Welcome to this week’s deep breakdown article. I know I haven’t been super consistent with articles so far, and I just wanted to let you know that from now on, we will have a news summary article every Monday and one deep breakdown every Thursday. Enjoy today’s article!

Most founders completely underestimate how useful a launch can be for a startup. They see companies that go from a basic MVP to 10k MRR literally overnight, and they think that they either:

- got lucky (tons of upvotes on Product Hunt/HN/Reddit/etc.), or

- had a huge audience beforehand (Instagram, LinkedIn, newsletter)

What I’m about to show you is the complete breakdown of how you can launch your startup and gain traction for your product extremely quickly.

Your product

The first step is actually evaluating your product. You’ve probably heard this advice 10,000 times, but your product needs to solve a painful problem.

You don’t need to change your idea; small variations to it are fine.

To give you an example, you’ve probably seen a million AI agents. For this example, we will specifically focus on coding agents.

I’ve seen so many coding agents online, but most of their launches fail pathetically.

On the other hand, I’ve seen companies like Freebuff and Bullet do insanely well. The only difference is that Freebuff has ads so it’s free, and Bullet is designed for max speed. That’s it—barely any change, but by focusing on one painful feature instead of trying to solve a million things, they increase their chances of a viral launch by 100x.

Ultimate launch schedule

I’m about to give you the ultimate schedule for your launch, step by step with detailed instructions. There are some steps in here that not everyone can do; if you can’t do them, just try your best or move on to the next step. However, this protocol is designed so that the important parts require zero existing audience and little luck.

1. Warm up your accounts and your audience

• Most launch platforms require you to have “karma” or a certain amount of days on the platform to even post. As a result, make your accounts ahead of time and spend a little bit of time engaging and contributing.

• If you have any kind of audience whatsoever—this can be 10 followers, 100, 1,000—just hint at them that you’re launching something big and try to suggest that they warm up their accounts on your launch platforms if they want to support you.

2. Test your product

• Before your launch day, test the virality of your product. Put it on a smaller subreddit or a smaller Discord server and just see how people react. If you find any serious issues, then go back to your product and fix them.

3. Build a hub/testing center

• In your product, you want to build some way that people can click on a link and quickly test your product without making an account (you can add a redirect after a little bit of testing to ask them to make an account or download).

• If your product is hardware, then try to make something interactive or a nice landing page with clear visuals.

4. Platform 1 — Product Hunt

• On Product Hunt, people with more “karma” are called top hunters. One upvote from a top hunter is worth 5–20 normal upvotes. You can look for top hunters on the Product Hunt website, then try to find their accounts on X/LinkedIn and send them a friendly DM asking them to check out your launch.
  

• This is the very first platform we are going to start the launch protocol on. At exactly 12:01 PST, you want to submit your startup to Product Hunt. The leaderboard resets every day, so this is the perfect time to maximize your chances. I would also go ahead and add this link to your product’s landing page so people from Hacker News and Reddit will also help on Product Hunt.

5. Your own audience

• If you have any audience (a newsletter, an X account, an Instagram) post on there about the launch. Even with small accounts, you can still help your launch.

6. Platform 2 — Hacker News

• Go onto Hacker News, type in “SHOW HN: your product description in one sentence” in the title bar, and then a description in the description field.

• I would do this immediately after your Product Hunt submission. Also, if your Hacker News post doesn’t take off, I would just hit a different URL on your website and submit again in a few hours (don’t keep spamming though).

7. Platform 3 — Reddit

• Hopefully you’ve gone ahead and scouted out the best subreddits ahead of time for promoting your product. Now go ahead and write and publish your posts (I’m not going to go into detail because launch strategy depends on the subreddit).

8. Platform 4 — Indie Hackers

• Indie Hackers is an article-based platform specifically for tech. Go ahead and publish an article on here. Try to include insights about building instead of just pure promotion.

9. Platform 5 — Community-based

• This is basically any chat-based option, like Discord, Slacks, etc. Essentially, just go ahead and send this out. Again, just like Reddit, be smart about where you send it and what the rules are in that server. For this one, if you continue it throughout the day and don’t get banned, you can truly change the outcome of your entire launch (but this one requires the most hustle).

—————————

Make sure to respond to any comments on your launch, fix any bugs that arise, etc. In general, make sure you have the whole day set aside for your launch.

By the end of the day, you should have gotten some form of traction. The beauty of doing all these platforms at once is they all help speed each other up. For example, someone on Reddit might see your project, upvote you on Reddit, but then go to your landing page and also upvote you on Product Hunt and Indie Hackers, and sign up.

If you want me to go into more detail about any of these platforms, I would love to. Just comment below. Also, I’m thinking about writing my next article about how to market a small startup. Let me know what you would be looking for in an article like that in the comments as well!

P.S. If you want my ultimate launch day checklist, subscribe to my Substack and I will send the link over!

AI changed how people train models, deploy applications, and scale infrastructure.

But now Nvidia might transform how companies build AI systems at scale.

This week, Nvidia announced its new full-stack platform strategy to become the operating system for AI. Rather than selling only processors, Nvidia is focused on software, firmware, data centers, and end-to-end AI platforms, positioning itself as the essential layer between raw compute and enterprise applications.

Nvidia shifts strategy to software and platform services beyond hardware sales

For the past few years, most Nvidia revenue has flowed from GPU sales to hyperscalers. Increasingly, the company is betting that the biggest opportunities lie in owning the entire stack that enterprises need to deploy AI in production.

The biggest AI infrastructure companies of the next decade may not look like pure hardware makers at all.

2. Robotics Just Got a Major AI Upgrade

When most people think about AI applications, they think about language models.

But robotics is increasingly becoming the new frontier.

This week, robotics startup Figure AI released new demonstrations of its humanoid robot performing complex warehouse tasks after integrating advanced AI reasoning systems. The robot can now adapt to unexpected situations and learn new tasks faster than previous generations.

Figure AI demonstrates next generation humanoid robots with advanced AI reasoning

Over the last five years, companies like Boston Dynamics showed what robots could do physically. As AI models improve, startups are shifting their attention toward what robots can actually accomplish when given better decision making systems.

The next generation of robotics companies may compete less on hardware innovation and more on the AI systems that let robots understand and adapt to the real world.

3. The Influencer Economy Just Got Weirder

When most people think about biotech marketing, they think about conferences and research papers.

But influencer culture is increasingly infiltrating science.

This week, reports emerged of biotech founders and researchers leveraging social media to build personal brands while promoting their companies, with some taking on vampire themed aesthetics and personas to stand out in crowded markets. The strategy blurs lines between authentic science communication and celebrity driven hype.

Biotech influencers use provocative personas to capture attention and funding

The first phase of biotech marketing relied on institutional investors and peer review. The next phase may depend on direct audience engagement and personal branding. Social media reach, personality, and narrative building are quickly becoming just as important as actual scientific progress.

I think this is a sign of the impending merge between biotech and tech. (more on that soon in a coming post, subscribe to read it!)

cool tweets

New rounds

Auradine — $110M Series A for energy-efficient AI compute silicon and software platforms.

Castelion — $1B Series C at undisclosed valuation for hypersonic strike missile manufacturing.

Muon Space — $250M Series C for satellite Earth observation and communications constellations.

Network Bio — $50M for disease-specific AI models using human biological data.

Rillet — $100M Series C for AI-native ERP and continuous financial closing.

Zenity — $125M Series C for autonomous AI agent governance and security.

Up until that point the way the military got its tech was different.

1. The military would make some specs and design requirements describing what they want, and assign a certain amount of money they are willing to pay for it

2. A few large incumbent companies would compete for this contract by presenting designs, timeliness, ideas

3. One of these companies would be selected by the military, they would be paid the predefined amount, and would deliver the equipment to the military

This model has many, many issues with it.

1. It’s very expensive for the military, in other markets companies compete to bring prices down, but in defense these companies just competed for contracts

2. These incumbent defense companies can’t attract top talent and aren’t software first (big tech which does have both of these isn’t interested in defense)

This is where Anduril comes in, their business model is completely different. Instead of trying to win a contract, they build things before the government even awards them the contract, taking all the risk on themselves. Once they have made the actual tech, then the military can choose to buy this.

Specifically what anduril builds is a variety of drones, fighter jets, submarines, etc. All of these plug into a software called LatticeOS which is an ai native platform to analyze all this data and coordinate all the hardware.

Their thesis is that as their product catalog increases, latticeOS becomes more and more powerful.

The economics of Anduril

So far Anduril’s Thesis has been performing extremely well, the company has been awarded many contracts, and they’ve been shooting up the setter 30.

In addition, LatticeOS has reportedly been performing extremely well, and on the hardware side Anduril has a massive factory called Arsenal-1 in Ohio.

The thing is, in order to see Anduril’s future we have to look past the current economics, and pay more attention to their acquisitions.

The future of Anduril

I said earlier that before Anduril basically all defense contracts were given to the same 5 companies, the truth is that hasn’t changed that much, while it is easier, the truth is its practically impossible for a tiny startup to get a contract.

But this introduces a paradox. The military needs innovative equipment (which startups. Are great at making, and incumbents aren’t) but they also need that equipment to be reliable and durable (incumbents are good at this, startups are bad).

So we clearly need a bridge, or an interface between these startups and the military. And this is what Anduril is, the strategy that Anduril has taken up, and that they will accelerate in the future has three steps.

1. Identify innovative defense tech startups

2. Acquire those startups

3. Polish and refine the tech, and then use the trust DoD has in Anduril to win a contract

   

Here is a list of examples :

Area-I (April 2021): ALTIUS - Agile-Launched, Tactically Integrated Unmanned System

Copious Imaging (September 2021): WISP - Wide-Area Infrared Sensing with Persistence Sensor

Dive Technologies (February 2022): Dive-LD - Autonomous Underwater Vehicle (AUV)

Adranos Energetics (June 2023): Solid Rocket Motors

Blue Force Technologies (September 2023): Fury - Group 5 Autonomous Air Vehicle

(Credits to this list go to Packy McCormick at not boring)

This business model is insanely good, it creates a beautiful flywheel for Anduril.

Acquire company -> refine tech -> sell it -> gain trust -> make latticeOS more powerful by adding the new tech in -> repeat the cycle with the trust and latticeOS helping the next cycle become even faster and more profitable

Comment down below what your biggest challenge is in starting your own startup/business

Lets help each other!

1. Travis Kalanick Is Betting Big on Industrial AI

AI changed how people write, code, and search.

But now it might transform how the physical world operates.

This week, Travis Kalanick’s startup Atoms announced a $1.7 billion funding round to expand its industrial AI platform. Rather than building another chatbot, Atoms is focused on factories, warehouses, logistics, and manufacturing, using AI to automate the systems that move physical goods around the world.

Atoms raises $1.7 billion to expand its industrial AI platform

The interesting part is not the funding round itself.

For the past few years, most AI investment has flowed into foundation models. Increasingly, investors are betting that the biggest opportunities lie in applying AI to traditional industries that make up the majority of the global economy.

The biggest AI companies of the next decade may not look like software companies at all.

2. The Next Space Race Is About Data

When most people think about the space industry, they think about rockets.

But rockets are increasingly becoming the easy part.

This week, satellite imaging startup Extellis emerged from stealth after raising more than $7 million to develop a new generation of Earth observation satellites capable of capturing images in conditions where traditional optical satellites struggle, including through clouds and at night.

Extellis raises funding to build next-generation Earth observation satellites

Over the last decade, companies like SpaceX dramatically reduced the cost of getting to orbit. As launch becomes cheaper, startups are shifting their attention toward what they can actually do once they’re there.

The next generation of space companies may compete less on launch vehicles and more on the unique data and services they can provide from orbit.

3. The Biggest Bottleneck in AI Is No Longer the GPU

The AI race has largely been viewed as a race to build bigger models and buy more GPUs.

That may no longer be true.

This week, AI infrastructure startup Eliyan raised $145 million at a $1 billion valuation to solve one of the industry’s fastest-growing problems: moving data between AI chips. Modern accelerators have become so powerful that communication between them is increasingly becoming the limiting factor.

Eliyan raises $145 million at $1 billion valuation to ease AI chip data bottlenecks

The interesting part is not the funding round itself.

It’s what it says about where AI infrastructure is heading.

The first phase of the AI boom was about compute. The next phase may be about everything surrounding compute. Networking, memory bandwidth, advanced packaging, and chip interconnects are quickly becoming just as important as raw processing power.

The companies solving these invisible bottlenecks may become some of the biggest winners of the AI era.

Other links

Moonshot AI temporarily paused new Kimi subscriptions after demand exceeded available compute while reportedly preparing for a Hong Kong IPO.

CuspAI raised $450 million and launched the AI Materials Foundry to accelerate the discovery of new materials for industries including semiconductors.

New rounds

Eliyan — $145M Series C at a $1B valuation for AI chip interconnect technology.

ChipAgents — $60M Series A extension to automate semiconductor design with AI agents.

Etched — $300M Series C at a $10.3B valuation to scale AI inference chips.

CuspAI — $450M to build AI systems for materials discovery.

1. Introduction: The Creator Economy Is About Distribution

2. A Brief History of Distribution

3. Manufacturing, Media, Search, and Social

4. The Shift to Creator-Owned Distribution

5. Audience as an Asset Class

6. Trust, Attention, and Compounding Audiences

7. The Unbundling of the Media Company

8. The New Economics of Consumer Businesses

9. Why Distribution Beats Production

10. Creator-Led Brands and Commerce

11. AI and the Next Creator Wave

12. Where Value Will Accrue

13. What We Can Learn

14. Conclusion: Audience Ownership as a New Business Asset

1. Introduction: The Creator Economy Is About Distribution

Every era of business has one scarce resource that decides who wins.

• Industrial age → capital (factories, machines, inventory financing)

• Mid-century → shelf space and broadcast slots

• Early internet → search rank

• Today → attention, captured through distribution

It’s tempting to call the “creator economy” a place where you make content → run ads/subscriptions → make money, But that misses the point.

In 2026, every new b2c startup, app, company, is strongly benefitted by having content.

The unit of value isn’t a video or a post. It’s an audience.

This piece argues:

• The creator economy is a distribution economy, not a content economy

  • its main use is leverage for any other kind of business

  • Audience is becoming an asset class, like real estate

    • it generates money by simply existing (similar to renting)

    • can be sold (like selling real estate)

    • But unlike real estate, it can also be used to launch other businesses and products of your own

• This reframing explains creator brands, the AI wave, and where value goes next

2. A Brief History of Distribution

For most of commercial history, distribution required expensive infrastructure — and expensive infrastructure meant a few gatekeepers controlled access to the public.

  PRODUCER/CREATOR  →  [ GATEKEEPER ]  →  AUDIENCE
                       (owns the pipe)

The pattern, era by era:

| Era | Scarce Resource | Gatekeeper |

| Industrial Capital | Physical distribution | Retailers, distributors |

| Broadcast media | License to broadcast | TV/radio networks |

| Search | Algorithmic rank | Search engines |

| Social | Feed relevance | Social platforms |

Each shift looked like democratization at first. but then it changed once the platform needed to monetize itself.

• Early YouTube/blogging rewarded anyone who showed up

• Then algorithmic feeds replaced chronological ones

• Reach became something won daily from an algorithm, not owned

The lesson is being early to a social media platform that is growing, is very rewarding.

3. Manufacturing, Media, Search, and Social

Same mechanics, different technology, each time:

🏭 Manufacturing era

• Scarce asset: trucks, warehouses, retailer relationships

• P&G/Coca-Cola won on decades of shelf relationships, not unbeatable products

📺 Media era

• Scarce asset: a broadcast license

• 3–4 channels set the national agenda by scheduling alone

• Advertisers paid networks, not creators

🔍 Search era

• Scarce asset: algorithmic ranking

• An entire industry (SEO) emerged just to manage this dependency

📱 Social era

• Scarce asset: feed relevance

• Crack the algorithm → fast reach

• But reach was rented, one algorithm update could erase it overnight

Manufacturing → Media → Search → Social → Creator-Owned
   (capital)   (license) (rank)  (feed)    (audience)

  In every prior era: an INSTITUTION owned the pipe.
  Now, for the first time: the INDIVIDUAL can own the relationship.

4. The Shift to Creator-Owned Distribution

“Own your audience” is a cliché — but the mechanism behind it is genuinely new.

Three things made it possible:

1. Low-cost direct channels

• Email lists, SMS, owned apps/communities

• No algorithm between the creator and 50,000 subscribers

• An email list ≠ a follower count — it’s reachable at will

2. Direct monetization infrastructure

• Substack, Patreon, Shopify, Gumroad, etc.

• Converts audience → revenue without an ad-selling middleman

• Audience becomes a customer base, not just eyeballs to resell

3. Platform-agnostic identity

• A creator’s name/face/voice travels across platforms

• A magazine’s brand dies with the magazine

• A creator’s brand survives the death of any single platform

   OLD:  Platform owns the relationship
         Creator ⇄ [Platform Algorithm] ⇄ Audience

   NEW:  Creator owns the relationship
         Creator ⇄ Audience
              ↑
        Platforms = discovery / marketing only

Bottom line: platforms are becoming billboards. The real relationship lives in the list, the community, or the direct product.

5. Audience as an Asset Class

If audience can be owned, it should be evaluated like any other asset:

• Size (Raw reach):

  • Bad proxy: Follower count.

• Quality (Specificity and engagement):

  • Bad proxy: Broad, generic reach.

  • Good signal: Niche focus paired with high engagement.

• Durability (Surviving platform and algorithm changes):

  • Bad proxy: Feed-dependent following.

  • Good signal: Owned list or paid memberships.

• Yield (Money earned per unit of audience):

  • Bad proxy: Impressions.

  • Good signal: Subscription or product revenue.

• Liquidity (Ability to transfer or extend the audience):

  • Good signal: Ability to move into new products or get acquired.

The better diligence questions:

• How much of this audience is owned vs. rented from a platform?

• Will this audience pay for something beyond free content?

• How concentrated is revenue in one channel, sponsor, or algorithm?

These are the same underwriting questions you’d ask of any cash-flow asset with channel-concentration risk.

6. Trust, Attention, and Compounding Audiences

Audience value doesn’t grow linearly — it compounds, and the compounding mechanism is trust.

  First-time viewer  →  little attention, ~no trust
  2-year subscriber   →  attention + credibility
                          ↓
                    converts to revenue at a
                    far higher rate per person

• Trust compounds like interest: slow at first, then a threshold where the audience will buy almost anything the creator recommends

• The strongest creators sacrifice short-term virality for long-term consistency — that’s a trust investment

Counterintuitive but consistent finding: niching down beats broadening out.

• Broad, generic audience → hard to monetize (no shared specific need)

• Narrow audience, strong shared interest → easy to monetize (creator becomes the definitive source)

Implication: the winners aren’t optimizing for the biggest audience. They’re optimizing for the deepest, most specific trust they can defend.

7. The Unbundling of the Media Company

Traditional media bundled five functions that each used to require serious capital:

 ┌─────────────────────────────────────────────┐
 │           TRADITIONAL MEDIA BUNDLE           │
 │  Production │ Curation │ Distribution │ Ads │ Brand │
 └─────────────────────────────────────────────┘
                        │
                 UNBUNDLES INTO
                        ▼
   Production → phone + editing software
   Curation   → a personal point of view
   Distribution → platform algorithm + owned list
   Ads        → self-serve ad platforms
   Brand      → the creator's own name

• This explains why legacy media struggled despite still making good content: they charged a bundled price for services a single motivated person can now assemble à la carte, cheaper and faster

• It also explains talent leaving legacy institutions — not because the content changes much, but the economics of owned distribution are just better

Second-order effect: successful creators eventually re-bundle — hiring editors, building sales teams, acquiring media properties.

The difference: the bundle now gets rebuilt around a creator’s trust relationship, not around institutional infrastructure.

8. The New Economics of Consumer Businesses

Old model:

  Product  →  Spend heavily on customer acquisition  →  Maybe find customers
              (ads, retail slotting, distribution deals — paid BEFORE revenue)

Creator-led model:

  Trusted audience already exists  →  Product  →  Near-zero CAC for first cohort

This is a structural advantage, not a marginal one:

• Changes which businesses are viable at small scale

• Changes how much capital a launch needs

• Solves the expensive half of the business (trust) before the product half

This is why creator-founded brands have proliferated in beauty, food & beverage, apparel, supplements, and software — not because creators are better product designers, but because they inverted the traditional sequence.

The real risk: over-dependence on one person’s continued relevance.

The fix: durable creator brands invest early in product quality and team/brand infrastructure that can eventually stand somewhat apart from the founder’s daily content output.

9. Why Distribution Beats Production

Production quality is easy to copy. Hire similar talent, license similar gear, eventually match quality.

Distribution is hard to copy. A trusted relationship built over years can’t be purchased — it has to be earned in real time.

      EASY TO COPY                    HARD TO COPY
  ──────────────────           ─────────────────────
   Camera quality                Years of consistency
   Editing polish        VS      Specific audience trust
   Production budget              Track record earned over time

Historical parallel: a dominant newspaper or network didn’t win because it always had the single best journalists — competitors often matched them. It won because it had a distribution relationship (subscriptions, appointment viewing, newsstand placement) built over decades that a new entrant couldn’t buy quickly.

Same principle today, at the individual level: the creator who wins usually isn’t the one with the objectively best content — it’s the one with the most durable, specific distribution relationship.

10. Creator-Led Brands and Commerce

Creator-led commerce is the clearest proof of the distribution-over-production principle.

• Creator launches a product → already has people who know, trust, and want to try it on day one

• Not always the best product on a blind test — but the hardest problem (getting known and trusted) is already solved

Shift in brand marketing budgets:

  Traditional ad spend  ──────►  Creator partnerships
  (rented attention)             (co-owned distribution asset)

• Some brands move beyond one-off sponsorships into equity/profit-sharing deals — treating the creator’s audience as a co-owned asset

• Creators increasingly launch their own branded products rather than promoting someone else’s — if the endorsement is valuable enough to move a third party’s sales, the creator captures more value by owning the product and margin directly

A new supporting layer has emerged to serve this: white-label manufacturers, fulfillment partners, and tools that let a creator with distribution launch fast — without building manufacturing or logistics themselves.

11. AI and the Next Creator Wave

AI gets framed mostly as a production tool — faster writing, faster editing, cheaper voiceovers.

But if distribution (not production) is the scarce resource, AI’s bigger effect is on distribution — pulling in two directions at once:

↑ Increases the value of existing trust

• AI lowers the cost of producing competent content → floods the market with generic-but-competent content

• More noise = trust and specificity become more differentiating, not less

• Established creators with real trust relationships: advantage widens

↓ Threatens the discovery layer

• AI assistants/agents start summarizing, filtering, and deciding on a consumer’s behalf

• Open question: does the AI agent become a new gatekeeper, sitting between creator and audience like search and feeds did before it?

   Commodity info / generic decisions  →  AI likely becomes new gatekeeper
   Trust, taste, identity, community   →  AI reinforces existing owned relationships

Likely answer: both, depending on category.

• Generic informational content → exposed to AI substitution

• Personal trust / relational value → much harder for AI to replace

12. Where Value Will Accrue

• Individual creators / small teams with owned audiences → growing share of attention and spend, especially where trust matters (beauty, wellness, finance, parenting, niche hobbies)

• Infrastructure serving creators → payments, subscriptions, fulfillment, community tools

• Discovery-only platforms → pressure to prove value, since reach via algorithm is rented, not owned

• Media/brand incumbents that don’t unbundle → keep losing ground to smaller, specific, creator-led competitors

• AI-native discovery/agents → the biggest wildcard: new centralized gatekeeper, or accelerant for owned distribution?

13. What We Can Learn

• Build owned distribution before you need it — don’t wait for a platform algorithm change to force the issue

• Optimize for depth of trust over breadth of reach, especially early — specificity is often the defensibility, not a phase to grow out of

• Treat audience as a balance-sheet asset, not an income-statement line — turning down a bad sponsorship to protect trust is often the correct long-term call

• Diversify distribution channels deliberately — single-platform dependence recreates the exact gatekeeper risk this piece describes

• Expect AI to raise the value of authentic relationships — generic content is exposed; trusted, specific relationships are comparatively protected, if actively maintained

14. Conclusion: Audience Ownership as a New Business Asset

 Shelf space → Broadcast license → Search rank → Feed algorithm → Owned audience

Every era’s scarce resource decides who accrues value — regardless of who makes the objectively best product.

Today, that resource is a specific, trusted, addressable relationship between a creator and an audience — and for the first time, the tools exist to own that relationship directly instead of renting it.

What this changes:

• Evaluate creator businesses like subscription/financial assets — durability, yield, concentration risk — not follower count

• New consumer brands succeed on distribution assembled before the product exists

• Media companies rebuild outward from individual trust relationships, not inward from institutional infrastructure

• The open question isn’t whether AI changes production — it will. It’s whether AI discovery becomes a new gatekeeper or strengthens owned, trust-based distribution

Either way, one principle holds:

Distribution, not production, is where value accrues. The audience relationship, not the individual piece of content, is the asset worth building.

For much of the AI boom, the assumption was that the frontier would remain concentrated among a handful of American companies.

That assumption is becoming harder to defend.

This week, Moonshot released Kimi K3 while Alibaba unveiled Qwen 3.8, two large models designed for coding, reasoning, and agentic workflows. While debate continues over exactly how they compare to models from OpenAI, Anthropic, and Google, the gap between Chinese and American AI systems appears to be narrowing.

What makes this particularly important is that many of these models are being released with far more open access than their Western counterparts. Developers and COmpanies are preferring these models since they can be self hosted, and are often much much cheaper than OpenAI/Anthropic.

The result is that AI will likely change the world, but there is a chance that fierce competition drives the cost of intelligence so low, that even though AI is world changing it might not be the best business model.

2. OpenAI Wants to Build the Next Computing Device

The internet had the browser.

Mobile had the smartphone.

AI may require something entirely new.

This week, more details emerged about OpenAI’s upcoming hardware project being developed alongside former Apple designer Jony Ive. Early reports suggest the device will move away from traditional screens and app-based interfaces, instead centering around AI as the primary way users interact with technology.

The significance is not the device itself. It is what it says about where the industry is heading.

Every major computing platform eventually produces new hardware. Microsoft dominated the PC era. Apple dominated the smartphone era. OpenAI appears to be betting that AI will become its own platform, one that cannot simply live inside existing devices forever.

Whether OpenAI succeeds where countless startups have died, is questionable.

3. China Is Building the Space Equivalent of a Data Center

This week, Shanghai Xingshu Tiansuan Space Technology launched the first satellites in a project that ultimately aims to deploy 1,000 computing satellites capable of processing data directly in orbit rather than sending everything back to Earth.

The interesting part is not the satellites themselves, It’s what they represent.

SpaceX spent the last decade proving that space can be industrialized at scale. Starlink transformed satellites from expensive one off projects into mass-produced infrastructure. China’s new computing constellation applies the same logic to computation, treating orbit as a place to deploy computing capacity rather than simply communications hardware.

Many people view the space race as a competition over volume of launches. Increasingly, it looks like a competition over infrastructure. Launch systems, communications networks, manufacturing capacity, and now computing resources are all moving into orbit.

Other links

• alliance to replace anthropic MCP

• Microsoft disentangles from external AI partners

New rounds

We communicate through speech, keyboards, touchscreens, and physical movement. Every command must pass through muscles before it reaches a machine.

But what happens when technology can connect directly to the brain?

That is the vision behind brain-computer interfaces (BCIs). By creating a direct communication channel between the brain and external devices, researchers hope to allow people to control computers, prosthetics, and even entire digital environments using only their thoughts.

What once sounded like science fiction is rapidly becoming reality. Advances in neuroscience, machine learning, microelectronics, and surgical robotics are making it possible to record brain activity with unprecedented precision. Companies and research institutions are now demonstrating systems that allow patients to type, move robotic limbs, and communicate despite severe paralysis.

Yet the road ahead remains challenging.

The human brain is the most complex system known to science. Building a device that can safely interpret neural signals, operate reliably for years, and deliver meaningful benefits to users is one of the hardest engineering problems ever attempted.

To understand where brain-computer interfaces are actually heading, we need to separate hype from reality and examine the science, technology, economics, and limitations shaping the field.

In this deep dive, we’ll cover:

• How brain-computer interfaces evolved from neuroscience experiments into commercial products.

• Why advances in AI are accelerating progress in neural decoding.

• The major technical and regulatory challenges still facing the industry.

• The companies competing to build the next generation of human-computer interaction.

How We Got To Brain-Computer Interfaces

The idea of connecting machines directly to the brain is older than many people realize.

In 1924, German psychiatrist Hans Berger recorded the first human electroencephalogram (EEG), proving that electrical activity inside the brain could be measured.

For decades, neuroscientists studied these signals primarily as a research tool. While they could observe brain activity, they lacked the computing power necessary to decode it into meaningful actions.

During the 1970s and 1980s, researchers began exploring whether neural signals could be used to control external devices. Early experiments demonstrated that both animals and humans could learn to influence computers through brain activity alone.

The term “brain-computer interface” emerged in the 1970s, but practical applications remained limited.

Major progress arrived in the 2000s through projects such as BrainGate, which showed that implanted electrodes could allow paralyzed patients to control cursors and robotic arms.

Today, advances in AI, semiconductor technology, wireless communication, and surgical robotics are enabling a new generation of BCIs that are dramatically more capable than earlier systems.

How Brain-Computer Interfaces Work

A brain-computer interface consists of four major systems:

Neural Sensors

Devices that detect electrical activity in the brain.

These may include:

• EEG headsets placed on the scalp

• Electrocorticography (ECoG) arrays placed on the brain surface

• Implanted electrode arrays inserted directly into brain tissue

Signal Processing

Raw neural signals are extremely noisy.

Signal processing systems filter, clean, and organize the incoming data before it can be interpreted.

AI Decoding Systems

Machine learning models analyze neural activity patterns and attempt to determine what the user intends to do.

Examples include:

• Moving a cursor

• Selecting letters

• Controlling a robotic arm

• Generating speech

Output Systems

Once intentions are decoded, commands are sent to an external device.

Examples include:

• Computers

• Smartphones

• Prosthetic limbs

• Wheelchairs

• Augmented reality systems

Why Brain-Computer Interfaces Are Improving

The biggest reason BCIs are advancing is that several critical technologies are improving simultaneously.

Progress is occurring across neuroscience, machine learning, hardware, and manufacturing.

Just as AI accelerated robotics, it is now accelerating neural interfaces.

Neural Recording

The quality of a BCI depends heavily on the quality of its neural data.

Earlier systems could only capture limited signals with low resolution.

Modern implants can record from hundreds or even thousands of channels simultaneously.

Researchers continue to improve:

• Electrode density

• Signal quality

• Wireless communication

• Long-term implant stability

More data allows AI systems to better understand what users are trying to communicate.

Artificial Intelligence

AI may ultimately be the most important driver of progress.

The brain does not generate simple commands like computer code.

Instead, neural activity appears as complex patterns distributed across millions of neurons.

Machine learning models excel at finding patterns within large datasets.

Recent AI systems have dramatically improved:

• Speech decoding

• Motor intention prediction

• Cursor control

• Text generation from neural activity

Several research groups have demonstrated systems capable of translating brain activity into text and speech with increasing accuracy.

As foundation models improve, neural decoding capabilities may improve alongside them.

Hardware Miniaturization

Brain-computer interfaces benefit from the same semiconductor trends that transformed smartphones.

Modern devices are becoming:

• Smaller

• Lower power

• More wireless

• More computationally capable

Advanced chips allow neural processing to occur directly on-device, reducing latency and improving user experience.

Surgical Robotics

Implanted BCIs require extremely precise placement of electrodes.

Recent advances in robotic surgery are helping reduce risk while improving implantation accuracy.

This is particularly important for next-generation systems that may contain thousands of recording channels.

Compounding Progress

The most important observation is that all of these advances reinforce one another.

Better sensors create more data.

Better AI extracts more information from that data.

Better hardware processes signals more efficiently.

Better surgical techniques make advanced systems more practical.

Progress in one area accelerates progress across the entire ecosystem.

Past Failures

Brain-computer interfaces have experienced multiple hype cycles over the past several decades.

Many early efforts underestimated the challenges involved.

Cyberkinetics

Cyberkinetics helped pioneer implanted BCI systems through the BrainGate project.

While the technology demonstrated impressive capabilities, commercialization proved difficult due to hardware limitations, regulatory hurdles, and the complexity of long-term implantation.

Consumer EEG Startups

Numerous startups attempted to build consumer brain-reading headsets during the 2010s.

Many products struggled because non-invasive sensors could only capture limited information, making it difficult to deliver compelling everyday use cases.

These experiences highlighted a recurring lesson:

Generating impressive demonstrations is much easier than building a scalable product.

Current Successes

A new generation of companies is attempting to solve the challenges that limited previous efforts.

Neuralink

Founded by Elon Musk, Neuralink is developing high-bandwidth implanted brain-computer interfaces.

The company combines custom chips, flexible electrode threads, and robotic implantation systems designed to improve signal quality while reducing surgical complexity.

Recent demonstrations have shown users controlling computers and digital interfaces through implanted devices.

Synchron

Synchron is pursuing a less invasive approach through a device called the Stentrode.

Rather than requiring open-brain surgery, the system is implanted through blood vessels, potentially reducing risk while still enabling neural communication.

The company has focused heavily on assisting patients with severe paralysis.

Precision Neuroscience

Precision Neuroscience is developing high-density neural interfaces designed to record large amounts of brain activity while minimizing invasiveness.

The company aims to create systems that can eventually support communication, control, and neurological applications.

Paradromics

Paradromics is building high-bandwidth neural interfaces focused on restoring communication for individuals with severe neurological conditions.

The company emphasizes data density and neural signal quality as key drivers of future performance.

What We Can Learn

The history of brain-computer interfaces suggests that technical breakthroughs alone are not enough. Many early projects proved that neural interfaces could work, but struggled to become sustainable businesses.

The most successful BCI companies will likely start in healthcare, where the technology delivers immediate value for patients with paralysis, ALS, spinal cord injuries, and other neurological conditions. These markets provide clear demand, strong economic incentives, and a practical path through regulation and reimbursement.

Over time, the biggest opportunities may come from building platforms rather than individual products. Just as smartphones created entire software ecosystems, future BCIs could support communication, computing, prosthetics, and other applications built on top of shared neural hardware and software.

Data may also become a major competitive advantage. Every implanted device generates neural data that can improve decoding models, creating a feedback loop where more users lead to better performance and stronger products.

Ultimately, the winners will probably not be the companies with the most impressive demonstrations, but the ones that can make brain-computer interfaces safe, reliable, affordable, and scalable. The challenge is no longer proving that BCIs are possible. The challenge is turning them into products that millions of people can actually use.

Future Outlook + Timelines

2026–2030

├─ Clinical trials expand.

├─ Communication tools for paralysis improve significantly.

├─ AI dramatically increases neural decoding performance.

├─ Regulatory approvals become a key milestone.

2030–2035

├─ Higher-bandwidth implants enter broader medical use.

├─ Speech restoration systems become increasingly practical.

├─ Hardware becomes smaller, safer, and more reliable.

├─ Competition intensifies among leading BCI companies.

2035–2045

├─ Brain-computer interfaces move beyond healthcare into broader applications.

├─ Neural control of computers becomes more seamless.

├─ AR and wearable computing begin integrating neural inputs.

├─ Long-term implants become increasingly common.

2045+

├─ BCIs become a major computing platform.

├─ Human-computer interaction shifts beyond keyboards and touchscreens.

├─ Neural interfaces merge with AI-powered digital assistants.

├─ The boundary between biological and digital systems becomes increasingly blurred.

Key Forecasts

Morgan Stanley

• Estimates that neurological applications could create a substantial long-term market for brain-computer interfaces as clinical adoption expands.

ARK Invest

• Predicts neural interfaces could become an important platform technology if improvements in bandwidth, safety, and cost continue.

Industry Consensus

• The winners will likely be companies that solve safety, reliability, manufacturing, and regulatory approval—not simply those that demonstrate the most advanced prototypes.

Conclusion

While the exact timeline remains uncertain, the direction is becoming clearer. Brain-computer interfaces are increasingly being viewed not as a niche neuroscience experiment, but as a potential new computing platform. Just as smartphones transformed how humans interact with information, advances in neuroscience, AI, and hardware could fundamentally change how people communicate with machines. The next decade will likely be defined less by futuristic demonstrations and more by the companies that can make brain-computer interfaces safe, reliable, affordable, and useful in the real world.

For most of the AI boom, companies have competed to build better assistants.

This week, xAI launched Grok 4.5, its newest model optimized for coding, agentic workflows, and knowledge work. Rather than positioning it as a general purpose chatbot, xAI is increasingly presenting Grok as infrastructure for software engineering and autonomous digital workers.

The bigger story is that AI companies are beginning to converge on a similar realization: the largest market may not be consumer chat at all.

The most valuable AI systems could be the ones that write software, conduct research, manage workflows, and complete multi step tasks with minimal human supervision. Grok 4.5 was trained alongside Cursor and is being integrated directly into developer workflows, reflecting a broader shift toward AI as a worker rather than simply an assistant.

In addition while many doubt that grok 4.5 is actually mythos class, it does perform similar to opus, for a fraction of the cost.

2. Palantir and NVIDIA Are Building “Sovereign AI”

The first phase of the AI boom was about building larger models.

The next phase may be about deciding where those models live.

This week, Palantir and NVIDIA expanded their efforts around Sovereign AI, a framework that allows governments, defense organizations, and critical industries to deploy advanced AI systems entirely within their own controlled environments. Rather than relying on public cloud providers, organizations can run AI infrastructure on premises while maintaining complete control over data and operations.

This move was likely triggered by various AI companies announcing they will begin making their own chips. This is obviously a huge threat for Nvidia, and we can see their response is the open source nemotron models.

3. China Just Approved the World’s First Invasive Brain Chip

For years, brain-computer interfaces felt permanently stuck in the future.

That future may be arriving faster than expected.

According to a recent analysis highlighted by MIT Technology Review, China has approved what researchers describe as the world’s first invasive implantable brain-computer interface, marking a significant milestone in the commercialization of neural technology. The approval is part of a rapidly expanding Chinese BCI ecosystem that now spans research institutions, clinical trials, and approved medical products.

Most people associate brain chips with science fiction or companies like Neuralink.

The long term significance may be even larger.

Historically, computers have required humans to adapt to machines through keyboards, mice, and touchscreens. Brain-computer interfaces reverse that relationship by allowing machines to adapt directly to human intent.

If the technology continues to improve, the future of computing may involve fewer screens and more direct interaction between minds and machines.

4. Drone Delivery Is Quietly Becoming a Real Business

For over a decade, drone delivery has existed mostly as a futuristic demo.

Now companies are beginning to scale.

This week, Irish startup Manna announced a major expansion into the United States, establishing operations and manufacturing facilities in Tulsa, Oklahoma. The company has already completed hundreds of thousands of deliveries and plans to use Tulsa as the foundation for a broader nationwide rollout.

Much of modern commerce is built around the assumption that moving physical goods takes time. Autonomous delivery networks challenge that assumption by dramatically reducing the cost and speed of local logistics. If successful, they could reshape everything from food delivery and pharmacy services to retail and e-commerce.

Other links

• Biggest breaches of 2026

• slow motion messaging app

New rounds

Every building, product, road, and supply chain ultimately depends on people performing physical work. But what happens when machines become capable of operating in the human world as naturally as we do?

That is the vision behind humanoid robots. By combining advances in artificial intelligence, computer vision, batteries, and mechanical engineering, researchers are creating machines that can walk, manipulate objects, learn new skills, and perform tasks once thought to require a human worker.

This future is beginning to take shape. Major technology companies and startups are investing billions of dollars into humanoid robotics, while recent breakthroughs in AI have dramatically expanded what robots can perceive, understand, and accomplish. Tasks that seemed impossibly complex just a few years ago are now being demonstrated in laboratories and pilot deployments.

But turning demonstrations into useful products is difficult.

The physical world is messy and unpredictable. A robot must safely navigate crowded spaces, handle unfamiliar objects, recover from mistakes, and operate reliably for thousands of hours. Even when the technology works, it must be affordable enough to compete with human labor.

To understand where humanoid robotics is actually heading, we need to separate marketing from reality and examine the technology, economics, and challenges driving the industry.

In this deep dive, we’ll cover:

• How robotics evolved from fixed industrial machines into general-purpose humanoid systems.

• Why advances in AI have sparked a new wave of optimism for humanoid robots.

• The major technical and economic barriers that still need to be overcome.

• The companies competing to bring humanoid robots from research labs into everyday workplaces.

How we got to humanoid robots

The idea of human like machines is far older than modern robotics.

For centuries, inventors imagined and built mechanical automatons capable of mimicking simple human actions.

The modern robotics era began in the 1960s with industrial robots such as the Unimate. These machines transformed manufacturing by performing repetitive tasks with speed and precision. However, they were highly specialized, operating only in controlled environments and following programmed instructions.

Over the following decades, robotics continued to advance, but most robots remained purpose built. Engineers could design machines to weld car frames, move boxes, or assemble electronics, yet creating a robot capable of handling a wide variety of tasks was not possible.

This led researchers toward humanoid robots.

Because homes, factories, warehouses, and tools are all designed around the human body, many believed the most versatile robot would share a human form. A machine with two arms, and two legs could operate in existing environments without requiring major changes to infrastructure.

One of the most significant milestones came in 2000 when Honda introduced ASIMO, a robot capable of walking, climbing stairs, and interacting with people. While its abilities were limited, ASIMO demonstrated that practical humanoid locomotion was possible.

Today, collapsing costs of intelligence, and in the electric stack are making this reality possible.

How Humanoid robots work

A humanoid robot has 4 systems that let it function.

1. Actuators - motors, servoes, devices that convert energy into motion

2. Control and Software - the brain of the robot

3. Power - batteries that supply the rest of the robot

4. Sensors - Lidar, microphones, cameras, anything that lets the robot perceive its environment

Why humanoids will become better

The biggest reason humanoid robots are becoming possible is that every major component of the system is improving at the same time.

A humanoid robot is not powered by a single breakthrough. It is the combination of better sensors, more capable AI models, improved hardware, cheaper computing, and stronger batteries that allows robots to move from controlled demonstrations into real-world environments.

Sensors

A robot can only interact with the world if it can accurately perceive it.

Early robots operated in highly controlled environments because they had limited information about their surroundings. Modern humanoids are equipped with increasingly advanced sensors, including cameras, LiDAR, microphones, force sensors, and inertial measurement units (IMUs).

Computer vision systems allow robots to identify objects, track movement, and understand spatial relationships. Force and touch sensors help robots determine how much pressure to apply when handling objects, while balance sensors allow them to maintain stability while walking or carrying weight

AI

Hardware alone does not make a robot intelligent. The biggest recent shift in robotics has come from advances in artificial intelligence.

Traditional robots were programmed with explicit instructions: move this arm to this position, repeat this action, avoid this obstacle. This worked well in factories but failed in unpredictable environments.

Modern AI allows robots to learn patterns from data. Vision models help them recognize objects, language models allow them to understand instructions, and increasingly sophisticated robotics models help translate goals into physical actions.

The emergence of “world models” could be especially important. Instead of simply reacting to what they see, future robots may build internal representations of their environment, allowing them to predict outcomes, plan ahead, and adapt to new situations.

Cost Curve

For humanoid robots to scale, they need to become cheaper. Early systems will be expensive, but costs could fall as manufacturing improves and production increases.

Key drivers of cost reduction:

• Falling component costs: Motors, sensors, batteries, and AI chips are benefiting from advances in adjacent industries like EVs and consumer electronics.

• Economies of scale: Mass production will reduce manufacturing costs and improve supply chains.

• Better software: AI-driven learning reduces the need for expensive task-specific programming.

If humanoid robots become affordable enough, they could enter a positive feedback loop: lower costs → more adoption → higher production volumes → even lower costs.

Compounding accelaration

The most important change is that these improvements are happening simultaneously.

Better sensors provide more data. Better AI learns from that data. Better control systems turn decisions into movement. Better actuators and batteries allow those movements to happen reliably.

This creates a positive feedback loop where progress in one area accelerates progress in another.

Humanoid robots are improving not because one technology has finally been solved, but because the entire system is advancing together

Past Failures

Humanoid robotics has gone through multiple waves of excitement, but many companies discovered that building a useful humanoid robot was far harder than expected.

SoftBank Robotics

SoftBank Robotics launched Pepper in 2014 as a social robot designed for customer interactions. Despite early hype and deployments in stores, Pepper struggled to find strong commercial demand due to limited capabilities and high costs. Production was eventually scaled back as the market failed to develop.

Honda

Honda spent decades developing ASIMO, one of the most famous humanoid robots ever created. While ASIMO demonstrated impressive mobility, it remained a research project rather than a commercially viable product, with high costs and limited real-world usefulness preventing widespread adoption.

Current succeses

After decades of slow progress, recent advances in AI and robotics have created a new wave of humanoid companies focused on commercial deployment.

Figure AI

Figure AI is developing general-purpose humanoid robots designed for industrial environments. The company has focused on combining advanced AI models with human-like hardware, aiming to deploy robots in warehouses, manufacturing, and other labor-intensive industries. Its partnerships with major companies have helped accelerate real-world testing.

Tesla

Tesla is developing Optimus, a humanoid robot designed to perform general-purpose tasks. Tesla's advantage comes from its expertise in AI, manufacturing, batteries, and large-scale production, which could allow it to produce robots at significantly lower costs if the technology matures.

Future Outlook

Future Outlook + Timelines

2026–2030

├─ Humanoid robots enter factories and warehouses.
├─ AI improves robot learning and task flexibility.
├─ Manufacturing scale becomes the main bottleneck.

2030–2035

├─ Costs fall through mass production.
├─ Robots expand into more industries.
├─ Companies compete on reliability and deployment scale.

2035–2045

├─ Humanoids become common in industrial environments.
├─ Robots take on increasingly complex physical tasks.
├─ Hardware and AI improvements accelerate adoption.

2045+

├─ Humanoid robots become general-purpose labor platforms.
├─ Robots operate across manufacturing, logistics, and potentially homes.
├─ Robotics becomes a core layer of the global economy.

Key Forecasts

Goldman Sachs

• Estimates the humanoid robotics market could reach significant scale as costs decline and adoption increases.

Morgan Stanley

• Predicts millions of humanoid robots could be deployed globally as AI and hardware mature.

Industry Consensus

• The winners will likely be companies that solve cost, reliability, and large-scale deployment — not just those with the most impressive demonstrations.

Conclusion

While the exact timeline remains uncertain, the direction is becoming clearer. Humanoid robots are increasingly being viewed not as a robotics experiment, but as a potential new layer of physical infrastructure. Just as computers transformed information work and automation transformed manufacturing, advances in AI, hardware, and robotics could transform how physical tasks are performed. The next decade will likely be defined less by impressive demonstrations and more by the companies that can make humanoid robots reliable, affordable, and scalable in the real world.

Most AI products today are designed for writing, coding, and general knowledge work. Anthropic is now making a direct push into science.

This week, the company launched Claude Science, a specialized AI workbench designed to help researchers analyze papers, manage complex workflows, and accelerate scientific research. The product is part of Anthropic’s broader effort to bring AI deeper into life sciences and healthcare.

The larger signal is that AI companies are beginning to move beyond “productivity tools” and into domain-specific expertise. The biggest impact of AI may ultimately come not from helping people write emails, but from helping scientists discover new medicines, materials, and technologies faster.

For decades, scientific progress has been constrained by human bandwidth. If AI can meaningfully increase the rate of discovery, it could become one of the most economically important technologies ever built.

2. World Models Could Be the Next Big Leap in AI

Large language models are excellent at predicting text. The next challenge is teaching AI to understand reality itself.

Researchers are increasingly focused on “world models”—AI systems that build internal simulations of how the world works and can predict the consequences of actions before they happen. These models could help AI reason, plan, and interact with physical environments far more effectively than today’s systems.

The significance goes beyond better chatbots. World models could become the foundation for autonomous robots, virtual testing environments, and AI agents capable of making long-term decisions.

Many researchers believe scaling language models alone won’t get us all the way to more general intelligence. If that’s true, world models may represent the beginning of AI’s next era: systems that don’t just describe the world, but understand it.

3. A Startup Is Betting People Want Their Keyboards Back

For over a decade, smartphone innovation has mostly meant bigger screens and more apps.

This week, Clicks showed off the final production version of its Communicator smartphone, a BlackBerry-inspired device that combines a physical keyboard with modern Android apps. The company plans to begin shipping the device later this year.

At first glance, this looks like nostalgia. But it may actually reflect something deeper: growing dissatisfaction with how modern devices compete for our attention.

A new generation of hardware startups is exploring a different question—not how to maximize engagement, but how to help people focus. As smartphones become increasingly commoditized, some of the most interesting innovation may come from rethinking the user experience rather than adding more features.

4. Former DeepMind Researchers Are Turning Poker AI Into Investment Returns

Some of the most valuable AI companies may never become household names.

This week, TechCrunch profiled EquiLibre Technologies, a startup founded by three former DeepMind researchers who previously built AI systems capable of defeating elite human poker players. The company is now applying similar technology to financial markets and has reportedly reached a valuation of roughly $500 million.

The connection makes sense. Both poker and investing require making decisions with incomplete information, adapting to uncertainty, and constantly updating strategies as new data arrives.

The broader trend is that AI is beginning to move from generating content to making decisions. While consumer AI products capture most of the headlines, some of the largest economic opportunities may emerge in finance, logistics, scientific research, and other industries where better decisions directly translate into better outcomes.

Other links

• OpenAI considering giving 5% equity to US government

• An Indian tech billionaire is spending $30M of his own money on an AI alternative to Microsoft Office, showing founders are starting to challenge software categories that once looked untouchable’

New rounds

For over a century, industry has relied on heat, chemicals, and fossil fuels. But what if we could manufacture materials the same way we brew beer?

That is the promise of biomanufacturing. By treating biology as an engineering discipline, scientists are programming cells to produce chemicals, medicines, foods, and materials more efficiently—and sometimes create entirely new ones.

This shift is already underway. McKinsey estimates the bioeconomy could generate up to $4 trillion in annual economic impact, and the rapid scale-up of mRNA vaccines showed what biological manufacturing can achieve.

But scaling biology is hard.

Moving an engineered microbe from a lab bench to a 100,000-liter bioreactor remains one of the toughest challenges in modern engineering. The industry faces major bottlenecks in fermentation capacity, and many highly funded startups have failed to make the economics work.

To understand where biomanufacturing is actually headed, we need to look beyond the hype and examine the science, economics, and infrastructure driving the field.

In this deep dive, we’ll cover:

• How humanity evolved from ancient fermentation to modern synthetic biology.

• Why fermentation is only one piece of the biomanufacturing landscape, alongside pharming and chemoenzymatic processes.

• The lessons of the 2010s biofuels boom and bust.

• The companies turning biological manufacturing into products people can buy today.

How we got to modern biomanufacturing

Long before humans understood the concept of a cell, let alone DNA, we were already running biomanufacturing plants.

Every time an ancient civilization brewed beer, baked leavened bread, or fermented cheese, they were using living organisms as biological factories.

We can consider this the 0th stage of biomanufacturing.

The transition from kitchen craft to industrial science began in the late 19th and early 20th centuries.

• Louis Pasteur demonstrated that fermentation was caused by living microbes, not chemical decay. This shifted the industry from guesswork to controlled microbiology.

• during the 1940s, scientists realized they could use fermentation to produce penicillin and acetone, driven by huge demand, these were the first industrial scale bioreactors (basically tanks filled with bacteria)

  • This design choice still influences modern biomanufacturing (not always in a good way)

This was the 1st phase of biomanufacturing.

The real paradigm shift occurred in 1973, when Herbert Boyer and Stanley Cohen successfully transferred a gene from one bacterium to another, inventing recombinant DNA technology.

In 1978, Genentech used this breakthrough to produce synthetic human insulin. Before this, insulin had to be painstakingly extracted from the pancreas glands of millions of slaughtered pigs and cows, which frequently caused allergic reactions in patients. By turning bacteria into insulin factories, Genentech proved that programmed biology could outscale and outperform traditional chemical and agricultural extraction. Again, using bioreactors to grow and produce the insulin.

We can consider this phase the 2nd phase of biomanufacturing.

We are now entering the 3rd jump, which is very similar to how computing exploded in the 80s and 90s. This phase started roughly 20 years ago, Technologies critical to biomanufacturing will collapse, and the cost curve for starting a biomanufacturing company will rapidly slope downwards.

At the same time, we will shift away from big steel tanks to alternative hardware for manufacturing. In this phase biomanufacturing will explode from a method used to make niche products, to a broad manufacturing method.

How the 3rd generation of biomanufacturing will work

the third generation of biomanufacturing will ultimately have 2 major characteristics.

1. it will be accelerated by crashing costs for biotech

2. it will involve a diversification of the manufacturing hardware away from big tanks

In this section we will look at which technologies are reducing in cost and how it will help, and what will be used in replacement of traditional bioreactors.

cost curves

AI driven design

Lets imagine you are attempting to manufacture chemical A, and so you are trying to find a protein B that will stick to the precursor of Chemical A (we will call this precursor X). historically you would likely try to read literature and studies to guess what proteins might be able to stick to precurosor X and turn it into Chemical A. After picking 100s or 1000s of proteins, you might use very simple protein docking simulations to narrow it down to a few thousand or a few hundred targets. then you would manually test each one.

However with AI driven design, you would simply take all of your targets and then enter their amino acid sequences into an AI model like alphafold3. Then this model will actually take the list of aminoa cids and predict what the proteins 3d structure will be. Other AIxBIO models can predict whether or not your protein will bind to others, what metabolic paths it will be involved in, etc.

This will allow companies to entirely skip large parts of the discover process, IN addition some models can actually design compeltly new biological components, that aren’t present anywhere in nature, this can unlock huge potential for biomanufacturing.

Synthesis costs

DNA synthesis is the core technology that enables biomanufacturing, it contains the actual isntructions for producing a material. Virtually every kind of biomanufacturing requires DNA synthesis at spome stage of development. The cost of synthesis is rapidly collapsing and this will also collapse costs for many biotech companies.

Its important to note however, that gene synthesis is still quite expensive, and out of these 3 cost curve technologies, this is likely the one that will take the longest to actually collapse. (likely via enzymatic DNA synthesis)

(TWST biosciences a popular silicon based DNA synthesis vendors stock YTD)

Lab Automation

today roughly 10% of a biomanufacturing companies costs can be attributed to paying technicians to complete simple, relatively manual work in a lab. Around 30% of a biomanufacturing companies costs is labor costs overall.

Companies like Opentrons are already starting to build systems to automate labs, and these systems are rapidly becoming cheap enough for most companies to afford.

AI agents also have high potential to replace more complex jobs that require more knowledge and experience. Companies like Gingko Biosciences have teamed up with OpenAI, using robots+AI agents to run fully automate dlabs, and they have seen relatively good results.

Moving past fermentation + tanks

Historically in order to biomanufacture, the target chemical that needs to be manufactured has the DNA sequence coding for it, inserted dinto either bacteria or yeast. Then these organism are placed in large steel tanks, where sugar is added and condition are carefully monitored. The microorganism release the target protein, and then machines are used to purify the product.

This works great for making things like insulin, but for biomanufacturing to become a major source of materials, it needs to branch out. Below are a few of the most promising, novel biomanufacturing methods.

Cell Free biomanufacturing

Cell free biomanufacturing removes the cell entirely, in a cell generally the DNA → RNA → target protein that’s being manufactured. What cell free biomanufacturing does is it removes anything that isn’t involve dint hsi process such a s the cell membrane. This purification makes cell free systems expensive, but this cost will come down, in addition cell free biomanufacturing does not require a lab to run, just DNA and the cell free powder, which allows for on point manufacturing (don’t ened a factiory, can be made where needed with a tube).

Chemoenzymatic cascades

Currently, biomanufacturing is really only used to make biological chemicals. things like insulin, drugs, etc. Traditional, industrial, non biological chemicals are usually made using petrochemical processes. These involve using precursors along with huge amounts of heat, pressure, and often toxic chemicals to manufacture.

What chemoenzymatic procceser due is simple, the steps of making a chemical that just involve a simple, clean, chemical reaction are kept. But steps that involve huge amounts of heat, or expensive and toxic chemicals, are replaced with enzymes (proteins that can catalyze reactions).

Pharming

traditionally the organisms used for producing the target chemical/material are microorganism like bacteria and yeast. However the issue with these is they require expensive bioreactors, align with feedstock and nutrients. An alternative to make this cheaper, is instead inserting the same genes for the target chemical into a Plant or Animal. To give you an example, a certain drug is manufactured using goats, the DNA is inserted into the goats genome (harmless, does not hurt the goat in any way), this results in the target chemical being produced in the goats milk and then can be purified.

Carbon Waste

another change in technology could be changing the actual source of energy and materials. In a bacterial or yeast absed biomanufacturing system, sugar + nutrients are usually used to feed the organism, with carbon feedstock system, carbon is used instead, often with organisms like algae. This way the manufacturing system is carbon negative and in some cases cheaper.

Past Failures

Amyris

Amyris was founded in 2003, early on they attempted to use genetically modified bacteria to produce precursors for anti-malarial drugs. This business was successful, and as a result they went public on NASDAQ, this si when they made their biggest mistake. They chose to attempt to produce renewable diesel and jet fuel, however scaling up bioreactors is not easy, humans have spent 100s of years figuring out how to get oil out of the ground as cheaply as possible. Afterwards Amyris tried to make consumer products such as Squalene (derived from sharks, used in cosmetics), however the company never managed to actually live up too its promises, and went bankrupt and then restructred.

Zymergen

Zymergen was founded in 2013, early on they attempted to use machine learning and robotics to automate the discovery of new microbes for industrial materials. This business pitch was wildly successful at attracting venture capital, and as a result they went public on NASDAQ with a multi-billion dollar valuation, this is when they made their biggest mistake. They chose to attempt to produce a bio-engineered optical film called Hyaline for flexible smartphone screens, however scaling up bioreactors is not easy, and convincing electronics giants to change their hardware supply chains is even harder. Just four months after their IPO, Zymergen had to admit the material faced severe technical scaling issues and would make zero revenue, causing their stock to crash 70% in one day. The company never managed to actually live up to its promises, and was eventually bought out by Ginkgo Bioworks for pennies on the dollar.

Lessons

• bioreactors + GMO bacteria are only suitable for producing high value, smaller scale commodities (drugs)

• In order to scale into other types of materials and chemicals, we must use alternative biomanufacturing methods

• Stay away from synthetic Biology companies, unless the unit economics are strong at a small scale, this way if the technology fails to scale, the company can still be profitable.

Current succeses

Solugen

Solugen was founded in 2016, early on they used an engineered enzyme to produce bio-based hydrogen peroxide for wipes. This business was successful, and they raised hundreds of millions of dollars, which is when they made their smartest strategic choice. Instead of using living microbes in giant fermentation tanks, they combined pure enzymes with metal catalysts in a cell-free system called the Bioforge. This approach bypassed the scaling bottlenecks that broke previous biotech startups, allowing them to make chemicals with zero waste. Afterwards, Solugen expanded to a massive plant in Minnesota, hit a billion-dollar valuation, and proved that biology can actually compete with fossil fuels. (as long as you use the right method)

Pivot Bio

Pivot Bio was founded in 2011, early on they used gene-editing to create microbes that replace chemical nitrogen fertilizer for corn crops. This business was successful, and they deployed their product across millions of acres of American farmland, which is when they made their smartest strategic choice. Instead of trying to manufacture bulk chemicals inside a factory, they engineered microbes that grow directly on plant roots in the soil, using the crop itself as the environment. This approach bypassed the massive capital costs of building traditional chemical plants, allowing them to scale through existing seed distribution networks. Afterwards, Pivot Bio raised hundreds of millions of dollars, hit a multi-billion dollar valuation, and proved biological farming can displace fossil-fuel fertilizers.

Future outlook + timelines

2026–2028

│

├─ Manufacturing becomes biotech’s primary bottleneck.

├─ Investors shift capital toward fermentation, scale-up, and infrastructure.

├─ AI-driven strain engineering becomes standard.

└─ a16z argues many products remain uneconomic below ~$10/kg production costs.

2028–2032

│

├─ Biology foundries and contract manufacturing platforms expand.

├─ Startups increasingly adopt a “fabless biotech” model.

├─ New bioreactor designs reduce scale-up costs.

└─ Infrastructure investment outpaces investment in individual products.

2032–2037

│

├─ Specialty chemicals become economically competitive.

├─ Biomaterials and industrial enzymes scale into major markets.

├─ Cost and manufacturing efficiency become the key competitive advantage.

└─ Biology is increasingly viewed as an industrial platform, not just healthcare.

2037–2045

│

├─ Biological production competes with portions of petrochemical manufacturing.

├─ Large-scale biofactories become critical industrial assets.

├─ Governments begin treating biomanufacturing as strategic infrastructure.

└─ Automation dramatically lowers development and production costs.

2045+

│

├─ Biology becomes a general-purpose manufacturing technology.

├─ Large fractions of chemicals, materials, and food ingredients are biologically produced.

├─ Factory deployment becomes more important than organism engineering.

└─ The bioeconomy increasingly resembles traditional industrial manufacturing.

──────────────────────────

KEY FORECASTS

────────────────────────────

McKinsey (2020)

• Biological applications could create $2–4 trillion in annual economic impact over the next 10–20 years.

Andreessen Horowitz

• Current fermentation economics struggle below ~$10/kg products.

• To account for scale-up risk, products ideally exceed ~$100/kg.

• The largest opportunity in synthetic biology is reducing manufacturing costs rather than improving genetic engineering.

McKinsey Future Arenas (2024)

• Industrial and consumer biotechnology could become a $340–900 billion annual industry by 2040.

Investor Consensus

• The next decade’s winners are likely to be companies that lower the cost of biomanufacturing rather than companies that simply create new organisms.

Conclusion

While the exact timeline remains uncertain, the direction is clear. Investors increasingly view biomanufacturing not as a niche biotechnology sector, but as the foundation of a future industrial platform. Just as advances in semiconductors enabled the software revolution, advances in fermentation, automation, and biological manufacturing infrastructure may enable a new generation of products made with biology. The next decade will likely be defined less by breakthrough organisms and more by the factories, tools, and infrastructure that make biological production economically competitive at scale.

As AI models become larger and more powerful, one challenge continues to grow alongside them: energy consumption.

This week, former Databricks AI executive Naveen Rao unveiled a new startup focused on radically improving AI efficiency. The company believes that current AI systems waste enormous amounts of computational power and that new hardware architectures could reduce energy costs by as much as 1,000×.

The timing is important. Training and running frontier AI models increasingly requires massive data centers, specialized chips, and huge amounts of electricity. While most discussions around AI focus on model capabilities, the long-term scalability of AI may ultimately depend on making intelligence dramatically cheaper to compute.

If successful, breakthroughs in AI hardware could prove just as important as breakthroughs in AI models themselves.

2. Patronus AI Raises $50 Million to Build Digital Worlds for AI

One of the biggest challenges in AI today isn’t building agents—it’s reliably testing them.

This week, Patronus AI announced a $50 million funding round to build large-scale simulation environments designed to stress-test AI systems before they are deployed in the real world.

The idea is straightforward: create realistic digital environments where AI agents can encounter difficult situations, unexpected edge cases, and adversarial conditions thousands of times before interacting with actual users.

The approach mirrors how aerospace companies use simulators before flying aircraft and how self-driving companies test vehicles before putting them on public roads.

As AI agents become increasingly responsible for tasks involving research, software development, customer support, and business operations, the ability to rigorously evaluate their behavior may become one of the most important parts of the AI stack.

3. NewLimit Raises $435 Million to Push Longevity Treatments Toward Human Trials

Longevity startup NewLimit announced a massive $435 million Series C this week, one of the largest financings in the anti-aging sector in recent years.

Founded by Coinbase CEO Brian Armstrong and other biotech entrepreneurs, the company is focused on epigenetic reprogramming—an approach that aims to restore cells to a more youthful state by modifying gene expression patterns associated with aging.

The new funding will help move NewLimit’s lead compounds into clinical trials targeting fatty liver disease, marking a major transition from research into potential therapeutic applications.

The broader trend is hard to ignore. After years of being viewed as a fringe area of biotech, longevity research is increasingly attracting serious capital, top scientific talent, and clinical development programs. Investors are betting that understanding and modifying the biological mechanisms of aging could become one of the largest opportunities in medicine over the coming decades.

Other Links

• GPT 5.6

I publish daily news articles on the site (not emailed, so no spam), and send out a weekly 5+ page deep dive via email. Subscribe if you want updates for both.

This week, Stripe, Anthropic, the OpenAI Foundation, Bill Gates, and several other backers launched Intercept, a new nonprofit focused on preventing respiratory infections like the common cold and flu.

The organization plans to fund broad spectrum vaccines, antiviral technologies, computational protein design, and even building air purification systems. One particularly interesting idea involves engineered proteins that could be sprayed into the nose to capture viruses before they infect cells.

The larger story is that some scientific problems remain unsolved not because they’re impossible, but because they lack strong commercial incentives. Intercept is betting that modern biotechnology has finally advanced far enough that eliminating a large fraction of respiratory infections is worth pursuing.

2. IBM Unveils a Sub Nanometer Chip Architecture

This week, IBM unveiled a new sub nanometer chip architecture, pushing semiconductor manufacturing closer to the physical limits of what can be built.

For decades, better computing came from making transistors smaller. Today, that approach is becoming increasingly difficult as engineers run into challenges involving heat, power consumption, and even the behavior of individual atoms.

The announcement highlights an important trend in AI, future breakthroughs may depend just as much on advances in hardware as advances in software. The race for AI leadership is often discussed as a competition between models, but increasingly it is becoming a competition between entire infrastructure ecosystems.

3. A $2.3 Billion Bet That Video Games Can Train Real World AI

General Intuition emerged from stealth this week with more than $2 billion in backing and a surprisingly simple idea. Virtual worlds may be the ideal training ground for future AI systems.

The company believes that realistic simulations can provide nearly unlimited opportunities for agents to learn planning, adaptation, and decision making. Unlike the real world, virtual environments allow millions of experiments to be run quickly, cheaply, and safely.

The concept is similar to how self driving companies rely on simulation before testing on real roads. If successful, General Intuition could help shift AI training away from static datasets and toward interactive environments where models learn through experience rather than observation alone.

4. NVIDIA Wants to Become the Operating System for Robots

This week, NVIDIA unveiled Halos, a safety platform designed for humanoid robots and other autonomous machines. Rather than building robots itself, NVIDIA is building the software, simulation tools, and safety systems that could power an entire generation of them.

The platform combines AI models, simulation environments, and testing frameworks to help robots learn and operate more safely in real world environments. As companies race to deploy humanoid robots in factories, warehouses, and eventually homes, ensuring that these systems behave predictably is becoming just as important as making them more capable.

The bigger story is NVIDIA’s strategy. Just as the company became the foundation of the AI boom by supplying the hardware and software behind modern AI models, it now appears to be positioning itself as the infrastructure layer for physical AI. If humanoid robots become widespread over the next decade, NVIDIA doesn’t necessarily need to build the winning robot, it just needs to provide the tools that everyone else uses, just like they are doing with GPUs right now.

Other links :

Eli Lilly is building an "App Store for biotech" where companies can share data and train AI models for drug discovery.

Researchers launched PaperTok, a TikTok-style feed for scientific papers aimed at fighting AI-generated content overload in research.

I publish daily news articles on the site (not emailed, so no spam), and send out a weekly 5+ page deep dive via email. Subscribe if you want updates for both.

This week, Y Combinator-backed startup Dispatch publicly revealed its plans to build the infrastructure layer for space manufacturing : permanent uncrewed factories in orbit paired with reusable vehicles that bring products back to Earth.

Similar manufacturing systems have already been demonstrated on the ISS, but costs remain prohibitively high.

The microgravity environment is perfect for manufacturing certain pharmaceuticals, semiconductors, biorpinting tissues, etc.

The company plans to launch mid 2027.

(image of them testing their satellite prototype)

2. Anthropic Launches Claude tag

This week, Anthropic launched Claude Tag, allowing users to bring Claude directly into Slack conversations as a persistent AI teammate rather than a separate chatbot.

The product itself is less interesting than what it signals. For the past two years, dozens of startups have been raising money to build AI coworkers.

This reflects a broader trend in venture capital: investors are becoming increasingly skeptical of AI wrappers and increasingly interested in companies with proprietary data, unique workflows, or deep domain expertise. If OpenAI, Anthropic, or Google can add your core feature in a product update, your moat may not be strong enough.

As frontier labs continue moving up the stack, founders may need to focus less on generic AI assistants and more on solving specific industry problems that large model providers are unlikely to tackle directly.

3. GLM 5.2 Pushes Close to the Frontier at Fractional Cost

This week, Chinese AI lab Z.ai released GLM-5.2, a frontier-scale open-weight model with a one-million-token context window under an MIT-style license, making it freely downloadable and modifiable.

The model reportedly matches or exceeds GPT-5.5-level systems on several long-horizon coding and agentic benchmarks while operating at significantly lower inference cost. Its efficiency comes from a mixture-of-experts architecture, where only a small subset of its ~hundreds of billions of parameters activate per token, keeping compute requirements closer to a much smaller model.

It also uses sparse attention techniques that allow extremely long context windows without linear cost explosion, making it useful for large codebases, research documents, and multi-step agent workflows. Training was additionally optimized around agentic tasks, where the model iteratively writes, tests, and debugs its own outputs.

The broader signal is less about one model and more about trajectory: frontier-level performance is increasingly appearing in open systems, compressing the gap between closed proprietary models and widely available alternatives.

(Z.ai stock YTD on hong Kong stock exchange)

4. Colossal Biosciences valued at 10 billion

Colossal Biosciences is continuing to scale its de-extinction platform, with recent funding pushing its valuation above the multi-billion-dollar range as it expands efforts to revive extinct species like the woolly mammoth and other Pleistocene-era animals.

The company is combining ancient DNA reconstruction, CRISPR-based genome editing, and synthetic biology techniques to engineer proxy species that resemble extinct organisms rather than perfectly restoring them. Initial milestones focus on gene edits in elephant cells and mammoth-like traits such as cold tolerance and altered fat distribution.

While full de-extinction remains far from reality, the bigger shift is that investors are now treating synthetic conservation as a legitimate biotech category rather than speculative science. This places Colossal at the intersection of conservation biology, synthetic genomics, and platform biotech.

The broader implication is that biology startups are increasingly moving from treating disease to actively redesigning ecosystems and species themselves, a shift that could redefine what “biotech” even means over the next decade.

(deep dive on colossal and de-extinction tech coming soon - subscribe to stay updated)

(de-extincted dire wolf)

Other links

agent loops - blog

meta is working on a prediction market - NYT

when will Sonnet 5 release - WaveSpeedAI

Many people start by looking for markets.

A better approach is often to start with a problem.

The strongest startup ideas tend to come from problems that:

• You have personally experienced.

• You genuinely care about solving.

• You can spend years learning about it without losing interest.

A useful test:

Would I still enjoy learning about this
if there were no company attached to it?

If the answer is yes, you’re probably looking in the right place.

Experience Before Entrepreneurship

Before building a company, build expertise.

Examples:

• AI startup → learn AI engineering.

• Biotech startup → work in a lab or learn computational biology.

• Robotics startup → build robots.

Many founders underestimate how much domain knowledge compounds over time.

Experience
    ↓
Insight
    ↓
Better Ideas
    ↓
Better Products

Business skills can be learned along the way.

Building vs Networking

Networking is important, but building comes first.

A common mistake is spending months networking before creating anything of value.

A better cycle:

Build
  ↓
Reach Out
  ↓
Build
  ↓
Reach Out

As progress accumulates:

Build
  ↓
Users
  ↓
Revenue
  ↓
Funding
  ↓
Scale

People become dramatically more willing to help once they can see evidence that you’re serious.

Progress is often the strongest networking tool.

Reaching Out

One recurring lesson throughout the book is that many opportunities come from simply asking.

Potential mentors, investors, customers, and experts are often more accessible than people assume.

Most opportunities are not missed because people were rejected.

They are missed because people never reached out in the first place.

The downside of sending an email is small.

The upside can be enormous.

Fundraising

Fundraising is easy to think about as a pass/fail test.

In reality, it is often a feedback loop.

When investors say no:

• Look for recurring concerns.

• Identify patterns.

• Improve the company.

• Return with stronger evidence.

Pitch
  ↓
Feedback
  ↓
Improve
  ↓
Pitch Again

Rejection is frequently data.

The most useful investor meeting is not always the one that results in a check.

Bootstrapping

Not every company needs venture capital.

For some businesses, revenue can fund growth.

Venture capital accelerates growth, but it also introduces:

• Dilution

• Expectations

• Pressure

• Additional complexity

Funding should be viewed as a tool rather than a milestone.

The goal is building a successful company.

Raising money is only one possible path.

Hiring

Hiring becomes increasingly important as companies grow.

A few recurring themes:

• Hire slowly.

• Look for mission alignment.

• Skills matter, but motivation matters too.

• A bad hire can consume enormous time and energy.

Great Hire
     ↓
Better Team
     ↓
Better Execution
     ↓
Company Growth

Team quality compounds over time just like technical progress.

Difficult Decisions

Many aspiring founders imagine that problems disappear as a company becomes successful.

The opposite is often true.

Growth introduces new challenges:

• Team conflict

• Product pivots

• Missed milestones

• Layoffs

• Firing employees

Core Themes From the Book

Pick a Real Problem
          ↓
Build Expertise
          ↓
Create Something Valuable
          ↓
Reach Out
          ↓
Gather Feedback
          ↓
Improve and raise funding
          ↓
Scale

A lot of my thinking around startups comes from building StartupWiki, a startup research platform focused on finding and understanding companies earlier.

Visit StartupWiki.tech to learn more.

Deep tech companies often spend years in R&D before making any money, so how do you separate the ones that will make money from the ones that won’t, before they even start generating revenue?

For a SaaS company, the main question is, “will anyone pay for this?” But for deep tech, the very first question to ask is, “Will this even work?

In order to evaluate this, you need to look for a genuine technical breakthrough.

Some good ways to evaluate this :

• Product is based off of peer-reviewed research → verifies the tech is possible, at least on a small scale

• Patents → Signal that the company has made meaningful progress towards their goal, also serves as a signal for deep tech companies still in stealth

• Proprietary data → creates a compounding advantage against other companies

Even though understanding and researching this can be difficult, it’s very important for a deep tech company.

The next factor is something that can be evaluated even if you don’t have domain expertise ; the founder.

Look for founders who have previous domain experience, and have built things before. One of the most important things to do when meeting a deep tech founder is to make sure they understand what they are making.

For example, during Theranos’s development, the company was inconsistent about what their tech was, starting off by saying it was a robot arm in a box, then switching to saying it was microfluidics.

Take 20 minutes beforehand to research and write thoughtful questions about how the tech works, and watch as the founder answers to see if they know what they’re doing.

Once you’ve evaluated these 2 things, you can move on to the next factor. Is this worth investing in? It’s one thing to verify that the product is possible, but is this the right time to build it?

AI exploded because the cost of GPUs went down. Space startups benefited from decreasing launch costs, and biotech startups benefit from low sequencing and synthesis costs. You want to invest right as the cost curve starts to make the technology economically viable.

Basically, make sure the category the company is working in is growing, and has a lot of future potential to become cheaper. This ensures the company is being founded at the right time.

The very last thing to look at is, will anyone buy it? For example, Nuclear Fusion → Huge energy market. Make sure that the product actually has demand.

A bonus for the company is if the future of the entire company isn’t resting on 1 technical breakthrough. For example, years ago, SpaceX was promising Mars colonization now, but they aren’t even close today. However, even though they didn’t achieve that, they did achieve the milestone of collapsing launch costs and getting revenue from Starlink, which is why their early investors made huge returns.

If you want to actually use this framework on real companies, I built a simple database of startups structured around these same filters:

StartupWiki (https://StartupWiki.tech)