operate the initial research fab, while SpaceX may take responsibility for scaling the larger industrial footprint. The full ownership and operating structure remains unresolved, reflecting the early stage of the project and the complexity of coordinating multiple entities within a single manufacturing system.

At the manufacturing level, the distinguishing feature of Terafab is not simply planned output but the attempt to compress the semiconductor iteration loop. Musk describes an Austin “advanced technology fab” intended to house the equipment necessary to make logic, memory, lithography masks, packaging, and testing within a single building. The value of a facility like this would not come only from ownership of capacity. It would come from reducing the time between design revision and measured silicon performance. The proposed loop is straightforward in principle: produce or modify a mask, fabricate the chip, test it, revise the design, and repeat without relying on the conventional long chain of external handoffs.

The relevance of that change increases as AI chips become more specialized and as packaging becomes more central to system performance. Intel’s own materials on advanced packaging underline how much of modern compute capability now depends on combining multiple dies within a single package. Performance, bandwidth, thermals, and density increasingly depend on packaging and integration choices rather than on logic scaling alone. That is one reason the repeated pairing of “logic, memory, and packaging” in the Terafab discussion is significant.

Recent disclosures also suggest that Terafab may extend further upstream than initially assumed. SpaceX is reportedly exploring the development of its own GPUs, indicating that the project may encompass not only fabrication and packaging, but also chip architecture itself. If realized, this would further tighten the integration between hardware design and deployment environments.

The project’s semiconductor logic also rests on the assumption that future demand will be internally generated at very large scale. Tesla already positions efficient inference hardware as central to its autonomy and robotics ambitions. It is already involved not only in model development but in inference chip architecture, silicon validation, and deployment into vehicles. xAI, by definition, is compute hungry. SpaceX contributes a different form of demand, less through cloud-style workloads than through the physical possibility of transporting compute into new operating environments.

Terafab is expected to produce two broad chip classes. One class is for edge inference, mainly for Tesla vehicles and especially Optimus humanoid robots. The other is higher-power compute designed specifically for AI data centers, including those potentially deployed in space. These imply different optimization problems, different thermal assumptions, different packaging requirements, and different notions of useful performance.

The emphasis on Optimus is especially important. Musk argues that humanoid robot production could eventually exceed vehicle production by a wide margin, sketching a future in which global robot output reaches between one and ten billion units per year, compared with about one hundred million vehicles. That number is speculative in the extreme, but it performs an important analytic function. It establishes the internal demand model under which Terafab is being imagined. In that model, the key semiconductor challenge is not a modest expansion of current automotive compute. It is the industrialization of inference at robot scale. Once that possibility becomes the denominator, a terawatt-per-year compute target becomes an attempt to size a manufacturing system against a radically different future workload distribution.

Musk has also been arguing that space may become an economically superior environment for AI infrastructure once launch costs and solar deployment costs fall far enough. Orbit offers continuous sunlight, avoids atmospheric attenuation and seasonal variability, reduces battery dependence, and allows lighter solar structures than terrestrial installations. AI in space could become cheaper than AI on Earth within only a few years after the relevant cost-to-orbit threshold is crossed. This should be treated cautiously, but it is a structurally important direction. It means Terafab is being designed not only for a different chip supply chain, but for a different geography of compute. Its long-term purpose is inseparable from the assumption that some meaningful share of future intelligence production migrates toward where power can be scaled more cleanly.

This line of thinking is increasingly echoed across adjacent efforts. Companies such as Aetherflux are exploring similar combinations of energy generation and compute in orbit, reinforcing the idea that future infrastructure may be co-designed across power and computation rather than optimized independently. In that sense, Terafab can be understood as one component of a broader shift toward integrated energy-compute systems.

That in turn explains why Starship is woven directly into the Terafab narrative. Reaching the desired compute scale would require around ten million tons to orbit per year at approximately one hundred kilowatts per ton. Musk talks about early AI satellites around the one-hundred-kilowatt level, with future systems expanding toward the megawatt range. In the same way that some industrial visions are constrained by rail capacity or electric grid interconnection, this one is constrained by orbit access. SpaceX therefore functions less as a separate business in the story and more as the transport layer for a compute architecture. The manufacturing thesis of Terafab cannot be separated from that launch thesis without losing much of its logic.

There is an important industrial precedent buried inside that structure. Musk has repeatedly tried to internalize strategic bottlenecks once they became large enough to determine the pace of company development. Tesla internalized charging infrastructure at scale through the supercharger network. SpaceX attacked launch economics through reusability. Tesla has spent years attempting to turn manufacturing itself into a product-level competence rather than a downstream operational necessity. Terafab extends the same instinct into the semiconductor stack. The motivation is a belief that growth in autonomy, robotics, and AI will be constrained by an upstream layer that the current market treats as external and rationed. In that sense, Terafab is consistent with the broader Musk pattern. It emerges at the point where a purchased input becomes too strategically important to leave fully outside the firm boundary.

Even with Intel involved, the competitive logic of Terafab is narrower than broad commentary often assumes. The project is unlikely to succeed by displacing Nvidia, TSMC, Samsung, Micron, and the wider semiconductor ecosystem across all fronts. A more plausible interpretation is that it competes through specificity and internal alignment rather than breadth. It serves internal demand first. It optimizes for custom inference and deployment environments where Tesla and SpaceX possess unusually precise demand visibility. It attempts to accelerate the design-fabrication-test loop in ways conventional supply chains do not naturally optimize for. It integrates launch capability and downstream hardware programs into the same planning boundary.

The most recent development that clarifies the broader system is the integration of Cursor into the SpaceX ecosystem. SpaceX has secured the right to acquire Cursor for approximately $60 billion or to enter a large-scale partnership focused on joint AI development. This introduces a new layer to the Terafab thesis. Cursor represents acceleration at the software and model-development layer, while Terafab targets acceleration at the hardware layer. The integration suggests an attempt to align both cycles—software iteration and silicon iteration—within a single coordinated system.

The significance is not that Cursor directly affects semiconductor manufacturing, but that it tightens the coupling between software demand and hardware supply. Faster model development increases pressure on specialized silicon, while faster silicon iteration enables more aggressive software experimentation. In this sense, the Cursor integration reinforces the central Terafab thesis: the limiting factor is no longer any single layer of the stack, but the speed at which all layers can co-evolve.

The broader significance of Terafab lies in what it says about the current stage of AI industrialization. For several years, AI was discussed largely at the model layer. Then attention shifted to GPUs and data centers. Terafab reflects the next move downward. It foregrounds packaging, iteration speed, energy sourcing, launch access, and manufacturing integration. It treats compute as infrastructure that must be designed, built, and deployed within a larger industrial system. Whether or not the specific Musk roadmap proves correct, that shift in emphasis is important in its own right. It suggests that the next phase of AI competition may depend less on model novelty alone and more on who can shape the physical substrate of intelligence production.

This perspective also helps explain why the project extends beyond the terawatt scale in Musk’s presentation. After Terafab comes a more speculative idea of “petafab,” tied to lunar infrastructure and mass drivers. Most of that material belongs more to long-range systems imagination than to current industrial planning. Still, it is not irrelevant. It shows that Musk treats Terafab as one rung in a ladder where silicon production, orbital deployment, and eventually lunar logistics form a continuum. The project is conceived more as an expandable compute-production infrastructure tied to successively less Earth-bound energy and transport systems.

Best,
Ana

At E1, we’ve been spending a lot of time on a question that used to feel academic and now feels existential:

Can the United States still build at scale?

This debate comes from Relentless, a deep tech podcast and an E1 portfolio company. It sits squarely inside a broader theme we’ve been exploring: the U.S. vs. China manufacturing divide.

What This Debate Covers (Briefly)

In this conversation, Aaron Slodov and Sam D’Amico unpack:

• How the U.S. went from roughly 30% of global manufacturing to nearly half that, while China climbed to almost 30%.

• Why “going to China” doesn’t just mean lower labor costs. It often means plugging into a fully integrated engineering and manufacturing ecosystem.

• The idea that manufacturing is a form of reinforcement learning, where real process knowledge only comes from volume.

• Why companies like Apple export engineers and operational expertise, not just CAD files.

• Whether reshoring is realistic without systemic changes to capital markets, regulation, housing, and industrial policy.

• Why the “electric stack,” including batteries, power electronics, embedded compute, and electromagnetics, may define the next industrial battleground.

• How humanoid robotics could accelerate the shift in global manufacturing power.

It is a candid look at what it actually takes to build complex physical systems at scale and where that capability resides today.

If you care about deep tech and the long-term competitive position of the United States, it is worth your time.

Watch the full debate here:

Aaron Slodov’s Blueprint for Reindustrialization

After the debate, Aaron published a long, stack ranked blueprint for how the U.S. can rebuild a Shenzhen-like manufacturing ecosystem. The core claim is simple: there is no physical law stopping America from building dense, fast supply chain clusters. What made Shenzhen work was policy, subsidies, and infrastructure density, and Aaron argues those can be built here if we move with urgency.

Key ideas from the article:

• Reindustrialization is a prerequisite for long-term stability. Local manufacturing supports self-reliance, and self-reliance supports national resilience.

• The bottleneck is not wages. The U.S. struggles most with high mix, low volume manufacturing, the fast-turn work that startups and defense programs need. That is the beachhead that can scale into flexible mass production.

• The real cost is friction. Utilization, changeovers, quoting, scheduling, qualification, and supply chain coordination drive lead times and cost more than raw labor.

• Tech-enabled factories are a new category. Aaron argues that software-driven yield improvement, steep learning curves, and proprietary process IP create a manufacturing asset class that behaves more like tech than traditional industrials.

• Pure defense is a trap for most companies. He makes a strong case for dual-use businesses with a commercial-first posture, targeting an approximate 75/25 commercial to defense revenue mix.

• Speed is the flywheel. Faster suppliers make every downstream builder faster, which compounds demand and accelerates the whole ecosystem.

• The endgame is “factories as the product.” Not a marketplace layered on top of legacy shops, but vertically integrated operations that collapse lead times from weeks to days.

Read Aaron’s full article here: https://x.com/aphysicist/status/2028308559088214208

In many ways, Aaron’s blueprint sharpens themes we explored throughout our Substack series last year. The manufacturing gap is better understood as a systems gap than a wage gap. It manifests in calendar time, coordination friction, fragmented supplier networks, and the gradual erosion of tacit process knowledge. When manufacturing is viewed as a learning system, density and repetition become strategic assets. Lead times shape competitive outcomes. Ecosystems that run more production cycles accumulate capabilities that are difficult to replicate through capital alone.

The deeper implication across both the debate and Aaron’s essay is that industrial capacity underwrites strategic flexibility. A nation that can reliably translate design intent into physical output preserves economic resilience and geopolitical latitude. When production capacity migrates, so do learning curves, supplier relationships, and the compounding advantages that follow. Rebuilding capability therefore requires coordinated action across capital formation, permitting, workforce development, and technology adoption. The challenge is less about discovering new physics and more about aligning incentives quickly enough to regenerate industrial density before the learning curve gap widens further.

American manufacturing inspires a strange mix of pride and dissatisfaction. The United States remains one of the largest manufacturing economies in the world, yet the sense of fragility persists. That tension stems less from aggregate output and more from the composition of that output. The structure of the sector favors long product cycles, high volumes, and stable designs, while struggling in areas where production runs are short, requirements change quickly, and lead time carries disproportionate economic weight.

A recent essay by Austin Vernon offers a perspective that complements and illuminates this pattern. Vernon observes that “U.S. manufacturing has hollowed out in intermediate goods precisely because the economic structure of manufacturing has evolved in a way that emphasizes scale over flexibility.” This framing draws attention to the intermediate tiers of the supply chain where much of the value added in modern goods actually resides. Vernon’s work challenges conventional narratives focused on trade deficits and tariffs, and instead underscores how industrial organization and transaction structures influence competitive performance.

The central contribution of the original essay lies in its focus on calendar time and soft costs. In low-volume manufacturing, the quoted price of a part captures only a fraction of the economic burden imposed on the buyer. Engineering time spent waiting for quotes, administrative back-and-forth, scheduling uncertainty, and coordination across subcontractors often outweigh the fabrication cost itself. When volumes are small, those overhead layers cannot be amortized across thousands of units. They accumulate per order, creating a system where friction dominates value creation.

Digitized, end-to-end manufacturers alter that equation by collapsing latency across the procurement and production workflow. Automated manufacturability checks, immediate pricing, and direct integration from customer files to machine instructions reduce both elapsed time and labor intensity in quoting and job setup. The productivity gains compound because higher utilization of capital equipment improves return on assets while shorter lead times expand demand. Buyers place more orders when the penalty for iteration falls. Engineers design more ambitiously when procurement does not stall development cycles.

The essay’s discussion of robotics reinforces the importance of context. Automation changes the composition of labor demand, increasing the need for programming, maintenance, and systems integration expertise. In a labor market where high-skill technical talent commands a premium and remains scarce, the substitution of hourly labor with skilled labor does not automatically improve economics. Decisions about automation therefore depend on the availability and cost of complementary skills, not simply on wage comparisons at the factory floor.

One policy dimension deserves further attention: the local taxation of capital goods in the United States. In several states and counties, business personal property taxes apply to machinery and equipment, and in some jurisdictions inventory is taxed as well. Firms that invest in additional CNC machines, laser cutters, or inspection systems may face recurring tax liabilities tied directly to those assets. For manufacturing models that depend on increasing capital intensity and driving higher equipment utilization, these taxes reduce the incentive to modernize. The contrast with effective policy postures in some Chinese industrial zones highlights how tax and incentive systems can shift the economics of capital deepening. Recognizing how local tax structures interact with industrial modernization adds an important dimension to the discussion.

Financing structures also shape the trajectory of this transition. Digitized manufacturers often carry higher fixed costs due to software development, systems integration, and centralized operations. Their economics improve with scale because marginal costs decline as volume rises. Traditional underwriting models for small job shops do not always account for this dynamic. Lenders accustomed to asset-based evaluation may underappreciate the value of integrated software systems and recurring demand generated by instant quoting platforms. As financial institutions adapt to these hybrid models, access to capital can accelerate consolidation and expansion.

The implications extend beyond commercial markets. In defense procurement and advanced capital goods production, the ability to iterate quickly and incorporate design changes without prohibitive tooling delays affects both cost and strategic resilience. Shorter production cycles enable more frequent upgrades and reduce the risk associated with committing to long, static runs. In environments where adversaries adapt rapidly, manufacturing flexibility contributes directly to operational effectiveness.

Regulatory timelines represent another significant variable. Many certification and permitting processes were built around assumptions of slow design change and long product cycles. When approvals require months of calendar time for procedures that take hours to perform, the delay becomes an embedded cost in innovation. Reducing calendar friction without compromising safety standards would amplify the benefits of faster manufacturing cycles and increase the productivity of high-skill labor across the economy.

The broader strategic context remains important. The United States benefits disproportionately from being the primary market for frontier technologies and from hosting a dense network of capital, talent, and research institutions. Manufacturing systems that support rapid prototyping and short lead times reinforce that advantage by lowering the activation energy required to bring new physical products to market. As more components and subassemblies become available on predictable timelines, hardware ventures can compress development schedules and reallocate engineering time toward innovation rather than procurement management.

The reconfiguration under way does not hinge on centralized planning. It depends on firms recognizing that soft costs and elapsed time are measurable, reducible inputs. As digitized models spread across additional verticals such as machining, casting, forming, and electronics assembly, larger firms with integrated systems can operate at higher utilization rates, invest more heavily in automation where justified, and provide broader service menus to customers. Smaller shops that do not adapt may find their traditional advantages eroded as buyers gravitate toward reliability and speed.

The trajectory appears favorable, but it is shaped by factors beyond immediate technology. Local tax policy, labor market constraints, capital availability, and regulatory timing influence how fully these models take root. The underlying demand for shorter lead times and reduced coordination costs is already visible. Whether that demand translates into sustained structural change will depend on how these complementary factors align over time.

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

At E1 Ventures, we spend a lot of time looking into infrastructure and energy, and we recently published a deep dive on orbital datacenters. From time to time, I also come across articles that I would like to share with readers of our Substack, and Terraform Industries recently published a blog that is absolutely worth your time. The piece argues that the only way to scale AI compute is to completely bypass the archaic electrical grid in favor of direct-current, solar-plus-battery “islands.” While Western utilities are currently flat-footed, trapped in regulatory amber, China is already playing this endgame at a staggering scale - pumping out a full terawatt of solar panels annually and pushing stationary battery prices below $70/kWh in 2025. At that price point, the traditional gas turbine no longer makes sense.

Casey Handmer, the founder of Terraform Industries and a polymathic refugee from the halls of NASA’s Jet Propulsion Laboratory, has spent the last few years becoming the unofficial architect of a new kind of industrialism. He is the engineer’s engineer. His latest thesis is a declaration of war against the 19th-century infrastructure that is currently strangling the 21st-century’s greatest ambition: Artificial Intelligence.

To understand his vision, you have to understand the sheer, grinding stupidity of how a modern data center works.

Right now, if you want to train a massive language model, you need tens of thousands of GPUs. They operate on low-voltage direct current (DC). But the world around them is built on high-voltage alternating current (AC).

We take power from a distant plant, step it up to massive AC voltages to move it across the country, step it down at a substation, move it to a data center, and then use massive, expensive, heat-spewing power electronics to convert it back into DC so the chip can actually use it. In this process, efficiency is bled out at every turn. We build massive transformers, inverters, and rectifiers. We deal with the “hum” of the 60Hz grid. It is a logistical nightmare that adds billions to the cost and, more importantly, years to the construction timeline.

The AI boom has hit a wall, and that wall is the power grid. If you want a gigawatt of power today, the utility company will laugh at you. They’ll tell you to come back in 2032 after they’ve conducted seven environmental impact studies and upgraded three thousand miles of copper wire.

The response from the new school of hardware industrialism is a classic piece of Silicon Valley iconoclasm: Delete.

In the world of hardcore engineering (the world of SpaceX or Terraform) the most powerful tool in the shed is the deletion of requirements.

The “Direct Current Data Center” is the ultimate expression of this. The proposal is deceptively simple: Pave the desert with solar panels. Connect those panels directly to batteries. Connect those batteries directly to the GPU racks.

No grid. No AC converters. No gas turbines. No 200-mile transmission lines.

A solar panel is a slice of silicon that absorbs photons and drives electrons uphill. A GPU is a slice of silicon that regulates current flow. They are a natural match. By operating the entire system on high-voltage DC (around 1000V), you can use the same electrons hundreds of times in series, slashing the need for expensive copper. You remove the “moving parts” of the energy world. A gas turbine is a temperamental, high-maintenance beast that requires a supply chain of fuel and specialized technicians. A solar panel just sits there. It has no moving parts. It asks for nothing but the sun.

But critics immediately point to the obvious flaw: The sun goes down.

This is where the engineering brilliance shines. In the legacy world, a data center must have “five nines” of reliability (99.999% uptime). This is a cult-like obsession that requires massive backup generators and redundant everything.

But for AI training, this is a multi-billion dollar mistake. AI training is a “flexible load.” A GPU doesn’t care if it works at 100% capacity at 2:00 PM and 30% capacity at 2:00 AM, as long as it gets its work done over the course of the month. The model introduces a “Governor” - a software layer that regulates the GPU’s power consumption based on the state of the battery.

The physics here are elegant. GPU power consumption is proportional to the clock frequency multiplied by the square of the voltage. By throttling the clock speed by just a tiny fraction, say, 3%, you can reduce power consumption by nearly 10%.

Instead of the data center going “dark” when the battery runs low, the “Governor” simply slows the thinking. It rations the spicy electrons to ensure the system never actually crashes. Simulations show that this approach can achieve 99.7% utilization without ever needing a single drop of natural gas or a connection to the grid. It’s better to think a little slower for an hour than to stop thinking entirely.

In the AI era, the only metric that matters is tokens per dollar.

The data shows a “bifurcation” in the energy market. On one side, you have the old world: pure gas power. It’s cheap if you don’t care about emissions, but it’s hard to scale because nobody can build turbines fast enough. On the other side is the new world: pure solar and battery.

Between 2024 and 2026, the price of batteries plummeted. Chinese manufacturing giants began shipping integrated battery storage for under $70 per kilowatt-hour. At that price point, the math flips. The moment the cost of a battery and a solar panel drops below 10% of the cost of a gas turbine, the turbine becomes a liability. Under current assumptions, it is actually cheaper to delete the gas system entirely than to keep it as a backup. The complexity of the hybrid system eats the savings.

The scale of what is being proposed is, frankly, insane.

To power a gigawatt-scale AI data center, you need about 15,000 acres of solar panels. That’s a lot of glass. But the United States has 150 million acres of unpopulated desert west of the Mississippi.

If we paved that land, we wouldn’t just be powering AI; we’d be generating 10 Terawatts of power. That is more than the total global electricity generation of the entire human race today.

This is where the vision shifts from “Silicon Valley startup” to “Kardashev Level 1 Civilization.” These DC data centers are the “seed” of a new energy abundance. For ten months of the year, these solar farms will be massively over-supplied with power. That “excess” power can be given away for near-zero cost to neighboring communities or used to power machines that pull CO2 from the air to create synthetic fuels.

However, just when you think the idea is grounded in the dirt of the Mojave, it looks up.

One of the most provocative sections of the thesis involves “Space-Based Inference.” If the regulatory hurdles on Earth, i.e, the permitting, the environmental lawsuits, the local NIMBYs, become too much, we simply take the GPUs to orbit.

SpaceX’s Starship is the enabler here. If you can launch a gigawatt of compute into a sun-synchronous orbit (SSO), you get 24/7 sunlight. No batteries needed. No night-time. No “Governor.”

“Space AI has a unified regulatory regime,” Handmer notes dryly. “The vacuum is a harsh mistress, but she doesn’t require a three-year environmental impact study to install a rack.”

While the radiation of space degrades GPUs faster and the thermal management is a nightmare, the “utility” of a watt of power in space is a trillion times higher than the value of a watt used to heat a home. It is the ultimate application of the “High Ground.”

In this new “Direct Current” world, the only limit to human progress is how much of the planet we are willing to pave with the future.

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

In The Future of Work Is Bigger, Wilder, and More Human Than We Think, we focused on the changes already visible across the labor market. Large language models were unbundling roles into tasks, compressing teams, and amplifying individual leverage. Knowledge work was becoming faster, more fluid, and in some cases more human as routine cognitive labor fell away. The question at that stage was adaptation. How do people and organizations reorganize when AI becomes a permanent layer in everyday work?

In the article that followed, we pushed the timeline outward. Universal High Income entered the discussion as a possible endpoint of a fully automated economy. That piece explored economic structure in a world where automation extends deeply into both cognitive and physical labor. It framed abundance as a supply-side problem and meaning as a human one. It asked what kind of economy could support broad security once work stopped being the primary mechanism for distributing income.

This piece lives in the middle.

It focuses on the phase that arrives before abundance and before income is decoupled from work, but already begins to strain the wage-based system. The phase where AI stops helping people do work and starts doing the work itself, inside the economy as it exists today.

This phase is emerging through human emulators. They are a practical response to how modern work is actually organized. Most work today does not happen in factories or fields. It happens inside software. Dashboards. Ticketing systems. Internal tools. Spreadsheets. Web applications. These interfaces encode rules, policies, permissions, and institutional memory. They are where economic activity is executed.

Human emulators operate directly on that surface. They replicate the operational role of a human actor within the digital environment itself. They see what a human sees on a screen. They interpret context. They navigate interfaces. They carry multi-step workflows across systems. From the point of view of the software, they look less like tools and more like workers.

Imagine a mid-sized insurance company processing claims. A human emulator logs into the same claims dashboard as a junior adjuster, reviews incoming documentation, cross-checks policy terms, flags inconsistencies, requests missing information, updates case status, and routes edge cases to a supervisor. Nothing about the workflow changes. No new software is installed. The emulator uses the same tools, follows the same rules, and leaves the same audit trail a human would. From the system’s point of view, work is simply getting done.

Human emulators are built for exactly this terrain.

When an AI system can take responsibility for an entire workflow, logging into systems, interpreting context, making decisions, executing actions, and producing outputs, the economic unit of labor changes. Output no longer scales linearly with headcount. Productivity increases without proportional hiring. Execution becomes elastic.

As the result, teams get smaller. Hiring slows. Entry-level roles disappear first. Supervisory and exception-handling roles become more important. Organizations produce more with fewer people involved in day-to-day execution. The signals show up quietly. Fewer job postings. Fewer backfills. Flatter org charts. Higher output per employee.

This pattern is familiar. Manufacturing automation made factories leaner before it made them smaller. Software reduced clerical headcount gradually rather than overnight. Human emulators follow the same trajectory, but they move faster because they operate entirely in the digital domain.

Human emulators scale with compute. They do not require physical deployment or new supply chains. They can be instantiated, replicated, and improved continuously. Their marginal cost trends toward infrastructure rather than wages. That means substitution pressure arrives faster than institutions can adjust.

This is where xAI comes into focus.

xAI’s work on human emulation sits right at the transition between cognitive augmentation and labor substitution. Large language models expanded what machines can understand. Human emulators expand what machines can do inside real organizations. They turn intelligence into execution.

In a recent interview, xAI engineer Sulaiman Ghori described an internal project called MacroHard as a kind of “digital Optimus,” a system designed to perform any keyboard, mouse, and screen-based job by operating the same software interfaces humans already use. MacroHard is framed as a human emulator that can log into systems, navigate dashboards, interpret context, and run multi-step workflows end to end, without requiring companies to adopt new software or rebuild processes around APIs. It is a direct bet that the UI layer is the true surface area of modern labor, and that an agent that can reliably work that surface becomes economically equivalent to a worker.

Ghori also ties the emulator concept to xAI’s obsession with speed and hardware leverage. The model is described as faster-than-human rather than maximally deliberative, optimized for throughput and operational reliability in real workflows. On the scaling side, he floats a vision where the cost of deploying emulators collapses because compute can be deployed unusually quickly, including the idea of tapping idle Tesla vehicle computers as a distributed substrate and paying owners to lease cycles. Whether or not that exact distribution model materializes, the strategic direction is clear: if execution becomes something you can scale like compute, then labor substitution arrives as an infrastructure story rather than a software story.

Human emulators create a mismatch between how value is created and how income is distributed. Output increases, but access to output remains tied to employment. The result is downward pressure on wages in affected roles, upward pressure on returns to capital, and widening divergence between productivity and participation.

Career paths change as a result.

Entry-level roles historically absorbed inefficiency in exchange for training. When execution becomes abundant, that tradeoff disappears. Pipelines thin. Ladders steepen. Access to high-leverage roles becomes more competitive.

As I mentioned before, work has long provided more than income. It has provided identity, structure, and meaning. As fewer people are required to sustain economic activity, societies must find new ways to organize participation and purpose. This challenge emerges before full automation, not after.

Human emulators accelerate that timeline.

They also bring governance questions to the surface. A virtual worker needs credentials, permissions, auditability, and accountability. Errors have consequences. In regulated environments, liability matters. Organizations must decide how much autonomy to grant, how to monitor performance, and how to intervene when systems fail.

These constraints slow adoption in some contexts, but they do not reverse the trend. Even partial deployment changes hiring behavior. Even imperfect substitutes reshape cost curves enough to influence decisions.

This is why xAI’s work matters beyond any single product announcement.

Human emulators mark the moment software stops being merely a tool used by labor and begins to function as labor itself. It is Phase 1.5 of the AI labor transition. It follows the unbundling of cognition and precedes the arrival of abundance. It is structural, cumulative, and difficult to reverse.

And it is already underway.

We will continue from here.

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

In a previous article, The Future of Work Is Bigger, Wilder, and More Human Than We Think, we focused on the nearer-term effects of artificial intelligence on jobs, organizations, and careers. That discussion was grounded in what is already visible. AI systems are unbundling roles into tasks, compressing teams, and amplifying individual leverage. Some jobs are disappearing, others are mutating, and entirely new forms of work are beginning to emerge. The central question in that piece was how humans adapt when AI becomes a permanent layer in everyday work.

That framing, however, implicitly assumes that work remains the primary gateway to economic security. If you extend the timeline far enough, that assumption begins to wobble. Once AI and automation mature beyond augmentation and into broad substitution across both cognitive and physical labor, the question stops being how work changes and becomes whether work needs to anchor income at all.

This is the context in which Elon Musk’s idea of Universal High Income should be understood. In late 2023, Musk wrote that in a “positive AI future,” there will be “universal high income, not basic,” and that scarcity would largely disappear aside from what society chooses to keep scarce. He followed that with a strikingly reflective addendum. Even if material needs are solved, he wrote, it is “less clear how we will find meaning in a world where work is optional.”

Taken together, those two thoughts outline a far-horizon vision that is both optimistic and incomplete by design. It imagines abundance as an engineering problem that may be solvable, while leaving meaning as an open human question. Universal High Income, or UHI, is presented as an end state that forces deeper questions about economics, ownership, and purpose in an AI-dominated world.

What would have to be true for Universal High Income to be possible? What kind of economy could support it without collapsing into inflation or scarcity elsewhere? How would it be funded? And how does it relate to the future of work rather than contradict it?

Universal Basic Income has been debated long enough to feel familiar. It is usually framed as a floor. A minimum standard below which no one should fall. The arguments around it are well rehearsed and tend to revolve around cost, incentives, and fairness. Even critics of UBI generally understand its intent.

Universal High Income is something else entirely. The shift from “basic” to “high” changes the scale of the ambition and the nature of the problem. “Basic” is about sufficiency. “High” implies comfort, security, and choice. It implies that income is now a platform for living.

This immediately raises definitional issues. What counts as “high” income depends heavily on geography, household structure, and local cost dynamics. A number that feels generous in one region can feel barely adequate in another.

More importantly, the move from basic to high forces a confrontation with a core economic truth that is often glossed over in income debates. Income is not output. Income is a claim on output. You can increase claims overnight by transferring money. You cannot instantly increase the supply of housing, healthcare, education, energy, or skilled services. If you raise purchasing power broadly without expanding real capacity, prices rise. The result is inflation, not prosperity.

This is why Musk’s reference to the disappearance of scarcity is the condition that makes Universal High Income coherent. Without real abundance across the sectors people spend money on, UHI becomes a redistribution engine that pushes demand into bottlenecks and rewards whoever owns the constrained supply. With abundance, higher income translates into real gains in living standards.

Universal High Income, then, is less about checks and more about whether AI and automation can fundamentally change the supply side of the economy.

In 2017, speaking at the World Government Summit, Musk argued that automation would steadily eliminate jobs and that “some kind of universal basic income is going to be necessary.” That framing was defensive. It described UBI as a stabilizer in a disruptive transition.

The later shift to “universal high income” reframes the endpoint. Instead of focusing on preventing social breakdown, it imagines a future where productivity has expanded so dramatically that a strong standard of living can be shared broadly. The emphasis moves from mitigation to upside.

By 2025, Musk went further, speculating that advanced AI and robotics could make money itself feel less central, while still acknowledging constraints like electricity and physical resources. That caveat signals that even in his most optimistic scenarios, abundance is not magic. It is bounded by physics and infrastructure.

Seen this way, Musk’s optimism is not careless. It is conditional. It assumes continued progress in AI, robotics, energy, and manufacturing. It also assumes that society chooses to distribute the gains rather than allow them to concentrate indefinitely.

The future of work and the future of income are related but not identical problems. In the near to medium term, AI reshapes work rather than eliminating it wholesale. Tasks are automated, workflows are compressed, and individuals gain leverage. Work becomes more fluid, more project-based, and in some cases more human because routine tasks fall away.

Universal High Income operates on a longer timeline. It assumes a world in which automation is no longer confined to cognitive assistance but extends deeply into physical labor. In that world, the connection between hours worked and economic security weakens. People may still work, but work stops being the primary mechanism for distributing income.

In a wage-based system, automation feels threatening because it undermines the link between effort and survival. In a UHI system, automation can feel liberating because security is decoupled from employment. The same technology produces radically different social outcomes depending on the distribution framework wrapped around it.

A crucial constraint on the UHI vision is that today’s AI revolution is still largely digital. It has transformed language, media, analysis, and code. That is significant, but it does not automatically make rent cheaper or healthcare more accessible.

The biggest expenses in most households are tied to the physical economy: housing, healthcare, education, childcare, transportation, food, and energy. Universal High Income only feels “high” if the cost of these goods and services falls or becomes more elastic.

This is where robotics and physical automation become decisive. Software alone cannot build houses faster, care for the elderly at scale, or deliver medical services cheaply. Intelligence has to be paired with dexterity, mobility, and reliability in the real world.

Musk’s optimism about UHI is inseparable from his focus on robotics and manufacturing. His worldview assumes that AI migrates into factories, construction sites, logistics networks, and service environments. If automation remains narrow, abundance remains narrow. If automation becomes broad, abundance can follow.

Housing illustrates the point clearly. Housing scarcity in many regions is driven by a mix of labor costs, regulatory friction, financing structures, and slow construction processes. AI can accelerate design and planning. Robotics can reduce labor costs and speed up builds. But policy barriers must also shift. Without supply expansion, higher incomes simply inflate prices.

Healthcare follows a similar pattern. Administrative automation and diagnostic support can reduce costs at the margin, but real abundance requires scalable delivery models that reduce dependence on scarce human labor. Universal High Income assumes that these systems evolve alongside AI.

An AI and robotics economy is energy-intensive. Compute requires power. Automation requires power. Electrified logistics requires power. Musk has repeatedly acknowledged that even in a post-scarcity vision, electricity and physical resources remain constraints.

This acknowledgment grounds the UHI discussion in reality. If energy remains scarce, expensive, or politically blocked, abundance caps out. If energy becomes abundant, cheap, and reliably deliverable, many other constraints loosen simultaneously.

In this sense, UHI is indirectly a thesis about infrastructure. It assumes societies can build grids, generation capacity, transmission, and storage at scale. It assumes that permitting and governance adapt to support expansion rather than impede it. Abundance is not just technological. It is institutional.

Even if AI and automation massively expand productivity, the gains do not distribute themselves. By default, they flow to the owners of capital: data centers, models, robots, factories, and intellectual property. Universal High Income is therefore a deliberate societal choice.

There are two broad approaches to making that choice. One is tax-and-transfer. The other is a social dividend model based on shared ownership or shared claims on productive assets.

The Alaska Permanent Fund Dividend is often cited because it demonstrates that universal cash payments can be politically durable when framed as dividends rather than redistribution. The fund’s payout mechanism ties distributions to long-term asset performance, smoothing volatility and reinforcing legitimacy. While Alaska’s dividend is not UHI, it shows how distribution mechanisms can be designed to endure.

Scaling such a model to an AI economy would require society to accumulate ownership or claims on the automated productive base, possibly through sovereign wealth funds, royalties, or hybrid tax structures. The details are complex and contested, but the principle is consistent. If the economy becomes more capital-driven, broader capital participation becomes one path to broader income.

At basic levels, universal payments can be debated as one program among many. At high levels, they become central to fiscal architecture. Funding UHI would require rethinking tax bases in a world where labor income may shrink as a share of total output.

Consumption taxes attract attention because consumption does not disappear when work changes shape. Value-added taxes are widely used globally and can raise substantial revenue. Paired with universal payments, they can preserve progressivity depending on design.

Capital and profit taxation becomes more relevant in an automation-heavy economy, but it is also more mobile and sensitive to avoidance. Automation or “labor equivalent” taxes are conceptually appealing but difficult to implement cleanly without discouraging productivity.

The most plausible funding path is a hybrid system that combines broad-based taxes with mechanisms that explicitly link distributions to the growth of the automated economy. The goal is not to punish innovation but to ensure that its gains are shared.

Empirical evidence from basic income experiments offers cautious optimism. Finland’s basic income trial found no significant employment effects while improving wellbeing and reducing stress. Alaska’s dividend has not reduced aggregate employment, though it has shifted some work patterns toward part-time roles.

These findings suggest that unconditional cash does not automatically erode motivation. They do not, however, prove what happens when payments are large enough to replace most wages. Universal High Income remains largely untested territory.

The optimistic case assumes that automation fills labor gaps as people reduce work hours, preserving output while increasing freedom. The pessimistic case assumes shortages emerge in human-dependent services, raising prices. The difference hinges on whether abundance reaches the sectors that matter most.

The most revealing part of Musk’s Universal High Income comment is not the income itself but the question about meaning. Work has long served as a source of identity, status, and structure. Removing its necessity creates freedom but also uncertainty.

A positive interpretation is that UHI allows people to redirect effort toward activities that are currently undervalued economically: caregiving, community building, creative work, learning, and exploration. It may also lower the cost of risk-taking and entrepreneurship.

But this shift would require cultural adaptation. Societies would need new norms and institutions that help people build purpose outside traditional employment. Just as universal education became foundational in industrial economies, a post-work society may require new forms of civic and creative infrastructure.

Musk’s framing is useful precisely because it does not pretend that abundance solves everything. It acknowledges that meaning remains a human project.

We will continue to explore this topic in greater depth in our upcoming articles.

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

At the same time, an entirely separate infrastructure stack has been growing above our heads. Tens of thousands of satellites now orbit the planet, generating enormous volumes of data and operating as increasingly autonomous systems that demand low latency, high reliability, and insulation from terrestrial disruption. These two trajectories are converging, pushing computation off the surface of the Earth and into orbit.

With SpaceX confirming interest in an IPO next year and speculation swirling around a valuation approaching $1.5 trillion, institutional analysts started treating space-based data centers as a plausible market category. Elon Musk’s public support for deploying data centers in orbit sharpened that focus further.

Interest in orbital compute is no longer confined to SpaceX. Several large organizations now behave as though it represents a legitimate technical roadmap. Google’s Project Suncatcher targets prototype launches in 2027 through a partnership with Planet Labs. OpenAI leadership has shown sustained interest in launch capability, with Sam Altman exploring the acquisition of a rocket company and Eric Schmidt purchasing one outright. Blue Origin has maintained a dedicated internal team working on similar technology for over a year. When multiple well-capitalized teams converge on the same problem, the shift from speculation to execution is already underway. Axiom Space, a porfolio company, has also positioned itself as a commercial orbital platform company and has indicated interest in hosting compute infrastructure as part of its broader strategy.

To understand why data centers are moving toward orbit, it helps to look at the history of computing infrastructure. Each major transition followed a consistent pattern. Centralized machines gave way to distributed systems. Scarcity gave way to scale. Abstraction followed physical constraint. Mainframes became servers. Servers became cloud campuses. Today, cloud infrastructure is colliding with hard limits around energy, cooling, and time. When infrastructure encounters those limits, it relocates.

Orbit changes the fundamentals of energy. Solar power there is continuous, predictable, and intense. A satellite in low Earth orbit receives sunlight for most of its orbital period, and structures placed in specific regimes can access near-continuous illumination. Solar energy arrives without attenuation from atmosphere or weather.

That advantage becomes clearer when examined through systems engineering. In a dawn-dusk sun-synchronous orbit, solar panels receive sunlight almost continuously. Without atmospheric filtering, irradiance increases. Continuous exposure combined with higher intensity yields a substantial increase in usable energy. Battery dependence shrinks. Redundancy systems simplify. Power generation becomes a primary architectural element instead of a constraint.

Heat rejection defines the other half of the equation. On Earth, data centers rely on air, water, chillers, and cooling towers. Cooling absorbs a significant share of total energy consumption, and modern AI hardware has pushed thermal density to extremes. Jensen Huang once noted that a two-ton GPU rack consists almost entirely of cooling-related mass. As compute density increases, cooling becomes the dominant design driver.

This is where much of the public discourse has gone wrong. Space cooling is often framed as a binary: either space is an infinite heat sink or cooling is impossible without air. Neither is accurate. In orbit, there is no convection, so the system lives or dies by radiating heat. That turns the problem into a tractable engineering trade between operating temperature, radiator area, and radiator mass.

Anchoring that discussion to a real platform helps. Using a Starlink V3-class spacecraft as a reference, a communications-optimized system at roughly 20 kilowatts today can be scaled toward a compute-optimized variant at around 100 kilowatts without running into a physics wall. First-principles radiator modeling in a dawn-dusk sun-synchronous orbit shows that solar array planform grows roughly linearly with power, expanding from about 250 square meters at 20 kilowatts to roughly 1,285 square meters at 100 kilowatts. Radiator planform grows in parallel, reaching on the order of 99 square meters at an operating temperature near 80 degrees Celsius. Even at that higher power level, radiators account for only about seven percent of total planform area, while solar arrays dominate the footprint.

Mass scaling tells a similar story, though with more uncertainty. Radiator mass depends heavily on areal density, which is an architectural choice rather than a physical constant. Lightweight radiator concepts approach two kilograms per square meter, conventional designs sit closer to five, and conservative system-level implementations can exceed ten. At 20 kilowatts, radiators are a minor mass component, roughly ten percent of a two-ton spacecraft. At 100 kilowatts, radiator mass rises materially but remains a minority share of a heavier compute-optimized system, on the order of fifteen to twenty percent of total mass, with solar arrays accounting for most of the increase. The International Space Station provides a useful reference point here, rejecting roughly 70 kilowatts of waste heat using ammonia-loop radiators with a deployed area of about 78 square meters and a system-level areal density near 14 kilograms per square meter. The ISS is not mass-optimized, but it demonstrates that rejecting near-100 kilowatts of heat in low Earth orbit is operationally routine.

The takeaway is not that radiators are trivial. It is that they scale predictably.

Energy availability and cooling architecture justify sustained attention on their own, but latency strengthens the case further. Modern satellites generate massive volumes of raw data. Imaging platforms collect terabytes daily. Communication constellations route traffic continuously. Earth observation systems depend on onboard inference to classify, compress, and act on data. Downlinking raw data consumes bandwidth and time that exceed acceptable limits for many applications.

On-orbit processing reshapes the system by moving intelligence closer to data generation. Orbital data centers aggregate, analyze, and redistribute information while minimizing ground contact. This mirrors the architectural shift that occurred inside terrestrial data centers when GPUs were colocated with storage and memory. The same principle now applies at planetary scale. Optical links through vacuum deliver speed advantages due to direct paths and higher propagation velocity, creating meaningful gains for workloads sensitive to latency.

Orbit changes the cost of power, the method of cooling, and the structure of data movement. In specific cases, compute gains value off Earth.

Understanding the hardware clarifies the economics. A satellite consists of two elements: the bus and the payload. The bus provides structure, power, thermal control, propulsion, and communications. The payload performs the mission. For a data center satellite, the payload is compute. Integration choices directly determine cost, performance, and iteration speed.

This distinction explains why vertical integration appears repeatedly in serious economic analyses. Every interface introduces margin. In orbit, mass and efficiency constraints amplify those costs. Paying margin at each layer erodes feasibility before launch.

SpaceX anchors this transition. Historically, launch economics forced space systems to be small, delicate, and irreplaceable. Frequent, reusable launch changes that equation. Space infrastructure adopts industrial characteristics. Systems grow larger. Failure becomes expected. Replacement becomes routine. Iteration accelerates.

Starship enables this shift by lowering the cost of lifting tens or hundreds of tons into orbit. That capability unlocks infrastructure classes that previously lacked economic justification. Large solar arrays scale. Radiators expand. Compute payloads modularize. Orbital assembly enters the design space.

Starlink already operates a laser-linked mesh that transports enormous data volumes independently of terrestrial fiber. Adding compute to that network follows naturally. Processing data inside the network reduces transport cost and latency. Musk has repeatedly acknowledged orbital data centers as part of SpaceX’s long-term vision, and investor perception has shifted accordingly. Connectivity evolves into compute. Compute supports launch cadence. The flywheel reinforces itself.

Starlink’s subscriber growth supports this trajectory. Musk has indicated that future Starlink variants will target data center workloads. The Starlink V3 concept includes extremely high-throughput laser links and requires Starship for deployment. These characteristics align with on-orbit compute requirements and signal production-scale intent. Musk has also outlined scale targets involving annual satellite mass, power per satellite, and aggregate compute capacity that exceed conventional industrial intuition. These figures reflect a mental model centered on factory-driven scale and cost reduction, reinforced by concepts such as lunar satellite factories and electromagnetic launch systems.

SpaceX’s infrastructure trajectory already aligns with orbital data center development.

Another company to watch closely is Starcloud. Recently, it launched a satellite carrying an NVIDIA H100 GPU and executed an AI workload in orbit. The hardware performed sustained computation in the space environment. This demonstration established that data center-grade accelerators function in orbit, with power delivery, thermal management, and networking reaching workable thresholds. The remaining questions concern durability and cost.

Starcloud’s architecture separates compute payload from satellite bus. The bus was sourced externally, trading integration for speed and validating baseline feasibility. This highlights the efficiency gains available through tighter integration. The long-term vision involves a mesh of orbital compute nodes linked laterally and scaled through launch cadence. Solar arrays supply power. GPUs execute workloads. Thermal loops carry heat to radiators. Optical links form a moving fabric equivalent to a data center network.

Aetherflux, another portfolio company,y has also taken note of the trend. Its origins lie in space-based solar power, but the company now connects that energy strategy directly to compute. The core concept places power generation and compute together in orbit. Galactic Brain describes a constellation of solar-powered satellites running AI workloads in space, with initial launches targeted for early 2027. Capacity scales with launches, allowing power and compute to grow together without reliance on terrestrial grids.

This strategy demands solutions across thermal management, radiation tolerance, networking, and lifetime economics. Success would establish an operational orbital data center node, with compute functioning as both workload and justification for orbital power infrastructure.

The economic reality becomes clearer when the numbers are examined. Andrew McCalip’s analysis frames the debate quantitatively. Large-scale orbital data centers powered by space-based solar are physically viable, but economic competitiveness presents the challenge.

For a 1 gigawatt system operating over five years, orbital architecture costs approximately three to four times more per watt and per megawatt-hour than a ground-based data center paired with a combined-cycle gas turbine. Under assumptions that include Starship-class launch at $500 per kilogram, Starlink-like hardware costs, and near-continuous sunlight in a terminator orbit, the orbital system reaches roughly $51 per installed watt with a levelized cost of energy near $1,167 per megawatt-hour. The terrestrial benchmark reaches $15.9 per watt and $426 per megawatt-hour. Achieving one gigawatt in orbit requires close to 40,000 satellites, approximately 4.6 square kilometers of array area, and roughly 29 million kilograms to orbit.

This framing reduces the problem to cost per usable watt delivered to GPUs under mass, launch, and thermal constraints. Competitiveness appears only under aggressive assumptions that require extreme vertical integration. Thermodynamics governs the system. All waste heat must be radiated. Radiator area scales to kilometer dimensions. Thermal equilibrium remains within modern accelerator limits under careful surface and orbit selection.

Short-term economics favor ground-based facilities. Orbital compute supports broader industrial effects when evaluated over longer horizons.

The promise of orbital data centers is immense, and it is matched by an equally formidable set of challenges that cannot be waved away. Thermodynamics dictates architectural choices. Heat sinks must be constructed. Radiator surface area expands with power density. Radiation steadily degrades hardware, making error correction, redundancy, and system-level resilience foundational. Replacement cadence becomes a defining economic variable. Networking requires optical systems capable of precision pointing and dynamic routing across constantly moving nodes. Maintenance follows fleet economics, and debris risk imposes strict requirements around collision avoidance and reliable end-of-life deorbiting.

Progress depends on coordinated advances across multiple fronts. Radiator efficiency must improve to reduce mass penalties. Materials must tolerate sustained thermal and radiation stress. Software architectures must assume faults as a normal operating condition. Optical networking must reach a level of predictability that supports data-center-class workloads. Launch cadence must continue to increase while costs decline. Replacement logistics must become routine rather than exceptional.

Despite the challenges, the benefits motivating continued investment are tangible. Energy in orbit is continuous and clean. Radiative cooling eliminates water consumption. Processing data where it is generated reduces bandwidth and latency. Compute infrastructure detached from land and borders introduces strategic flexibility. Over long operational lifetimes, carbon impact improves as launch emissions are amortized and power remains solar. Deployment timelines compress as capacity additions depend on manufacturing and launch rather than permitting.

Thermodynamics provides the most honest lens. The viability of orbital data centers comes down to how much computation can be delivered per kilogram of launched mass once power generation, heat rejection, and structural requirements are accounted for. Progress unfolds incrementally, with prototypes giving way to space-native workloads and broader adoption following economic convergence. In that future, orbital data centers complement terrestrial infrastructure, handling classes of workloads that Earth struggles to support.

What distinguishes this moment is realism. The ambition remains large, but it is increasingly paired with explicit assumptions, quantitative modeling, and measurable targets. The conversation has finally moved from vision toward execution.

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

Factories, transportation, telecommunications, and digital infrastructure all expanded because energy was abundant, cheap, and dependable. The grid grew in lockstep with demand, absorbing economic shocks and weather events with remarkable resilience.

Then, sometime after the turn of the century, the system quietly stopped expanding. Total power generation flatlined. Coal plants disappeared faster than new plants could replace their output. The grid aged while regulatory layers multiplied and large-scale transmission projects slowed to a crawl. This stagnation remained masked behind efficiency improvements, steady natural gas output, and an era of relatively predictable demand.

Today that era is over. The combination of electrification, industrial reshoring, extreme weather, and especially artificial intelligence has thrust the United States into the opening chapter of a new energy race. The grid, built for a different century, is facing pressures it was never designed to withstand.

The next decade will determine whether America emerges stronger, with a modernized energy backbone, or suffers a period of rolling shortages, rising prices, and missed technological opportunities. Let’s go through the forces reshaping the system and the choices ahead.

The United States operates what is often called the largest machine on Earth: a vast web of wires, substations, transformers, and control systems spanning more than six million miles. But this machine is showing its age.

A large share of transformers, breakers, and distribution lines has exceeded or is approaching the end of their intended service life. Many components date back to the 1960s and 1970s. The grid was built for a world where power mostly flowed from large fossil plants outward to homes and businesses. That world has vanished.

Modern demand is different. Digital industries run around the clock. Data centers draw the energy equivalent of cities. Vehicle charging, electrified industrial loads, rooftop solar, and battery storage all complicate patterns of consumption. Despite this complexity, the basic rule of electricity still applies: the system must balance supply and demand every second.

Yet the United States does not have a single unified grid. Instead, it is partitioned into four large regions with limited capacity to move power between them. That fragmentation creates vulnerabilities. Regions with abundant low-cost renewable energy often cannot move that power to where it is needed. Regions with soaring demand cannot import enough energy during peak hours.

Interconnection upgrades remain scarce, and the backlog for new transmission projects stretches years. The approvals process, involving federal agencies, state commissions, regional operators, and local permitting bodies, has become one of the most significant constraints in American infrastructure.

The result is a structural mismatch between where power is generated and where it is needed. This mismatch is about to collide with demand growth unlike anything the grid has ever experienced.

For decades, energy planners assumed flat or modestly rising electricity demand. Efficiency gains in appliances and buildings suppressed growth, masking the lack of expansion in generating capacity. This assumption collapsed around 2023 as hyperscale computing began its exponential climb.

AI workloads require dense, continuous power. A single advanced data center may draw hundreds of megawatts, and larger campuses can top a gigawatt. To put that in perspective, one gigawatt can power hundreds of thousands of homes. The buildout of such facilities across Virginia, Ohio, Texas, and the Midwest is already straining local grids.

National studies now project that the United States will need roughly 2,000 additional terawatt-hours of electricity within the next 15 years. That is nearly seven times the net new electricity added in the previous 15 years. Importantly, this is not just about AI. Manufacturing is returning to U.S. soil. Electric vehicles are becoming mainstream. Heating systems are shifting from combustible fuels to heat pumps.

The grid is being asked to do far more than it was ever designed for, at a pace that defies past norms.

Extreme weather events today are more frequent, more severe, and more expensive. Hurricanes flatten poles and substations. Wildfires destroy long-distance transmission routes. Freezes lock gas plants and pipelines. Heat waves overwhelm local distribution networks as cooling demand surges.

The 2021 Texas winter storm, which halted power for millions, caused hundreds of deaths and tens of billions in damage. Similar patterns have appeared in California fire seasons, the Midwest during polar vortex events, and the Southeast during hurricane seasons.

The irony is that America has plenty of generating capacity most of the time. The problem occurs during rare peaks: the hottest week of summer, the coldest mornings of winter, or the moments when a storm knocks out multiple critical components simultaneously. The grid must be sized not for average conditions but for the most extreme. Right now, it is not.

Even if regulators were to approve every transmission project tomorrow, the supply chain cannot keep up. Gas turbines, the fastest source of firm power, have order backlogs stretching into the end of the decade. Transformer costs have soared across all categories, and the United States relies heavily on imports.

Critical minerals pose another constraint. Lithium, cobalt, nickel, and rare-earth processing is dominated by a few nations. Even though raw materials may be abundant globally, refining and processing capacity is concentrated, which introduces geopolitical risk and price volatility. As demand for batteries and electrified systems grows, these bottlenecks will shape deployment timelines.

These constraints are not insurmountable, but overcoming them will require both industrial policy and substantial private investment.

Despite short-term challenges, the United States holds extraordinary long-term advantages. It has vast natural gas reserves, world-leading solar and wind potential, extensive land availability, and deep capital markets. American universities, laboratories, and private companies continue to propel innovation in nuclear, storage, and next-generation grid technologies.

The combination of abundant resources and technological sophistication positions the United States to lead the next global energy cycle. But capturing that position will require rethinking how power is generated, transmitted, stored, and regulated.

Solar has evolved from a niche technology into the world’s dominant form of new energy generation. Costs have fallen dramatically due to global scaling, manufacturing improvements, and material innovations. The speed of decosting has surprised even optimistic analysts. Every year, global installations exceed forecasts, establishing a pattern of compounding growth.

Batteries are following a similar trajectory. Utility-scale storage in the United States has increased more than twentyfold since 2020. Lithium-ion remains the dominant chemistry, but alternatives such as sodium-ion, solid-state, and flow batteries are gaining traction.

This growth matters because it solves the core challenge of solar intermittency. Storage handles intraday mismatches between peak generation and peak consumption. In places like Texas, batteries already contribute significantly to grid reliability and frequency regulation. As storage costs continue to decline, their role will expand dramatically.

Beyond utility scale, residential and commercial storage is becoming more common. Rooftop solar paired with home batteries transforms consumers from passive customers into active participants. This decentralization mirrors what happened with telecommunications when mobile phones leapfrogged landline infrastructure. The distribution grid is beginning to shift from one-way power flow to a dynamic, multi-directional system.

While renewable generation is rising rapidly, natural gas remains the backbone of reliability. Combined-cycle gas plants offer high efficiency, ramp quickly to meet peaks, and provide firm baseload support. Because many nuclear reactors are aging and coal plants continue to retire, gas will remain essential for decades.

Long-term forecasts suggest that natural gas will support everything from electricity supply to industrial heat and manufacturing feedstocks. Even with aggressive renewable buildouts, the grid cannot meet surging demand without substantial gas capacity, especially in regions with variable wind and solar resources.

The interplay between gas and renewables is not a competition but a partnership. Together, they form the backbone of a cost-effective and resilient power system during the transition phase.

Nuclear energy offers a stable, carbon-free source of baseload generation. But nearly all U.S. reactors were built decades ago, and bringing new units online faces regulatory, economic, and political hurdles. Small modular reactors have generated excitement because they promise safer, smaller-scale deployment. Government agencies expect pilot SMRs to begin operation within a few years.

Yet even under optimistic scenarios, meaningful nuclear expansion will take decades. Massive reindustrialization of American nuclear manufacturing and construction capacity would be required to rebuild a fleet at scale. For now, nuclear serves as an essential part of the reliability mix rather than a short-term growth engine.

One of the most fascinating trends is how hyperscalers are responding to grid constraints. Many are now building their own dedicated power plants. Some connect to the grid in limited ways for backup or balancing. Others operate almost entirely independently.

In sunny regions, solar-plus-storage systems can support most compute demand with overbuild. Gas turbines handle the rare periods when batteries or sun fall short. Captive power plants can even supply excess energy back to local grids when demand drops. In effect, large data centers are evolving into hybrid utilities.

This shift will alter power markets. Utilities, once the sole providers of large-scale energy, may increasingly partner with or compete against vertically integrated tech firms. Regions that embrace data center development may unlock new sources of resilient generation if policies encourage bidirectional energy flows.

Federal legislation over the past few years has attempted to tackle different parts of the energy challenge. Some policies emphasize grid modernization and transmission upgrades. Others support domestic manufacturing of solar, batteries, and critical minerals. Still others expand fossil fuel leasing in the name of affordability and reliability.

In late 2025, the federal government launched an ambitious initiative aimed at accelerating innovation in nuclear technologies, fusion research, and grid optimization using artificial intelligence. This strategy recognizes that the most important competitive advantage in the next generation of energy systems will be speed. Whoever can discover new materials, optimize grid flow, and build infrastructure faster will set the global pace.

However, no policy will succeed alone. Coordination across states, utilities, regional grid operators, and private capital markets is essential. Transmission reform remains one of the most urgent gaps. Without it, clean energy cannot reach demand centers and reliability risks will increase.

To navigate the next decade successfully, the United States must solve several interconnected challenges:

1. Accelerate transmission development by shortening permitting timelines and modernizing approval processes.

2. Expand and diversify manufacturing of key grid components including transformers, turbines, and conductors.

3. Scale domestic battery production and mineral processing to reduce geopolitical risks.

4. Deploy distributed energy resources that ease local stress and create resilience at the grid edge.

5. Invest in flexible gas capacity to stabilize supply as renewables accelerate.

6. Modernize market structures so prices reflect real-time supply and demand, encouraging investment in flexible resources.

7. Prepare for extreme weather with hardened infrastructure and improved forecasting.

These changes require large-scale collaboration. But they also present an opportunity to rebuild the American energy system around resilience, abundance, and strategic advantage.

The next ten years pose the greatest risk to American grid reliability since the early days of electrification. But they also form the greatest opportunity in generations. Demand growth from AI and industrial expansion will force innovation. Supply chain strains will catalyze new domestic industries. The need for resilience will accelerate investment in advanced technologies.

In the long run, the United States has the potential to build an energy system that is cleaner, more reliable, and more affordable than any that came before. But the transition will not be smooth. The decisions made today about transmission, storage, natural gas, nuclear innovation, and distributed energy will define the contours of American prosperity for decades.

Energy abundance is not merely an economic issue. It is a national security imperative, a foundation for technological leadership, and a prerequisite for continued growth. The nations that secure reliable, low-cost energy will lead in AI, manufacturing, research, and defense.

America has all the ingredients to lead this next chapter. The question now is how quickly it can adapt.

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

Most conversations about American manufacturing happen far from the places where things are actually made. Analysts, policymakers, and tech commentators debate the future of industry from conference rooms and podcast studios, but very few people bother to walk into a shop and see what’s happening on the ground. That’s why Zach Labman’s work is so valuable. Zach spent six months traveling the country, visiting nearly 75 factories and documenting the reality of modern American production. His original reflections, shared first in a viral X-thread, provide the kind of firsthand insight that’s increasingly rare in a discourse dominated by abstractions.

This essay also builds directly on themes we’ve explored in our previous Substack pieces about U.S.–China industrial competition and the manufacturing ecosystems that underpin national power. China’s strength didn’t emerge from a single master plan but from decades of compounding capacity - skills, suppliers, OEMs, financing mechanisms, and an education system aligned toward production. Zach’s findings make clear that America’s challenge isn’t a lack of talent or ambition, but a misalignment of the systems that support industrial work. Where those systems line up - education, capital, policy, technology, and industry - you see small factories thriving. Where they don’t, you see slow decline.

One of the most striking aspects of Zach’s observations is how uneven the landscape truly is. He encountered tiny machine shops that are aging out with no succession plan and mid-market manufacturers quietly doing $50–100M in revenue on home-built tooling. He found companies with world-class throughput operating next to competitors that still run on whiteboards and tribal knowledge. This variation reveals the central tension in American manufacturing today: success is highly local and highly dependent on coordination. Without aligned incentives, even the most capable shops get trapped in defense-sector dependency or struggle to scale beyond a handful of customers.

What we found especially resonant is Zach’s argument that reindustrialization isn’t a single policy lever but a systems problem. Apprenticeships matter. Capital allocation matters. Regulatory predictability matters. The availability of OEMs matters. And importantly, cultural narratives matter: when younger generations are told that meaningful work only happens in software or finance, it’s no wonder the talent pipeline breaks. Zach’s work reinforces a truth that often gets lost in political rhetoric - America doesn’t lack people who want to build; it lacks the structures that make building viable. His observations are a reminder that revitalizing U.S. industry requires rebuilding not just factories, but prestige, pathways, and purpose.

Ultimately, Zach is optimistic, but with clear-eyed realism. He sees a future where the U.S. can compete again, not by replicating the past but by combining advanced technology with renewed industrial capability. We share that optimism. The early signs of a new industrial era are already visible: in the nuclear startups, the machining clusters, the robotics OEMs we hope to see emerge, and the growing recognition that national strength depends on making real things again. This essay adds an essential layer to that conversation. Check out Zach’s work on X below and Happy Thanksgiving to all!

https://x.com/zachglabman/status/1944802091459076249?s=20

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

A few days ago, the internet lit up with a familiar clash. Congressman Ro Khanna, the Democrat who represents much of Silicon Valley, posted about the controversy surrounding Call of Duty: Black Ops 7. Players had discovered that the game shipped with dozens of AI-generated icons, posters, and achievement badges. Artists quickly pointed out the telltale signs of machine-made slop. Khanna jumped in, calling for regulations to stop companies from using AI to eliminate jobs just to pad profits. He suggested taxes on mass displacement and a real voice for workers before tools get deployed. As of this week the original thread has surpassed 40 million impressions, and Khanna now cites recent warnings that without policy intervention the top ten percent of the cognitive distribution could capture the overwhelming majority of AI-driven productivity gains this decade.

David Friedberg, the co-host of the All-In podcast, fired back almost immediately. AI does not just cut costs. It unlocks entirely new categories of work that pay better and demand more skill. The exchange lasted only a handful of tweets, but it captured the central tension of our moment: one side sees AI as a predator stalking livelihoods; the other as the greatest job-creation engine since the internet.

Both sides have real evidence on their side. The destruction is happening faster than almost anyone forecasted two years ago. But the creation side is accelerating too, and the net picture is more nuanced than either camp admits.

Through mid-November 2025, tech companies announced over 180,000 layoffs, although only about 28 percent were explicitly tied to AI. Microsoft cut thousands while touting that up to 30 percent of its code is now AI-written. Intel slashed 15–20 percent of its workforce amid massive AI investments. Amazon eliminated 14,000 managerial roles to “flatten” the organization and redirect capital toward AI. Overall U.S. layoffs topped 1 million in 2025, with AI explicitly blamed for tens of thousands.

White-collar professions once thought untouchable are suddenly exposed. A St. Louis Fed analysis released in August 2025 found a striking 0.47 correlation between an occupation’s AI exposure and its rise in unemployment from 2022 to mid-2025. Computer and mathematical roles, with exposure scores near 80 percent, saw some of the sharpest jumps.

Creative fields are bleeding too. Duolingo quietly off-boarded chunks of its contract translators after shifting lesson creation to large language models. Klarna’s AI assistant, launched in early 2024, handled the workload of 700 human agents within a month, contributing to the fintech’s decision to shrink overall headcount by 40 percent by mid-2025. But by November, Klarna had begun rehiring humans into new AI-escalation and conversation-design roles after the bot increased churn on complex cases, and net headcount is now higher than before the rollout, with remaining employees earning more.

Walk into any agency or studio and the stories are grim. Junior designers watch clients cancel retainers because Midjourney spits out mood boards in minutes. Anthropic CEO Dario Amodei went further in multiple 2025 interviews: he warned that AI could eliminate up to half of all entry-level white-collar positions within five years, potentially driving unemployment into the 10–20 percent range for the first time since the Great Depression.

This is not hype. It is happening right now.

But here is where the story takes an unexpected turn.

For every job AI kills, it quietly births two or three others that didn’t exist five years ago. Prompt engineers, AI-ethics auditors, synthetic-data curators, human-AI collaboration specialists, and model-safety testers are now line items on real org charts. Entire secondary ecosystems are emerging as well: AI-maintenance technicians on factory floors, robotics-integration specialists, automated-quality-control supervisors, edge-compute hardware operators, AI-enhanced logistics coordinators, and digital-twin simulation engineers in advanced manufacturing. And behind the scenes, an even larger wave of AI-infrastructure jobs has exploded—data-center architects, liquid-cooling and power-management engineers, GPU-cluster technicians, chip-packaging specialists, model-serving reliability engineers, and grid-integration experts who keep AI’s energy-hungry workloads alive.

The spillover is just as significant. Venture capital firms are standing up AI-focused funds; startups are staffing up with go-to-market strategists, AI-tooling product managers, and regulatory-navigation experts; and a new class of founders is building companies around everything from evaluation frameworks to inference-optimization hardware. LinkedIn reports that job postings requiring AI skills have grown more than sevenfold since 2022, and recent data shows that “prompt engineering” has already been overtaken by newer skill sets such as AI-system design and human-AI orchestration.

Companies that adopt generative AI most aggressively are not shrinking their headcount overall. They are shifting it upward. Newer studies show that firms in the top quartile of AI adoption are now increasing employment by almost 5 percent year-over-year, and wage premiums in high-AI-exposure roles have climbed to nearly 30 percent.

And here’s the part almost everyone seems to miss: a significant amount of the panic is driven by early, noisy data being interpreted as destiny. It’s true that employment among very young software developers dipped in the past couple of years, but that shift comes from an extremely narrow age band that represents only a small slice of the profession. When you zoom out and look at developers across the full age distribution, the overall number of software jobs has continued to grow, even through the release and adoption of tools like ChatGPT and Copilot. Recent Bureau of Labor Statistics data shows that software developer employment has reached an all-time high of about 1.71 million. The 22–26 cohort remains down, but every older cohort is sharply up.

The historical parallel is almost too neat to believe. When ATMs rolled out in the 1970s, bank-teller jobs were supposed to vanish. Instead, the cost of operating branches plummeted, banks opened more locations, and the number of tellers actually rose for two decades. They just did higher-value work: relationship management, complex problem-solving, sales. The same pattern played out with spreadsheets and accountants, search engines and travel agents, e-commerce and retail clerks.

AI is doing that on steroids.

That brings us back to Friedberg and Khanna.

Friedberg is correct that taxing automation would slow progress and keep people trapped in low-productivity roles. History is brutal on this point. Countries that tried to preserve jobs in declining industries (think British coal miners in the 1980s or French farmers fighting supermarket chains) ended up poorer and less competitive.

Yet Khanna has a point no serious person can dismiss. The transition is brutal for the individuals caught in it, and the gains are not flowing evenly. A junior artist laid off today does not magically become a prompt-engineering wizard tomorrow. Regions outside the coastal tech bubbles risk being left behind entirely. Without deliberate policy, AI could super-charge the inequality that already tears at America’s social fabric.

And there is a deeper layer: what ultimately determines whether this transition becomes a crisis or a leap forward isn’t the raw number of jobs destroyed or created, it’s the speed and capacity of workers to move between them. Education systems, mobility, and labor-market flexibility matter far more than whether AI eliminates a specific task. The core question is whether younger workers, displaced workers, and workers far from tech centers will have the pathways, safety nets, and incentives to pivot quickly enough. If they do, the job churn becomes an engine of upward mobility. If they don’t, it becomes a trapdoor.

The answer is not to slow the technology. It is to accelerate human adaptation.

Companies that succeed in this new world are the ones treating AI adoption like the Manhattan Project for talent development. Amazon pledged $1.2 billion to retrain 300,000 employees, and has since raised its commitment to $4 billion over the next four years. JPMorgan runs mandatory AI literacy courses for every new hire. Governments must help too.

But even this focus on retraining and adaptation misses a larger transformation that is quietly unfolding. Most discussions about AI and jobs assume that the basic structure of work will stay intact: that people will still join large companies, slot into formal roles, move up predictable ladders, and rely on institutions that look more or less like they always have. Yet the deeper shift already underway is not just about what jobs survive, but about the very unit of economic organization.

I have built crypto startups before, and I have been following decentralized networks and the early experiments around Decentralized Autonomous Organizations for a long time. What always fascinated me was the idea that coordination, governance, and productivity might eventually happen outside of traditional corporations. Balaji Srinivasan’s work on network states pushed this further and treated communities as potential economic and political actors. For years this felt like a fascinating but premature vision. Then AI arrived, and suddenly the underlying assumptions began to change.

The corporation dominated the last century because scale required layers of managers, administrators, planners, and operators. Productivity flowed from hierarchy and specialization. A small group could not compete with a giant firm because the operational muscle simply was not there. AI breaks that logic. When a small team can rely on AI systems to handle research, coding, design, analysis, logistics, financial modeling, and customer interaction, the gap between big and small narrows dramatically. A handful of aligned people can now produce the output of a far larger department. This alters the gravitational pull of work itself. People may begin to choose affiliation and mission over institutional prestige, because the advantages of scale are no longer monopolized by corporate structures.

This is where the idea of the network state shifts from political theory to economic reality. AI gives small, mission-driven groups the coordination power that only large institutions used to have. A few researchers with powerful models can run what looks like a miniature research lab. A tiny media collective can produce the output of an entire newsroom. Local communities can handle planning, education, or social support with a level of sophistication that once required bureaucratic machinery. Careers begin to look less like ladders and more like evolving networks of collaboration.

And this leads to a strange but increasingly plausible possibility. AI may allow people to organize their working lives around their interests rather than around the jobs that happen to be available. For decades we encouraged people to pursue their passions, but economic reality usually pushed them into credentialed, institutionally defined paths. AI erodes those constraints. When individuals and small groups have the leverage that only big organizations once possessed, interests can become productive rather than merely recreational. A group of biologists, gamers, amateur historians, or robotics enthusiasts can become something much closer to a functioning enterprise or research community than ever before. What begins as a hobby or an ideological alignment can gradually turn into a self-sustaining economic structure.

This does not mean that everything becomes utopian. Our entire system of benefits, accreditation, healthcare, and worker protection grew around the assumption of long-term corporate employment. If AI fragments the workforce into thousands of small and fluid entities, society will need to rethink how people build skills, maintain stability, and gain access to essential services. The rise of networked work will require new norms of trust, new forms of belonging, and systems that can support people who move between many affiliations rather than staying tied to one employer.

Still, the direction of travel is becoming clearer. The future of work may not be defined by climbing corporate ladders but by participating in communities that reflect one’s values, interests, and ambitions, each supercharged by AI. Instead of asking which careers are safe, people may begin asking which networks they want to help build.

The machines are coming. Some jobs will die screaming. Many more will be born in places we cannot yet imagine. And people should also prepare for a world where work is no longer anchored to the institutions of the past, but to the communities and networks that will define what comes next.

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

Foreword:

In my previous piece, readers were given a glimpse into what a ‘Small’ Factory looks like in China. Now, we step into the neighborhood that manufactures. This piece dives into three factory systems, what goes on behind their doors, and what this means for U.S reindustrialization.

- Noah

From my hotel window, I watched a man welding his motorcycle in the middle of the sidewalk - the same place where families biked and walked home from dinner. Naturally, I went downstairs to have a look. On a 20-minute walk from my hotel, I passed machine shops, barbers, grocery outlets, and convenience stores. On one corner, a small shop sold household staples and Chinese snacks; next door, 10 CNC machines whirred through stainless steel.

This version of Dongguan is a walkable community that many Americans dream of living in. But it’s also a city with factories and machine shops sprinkled everywhere. One could call it an industrialist’s heaven.

Dongguan is a major manufacturing hub, often called “the world’s factory.” It sits directly north of Shenzhen and is part of the recently coined Greater Bay Area, a place some call the greatest city in the world. Given its role in global supply chains, a visit here was an essential stop on my pilgrimage through industrial China.

In 2018, Dongguan counted 10,000+ industrial enterprises.1 For perspective, here’s a photo showing just the Chamber of Commerce members from Dongguan city and neighboring Fujian province.

I spent three days in Dongguan, wedged between a Taiwan trip and the Canton Fair - one of China’s largest import and export fairs. Initially, the plan was to visit five factories, but I accidentally stumbled into five more.

We’ll now dive into three different factory systems and what each reveals about how China builds.

System #1 — The Gen Z Factory

The first factory I visited was founded by a 22-year-old in 2018. Here, they produce products such as PC fans, liquid-cooling systems, and CPU heatsinks. This factory was effectively an OEM “wrapper”. They didn’t manufacture any of the components themselves, but the company thrived by bundling everything into a finished product for resale.

The entire production floor was filled with assembly lines. These lines encompassed many tasks, including soldering, quality assurance, packaging, and plastic wrapping. These were simple, repeatable tasks that could be completed in less than 30 seconds each. Some tasks were done directly by humans (e.g., soldering), whilst others were done by machines and checked by humans (e.g., verifying the results of an autonomous screwing machine).

My main takeaway here wasn’t about assembly, though; it was that this factory felt like somewhere that would be enjoyable to work at.

For example, they had a Christmas tree in their office. Apparently, many in China’s younger generation have started celebrating this holiday, and an employee even mentioned receiving gifts from their boss.

As I mentioned earlier, the founder founded the company at 22. This different management style had some tell-tale signs:

• There were 3 cats running around the office

• The founder had a cycling machine, a massage chair, and a giant curved gaming monitor in his office

• Workers had their own desks, kitted out with Rubik’s cubes and custom mechanical keyboards.

And this factory didn’t even operate 24/7. After stepping out of a meeting with an employee, I noticed that many of the office lights were off. In the distance, a group of smiling employees could be seen rushing into an elevator — off for their 1.5-hour lunch break. This was the opposite of the other factories I had visited, where, even if there was only one shift, they would never shut down their production line mid-day. As China continues to develop, it’ll be interesting to see how workplace conditions and norms evolve.

This “Gen-Z factory” was much less cutthroat and lacked vertical integration. Yet they were cranking out new products left and right. It seemed a bit too disjointed with so many product offerings, but from a worker’s standpoint, this place would be a dream compared to others I’ve set foot in.

After the tour, I stepped out of the elevator and found myself facing a giant injection-molding plant. I tried to take photos, but I was promptly pushed away. “No photos,” an employee said, despite the place being surrounded by wide-open glass doors. Disappointing.

System 2: Scaling Amazon Land

If you order anything on Amazon, it’s most likely from China. Amazon’s Marketplace is essentially a wrapper around Chinese factories. Platform vendors will purchase wholesale products from China and resell them to American consumers at small margins.

This company didn’t advertise that it was part of Amazon’s backbone, but when I talked about reselling strategy with them, they casually mentioned supplying several of Amazon’s top 10 scale vendors.

This company exemplifies my view of the typical Chinese factory.

• They produced some of their goods in-house(via sheet metal fabrication)

• They bought many of their components from other local factories (e.g., precision sensors)

• They had many of the same copy-and-paste semi-automated assembly lines seen in other factories.

• They had plans in motion to automate assembly lines in their factory.

This company shared a cafeteria with nearby factories. Since this was an industrial district, multiple factories could share the same facilities. This is one of the many compounding aspects of China’s factory density. Many factories I’ve visited in America are giant compounds in the middle of nowhere.

You can imagine factory workers sitting down for a bite to eat and sharing knowledge, horror stories, and jokes from the day. Yes, factories can save on employee meal costs by sharing facilities, but the real benefit is the knowledge that flows across companies. If one company discovers a new technique to make its production more efficient, it will inevitably be adopted by all the other factories in the area.

Later, I asked about the trade-off between quality and quantity. Their answer was blunt: higher-quality products earn bigger margins but sell in lower volumes; cheap, high-volume items sell a lot more but make less per unit. The point was clear: quality and quantity live on a sliding scale. There is a pervasive stigma in the United States that China only produces low-quality goods; however, this belief is no longer true. Many of these factories can produce higher-end products, but market incentives on platforms like Amazon push them toward volume.

Stepping off the elevator onto the second floor, we were met by three large assembly lines. About 25 workers sat on stools along each line, performing a specific task repeatedly. Workers sat behind semi-automated machines, either helping operate them or supervising the results. Much of the work had been diverted from manual labor to machine-oriented, with the human at the end of the loop for QC (quite similar to vibe coding nowadays).

Similar to the PC component manufacturer, this factory operates ~10 hours a day with no night shift.

They told me that they were going to install an entirely autonomous scale assembly line on the 4th floor. They showed me their blueprints and a video of the line in operation at an unknown location. They mentioned that they would send a video when it was complete.

They had “partnered” with another company to develop this assembly line. Who was responsible for doing this? I want to meet them. And the final cost? $1 million.

“Next year, the entire line will be automated,” they said.

“Wow, this is so cheap”, I told them. “You can easily raise a seed round in the U.S. for this. They laughed when I said this and responded, “If you think it is cheap, you can buy our products”.

“Machines do not miss, but people will make mistakes.”
— Factory manager

They said they don’t plan to fire the workers and keep the other assembly line in place, though time will tell what happens in the long run.

System #3 — The Injection Molding Industrial Complex

Injection molding is the most common method for producing low-cost plastic parts in large volumes. Chances are, if it’s made of plastic, it’s made in an injection molding machine, including many toys, household items, and automobile parts. 3D printers are starting to eat into their market share, but injection molding still reigns supreme.

Given that injection molding has existed for over 150 years and that Dongguan is one of the top cities in the world for this industry, I was interested to see how the actual factories operate. You could call them the lungs of Dongguan — breathing in molten plastic, and exhaling finished parts.

In Dongguan, I visited two mid-sized full-service injection molding factories.

Injection Molding Factory #1:

The first factory had 19 injection molding machines and produced parts for many Fortune 500 companies, including famous American tool companies and large Chinese automobile brands. Hot molten plastic was pumped into their custom-made, CNC-machined molds, hardened into plastic parts, and then dumped onto a conveyor or into a bin.

A few of the lines with simpler parts could operate automatically. But most of the machines had a worker stationed alongside, performing simple tasks such as QA or part refinement with X-acto knives.

Despite producing components that are in the garages of many American consumers, this wasn’t one of China’s famous dark factories. There was still a great deal of manual labor ripe for automation.

And despite the Made in China 2025 push, many of the injection molding machines were from Japan. However, you could easily foresee this changing over the next 25 years as older machines are phased out.

They mentioned that employees worked from 9 a.m. to 9 p.m., with an hour for lunch and another for dinner. Then came the night shift from 9 p.m. to 9 a.m. That’s when it finally clicked why 996 is so common in Chinese factories: it keeps production running for 24 hours a day.

Night shifts are something I keep thinking about when comparing China to the U.S. The pandemic wiped out most 24-hour fast-food franchises back home. I don’t believe that many American workers will put up with working night shifts in a factory. It’s humorous to watch Americans complain about the 9-to-5 or campaign for four-day workweeks. Many don’t even want to go into the office in the first place. Yes, this is a comfortable lifestyle, but it’s not a winning strategy. China is moving in the opposite direction.

Injection Molding Factory #2

The second factory had 14 injection molding machines and followed a similar pattern to the first. It felt eerily similar, as if they had copied and pasted their business model and factory layout. They had many of the same machines, used the same materials, and used the same conveyor belts. They also had cafeterias, multi-floor factories, and offices of English-speaking Chinese sales representatives. It’s like everyone took the same “factory 101 course”.

Rows of workers sat at their stations, repeating the same task thousands of times (Yes, this still happens in 2025). One woman sat on her stool, hammering each injection-molded case shut. At another seat, a young woman soldered wires to PCBs. This same woman in America would be working at Starbucks or still in college, but in this factory, many were found on the lines. Gender balance was surprisingly even, which makes sense now that much of the most strenuous physical labor is automated.

Like many factories, this place was a one-stop shop. They offered to sell their existing inventory, design and produce custom molds, and even maintain those molds for free if you made them in-house. They were willing to do almost anything to close a deal. Five different reps messaged me late at night offering free design reviews, material suggestions, and process advice. They were willing to do the grunt work — even some of my work — to build the relationship.

“Our goal is long-term partnerships”
— Factory Owner

They weren’t worried about short-term costs - only if you were visiting other factories.

This factory followed the same family-run pattern, with the owner at the top, the son as manager, and the spouse in the office—one big, happy injection-molding family.

We ended the visit by discussing tariffs. Each factory I’ve spoken with has had its own opinions and strategies on tariffs, but all have mentioned a pullback by American companies in the past year.

Seeing these factories in operation demystifies them. Instead of the plastic components inside your home feeling foreign, they start to feel familiar. You understand where they come from and how they’re made.

After stepping foot into just two injection molding factories, I am confident I understand enough to get started. The hard part is the knowledge that isn’t documented online. The know-how of how to make quality molds. The ability to hire employees who understand injection molding. Being able to source quality PPE, PLA, and other materials for the machines. And being able to sell the parts at a competitive cost. Maybe American companies - drowning in VC money and balance sheets - can acquihire teams of Chinese engineers for a few million each. A small price to pay in the grand scheme of things, although Beijing would undoubtedly attempt to block this.

Even if you can’t visit these factories yourself (though you should), Alibaba is a goldmine. Most of these companies post videos, VR tours, detailed pricing, machine counts, employee numbers, and even their certifications. It’s basically LinkedIn for factories — you can find out everything about a manufacturer: what they make, how they run, when they started, and how big they are.

Final Thoughts

1. Density wins. When factories, tool shops, and suppliers are within walking distance, knowledge spreads faster than any textbook. This creates advantages that isolated factory compounds in America lack.

   The United States must double down on the hub-like strategy that China employs. El Segundo is a resurgent hub for defense tech. The Boston area is a hub for research and maritime technology. Silicon Valley is a hub for artificial intelligence and robotics. We must welcome business activity in these places with open arms and give founders the tools they need to succeed.

2. Even in 2025, automation is pragmatic, not ideological. SMBs automate where ROI is clear; everywhere else, humans stay in the loop for quality and edge cases.

   This same pattern can even be observed in an American Tesla factory tour; the welding, painting, and stamping are automated, but final assembly and QA are still manual.

3. Labor & regulation are strategic levers. China’s mix of employer housing, longer shifts, and looser local friction lowers operating costs. The U.S. can’t out-compete on this basis alone; it needs policy, subsidies, or automation-first strategies to bring manufacturing back.

   I’m unoptimistic that the U.S. will bring back the same manufacturing jobs that it had during WW2. Instead of 22-year-olds working on assembly lines, they will likely be in charge of repairing robots or supervising machines.

   American labor laws are too strict, and this will not change anytime soon. Replacing these jobs with robots is the only way for America to regain its manufacturing prowess, unless civil liberties are lost in the process.

4. Undocumented Process Knowledge: If you want to build, the most complex parts aren’t the machines; it’s the undocumented process knowledge and know-how: mold design, supplier relationships, repair culture, and people who can run the shop. Alibaba and factory reps can get you most of the specs, but much of this takes time.

   Policy around high-skilled immigration from China is incredibly tense right now — and rightly so — but those with a voice must continue to speak up. The easiest way to gain process knowledge is to capture it from those who already have it. Money talks.

Reindustrialization cannot happen overnight. It’s a national program in which policy, capital, and culture must move in the same direction; this will take many years or decades. Short of this, the fastest route for outsiders is to buy capability (e.g., by acquiring teams of Chinese engineers) rather than building entire factories from scratch.

I’ll end this with the words of Elon Musk, whom the Chinese seem to adore.

Be optimistic about the future. There’s no point in being pessimistic. It’s just too negative

Up next: notes on automation from inside the doors of global behemoths

Foreword:

In my last piece, I wrote about America’s struggle to rebuild its capacity to make things — and the urgency of learning from China’s speed, discipline, and obsession with execution. This week, Noah Olsen takes that conversation from theory to practice. A former VC and engineer who left San Francisco to see how factories actually run, Noah’s first dispatch from Fujian is not a think tank report or a policy brief — it’s a firsthand look at what “moving fast” really means when the world’s supply chains are humming.

- Ana

Noah Olsen

By 9 a.m., the heat had already turned the factory floor into a furnace. Yarn spun off spools like spider silk. Shirtless men in flip-flops rolled carts between machines. Rolls of fabric slid down tracks outside the building into awaiting trucks. This was my first glimpse into how China actually builds.

The factory I visited was in Quanzhou, Fujian province, which is historically known as one of the starting points of the Maritime Silk Road. Marco Polo and other traders once passed through here, helping connect China with regions such as Japan, India, and beyond. It is only fitting that, hundreds of years later, this place remains one of the beating hearts of the textile industry.

I was introduced to this factory through a warm connection from a local friend. It has been in operation for over 26 years and is entirely family-owned. The friendly couple who run it have built up production steadily over the years without taking on any debt. This factory takes in spools of yarn as the raw material and uses industrial-grade sewing machines to produce rolls of fabric. This fabric is later used to make sweatshirts, exercise bands, shoelaces, and other materials for predominantly fast-fashion brands such as Shein and Zara. This factory isn’t fully vertically integrated in producing finished garments such as sweatshirts, but it plays a crucial role in the supply chain for the end product.

As I was escorted into the factory, I was awestruck by the sheer size of everything. This wasn’t a single-story industrial warehouse like the ones I’d grown used to seeing in the United States. It reminded me of China’s pig skyscraper factories.

The newly built factory complex stretched across three six-story buildings, in addition to adjacent plots that were under development. The owner told me the entire process had taken about 15 months, including planning, but the actual construction had been completed in only five. Can you even get an environmental review done this quickly in America?

While the warehouses and factory floors were fully operational, the office spaces were still unfinished. The corporate staff were temporarily stationed on the bottom floor of the main building. They hadn’t waited for their air conditioning or polished offices to be complete before starting production. This portion reminded me of many scrappy Silicon Valley Startups with newly acquired office space.

Just one floor above, construction workers mixed cement and paint with drills. Concrete and building materials were scattered across the ground. It was an untidy, noisy, and messy environment, but it was music to your ears and eyes. Things were being built in real time.

The factory didn’t wait for building materials to be cleaned up before starting production. They didn’t wait for office spaces to be finished or for air conditioning to be installed. They set up the real revenue generators — the machines — first, and dealt with the rest later.

At this point, the factory was incredibly barebones. For example, the smoking area consisted of a bucket of murky water in the stairwell, next to a cracked-open window. I doubt this would pass inspection in America. Things felt scrappy here, as if many rules simply didn’t exist.

The elevators weren’t fancy or lined with trim, but they served their purpose: carrying people and equipment from one floor to another. And like I mentioned earlier, the building didn’t even have air conditioning. As we walked through each floor, even the owner was sweating profusely. Blood, Sweat, and tears were being poured into the development of this factory, even at the highest levels.

They also had a cafeteria. When I’d asked earlier to try the food, I was told it didn’t exist. Perhaps they didn’t want me to dine with the working folk. From my experience, these cafeterias usually only exist at posh tech companies in the U.S., such as Google or Anthropic. Yet even at a small factory in China without air conditioning, an auntie was there cooking hot meals for the employees.

Workers didn’t have to leave the building to eat, which kept production moving more efficiently. If an issue came up, they could walk straight from lunch to the floor to resolve it. One of the employees told me the company also provided optional housing. So they were providing their workers with both food and shelter. Perhaps you could call this an all-inclusive resort.

The meals and housing did not come easily, however. Here’s a screenshot of a conversation I had with one of the employees:

To get this straight and add more context: I was told that each employee has a choice of 2 days off per month, with the option to forgo those days in turn for additional wages. I confirmed this multiple times because I couldn’t believe it. They received two days off per month.

I was also told that employees worked anywhere from 10 to 14 hours a day, depending on demand. This takes the standard 996 — a term popularized by Jack Ma— to another level. The only people I know who work 996 in America are some startup founders in SF and investment bankers.

As we continued through the factory, the CEO brushed past many of the floors and machines as if they were routine. Yet as someone who hasn’t seen many of these things before, they were incredibly fascinating. I wanted to stay on each floor for hours, but we kept moving.

Some employees walked around without shirts or in slides, but nobody seemed to bat an eye — it was simply part of the process. Shirtless employees can even be seen in company marketing materials, which makes me wonder if this is a common occurrence. I was told that, since they were manufacturing textiles, installing large fans would mess up the product. I’m told that air conditioning will be installed next year. The lesson was clear: they don’t wait for comfort before they build.

The closest American parallel I can think of to this is Amazon. The company faces constant outrage for worker efficiency practices that, in China, barely register. Whether due to censorship or culture, people here seem to accept their role, do their job, and go home.

These employees work long hours but receive meals and housing in return. I wasn’t able to speak with any of the workers directly, but I didn’t feel an aura of misery while being around them. The employees glancing at me had smiles or curious looks on their faces, even when the boss was not around. I think my main takeaway from this factory visit is the labor force. People in America are not willing to endure these conditions. Is it ideal? I don’t think so. But they are not inhumane. They’re part of the grind to build great things.

The mountain of regulations and legislation preventing things like this from happening in America is staggering, due to the number of environmental reviews, OSHA compliance requirements, labor laws, and more. Yes, many of these things have a reason to exist, but they have made it impossible to move fast. Is America’s only fighting chance to become globally competitive in manufacturing to replace laborers with robots? Good luck.

As I roamed the floors with the owner for the first few hours, I was guided along very quickly. What was incredibly fascinating to me — hundreds of shelves of fabric rolls in a dark warehouse, 50+ industrial-grade sewing machines per floor, or other weaving apparatuses— was second nature to the owner. This left me incredibly awestruck by many of these things. I had never seen so many machines in one place.

Me: “Is this the largest textiles factory in China?”

Owner: “This is known as an SMB — a small to medium-sized business. We are very small compared to other factories inside China.”

He repeated this point several times, humbly. What I’d naively assumed was one of the largest textile plants was, in reality, a child amongst giants. One floor alone held more than fifty machines, all dedicated to making drawstrings.

I was told that they were in the middle of an end-of-month shift change, so many of the machines were not in operation. Two nights later, the owner sent a WeChat video showing all the machines in operation.

Was this one of China’s famous dark factories? No. However, there was little back-breaking labor left — only repetitive, tedious, and ultimately automatable tasks: cleaning the machines, repairing them, collecting the product, and so on. Most of the manual work was handled by machines, forklifts, or, in this instance, gravity:

There were no humanoid robots or robotic arms in sight. I was told by the owner that he has plans to bring in more automation over the next decade, aiming to increase revenue fivefold.

After the initial shock wore off, I started piecing together how all of this actually worked. It boiled down to a few essentials:

1. Sourcing the raw material — hundreds of thousands of rolls of yarn.

2. Acquiring the machines, many built by another company in Fujian

3. The workforce

4. Securing the real estate and factory infrastructure

5. Operations.

6. Legislation

You can distill this down to four things: equipment, workforce, legislation, and operations.

Ultimately, everything came down to execution. The R&D was already done; they had the machines, materials, and people in place. They were simply left with making sure everything was running smoothly. Now, of course, this doesn’t apply to new R&D initiatives, but at the most basic level, the formula was simple: manufacture the product and move it out the door.

This means that any friction lowers margins and slows productivity. Energy costs, labor costs, construction delays, regulation — the list goes on. Reducing each of these barriers gives Chinese factories a competitive advantage over the rest of the world.

Can America do the same thing? Here are a few barriers that come in the way of this:

• Energy in the U.S. is more expensive (and less abundant)

• Labor is more expensive

• Few American workers are willing to be subjected to these working conditions, such as working longer hours for the same pay.

• Regulatory oversight, environmental concerns, and OSHA slow down construction and production to a crawl.

The conclusion I have come to is that manufacturing textiles in America does not make economic sense unless it’s subsidized. I know of one startup, Anatar, that is working on producing textiles in America. Many of their target customers are governments and other institutions that are willing to pay a premium. Without this premium, we cannot create textile factories in the U.S., because the raw inputs lead to the U.S. losing every time.

Labor, energy, and regulation are a losing battle. Businesses fundamentally cannot erode these costs — despite their best efforts — without policy change. It’s up to the government to help solve these problems. This is actually a very disappointing and disheartening conclusion that I have come to. Unless you are competing in subsidized industries or can pave a new path forward with industrial automation, you’re setting yourself up for failure.

Reindustrialization in America will require coordination between industry and government, alongside repeated pressure from citizens who still believe in building for democracy. The “government” isn’t some distant machine; it’s people making decisions, or avoiding them. If we want to compete, these decisions have to start breaking in favor of production.

The reindustrialization movement in America is a start, but it’s just that. Getting our industrial base back online will take time, capital, and a change in mindset. It will not happen through a single company or policy. It will take a generation of people who decide that building here, at home, is worth the cost. The next decade will decide whether that belief holds or fades.

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

Dan Wang’s Breakneck: China’s Quest to Engineer the Future and Patrick McGee’s Apple in China: The Capture of the World’s Greatest Company are not companion volumes in any formal sense, but they rhyme. Wang gives a clear frame for the country that moves first and argues later. McGee maps how a single American company turned a faraway factory map into the beating heart of its business. Taken together, they are a live wire running across the U.S.–China competition. They did not flip my priors. They hardened them. I, along with many hard tech founders and investors, have long believed the United States needs to rebuild its capacity to make things and to fund the people who bend atoms rather than just move bits. These books deepen that conviction and sharpen a more uncomfortable question I keep asking myself lately. Is it already too late to compete with China? I will answer that, but first the ground we are standing on.

Wang’s core claim is not a provocation for the sake of it. He argues that China is run by engineers and the United States is run by lawyers. That one sentence gets you far. If the people at the top are trained to build, they will use building as their instrument for legitimacy and control. If the people at the top are trained to litigate and negotiate, they will reach for process as their main tool and for caution as the default setting. On Ashlee Vance’s Core Memory podcast, Wang makes the point without theatrics. Chinese leaders, in his telling, treat metros, bridges and transmission lines as expressions of national intent. American leaders treat big projects as problems to be processed through a dense thicket of permissions. You can hear Ashlee nodding because you cannot sit in California for long, look at the world’s most bountiful technology economy, and square it with how little actually gets built around you.

McGee’s book is the more surgical instrument. He shows in steady detail how Apple did not simply outsource assembly to China. It poured money, people and knowledge into the place. He cites a quiet pledge by Tim Cook in 2016 to spend a quarter trillion dollars in China over five years. The number has been reported and debated for years, but the magnitude matters less than the habit behind it. Apple was already spending at a rate that could justify a promise of that size. The company sent its engineers to work directly with suppliers, made sure everyone followed the same methods, and trained an enormous number of Chinese workers - so many that it’s more than the entire workforce of California. When you put that much money and know-how into a place, you change it, and it learns how to work without you. McGee’s point is tough but makes sense: Apple built a system in China that now limits what it can do. At first, Apple had leverage; then, it became dependent. Now, it’s trapped.

For a real-world example, look at Zhengzhou, “iPhone City,” in fall 2022. The city tried to keep making iPhones even as Covid lockdowns kept people stuck in their homes. Workers protested, and videos showed chaos at factories. Apple fought to keep production going - it didn’t totally stop, but it was clear the company wasn’t really in control. In China, the government sets all the rules, so foreign businesses are always renters. You can see this in export statistics and in how Apple keeps balancing promises to Beijing with new factories in India and Vietnam. Apple is spreading out some final steps of its manufacturing, but the core work is still rooted in China’s way of doing things.

Wang’s frame and McGee’s reporting reinforce each other. One explains how China decides. The other shows what happens when a Western company builds its own right arm into a system that thinks in projects, not debates. It is tempting to stop there and declare the race fixed. It is not. The two books also surface the vulnerabilities under the surface swagger. China’s engineering instinct extends past dams and tunnels into the social sphere. The same confidence that builds cities builds policies like the one-child era and Zero-Covid. Those policies carry costs that take a generation to fully expose. The price can be paid in demography, in trust, in the small ways people decide to stop trying so hard because the system no longer rewards risk. Wang knows this too. He lived through the lockdowns, watched young people chant in Shanghai streets on a cold night, and kept writing when censors took aim at his site. He is an admirer of the capacity to build and a critic of what happens when you start treating people as units in a project plan.

The United States has its own brand of failure. It is not a shortage of intelligence or imagination. It is an absence of institutional urgency. If you want a depressing statistic that does not fit on a motivational poster, read the NYU Transit Costs Project. The team has cataloged rail projects across the world and made the grim arithmetic public. The average subway kilometre in comparable democracies tends to land in the tens or low hundreds of millions of dollars. New York routinely builds at a price tag ten times that and sometimes more. Bloomberg has reported that the first phase of the Second Avenue Subway cost roughly two and a half billion dollars per mile. The next phase is projected to be even higher by some accounts. It is not a law of physics. It is a choice stack. And each choice adds time and cost until the ambition congeals into a case study in why we cannot have nice things.

California’s high-speed rail has become the other talisman of aspiration without execution. Voters approved it in 2008. The project has been chopped and resized and rescued and threatened. There is guideway in the ground. There are bridges you can drive under. There are contracts and invoices and jobs. There is also an Inspector General saying the Merced to Bakersfield starter segment may slip to the early 2030s, a federal critique that stopped just short of a eulogy, and a cost range that no longer fits inside the political rhetoric that launched the idea. The right interpretation is not that America cannot build. It is that America will not build through its current versions of process and politics. In this context Wang’s phrase about the country of lawyers sounds less like a snark and more like a diagnosis.

I am a deep tech investor by choice and by temperament. Most of my time goes to people who build better tools, weld stronger frames, invent new materials, and push processes to deliver more - even when those processes fight back. These books did not convince me that manufacturing matters. They pressed the argument into the table. They also sharpened a question I have been asking with more frequency this year. Is it too late to compete with China? The country does not just build phones at scale. It builds drones, battery factories, and the machine tools that make your machine tools. It is learning to make cars that people will buy in Europe in numbers large enough for politicians to notice. It fills entire shipping schedules with cars and then figures out how to assemble them locally when tariffs arrive. Overcapacity at home pushes firms to chase sales abroad. Strategy has a way of finding an outlet.

There is another part of the Chinese story we gloss over too easily. The country took on the pollution that the rich world did not want. Rare earth processing is one example. It produces toxic waste by the truckload, sometimes with radioactive traces, and it leaves behind the kind of environmental debts you can see from space if you know where to look. It is not romantic. China decided that being the stomach for the dirty parts of the clean future was worth the cost. That choice delivered leverage and scars. Both matter in the years ahead.

So again. Is it too late? The short answer is no. The longer answer is that winning will not look like a return to the past. It will look like a country that relearns how to take itself seriously in the places where it still holds unfair advantages and that chooses to invest its political capital in building things that last. It will not come from a speech. It will come from a chain of very specific moves and a run of years where the execution stories outweigh the X sermons.

The first part of the positive answer is that the United States still holds the upstream of a lot of technologies that matter. We are strong in foundational research. We lead in many of the software stacks that orchestrate hardware. We attract skilled people who want to innovate, not just fill jobs. While factories are essential, strong research and top talent help them thrive. When the U.S. builds up a new industry, the entire support system - money, talent, and clear laws - grows around it. Our global connections also help link ideas and businesses worldwide. But relying only on research and innovation isn’t enough; we need strong manufacturing and hands-on expertise to support our strengths. That’s where we need to improve.

The second part is that China’s strengths come with soft tissue. The engineering state moves fast when it wants to, but it pays in social trust when it overreaches. Zero-Covid worked until it did not. It controlled for the virus and then failed to control for people. The country reversed course hard and late. That kind of shock introduces a wobble that does not go away in a quarter. Demography is the other drag. The one-child era left a shape that no policy can resculpt quickly. Labor markets change. Consumption patterns swerve. Most importantly, the mood shifts. You can see it in the drift of young talent into quieter lives, in bursts of entrepreneurial energy that look tired around the edges, and in the unsteady treatment of big tech companies that used to set the tone for a generation. It is not collapse. It is the kind of noise that complicates an otherwise clean graph of progress.

The third part is to look hard at where U.S. execution actually works - and scale those instincts like it matters. The CHIPS program is a small but necessary start: it forces coordination around what no single firm can handle alone. But sovereignty is won upstream - in the toolmakers, specialty chemical firms, metrology experts, and power-electronics providers that make the glamorous fabs possible. America’s way back is an obsession with those “boring,” vital layers.

We also need to reckon with law and time. Founders building factories tell you the same story: permits, reviews, hearings, appeals, the appeals to the appeals. By the time the bureaucracy catches up, the factory has moved or the investor is gone. The answer is not to erase environmental protection or community consent - but to route around paralysis. Fast-build countries identify strategic projects, slot them into corridors where rules are clear and timelines binding, and hold the public side accountable to delivery. The U.S. already has pockets of that logic; we need more. We must treat schedule certainty as national advantage, anchoring time limits in months, not election cycles.

During Trump’s second term, the administration enacted a series of executive orders and policy actions focused on strengthening U.S. industrial and manufacturing capacity, national security, and supply chain resilience. These included modernizing defense procurement to encourage innovation and speed, prioritizing domestic production of critical minerals for advanced manufacturing, and reducing regulatory barriers to accelerate approval of pharmaceutical and energy infrastructure projects. Major trade interventions, such as a universal 10% tariff and harsher reciprocal tariffs, were introduced to address the trade deficit, alongside regulatory reforms targeting a revival of domestic manufacturing in sectors deemed strategically vital. Additional measures emphasized revitalizing the U.S. nuclear industry and securing supply chains for essential pharmaceuticals, all signaling an intent to realign industrial strategy with national security imperatives and reduce dependencies on foreign producers.

These examples show how the administration is attempting to tie high-level strategy (tariffs, trade, emergency designations) to industrial execution (minerals, medicines, supply chain resilience). The risk is real: policy announcements are easier than building the toolmakers, qualifying suppliers, or cutting through local red tape. But the signal is clear: sovereignty may live upstream, but speed is becoming strategy in Trump’s second term.

While these policy moves aim to close the gap between government strategy and hands-on industry results, the story of drones illustrates what’s at stake when practice and production aren’t kept close to the source of innovation. DJI is a Shenzhen company that grew up in the slipstream of Apple’s supply base and the city’s allergy to bureaucracy. It had the good sense to focus and the nerve to ship before the Americans could finish a white paper. It dominates the consumer market by a margin that would embarrass a monopoly case. The lesson is not that we cannot build drones. The lesson is that when you do not keep the practice of building close, you cede the learning curve to someone who does. When that happens, even a ban does not give you a domestic champion. You end up with a vacuum and a lot of speeches.

The American positive answer is specific and a little boring, which is why politicians have a hard time selling it. We need to back the people who make the machines that make the things. We need to harden a handful of foundational sectors where we must not be at the mercy of other people’s factories. We need to redesign the way the public sector handles projects so that a decade is not the unit of account. We need to tell a better story about why building matters. Not a nostalgia story and not a fear story, but a story about competence and pride and dignity in doing hard things that stay done. And we need to accept that winning will look like a long string of medium-sized successes rather than one grand slam that fixes the mood.

We could set up a federal group that acts more like a mission-driven team than a grant office. We used to do this. We can do it again. It would have engineers and program managers who stay long enough to accumulate scar tissue. It would have the legal authority to cut through disputes rather than escalate them. It would publish schedules and then live with the humiliation of missing them if that is what happened. It would pick factories and corridors and then defend them in court with the same vigor it uses to send a press release. You can call it an ARPA for Hard Things if you want. The name matters less than the posture. The unit exists to get physical projects over the line. The win condition is concrete poured and machines humming, not a ribbon cut on a website.

It would help to change our diet too. If you feed people a media stream of apps and crypto and influencer scandal, you end up with a country that thinks the real economy sits inside a phone. We should put factories on television again. We should send school trips to the places that smell like flux and oil. We should pay machinists as well as we pay salespeople. We should let kids see that a robot arm is not an enemy but a companion that needs someone who understands how to keep a gearbox alive. There is a culture piece to this. It is not enough to say the word industrial policy and assume the culture will follow. Cultures are trained. We trained ours to think that moving fast online counted as motion. We can retrain it to recognize motion in the world.

The last piece is the one that keeps me optimistic, even after reading two books that could have been written as obituaries if their authors were less careful. The United States has a gift for reinvention that rarely looks heroic in the moment. Our best stories of recovery are usually boring for a long time and then obvious all at once. That is how the semiconductor industry found its balance the last time it stumbled. That is how the shale patch learned to drill sideways. That is how the space business stopped talking about flags and started counting launches. You get a many-year window where nothing looks right and the critics are loud. Then a cluster of firms finds a groove and the numbers shift. If we decide to take manufacturing seriously as a national project rather than as a nostalgia trip, the same pattern is available.

This is where I land after reading Wang and McGee and spending a year listening to founders complain and then walk back into the shop to fix another problem before the sun comes up. It is not too late. It is late. There is a difference. Late means we cannot afford another cycle of performative policy that funds headlines and not weldments. Late means we schedule projects like adults, fund the upstream that no one sees, and accept that sometimes the greenest thing on earth is a factory that runs at yield because it wastes less. Late means we cannot keep exporting our pollution and our guilt and call it virtue. It means we take responsibility for the rare earths and the chemicals and the reactors that make an electrified life possible without frying the planet. It means we get serious about transmission and transit and the invisible work that turns a slogan into a commute.

I will keep backing the teams that want to put more metal in the world in smart ways. I will keep arguing for the version of America that knows how to build without forgetting why rights matter. I will keep telling myself that we get to choose whether we learn from the bargain Apple made and from the confidence China carries into its projects. The alternative is a future built by other people, delivered to our ports, and argued about by our lawyers. I would rather live in the version where our factories hum and our machines make machines and our kids grow up knowing that the sound of progress is not a keynote. It is the rhythm a place takes on when it remembers that ideas are only as strong as the hands that turn them into things.

You can call that faith if you want. I call it a plan that starts tomorrow morning. The rest is on us. 🇺🇸

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

Let’s give Tesla Optimus a bit of a break. While it’s busy sparring with Jared Leto and capturing headlines, an interesting robotics battle is unfolding quietly in the private markets. On one side stands Figure AI in San Jose. On the other, Apptronik in Austin. In my view, this is the matchup that matters.

Both are trying to do what generations of engineers and dreamers could not: build robots shaped like us, moving like us, capable of working in the same cluttered, unpredictable spaces we occupy. Both companies have money, talent, and momentum. Both have engineers who can talk torque curves one minute and neural networks the next. Both are sprinting toward the same vision — a humanoid labor force that can clock in for shifts in warehouses, factories, and, one day, our homes.

That’s roughly where the similarities end.

Apptronik has spent over a decade grinding through prototypes, learning hard lessons inside NASA’s Valkyrie program, building exoskeletons with mythological names, and failing just enough to know where the traps are. Its flagship robot, Apollo, is the product of that long slog.

Figure AI, by contrast, is pure Silicon Valley energy — a company born from blitz-scale ambition. Founded in 2022 by Brett Adcock, who had already built and sold two startups, Figure raised billions in under three years. Where Apptronik is patient, Figure is breathless. Where Apptronik speaks in field tests and spreadsheets, Figure speaks in compounding data curves and internet-scale learning.

The Apollo Lineage

Apptronik’s culture is steeped in NASA engineering discipline. The team’s Valkyrie robot, tested at the Johnson Space Center, taught them that humanoids fail in the least forgiving ways: they tip over at the wrong angle, cables fatigue, actuators burn out. That humility shows up in Apollo’s design.

One former NASA engineer who joined Apptronik in 2018 recalled the day Valkyrie face-planted during a DARPA test. “The machine just folded at the knees like a marionette whose strings had been cut,” he said. “We were embarrassed, but more importantly, we realized that actuators were the entire ballgame.” That moment seeded the obsession that defines Apollo today: actuators that can be mass produced, serviced, and trusted.

Apollo’s actuators are linear electromechanical drives with carefully tuned gear ratios. Unlike rotary actuators, which often require expensive harmonic drives, these linear actuators generate force along a straight axis, then convert it through linkages into rotational joint motion. Apptronik engineers talk about force density (newtons per kilogram) and thermal dissipation as the two numbers they watch most. The thigh actuators have embedded strain gauges and thermistors, which feed into a joint-level control loop running at hundreds of hertz. When a unit overheats, torque output is throttled automatically — but the whole assembly is designed to be pulled and swapped in under ten minutes. “Every actuator has to be thought of like a car brake pad,” an Apptronik engineer said. “Cheap, strong, replaceable.”

Apollo 1 in 2022 was tethered, upright, and mostly a lab prop. Apollo 2 in 2023 introduced hot-swappable batteries, letting the robot run close to twenty-two hours a day if you kept enough packs cycling. Apollo 2 is the company’s first real product: five foot eight, 160 pounds, with seventy actuators that can haul 25 kilograms safely. Apollo 3 in 2027 is slated to bring advanced multi-finger hands. Apollo 4 by 2029 or 2030 is meant to be a generalist worker, built in the tens of thousands, priced to compete with human labor.

Walk through Robot Park in Austin, and you get a sense of the company’s ethos. It is not a glossy demo center. It looks more like a training ground. Engineers replicate environments so Apollo can practice before deployment. One corner has a mock warehouse aisle. Another has supermarket racks. The academy trains operators to work with imitation learning and diagnostics.

Behind the scenes, Apollo already has pilot clients. In 2025, a logistics company in Texas ran a fleet of six Apollos for pallet loading and box stacking. The robots worked side by side with union laborers, who were skeptical at first but warmed when they realized the machines handled the most back-breaking jobs. In another case, a large food distribution center on the East Coast trialed Apollo for freezer work. Human turnover in freezers is notoriously high. Apollo’s ability to swap packs and work uninterrupted made managers sit up.

Apollo’s perception stack uses a fusion of Intel RealSense depth cameras and LiDAR for near-field navigation, coupled with IMUs in each limb for balance. The GPU inside is sized to run transformer-based perception models at about 20–30 frames per second. Training happens through human demonstrations captured via teleoperation rigs, then refined in NVIDIA Omniverse simulations. Policies are transferred with reinforcement learning and validated on hardware. Safety is handled with force-torque sensors at the wrists and ankles and compliance control that kicks in when unexpected forces exceed a threshold.

Apptronik’s business model mirrors Tesla’s parity math. Warehouse workers cost about $25 an hour fully loaded. Apollo as a service costs $21 today, dropping toward $17 by the end of the decade. Buy outright and the cost converges with human labor in three years. The company does not pitch Apollo as a marvel of engineering. It pitches Apollo as labor, cheaper and tireless.

Critical to this scale-up effort is Apptronik’s strategic partnership with Jabil, a global leader in electronic manufacturing services. In early 2025, Apptronik and Jabil announced a collaboration to mass-produce Apollo humanoids at Jabil’s advanced production facilities. Jabil not only manufactures Apollo robots for Apptronik’s clients worldwide, but also tests and deploys Apollo within its own factory lines, using the robots for material handling, kitting, sub-assembly, inspection, and lineside delivery. This allows Apptronik to validate real-world use cases and optimize Apollo’s AI models in a world-class production environment. Jabil’s supply chain expertise lets Apptronik unify procurement, accelerate time-to-market, and manage inventory and logistics globally. The partnership is designed to enable rapid, cost-effective scaling, making Apollo robots more affordable, easier to maintain, and ready for deployment in new settings such as retail, elder care, and home environments. By integrating Apollo robots directly into its manufacturing processes, Jabil serves both as Apptronik’s customer and its production partner, providing the flexibility and resources for expansion.

In September 2025, Apptronik’s strategic bet got a jolt. DeepMind announced Gemini Robotics 1.5 and ER 1.5 (based on Gemini 2.0), models that can reason, plan, and act across embodiments. Skills trained on robotic arms like ALOHA2 can transfer to humanoids like Apollo. The models can even query tools like Google Search before executing tasks. Apptronik was named as an early partner. Suddenly Apollo is lashed to one of the most ambitious AI roadmaps in the world. That means Apollo’s brain will evolve with Google’s cadence, absorbing skills far faster than Apptronik could develop in-house.

Figure AI and the Helix Gambit

If Apptronik’s story is patience, Figure AI’s is velocity.

When Adcock started Figure, he toured universities and AI labs looking for recruits. A Stanford PhD student who turned him down later said, “Brett had this look in his eye. He didn’t want to build a better actuator or a clever demo. He wanted to build a company at the scale of Tesla, but for robots.” In September 2025, Figure hit a $39 billion valuation. The number made old robotics engineers laugh out loud. But it let Figure build in parallel: the brain (Helix), the body (Figure 03), the battery (F.03), and the factory (BotQ).

Helix is the brain that defines the company. Unveiled in February 2025, it is a vision-language-action system that combines a large semantic backbone with a fast visuomotor controller. System 2 parses language and images at 7–9 hertz, condensing them into latent plans. System 1 runs at 200 hertz, turning those plans into continuous joint commands across wrists, fingers, torso, and head. Helix was trained on about 500 hours of teleoperation but has generalized far beyond. Two robots running the same weights collaborated on grocery storage, handing items back and forth while adjusting gaze and posture in real time.

Helix’s technical stack has features that engineers rave about. A vision memory module lets it “remember” partial glimpses of barcodes or objects, so it does not repeat motions. A state-history buffer keeps track of limb positions across action chunks, improving continuity. Force feedback acts as a proxy for touch, giving Helix the ability to adjust grip in real time. Ablation studies showed each feature cut seconds off package handling. The policy runs on dual embedded GPUs, splitting System 2 and System 1 across them for real-time inference.

The compounding effect of Helix became clear within months. In the spring, it sorted padded envelopes on conveyors, patting them flat for barcode scans with 95 percent accuracy. By August it folded towels, tracing edges with its thumb. By September it loaded dishwashers, singulating plates and reorienting glasses. No code changes. Just new data.

At BMW’s Spartanburg plant in South Carolina, Figure robots were tested on inserting sheet metal into assembly lines. The deployment was small, and insiders argue over how many robots were really used. But the fact remains: BMW, a German automaker famous for caution, wrote a check.

UPS has also sniffed around. In mid-2025, executives from the parcel giant visited Figure’s lab to see if Helix could be trained for sorting. The problem was clear: UPS loses money when packages pile up on conveyors during peak holiday weeks. If humanoids can keep pace in those periods, they pay for themselves.

Figure approached walking as a software problem. They trained thousands of simulated robots to walk across randomly generated terrains, forcing the control system to handle uneven ground, obstacles, and unexpected slips. This technique, called domain randomization, makes the software more adaptable when transferred to real hardware. The robots’ control loops run at very high speeds (measured in kilohertz), which helps keep balance in real time.

The result: when the trained policy was installed on physical robots, it worked immediately — no manual tuning required. Ten different robots were able to walk identically “out of the box,” a milestone that shows the training scaled reliably.

At the same time, Figure redesigned the robot’s body for manufacturability. Early prototypes were made with custom-machined parts, which are slow and expensive. Figure 03, their latest version, switched to cast, molded, and stamped components. These are standard mass-production methods that reduce part production time from days to seconds.

The company’s BotQ factory is built to eventually produce up to 12,000 robots per year. Inside, robots are already assisting in building other robots — a feedback loop of automation.

Each component (joints, circuit boards, etc.) is tracked in a manufacturing execution system (MES) that integrates with the company’s design and supply chain tools (PLM, ERP, WMS). This means if a defect shows up, engineers can trace it back to the exact supplier batch within minutes.

To ensure durability, reliability teams run robots through accelerated life tests (e.g., exposing them to heat in ovens to simulate long-term wear). Data from failures feeds directly back into the design process, tightening the loop between testing and improvement.

Power is handled by the F.03 battery, a torso-integrated 2.3 kilowatt-hour pack with active cooling, fast charging, potting, and flame arrestors. Fault injection tests show runaway cells failing safely without propagation. Figure claims it is the first humanoid battery on track for both UL and UN certification. The pack is assembled through a laser ablation and wire bonding line, followed by in-process quality checks, potting for flame retardancy, and a final BMS install. The MES logs every cell’s genealogy. If a failure occurs, engineers can trace it to the supplier lot in minutes.

And then there is Project Go-Big, the data engine. In partnership with Brookfield, Figure is planning to capture egocentric human video across 100,000 apartments, 500 million square feet of offices, and 160 million square feet of logistics space. The first milestone was navigation: Helix, trained only on human video, followed commands like “go to the fridge” in cluttered apartments. If the program scales, it could become the YouTube of embodied data, the missing corpus for humanoid training.

Adcock himself frames Go-Big with the swagger of a founder who has played this game before. “Robots will learn from how humans actually live, not from staged demos,” he told an audience in Palo Alto. “That’s how you scale.”

The Arrival of Figure 03

While I was finishing this piece, Figure unveiled Figure 03, and it reads like a ground-up redesign around three priorities, Helix intelligence, safe real-world home use, and truly scalable manufacturing, with the company’s own launch note laying out the core changes, a next-generation vision stack, new tactile hands, higher bandwidth data offload, softened exterior and audio upgrades for in-home interaction, plus a manufacturing flow that trades CNC for high-rate tooling so units can be built in volume rather than as prototypes. The vision system is explicitly engineered for high-frequency visuomotor control, each camera now delivers roughly twice the frame rate at one quarter the latency with a 60 percent wider field of view compared to Figure 02, and the optics have an expanded depth of field so Helix receives a denser and more stable perceptual stream that holds up when the scene gets cluttered or the robot’s head view is partially blocked. And perhaps most striking of all, Brett Adcock claims that nothing in Figure’s launch demo was teleoperated — a bold statement in a field still dominated by staged autonomy.

Figure 03 also embeds wide-angle palm cameras in both hands so the robot keeps close-range visual context during grasps, which matters when the head cameras are occluded by the robot’s own arms while reaching into cabinets or working inside tight spaces, and this hand-eye redundancy is paired with 10 Gbps mmWave offload so fleets can upload terabytes of sensor data for continuous learning at dock. The new fingertips are softer and more compliant to increase contact area on irregular objects, and Figure moved tactile sensing in-house after finding commercial options fragile in practice, the first-generation fingertip sensors register forces as low as about 3 grams so Helix can feel impending slip before it happens and modulate grip in real time on deformable, fragile, or moving items.

Safety and daily usability get equal billing, external hard parts are wrapped in washable soft textiles over strategically placed multi-density foam to eliminate pinch points, overall mass drops by 9 percent with a noticeable volume reduction versus 02 for easier maneuvering in domestic spaces, soft goods are fully removable without tools, and the company highlights clothing options from normal knitwear to cut-resistant fabrics along with side-mounted screens for ID and fleet management.

The audio stack is rebuilt for natural conversations, the speaker is about twice the size and nearly four times as powerful as before, the microphone is repositioned for clarity, and the robot supports real-time speech-to-speech, which matches the product’s explicit aim at home interaction rather than lab demos.

Power and autonomy are anchored by the F.03 battery, and Figure 03 adds 2 kW wireless inductive charging through coils in the feet so it can step onto a stand to top up while also wirelessly offloading data. Figure says 03 abandons prototype-grade CNC-heavy builds for die-casting, injection molding, and stamping, trims part counts and assembly steps across almost every subsystem, vertically integrates critical modules like actuators, batteries, sensors, structures, and electronics, and stands up a purpose-built supply chain alongside the BotQ facility backed by a tightly instrumented MES for full component and process traceability. All of this is built to serve Helix. For clarity on headline specs, Figure’s product page lists height 5′8″, weight 61 kg, payload 20 kg, runtime 5 hours, and walking speed 1.2 m/s, which sets expectations for early home and light-industrial tasks while the company keeps consumer pricing and a firm retail date unannounced and third-party reviews note that near-term deployments will likely remain pilot-oriented rather than broad home rollouts.

Cultures in Collision

The cultural gap between the two companies is stark. Walk into Apptronik’s Robot Park and you get the sense of a company that builds confidence by grinding through hours of uptime and acceptance tests. Walk into Figure’s HQ and you feel like you are inside a software company disguised as a factory, with engineers running ablation studies on memory modules while robots ferry components across the floor.

Apptronik speaks the language of labor parity, field serviceability, and contracts. Figure speaks the language of compounding curves, internet-scale data, and aggressive iteration. Apollo depends on DeepMind to keep its cognition competitive. DeepMind brings the cognitive horsepower that few research labs on Earth can match — its Gemini Robotics models are already capable of perception, reasoning, and motor planning across multiple embodiments. By linking Apollo’s body to DeepMind’s brain, Apptronik effectively plugged into a global AI infrastructure trained on multimodal data and high-fidelity physics simulation. The partnership allows Apptronik to skip years of cognitive R&D and instead focus on reliability, manufacturing, and deployment. In practice, this means Apollo can inherit the same reasoning stack that powers DeepMind’s embodied AI systems — models that can learn from video, reason in 3D space, and generalize across tasks. For a robotics company that has always led with engineering discipline, not software flash, that’s a remarkable advantage.

But there’s also danger in this arrangement. DeepMind is not a neutral supplier — it’s an AI empire inside Alphabet, with ambitions that may one day overlap or even compete with its partners. If DeepMind decides to spin out a unified cognitive platform for humanoid robotics — something akin to an Android for embodied AI — then Apptronik may find itself as the proving ground for that system’s first iterations rather than its exclusive beneficiary. The current collaboration gives Apollo early access to Gemini’s embodied reasoning models, but it also ties the company’s progress to DeepMind’s research cycles and priorities. Should Alphabet shift focus, restrict access, or commercialize an independent robotics stack, Apptronik could be left dependent on software it doesn’t control.

For now, the partnership looks symbiotic. As 2026 approaches, the industry will be watching closely to see whether DeepMind remains Apptronik’s ally, or whether Apollo becomes just the first host in a larger ecosystem that DeepMind ultimately owns.

In contrast, Helix is messy but entirely Figure’s own, a living architecture built, trained, and refined in-house. While Apollo borrows cognition from a cloud-scale research lab, Figure builds it directly into the feedback loop of its machines. Helix is not a model handed down from a corporate parent on a set release schedule; it evolves continuously, retrained on data streaming from Figure’s robots in factories, labs, and test apartments. This autonomy makes the system chaotic in the short term because parameters shift weekly and architectures change mid-deployment, but it also means the company controls every layer of its intelligence stack.

Who Wins?

Apptronik has the discipline of NASA engineers. Its actuators are clever, its pricing model will appeal to buyers, and its DeepMind partnership gives Apollo a powerful cognitive ally. Apollo 2 is a robot a warehouse manager can pilot today. Apollo 3 and 4 are mapped clearly.

Figure has the integration. Helix improves with each dataset. BotQ will manufacture at industrial cadence. The F.03 battery closes safety traps early. Project Go-Big feeds human behavior directly into the model. The culture is one of parallel loops and relentless updates. Before 03, the company’s advantage was theoretical: integration, speed, and capital. After 03, it is tangible. The robot’s design shows that Figure has solved, or at least systematized, three of the hardest problems in robotics — perception, dexterity, and manufacturability — inside a single, production-ready architecture. The valuation is frothy, but it reflects a system that learns faster than it breaks.

Both companies deserve credit for dragging humanoids out of labs and into the realm of contracts. When the world stops rewarding demos and starts rewarding deliveries, the edge will belong to the team whose loop spins fastest. Right now, that seems like Figure.

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

Tesla’s humanoid robot started life as a provocation, a person in a black-and-white suit walking on stage in 2021 while the crowd laughed and filmed. The joke did not last. By the fall of 2022, an actual prototype wobbled onto a stage in Fremont with a tangle of exposed wires and metal bones. That night established three truths at once. First, Tesla showed it could turn a half-formed idea into working hardware with remarkable speed. Second, the company would try to do as much as possible in house, from actuators to controllers. Third, this would be a decade-long fight with physics, supply chains, and human expectations.

The second act came in December of 2023, when Tesla posted a crisp video of a slimmer robot. The new machine, which the company called Gen 2, showed off clever feet with force and torque sensing, toes that move like the real thing, a neck that allows the head to track hands, and hands that finally looked like tools rather than props. The fingers carried tactile sensors. The robot walked faster than before and handled delicate objects, including the now famous egg trick. You could argue about how much of that scene was choreographed, and many did, but the hardware story was no longer hand waving. Tesla had taken weight out, friction out, and guesswork out, and had put sensing and controllability in. In public writeups of that Gen 2 reveal, Tesla highlighted a roughly ten kilogram weight reduction, all-new Tesla-designed actuators and sensors, faster walking, and brand-new hands with eleven degrees of freedom per hand and fingertip tactility on every finger.

A month later the internet saw Optimus fold a T shirt and went predictably crazy. Musk admitted the truth the same week. That sequence was not autonomous, not yet. The point of the clip was to show that Tesla’s hands and controls could trace a human-in-the-loop behavior with enough accuracy to be worth training. The admission matters because it frames what came next. Tesla started changing how the robot learns.

By late summer of 2025, the company began to push a new training regime that feels like a remix of its camera-first philosophy from cars. Instead of leaning on motion capture suits and VR teleoperation sessions, Tesla started strapping helmet and backpack rigs to workers and filming the boring, repeatable movements that make up most useful work. Five in-house cameras, lots of viewpoints, lots of mundane repetition. The footage becomes the raw material for imitation learning and policy distillation. The approach scales because you can put a camera on a person fast, and because the world is already full of tasks people perform without thinking about them. This is the simplest reading of what you tell a robot if you want it to generalize. Watch a thousand human micro-maneuvers. Find the invariants. Then try it yourself. Tesla has now documented the helmet-plus-backpack approach as five in-house cameras recording ordinary workflows for imitation learning.

Alongside that camera firehose is a second stream. Tesla has leaned into heavy simulation and what one well-circulated explainer called digital dreams. The idea is to generate or stage thousands of near-miss versions of a task, recover the implied action sequences, and bake the lessons into a policy that transfers back to the real robot. It is a variant of the sim-to-real tradition that robotics labs have chased for years, scaled up by Tesla’s taste for big datasets and cheap experimentation. It saves hardware from punishment. It compresses time. It also fails in interesting ways when reality does something unglamorous, like introduce dust or glare or a wobbly table. That is where the camera rigs help, because they collect the long tail of little problems that break clean demos.

Now to gold Optimus. A gleaming, gold-colored prototype surfaced, touched off a rush of hot takes, and briefly convinced some people that Tesla had shown a full third generation. But Musk confirmed that the gold unit is an interim build, a 2.5, not the third generation. The look drew eyes, which was surely the point, but the substance came from two places. First, the company showed that Grok, xAI’s model, can ride inside the demo loop and give the robot a voice and a way to parse casual requests. Second, the gait and body balance still looked like an active work in progress rather than a solved problem. It is important to hold the two ideas at once. A better language model lowers the friction of telling a robot what to do, but the hard part remains what your hands and feet do once a command meets the floor. A recent demo with Marc Benioff shows Grok handling spoken requests, but the responses and motion still lag in a way that proves the interface while reminding us the body is the hard part.

Tesla has been signaling a new compute module for both cars and robots that carries a much fatter memory subsystem and operation-specific accelerators for the kind of math that runs large policies at low latency. If Optimus is going to take the flood of video training, learn from simulated variations, and then close the control loop with tactile feedback and vision at high frequency, it needs headroom. That headroom shows up as bigger models, higher frame rates, and less waiting while a control policy catches its breath between frames. Put plainly, the brain cannot be a bottleneck if the hands are finally getting good. Tesla’s silicon roadmap makes this explicit. AI5, expected to appear first at TSMC around 2026, is a transitional chip that unifies training and inference for cars and robots. AI6, tied to a sixteen-and-a-half billion dollar Samsung deal with production targeted in Texas through the 2030s, is the true platform for scale. Musk has described it as delivering roughly eight times the raw compute, nine times the memory, and five times the bandwidth of current boards, with efficiency improvements up to forty times in some tasks. Where AI5 bridges the gap, AI6 is meant to be the backbone of both vehicles and Optimus at fleet scale. Texas just backed the Taylor site with a record state grant, which is one more reason to treat AI6 volume in the United States as a real plan rather than a talking point.

Let’s dig further. Tesla’s Gen 2.5 hands matter because they move well beyond earlier tactile prototypes, closely mimicking human anatomy in both form and function. For this generation, the hands themselves offer a combined total of twenty-two degrees of freedom - eleven per hand - distributed across individually actuated fingers and a dexterous wrist, making them one of the most advanced humanlike robotic hands seen in production settings so far. The fingers are not just equipped with an array of tactile sensors; their geometry is more natural, and wrist articulation is notably more fluid, so ordinary gestures come off as less mechanical and more intuitive. The feet incorporate enhanced balance control, allowing the robot to recover from stumbles and adapt quickly to uneven terrain. The head and neck have been redesigned with a sleeker profile and improved sensor alignment, ensuring that the robot’s eyes can reliably track hand movement during close-in tasks near the torso. Collectively, these upgrades give Optimus Gen 2.5 a level of subtlety in motion and grip that takes it much closer to the unconscious, continuous micro-adjustments humans make.

The Gen 2.5 upgrade is focused more on refinement than radical change. Optimus is still built close to human height (about 5’8”) and weighs near 125 pounds, but now features recalibrated actuators and quieter, smoother brushless motors - an important improvement over Gen 2’s audible gear chatter. Fingertip sensors remain highly sensitive, registering forces in fractions of a newton - essential for tasks that demand both delicacy and the ability to handle robust objects. The power pack is still a chest-mounted array, composed of Tesla’s 4680-type cells and rated at about 2.3 kWh (sufficient for roughly 8 hours of mixed robot work).

Not all the progress shows up in show demos. China’s 2025 restrictions on rare-earth magnet exports have slowed production of some actuators and motors. Elon confirmed that export licensing delays can disrupt manufacturing schedules for weeks at a time.

Meanwhile, the control system is steadily maturing. Tesla reports contact-sensing rates close to 1,000 Hertz, with central motion planning updating at tens or low hundreds of Hertz, resulting in walking that’s visibly steadier with fewer stops when obstacles appear.

For investors and analysts, Optimus is rapidly becoming a hinge for Tesla’s future value. Musk has repeatedly estimated that robots like Optimus could eventually be worth 80% of Tesla’s value and the company continues to target a per-unit price in the $20–30K range. As recent quarters have shown, Tesla’s vision has already moved beyond cars: it’s now a platform company, where energy, automation, and AI-driven robots are key leverage for scale and profit.

Of course, this is still Tesla, which means there is always a counter-chorus. The latest demos drew fire for stiff walking and awkward response times when the robot was asked to do something as simple as fetch a Coke. Critics pointed out the clunks in the gait and the mechanical noise that betrays a machine fighting its own friction. But the skepticism is healthy.

Where does the story go from here? Tesla has hinted that a true third generation is coming and that the company has not shown it yet. During the recent All-in Summit, Elon said they are finalizing the design for Optimus Version 3, and that it will have “essentially the manual dexterity of a human, an AI mind that can navigate and comprehend reality, and it will be made in very high volume.” He emphasized that the biggest remaining challenges are electromechanical problems especially in the hands (actuators, gearboxes, power electronics), and that none of those parts exist off the shelf in the required configuration so Tesla has had to build much of it from first principles. He also said that supply chain and production issues are real constraints -that there is no existing supply chain for many humanoid robot-components, so vertical integration is essential.

The safe bet is that Gen 3 brings stiffer and lighter joints with better power density, denser tactile arrays in the hands, higher-bandwidth perception, and a control stack that treats language as a soft interface for task setup rather than a crutch during execution. It will also ride on AI5/AI6. The better questions are all ugly and specific. How often does it fall? How long does it take to fix? How much power does it draw to move twenty kilograms from here to there and back again?

The cleanest way to read Optimus right now is to say it finally grew teeth. The body is no longer an excuse. The hands can feel. The feet can balance through contact instead of prayer. The head can follow the work. The brain is getting bigger. The training diet is moving from lab tricks to a flood of human footage and the synthetic variations that let a robot try and fail in private. The company is promising deployments inside its own factories. That is how real robots enter the world.

Musk has already tied expectations for Gen 3 to real factory deployment, not stage demos. Thousands of units in Tesla’s own plants are the near-term goal, though delays mean 2026 is the more realistic target for scaled production. By then, Optimus will need to prove that AI6 can run heavier policies at low latency, that its custom actuators can endure long duty cycles without seizing, and that its expanded camera array can handle real industrial lighting without breaking perception. The compute targets discussed publicly are eight times current raw compute, roughly nine times memory, and significantly higher bandwidth, which in practice translates to higher video ingest rates and tighter perception-to-action cycles on the edge.

For the home, 2027 is the earliest horizon. Elon has teased applications like babysitting, cooking, and household chores, but no public demo has shown the robot in a truly uncontrolled domestic space. If Tesla delivers Gen 3 into factories in 2026, the spillover into homes could follow within a year or two, depending on regulatory approvals and reliability metrics.

The future of Optimus is about proving, line by line, that a humanoid can hit industrial benchmarks and keep improving with each fleet update. If that happens in 2026, the story shifts from potential to inevitability. And Tesla will be the clear winner.

Stay tuned for the Part 4: Figure AI vs. Apptronik.

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

Long before anyone had seen silicon or servo motors, there were visions of artificial people brought to life. Ancient myths described Hephaestus forging golden handmaidens. In the Renaissance, Leonardo da Vinci sketched a mechanical knight. By the 18th century, clockmakers in Europe built delicate automata that played violins or wrote poems, not because they solved problems but because they delighted and unsettled. The 20th century carried these fantasies into science fiction, giving us Asimov’s robots with rules, cinematic machines that turned against us, and engineers at Honda and Waseda building real prototypes that walked stiffly across labs.

This history matters because it shows that humanoids have always lived in a liminal space, halfway between spectacle and utility. Asimo, Honda’s white-suited creation from the early 2000s, was one of the most beloved robots of its time. It could climb stairs and wave at schoolchildren, but it never stocked a shelf or delivered a package. Boston Dynamics’ Atlas leapt onto YouTube feeds with parkour routines and backflips, stirring awe and anxiety in equal measure, but it too remained a research showpiece, more gymnast than employee. These machines were incredible feats of engineering but never cracked the code of usefulness.

Now, in 2025, the field has reached a new epoch. Humanoids are being pitched not as mascots of technological progress, but as answers to labor shortages and productivity plateaus. Startups are raising hundreds of millions of dollars on the promise that robots will soon move boxes, screw bolts, and one day help in homes. Factories from Texas to Shenzhen are retooling to mass-produce bodies that look more like us than like forklifts. The humanoid dream has crossed over into commerce.

Unlike traditional robots chained to fenced-off cells, today’s humanoids are being designed to slot into the world as built for people - walking bipedally, manipulating the same tools, sensors registering the physical world, and AI architectures mapping tasks to intent. Proponents highlight how these robots offer operational flexibility and faster adaptation to fluctuating factory demands or high staff turnover. They can now not only cooperate safely alongside human workers but - with recent advances in dexterous manipulation, mobility, and sensory perception - tackle detailed tasks from wiring to components assembly, navigating tight storage aisles, or using human-grade tools with minimal workspace redesign.

In the United States, three names dominate. Figure AI is the poster child of Silicon Valley audacity. Founded in 2022, it has already unveiled humanoid robots powered by Helix, a vision-language-action model, representing a fusion of vision, language understanding, and robotic control into one unified, powerful system. Founder Brett Adcock speaks with unshakable conviction that humanoids will soon fold laundry, unload groceries, and eventually build themselves in factories. Figure’s pitch is that intelligence and form factor, married together, will lead to ubiquity.

Figure is now in formal talks for a $1.5 billion Series C round, which would value the company at approximately $39.5 billion - a 15-fold leap from its $2.6 billion valuation just a year earlier. The company has rapidly accelerated funding, and boasts a manufacturing facility, BotQ, designed to produce 12,000 robots per year so that robots can build robots.

Apptronik, born in Austin, takes a different route. Its Apollo robot is pitched less as a sci-fi savior than as an industrial partner. The company has struck deals with Jabil, envisioning Apollo as a humanoid capable of operating inside real factories. The rhetoric is less about taking over households and more about filling gaps on assembly lines and logistics floors. CEO Jeff Cardenas compares the humanoid race to the space race, casting Apollo as a test of American engineering grit. Where Figure leans into vision, Apptronik leans into pragmatism.

In early 2025 Apptronik raised $350 million through a round led by B Capital and Capital Factory, with participation from Google.

Tesla Optimus is perhaps the boldest bet of all. Elon Musk has gone on record saying that Optimus could account for as much as 80 percent of Tesla’s future value - an astonishing claim that places a robot above cars, energy storage, and even self-driving software. Optimus is supposed to be priced at twenty to thirty thousand dollars, cheap enough that factories and perhaps even households could buy them by the thousand. The vision is to scale production into the millions annually by 2030.

Yet reality bites. Tesla has already faced leadership turnover, with the head of Optimus departing in mid-2025. Export restrictions in China have slowed access to rare-earth magnets essential for production. Musk himself admitted the training pipeline remains in flux, shifting away from motion-capture suits toward vision-only learning.

In May, Musk said Tesla expects thousands of Optimus robots to be working in its factories by the end of 2025, with a long-term goal of one million units per year by 2030. A recently unveiled gold-colored Optimus 2.5 prototype - featuring Grok voice integration - was meant to signify progress, but the on-stage stumble and limited capabilities exposed the project’s fragility. Musk has confirmed that Optimus V3 is currently being designed.

Moreover, Tesla is reportedly lagging behind on its production target of 5,000 Optimus units for 2025, with actual output still in the low hundreds. A lighthearted demo video of Optimus serving popcorn at the Tesla Diner went viral - but even there, critics pointed out mechanical awkwardness and limited autonomy.

Elon Musk has pointed out that the hardest part of Optimus is not walking or balance, but the hands. Building hardware that can match human-level dexterity requires actuators, tendon-like transmissions, and forearm mechanics that can perform a wide range of precise tasks. Because there is no existing supply chain for these components, Tesla has to design motors, gearboxes, and control electronics entirely from scratch. Until that problem is solved, software will remain secondary.

This challenge is not unique to Tesla. Every robotics company working on humanoids runs into the same barrier: developing hands and forearms capable of reliable, repeatable manipulation. Progress in dexterity will determine whether these machines move from promising demonstrations to useful tools in everyday environments.

Against these contenders stands Agility Robotics, maker of Digit, the first humanoid to find real commercial deployment. Amazon has trialed Digit stacking totes in warehouses in Georgia. For that alone, Agility deserves recognition. Yet critics argue that Digit feels clunky, limited to repetitive demos of box moving, more incremental than transformative. In an industry addicted to spectacle, it risks being seen as competent but uninspiring.

On top of the names we mentioned above, E1 Ventures also met Persona AI, a newer team building heavy-duty humanoids for hazardous industrial work. Persona is led by CEO Nic Radford, who previously helped build NASA’s Valkyrie and later founded Nauticus Robotics, and by CTO Jerry Pratt, a longtime IHMC roboticist and former CTO at Figure; the company operates out of the Ion in Houston with a sister footprint in Pensacola, Florida.

Persona’s strategy is to target shipyards and similarly punishing job sites first. In spring 2025 it announced a collaboration with HD Korea Shipbuilding & Offshore Engineering, HD Hyundai Robotics, and Vazil to co-develop humanoids for complex welding and related tasks, with early prototypes expected by late 2026 and initial deployments discussed for 2027.

It is a credible founding team going after a real bottleneck, and the visit left the impression of promise paired with early-stage realities. Timelines are ambitious, technical risk is nontrivial, and the price whispers around the round sizes imply a valuation that feels generous for the company’s current maturity. That mix - experienced talent, a tough industrial niche, and a valuation curve running ahead of milestones - puts Persona in the promising-but-proving category for now.

Across the Pacific, China has made humanoids a national priority. Unitree Robotics, once known for its nimble robot dogs, has entered the humanoid arena with astonishing speed. Its R1 humanoid, priced under six thousand dollars, has made headlines for affordability. The H1 showed off cartwheels on national television during the Spring Festival Gala. The G1 has appeared in Austin wearing cowboy hats, part marketing stunt, part proof of presence. At the 2025 World Humanoid Robot Games in Beijing, Unitree humanoids sprinted 1,500 meters among hundreds of competitors. And yet, behind the spectacle, Unitree faces scrutiny. U.S. officials have raised questions about potential military ties and hidden access channels. Engineers doubt that a humanoid priced at the cost of a used scooter can meaningfully contribute to serious industrial tasks. Still, Unitree embodies China’s strategy: scale first, refine later. Put thousands of robots into the world, gather data, and iterate ruthlessly.

Unitree is now preparing for a Shanghai STAR Market IPO that could value the company as high as 50 billion yuan (about $7 billion). With investors like Alibaba, Tencent, and Geely - and profitability already in hand - the IPO could be one of China’s largest tech listings in 2025.

Yet beneath all this excitement lies the most stubborn truth of the field. The humanoid robot “revolution” is running into an inconvenient truth: real-world dexterity is brutally hard, and data - the fuel that supercharged large language models - is orders of magnitude scarcer in robotics than in text-based AI. As UC Berkeley roboticist Ken Goldberg argues, LLMs flourished by ingesting what amounts to roughly 100,000 years of human-readable internet text, but nothing comparable exists for hands-on motor skills; even attempts to mine videos or simulate manipulation fall short because fine-grained 3D motions, contact dynamics, and friction are painfully difficult to infer or faithfully replicate at scale. This is why seemingly trivial tasks - picking up a wine glass, changing a light bulb - remain unsolved at reliable, generalizable levels, despite headlines promising robots that will soon outshine surgeons, replace factory workers, or moonlight as in-home butlers. The field is now split between a data-maximalist camp betting that more demonstrations and teleoperation logs will close the gap, and a good old-fashioned engineering camp emphasizing physics, math, control, and carefully modeled environments. The pragmatic path forward likely blends both: engineer narrowly useful, robust systems that actually work in production, then use those deployed robots to bootstrap the datasets needed to improve them - much as Waymo’s autonomous fleets and Ambi Robotics’ warehouse pickers iterate through real-world feedback loops. For labor markets, this translates to a counterintuitive forecast: blue-collar trades - plumbers, electricians, skilled construction - are comparatively safe in the medium term because their core is dexterity under uncertainty, while process-heavy white-collar roles such as form intake, standardized customer workflows, and some back-office functions are more exposed to automation by software-first AI. The right move for builders and investors is to reset timelines, avoid hype cycles that risk a backlash, and fund painstaking integration of controls, sensing, and task design that yield commercial footholds today - letting the data follow the deployment, not the other way around.

The opportunity landscape is also now shaped by new demands on manufacturers and component suppliers. Building commercial-ready humanoids means advances in custom actuators, compact thermal-integrated battery housings, and new strategies for mounting dense sensor arrays. The companies that can rapidly iterate, prototype new mechanical designs, and work within tight tolerances - delivering parts for rugged, articulated structures or sensor-packed heads - will be pivotal.

Progress is costly. The “demo” robots of today may cost anywhere from $80,000 to over $300,000 apiece, without integration. Powering dozens of actuators and sensors means battery life is typically under eight hours (usually less in demanding tasks). Safety, too, is more than a slogan; while collaborative intent is built-in, it will take years of phased trials in real settings before safety can be assumed.

The future of humanoids will be shaped by how companies navigate this bottleneck. Some will continue to chase general-purpose fantasies, promising household helpers that never arrive. Others will embrace narrower deployments, succeeding in warehouses, factories, and other constrained environments where utility outweighs novelty. The winners in the United States are most likely Figure, Apptronik, and Tesla. Figure may pioneer the intelligence architecture. Apptronik may deliver the first credible industrial deployments. Tesla, if it can weather internal turbulence and external skepticism, may eventually brute-force scale in a way no competitor can match.

But it is also possible that the first humanoids to achieve mass adoption are not American. Unitree, with its aggressive pricing and state support, could flood the market with tens of thousands of units, building data flywheels that U.S. companies cannot match. That possibility keeps American investors awake.

Chris Camillo, the investor and raconteur, shared this on our recent conversation:

“I don’t think we’re going to see commercial scaling of humanoids to any large degree until 2030. We might start to see some small scaling in 2028, but real scaling probably starts in 2030 for commercial applications. In terms of home use, I think we’re seven to ten years away, maybe even a little longer, before we see meaningful scaling of humanoid bots for consumers.”

Here is our full conversation:

Looking forward, the humanoid industry is likely to unfold less like a sci-fi blockbuster and more like a slow-burn industrial revolution. In the short term, humanoids will remain fragile, expensive, and limited. In the medium term, they will find footholds in warehouses, logistics centers, and controlled factories, collecting the data that will make them better. And in the long term - perhaps a decade away - they may inch closer to the original dream of versatile humanlike helpers. The trajectory will not be linear. It will be fits and starts, hype and backlash, breakthroughs and humiliations.

History offers a useful guide. Automobiles were invented in the late 19th century but remained luxury toys for decades before mass adoption followed Ford’s assembly line. Industrial robots appeared in factories in the 1960s, but it took more than twenty years before they became ubiquitous in automotive plants. Personal computers arrived in the 1970s as hobbyist kits before spreading through offices and homes in the 1990s. Each of these technologies followed a curve: a burst of hype, a long valley of disillusionment, slow adoption in narrow niches, and eventual ubiquity once costs fell and utility proved undeniable.

Humanoid robots are likely to trace the same curve. Today’s demos are the horseless carriages of their era: wondrous, clumsy, and limited. The next few years will be the valley, when failures and frustrations test both investors and founders. But if companies survive long enough to accumulate the data, refine dexterity, and prove reliable in at least one valuable niche, the curve will bend upward. By 2030, humanoids may not be in every home, but in thousands of warehouses and factories, quietly moving the economy forward. The question is not whether humanoids will come, but whether today’s companies can survive long enough to deliver them.

Coming up next: a deep dive into Tesla Optimus (Part 3), followed by a head-to-head comparison of the private heavyweights in humanoid robotics - Figure AI versus Apptronik (Part 4). Be sure to subscribe so you don’t miss it.

Ana

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.

The teams we support are not building simple apps; they are building fusion reactors, humanoid robots, defense autonomy, space infrastructure, and advanced models that redefine what is possible. The stakes are higher, the timelines longer, and the risks more complex. That raises a central question for our craft: not only who to back, but how much to commit.

At E1 Ventures, we have built a methodology to answer that question. It is designed for frontier technology, where uncertainty is enormous but the upside can reshape entire industries or even societies. We wanted more than instinct, more than the adrenaline of a hot deal. We sought a framework rigorous enough for LPs, fair to founders, and flexible enough to adapt to markets that do not yet exist.

We begin with a base check of $250,000. That is the smallest allocation we will put into a company once it passes our filters. From there, we scale based on measurable signals that history, research, and our own datasets show matter most in high-risk, high-impact technology.

The first lever is founder quality. In frontier tech, the founder’s track record is often the most predictive variable. A prior exit adds weight, not because we expect them to repeat the same success, but because it shows they can navigate chaos. A unicorn exit increases conviction further. Peer-reviewed research and granted patents matter enormously: they are not résumé points, but evidence of scientific credibility in domains where reputations are earned. A balanced founding team, pairing a scientist with a seasoned operator, reduces execution risk. Network reach is quantifiable too. Introductions into defense agencies, energy utilities, regulators, or Tier-1 labs create leverage that accelerates adoption and lowers risk.

The next lever is execution. For SaaS, investors look at ARR or churn. In deep tech, we look at proof of feasibility. A propulsion system test with clear performance benchmarks. FDA milestones with documented trial progress. Defense contracts signed with delivery schedules. A DARPA grant or Space Force pilot. These datapoints allow us to increase conviction systematically. When available, unit economics feed into the same logic: LTV-to-CAC ratios, payback periods, or margin profiles. Early dual-use revenues, combining commercial and defense applications, add further validation.

The third layer is context. We analyze markets with the same rigor: size, growth rate, concentration, and white space. A massive market in climate technology or energy transition earns a higher multiplier than a narrow application. Business models are scored too. A high-margin software platform is not weighted the same as a hardware-heavy robotics play with long production cycles. The fundraising stage also matters. At pre-seed, where scientific risk is high, we size conservatively. By Series A, when prototypes are validated and contracts are signed, we scale checks upward. Political and regulatory dynamics feed directly into this model. A company benefiting from the semiconductor incentives, or defense procurement priorities receives a positive adjustment. Exposure to tariffs or regulatory hostility reduces allocation.

Portfolio construction adds the final layer. Our system does not just optimize around a single deal; it optimizes across the fund. We cap exposure at each stage, diversify across subsectors such as autonomy, space, energy, and biotech, and adjust for LP mandates. If we are overweight in defense, our model reduces allocations. If we are underweight in manufacturing or energy transition, it increases. If a company aligns with a sovereign LP’s strategic priorities, that alignment is captured in the model.

The result is a formula: base check plus founder and company adders, multiplied by market, model, stage, political, and regulatory context, then adjusted for portfolio construction. The outcome is not just a number. It is a story supported by data. A story we can share with LPs, explaining exactly why we sized the bet the way we did. A story we can show founders, proving that their work in research, traction, or partnerships translates directly into capital commitment.

Deep tech will never be only about numbers. No spreadsheet can capture the audacity of a team trying to power the planet with fusion, reprogram biology to cure disease, or build a space economy. But in a sector where uncertainty is highest, structure matters most. A disciplined methodology helps us resist bias, commit capital with clarity, and focus on the technologies that push the frontier forward.

At E1 Ventures, our approach blends data with judgment. It gives us clarity when the path is uncertain. It builds trust with our LPs, with our founders, and with ourselves. Because in deep tech, you do not just write checks. You write the future.

Ana

At E1 Ventures, we aim to make this happen. We build channels, share stories, and connect visionaries. We help founders take their ideas out of research labs and into spaces where investors, partners, policymakers, and communities come together. That’s where belief grows and momentum starts.

We understand this because we have experienced it ourselves. Before we became investors, we were founders. We faced tough problems, spent long nights, and felt that no one outside our group could truly understand. We know that progress speeds up when a community of supporters forms around an idea that seems impossible.

This is our role: to make sure deep tech founders are not only creating breakthroughs but also fostering the belief that makes those breakthroughs possible.

The Story as a Catalyst

At the beginning of 2025, we connected with Ashlee Vance, the journalist whose storytelling has captured the rise of Elon Musk and countless other frontier technologists. Early this year he launched Core Memory, a podcast dedicated to telling the stories of founders tackling the hardest problems in technology. We saw in Core Memory something deeply aligned with our mission - and so we chose to support it. Ashlee brings the craft of narrative; we bring belief in the importance of giving deep tech stories a platform.

A few months later, we met Jason Carman, founder of Story Inc., and the creative force behind Saturday Startup Stories (S3). Jason is, in the truest sense, a chronicler of entrepreneurial journeys. His work gives early-stage founders the ability to showcase not just their companies, but their struggles, their philosophy, their “why.” Jason told us about his idea around Lighthouse, a new video series exploring some of the hardest questions in technology and society. Again, we recognized the resonance - and chose to back Lighthouse as a way to bring bold, necessary stories to life.

Education as Community-Building

Why do we care so much about storytelling? Because we believe education is the cornerstone of community. The more people understand what deep tech is and why it matters, the more investors, policymakers, and everyday citizens will rally around it.

This is why we’ve also become regular partners of Deep Tech Week in New York and San Francisco. These gatherings aren’t just conferences - they bring awareness. They create spaces where scientists can meet capital, where investors can feel the energy of frontier innovation, where the public can glimpse futures not yet written. And in the same spirit, we’ve supported hackathons like the USS Hornet Defense Tech Hackathon, where bold ideas are tested, challenged, and accelerated in real time.

Our mission is not just to fund companies. It is to expand the circle of belief. Because belief attracts capital, and capital enables experimentation, and experimentation builds the technologies that redefine the world.

Longevity as a Lens

Last week, both Ashlee and Jason released new content about longevity. Ashlee interviewed Bryan Johnson.

Jason, meanwhile, released the first episode of Lighthouse series about The Secret to Accelerating Longevity.

This wasn’t coordinated, but it was telling. Longevity is not just about living longer; it’s about reshaping the architecture of human possibility. It is a perfect mirror for deep tech itself: audacious, uncertain, riddled with challenges, but profoundly consequential.

For us, longevity captures the essence of our mission as investors. We back companies not because their path is clear, but because their potential to transform human life is immense. And we support platforms like Core Memory and Lighthouse because they invite the world to understand, reflect, and care about those possibilities. By amplifying these stories, we help build the awareness and belief that founders need - and the conviction that LPs seek when they choose to back this space.

The Value E1 Brings

We support the channels, the networks, and the stage that allow bold ideas to resonate far beyond the lab or the whiteboard.

Our role is to create that ecosystem - one where extraordinary people are heard and connected to those who can help turn their ideas into reality.

We also believe in the power of storytelling. Through podcasts, videos, conferences, and every conversation that carries a founder’s vision into the world, we help educate and inspire. In doing so, we aren’t only investors; we are participants in the human project of pushing the frontier forward.

Deep tech is not just about technology - it is about belief. And belief spreads through stories. That is the capital we bring: the possibility of being heard, understood, and supported as the world’s boundaries are being reshaped.

Best,
Ana

After four months embedded in the chaotic heart of the humanoid robotics revolution - touring factories at Figure AI, Persona AI, and several other labs pushing the frontier, meeting teams across the country, and debating late into the night with leading roboticists - we’re ready to pull back the curtain. This series doesn’t just summarize; it kicks off with hands-on reporting, in-depth interviews, and stories from the front lines.

Humanoid Robotics 2025: Hype, Hope, and Hard Truths

Humanoid robotics is everywhere in the media: shiny robots strutting on stage, CEO boasts about labor disruption, billion-dollar deals inked at warp speed. Yet behind the viral clips and VC decks, a far more complicated picture is emerging.

Here’s the reality our research uncovered:

• Most of the promising demos you see - robots folding laundry, running marathons - are, frankly, part show and part promise. Under the hood, autonomy remains partial, with teleoperation and “Wizard of Oz” tactics still common, especially in high-risk or high-stakes environments.

• Mass-market robots for home use? Don't hold your breath. Industry insiders expect years before a safe, affordable, truly useful home humanoid arrives.

• Where the real action is? Industrial, logistics, and arena applications - where humanoids fill labor gaps in dangerous or repetitive settings, and can be rapidly iterated alongside humans. But the question remains: do we truly need robots that look and move like us, or could simpler, specialized robots handle these tasks just as well?
  Recent analysis highlights that specialized robots, built for specific tasks, often outperform humanoids in efficiency and cost-effectiveness, especially for routine or repetitive industrial jobs. Their dedicated design allows for greater precision, speed, and lower operational costs compared to humanoid robots, which may be physically limited or require complex control systems. However, humanoid robots offer the advantage of navigating spaces designed for humans without major modifications - and in environments where flexibility and interacting with people or using common tools is required, their form can make a difference.

• The safety debate is very real. New IEEE frameworks are coming, but the road is bumpy.

But step back, and the progress is staggering. For the first time, robots are performing useful, revenue-generating tasks inside live workplaces - and top investors, from “big auto” to deep-tech venture, are in heavy pursuit. The labor market opportunity - estimated in the tens of trillions - has both startups and giants banking on humanoids as the next industrial revolution.

If you’re at a robot conference or hanging out in a fabrication lab, you’ll sense it: What was once hype is quickly turning into field trials and practical deployments. But the shakeout is coming - there are over a hundred humanoid teams globally, and not all will make it past the first real-world hurdles.

Leading Into the Arena

Amid a surge of innovation and uncertainty in robotics, Ashlee Vance takes us deep into the heart of San Francisco’s “Robot Fight Club,” where tech’s sharpest minds transform engineering into both sport and spectacle. In this subterranean arena, the rules are being rewritten on the fly: startup renegades and hardware hackers pit their remotely-piloted, human-sized robots against each other in full-contact, MMA-style brawls - turning wild science fiction into a reality.

One of the most cutting-edge ventures fueling this movement is REKrobot, led by XR pioneer Cix Liv. Unlike anything you’ve seen before, REK merges virtual reality gaming with the thrill of actual robot combat. It starts in the digital world: players hone their skills by competing in a VR robot fighting league. For the very best, the game becomes reality - they’re invited to remotely pilot genuine humanoid robots and battle them in front of live crowds. Every move in the VR game translates directly to robot actions, blurring the line between gamer and roboticist.

REK acts as both an electrifying entertainment platform and a living laboratory for remote robot operation - making advanced robotics accessible, even irresistible, to a new generation of tech enthusiasts. As the gap narrows between virtual skill and physical mastery, the potential for new benchmarks in competitive robotics becomes clear.

This idea took center stage in E1’s recent conversation with Cix. It zeroed in on the urgent need for standardized benchmarks in robotics: concrete measures that could fairly assess pilot skill, compare robot performance, and track real progress in a rapidly evolving field. By championing these benchmarks, Cix sees an opportunity - not just to level the playing field for competitors, but to accelerate innovation and collaboration across the entire industry.

Now, as Ashlee Vance reveals in his exclusive, San Francisco’s Robot Fight Club is more than a spectacle - it’s a proving ground for the next chapter of robotic sport, skill, and ingenuity.

For the past few months, Cix Liv – real name – has been operating his company REK out of a no-frills warehouse space off Van Ness in San Francisco. The office has a couple of makeshift desks with computers and a bunch of virtual reality headsets on some shelves. More to the point, REK also has four humanoid-style robots hanging from gantries, and they’ve been outfitted with armor, boxing gloves, swords and backstories.

These machines represent the start of a robot fight club for Liv and his small REK team. They’re at the vanguard of a movement taking place in San Francisco to create a new sport in which robots piloted remotely by people will do battle inside of cages. This sport would mix flavors from mixed martial arts, pro wrestling, the tech world and anime, and, in so doing, would intermingle skill and theater in equal doses. “This is going to be the next UFC,” says Liv. “When this guy's walking around and he has full swords, you can feel the pounding in the ground. You know deep in your soul that this thing could kill you. It’s like when you see a lion or something and the hairs go up on the back of your neck. Once people can really feel this and see this, it’ll be fully mainstream.”

We’ve obviously had robot competitions for years. BattleBots started back in the 1990s, giving hardware nerds a chance to show off their cool contraptions. Those bots, though, were mostly ground dwellers and gimmicky. Now, however, we’re amid the rise of humanoid robots being built by lots of start-ups in the U.S. and, more notably, China. The humanoids bring with them the chance to create a combat sport that looks more familiar to mainstream viewers and the opportunity for better storytelling as we anthropomorphize the bots and develop tales for their pilots.

Some early stabs at these robot battle competitions have taken place recently in San Francisco. Two underground, invite-only events have been held in the parking garage of a downtown building as part of the Ultimate Fighting Bots (UFB) league. Pairs of robots squared off against each other, while a couple hundred people cheered them on from the sides. During intermissions at the first fight, humans stepped into the ring to chase each other with tasers while the fighters prepared their bots for the next battle. (Yes, really.)

The people behind UFB are Michael Cho and Xenia and Vitaly Bulatov, who are married. Cho is a longtime entrepreneur who has been working in the robotics field for years and originated the UFB concept. The Bulatovs also work in the robotics realm and have been organizing the events in San Francisco and pumping social media full of clips from their contests.

The scene is reminiscent of the early parts of the movie Big Hero 6 in which people of all types turn up in shady lairs with their battle bots while onlookers imbibe and gamble. Only the backdrop for that movie was San Fransokyo, and San Fransokyo was edgier and had better bots. These first competitions have been somewhere between exciting and farcical with the robots overheating and bumbling around in between their moments of ferocity. Liv has been the burgeoning star of these fights, winning the first competition and then shifting into a sort of Joe Rogan announcer role for the second, and he thinks he has a plan to uplevel the competitions.

LIV IS a large, brawny man who favors muscle tees that promote his biceps. He grew up in the worlds of online gaming and virtual reality. His name is, in fact, a gaming handle, and he made it legal after an identity theft incident. “I called some agency and they said that I could have my credit frozen for the rest of my life or legally change my name,” he says. “They meant the second option as kind of a joke. But, I was like, ‘Fuck it. I’m Cix now.’”

The Liv part is a nod to a company Cix founded in 2016 that let people livestream their virtual reality sessions. The technology developed by the company (LIV) stood as one of the first major efforts to transport the action taking place inside a VR headset out onto a screen for other people to see. Many people consider LIV videos to be the reason that the game Beat Saber took off as a viral success. And VR is now key to REK’s robot combat sport plans.

In REK’s idealized vision, pilots will don VR headsets, slide their arms into combat controllers and enter a virtual fighting cockpit. The pilots will then be able to initialize a series of attacks that are translated by software into movements carried out by the robots. We’re talking full-on punches and slaps and swings, and we’re talking about sword-wielding robots trying to butcher each other as god intended.

The technology required to make all this work is somewhere well beyond daunting. REK has already started training AI models on fighting moves gathered from existing data sets and videos and has been converting that training into maneuvers that can actually be performed by the bots. It’s also built an early prototype of its VR software that gives the pilot a holographic view of a robot’s body, surroundings, health and other performance metrics.

One problem for REK and anyone else that wants to get into this sport is that the current humanoids on the market aren’t really made for fighting. The ones being used most often in the San Francisco fights so far come from Booster Robotics. They’re the size of children and weigh 30kg each. The bots do have some balancing and fighting skills built in but tend to overheat when their actuators are fired in quick succession for, say, a flurry of jabs. To get around this overheating issue, the robots lower their torque automatically, which then lessens the power of the jabs. These robots also can recover pretty well when they’re falling forward but do a bad job of recovering when they’ve been pushed back.¹

REK has focused on using bots made by Unitree Robotics. It has two mid-sized bots and then two larger bots that weigh about 90kg each. These robots have a wide range of motion and can even come with some built-in boxing skills. Still, they’re not general-purpose fighters either and suffer from balancing issues as well. The big ones cost up to $100,000 each, and they’re usually found in university and corporate settings where they’re being used for research rather than as fodder for combat sports training.

Both Booster and Unitree are based in China, which is outpacing the U.S. when it comes to humanoid robots. The key technology on these systems is the actuators that control the movements in the limbs, and China is the actuator capital of the world. When REK wants to make a software change to the bots or order a replacement part, it’s often gated by dealing with overseas engineers and shipping schedules. “It would help if I knew Mandarin,” Liv says.

What REK really has going for it is its vast virtual reality experience. There’s Liv and then the company’s chief technology officer Amanda Watson. Before REK, she worked at Oculus/Meta for more than seven years, spending some of that time sitting alongside John Carmack. Watson led the development of much of the Link technology that made it possible to connect a VR headset to a gaming PC via Wi-Fi, which let players stop being tethered via cables to their computers. Also on the REK team is Nima Zeighami, who has spent more than a decade working on virtual reality, augmented reality, and game engines, and did a stint at Leia building cameras and light-field displays.

Watson has a track record of approaching the latency problems that can plague VR in novel ways, which could be key here. REK must make the pilots’ moves feel natural and real-time if the fights are to resemble true combat sports. “If you know the real issues associated with latency, then you can control for them,” Watson says. “A good programmer can make it look as though things are happening very responsively.”

IN JULY, Liv posted a video in which one of REK’s robots went nuts. The team had sent a full-body command to the robot, but its feet weren’t touching the ground at the time because it hung from a gantry. The robot could not depend on its usual stability mechanisms and compensated by flailing around widely. Liv had genuine fear and panic on his face as he approached the hulking mass and tried to decide how to stop it from wrecking itself and the REK office and its humans.²

This video told me two things. First, it’s obvious that the journey toward a robot fighting league will be an arduous one. The humanoid robots are not yet being mass produced, so they’re expensive and the costs for experimentation – accidental or otherwise – will be high. Add on all the software and virtual reality work that needs doing, and the slog becomes very real. Second, it told me that Liv might just have the media savvy needed to create a robot fighting league. He posted this low moment for REK on purpose, knowing it would go viral and generate interest in the cause.

While Liv has been participating in the Ultimate Fighting Bots events, it’s unclear if he and REK will stay linked to UFB or go off and create their own league. (We’ll have more on the UFB backers soon.) At the moment, Liv is convinced that REK is the only company working on the underlying technology components required to turn robot battles into a popular, spectator sport. “We’re building all this really complicated tech to make the mass consumption of this possible,” he says. “Our method of controlling the robots will be more capable than what anyone else is doing.” In one scenario, REK creates the software and hardware foundation of this new sport and sells its technology to others.

So far, REK has been mostly self-funded. “I’d made some money from a prior start-up and was deciding whether to get a mortgage on a house or have robots,” Liv says. “I chose to do this instead of having a house.”

During my two-day stint hanging out at the REK office, I would see Liv light up again and again when talking about what future fights could look like. His mind goes to patriotic storylines. “You could have Japanese fighters show up with a samurai and battle against a robot outfitted in chainmail from the U.K.,” he says. “Or you could have Tesla’s Optimus versus Unitree. If Elon was taking on China, it would be broadcast to the whole world, and you would literally have state-level engineers trying to ensure their country wins.”

Liv also likes the opportunities this type of sport offers for unconventional competitive pairings. Men versus women. Kids versus adults. The young versus the old. “It would be the first combat sport where you have parity and everyone is competing on a level playing field,” he says.

REK has already been building out the backstories for its bots. One of them is Derek – the robot who doesn’t want to fight because he’s peaceful at heart but has to fight to earn his freedom. He’s the one that spun out of control. “Poor Derek just wants to be free,” Watson jokes. Another machine is named Rambot and has dog tags belonging to John Rambo.

China is clearly ahead on producing the hardware that these bots need but, so far, has not embraced the full-on fighting aspect. Companies there have been doing sedate demos. Liv is convinced that San Francisco and the U.S. will be the robot battling leaders. “China is ahead on production and engineering,” Liv says. “But the U.S. is still the cultural battleground for the world. The winner of the robot fights will probably be an American using Chinese robots – at least for now.”

In its current incarnation, the technology is cool but unpolished. The robots can look goofy at times, and they’re far from executing long series of sophisticated moves. The scene itself, though, is exciting, and there is a feeling of inevitability to all of this. Like, obviously we’re going to have humanoids try and slaughter each other. And obviously people will want to be jacked in controlling these things as the slaughter takes place. If nothing else, it seems better to have bots damaging themselves than humans damaging themselves. “Gen Z doesn’t want to get punched in the face anymore,” Liv says. “Parents don’t want their kids getting concussions. The future of combat sports is robots.”

Of course, the bots will probably have their own opinions on all of this soon enough.

Stay tuned - Part 2 will go deeper into our own field notes, quantitative value chain analysis, and practical lessons for builders and investors in this new age of robots.

Best,
Ana

Views expressed in posts, including articles, podcasts, videos, and social media, are those of the individual E1 Ventures personnel quoted in them and do not necessarily reflect the views of E1 Ventures, LLC or its affiliates. The posts are not directed to any current or prospective investors. They do not constitute an offer to sell or a solicitation of an offer to buy any securities and may not be used or relied upon in evaluating the merits of any investment.

The material available here, as well as on any associated distribution platforms or public E1 Ventures social media accounts and sites, should not be interpreted as investment, legal, tax, or other professional advice. You should consult your own advisers regarding legal, business, tax, and related matters concerning any investment. Any projections, estimates, forecasts, targets, or opinions contained in these materials are subject to change without notice. They may differ or be contrary to opinions expressed by others. Any charts or data included in these materials are for informational purposes only and should not be used as the basis for any investment decision. Certain information contained here has been obtained from third-party sources. While such information is believed to be reliable, E1 Ventures has not independently verified it and makes no representation regarding its accuracy or suitability for any particular situation. Some posts may include third-party advertisements. E1 Ventures has not reviewed such advertisements and does not endorse any of the content contained in them. All posts speak only as of the date indicated.

Nothing contained on Substack or on associated content distribution platforms should be construed as an offer to purchase or sell any security or interest in any pooled investment vehicle sponsored, discussed, or mentioned by E1 Ventures personnel. Nothing here should be interpreted as an offer to provide investment advisory services. Any offer to invest in an E1 Ventures–managed vehicle will be made separately and only by means of the respective fund’s confidential offering documents. Those documents must be reviewed in their entirety and are made available only to individuals or entities that meet specific qualifications under federal securities laws. These investors, which include only qualified purchasers, are considered capable of evaluating the merits and risks of potential investments.

There can be no assurances that E1 Ventures’ investment objectives will be achieved or that its investment strategies will be successful. Any investment in a vehicle managed by E1 Ventures involves significant risk, including the possible loss of the entire amount invested. Any investments or portfolio companies described or referred to here are not representative of all investments made by E1 Ventures, and there can be no assurance that such investments will be profitable or that future investments will achieve comparable results. Past performance of E1 Ventures’ investments, pooled vehicles, or strategies is not necessarily indicative of future results. Some investments, including certain publicly traded assets, may not be listed if the issuer has not granted permission for public disclosure.

E1 Ventures may invest in various tokens or digital assets for its own account. In doing so, E1 Ventures acts in its own financial interest. It does not represent the interests of any other token holder and has no special role in managing or overseeing any such projects. E1 Ventures does not undertake to remain involved in any project beyond its position as an investor.