Leopold Aschenbrenner’s Situational Awareness fund just gave us a 2026 preview: a long AI-infrastructure, short-software book, run at roughly 4x leverage, that went from $45 billion to a Citadel fire sale in days when chip stocks lost over a trillion dollars in value. The first data point in a much bigger one?

Here’s what the leverage actually looks like once you zoom out from just one Wunderkind fund.

#1 Wall Street Consensus Bet

Situational Awareness, despite its name, was running the trade. Goldman's prime brokerage data shows hedge funds entered the second quarter of 2026 with a record 10% weight in semiconductors, the highest on record, while software weight fell to 6%, the lowest since 2019. Software shorts hit their highest level since Goldman started tracking the data in 2016. Hedge funds made roughly $24 billion collectively shorting software this year even as the sector lost nearly $1 trillion in value. Point72 and Whale Rock were running versions of the same book. This is the same shape as 1998: LTCM didn't blow up because one fund was leveraged, it blew up because half of Wall Street was running the same convergence bet, so when it broke, there was no uncorrelated buyer left standing on the other side.

#2 Circular Finance

Then there’s the thing with circular financing within the AI Money Machine. Nvidia invests in OpenAI and CoreWeave while selling both of them the chips that make the investment thesis work, is reportedly discussing guarantees on OpenAI purchases worth hundreds of billions, and remains CoreWeave’s buyer of last resort for unsold capacity through 2032. Each company’s revenue depends partly on the other’s continued spending. That’s not the same risk as one hedge fund’s leverage, it’s closer to LTCM’s actual problem: interconnection. If one node’s assumptions don’t hold, the stress doesn’t stay put.

Add it up and AI-linked investment-grade debt now sits at roughly $1.4 trillion, about 15% of the entire US credit market. AI-linked stocks make up something like 45% of S&P 500 market cap, the highest single-sector concentration since just before the 1929 crash, per Deutsche Bank research. Aschenbrenner's fund was one leveraged bet sitting inside a system that is itself running startlingly leveraged.

#3 Infrastructure Buildout

What is actually being funded is the largest infrastructure buildout in history, and it’s already showing up on Big Tech’s balance sheets. Alphabet’s free cash flow turned negative in the second quarter of 2026, first time since its 2004 IPO. Meta’s collapsed 91% to $784 million as AI spend hit $31 billion in a single quarter. Amazon is projected negative $17 to $28 billion for the year. Investment-grade bond issuance backing AI infrastructure hit roughly $218 billion through early July 2026, already blowing past all of 2025’s $80.5 billion, with Morgan Stanley projecting $250 to $300 billion more from hyperscalers this year alone. Oracle is carrying close to $96 billion in debt after an $18 billion bond and a $38 billion loan, and some of that Oracle-backed data center paper has already needed to offer investors higher yields to move, an early stress signal. Meta closed the largest private credit data center deal in history in 2025, $30 billion for a single Louisiana facility. CoreWeave’s GPU-backed loan facility got rated investment-grade by Moody’s this year, which is interesting to look at: it means pension funds and insurers, the money managing our retirement, are now eligible buyers of AI infrastructure risk that didn’t exist as an asset class three years ago.

#3.1 ENERGY!

Underneath the balance sheets sits a physical constraint no financing structure can fix. The buildout needs energy that is fast enough to bring online, abundant enough to meet demand, and cheap enough to keep the economics working, and the US currently has none of the three in sufficient supply. Nearly 2,300 gigawatts of capacity, more than the entire existing US power base, sits stuck in interconnection queues with waits stretching past five years, and half of global data center projects are already delayed by power and grid shortages. Leverage can finance a data center. It cannot finance a faster transformer.

#4 Will These Assets Actually Hold Their Value? (aka China’s Open Models)

Even if the financing holds, the underlying product is getting cheaper faster than the debt assumes. China’s open-weight models are collapsing the price of intelligence: DeepSeek V4-Pro runs about $0.88 per million output tokens against $25 to $30 for the leading closed models, and GPT-4-class inference has fallen roughly 95% in two years. This is the telecom fiber problem from 2000 in a new costume, the infrastructure was real, it simply wasn’t worth what the debt assumed it would be. It’s a different risk than everything above it, not a solvency question but a durability one: will the revenue survive commoditization before the capex gets repaid.

#5 Uncle Sam’s Fiscal Space Problem

Here’s the part that makes this different from 1998. When LTCM blew up, the US government had real balance sheet to absorb a shock: federal debt held by the public was well under 50% of GDP, and interest costs weren’t a binding constraint on anything. Today, the FY2026 deficit is projected around $1.9 to 2 trillion, close to 6% of GDP. Interest payments on the debt alone are around $1 trillion this year and, per CBO, are on track to more than double by 2036. Debt held by the public is already above 100% of GDP, heading toward 120% within a decade. The Committee for a Responsible Federal Budget has flagged that the average interest rate on that debt could exceed GDP growth starting around 2031, the textbook setup for a debt spiral.

#5.1 Treasury Yield, all time HIGH!

The 30-year Treasury yield closed at 5.28% on July 31, the longest stretch above 5% since 2007. Above 5%, Treasuries become real competition for capital, a pull that grows every time a signal like the Aschenbrenner news lands. Normally that rotation cushions the fall. It might not work that way this time. Stocks and long bonds have been selling off together in 2026, the correlation the 60/40 portfolio depends on has flipped positive, closer to the UK gilt crisis of 2022 than a normal correction, and China's Treasury holdings have fallen roughly 50% from their 2013 peak. But then there are additional triggers: Iran and the Strait of Hormuz pushing oil higher, a newly hawkish Fed under a less predictable chair, three regional presidents dissenting for a hike in July, the first unified hawkish dissent since 2016, and the government has less room to catch a shock than it did last time.

Cry Over Leverage

Unbundled, it's one leverage problem wearing five costumes: #1 concentration, a crowded trade with few uncorrelated buyers left; #2 circular finance, the hedge that wasn't, showing up in Aschenbrenner's book and in the financing web both; #3 the infrastructure dream, capex and negative cash flow racing a physical grid that can't move faster; #4 depreciation, the durability question China's models are forcing on every AI valuation; and #5 debt, the backstop with less room and a bond market already pricing the difference.

EXCITED

There has never been a more exciting time to invest. Platform shifts like this happen once, maybe twice in a career, and the breakthroughs are real. I am extremely grateful to get to invest through this. Looking back a decade from now, this correction will for sure read as a rounding error against how much value actually gets created. But being right about the long term and surviving the short term are two different skills. Leverage always looks like conviction right up until it looks like a margin call.

P.S. — what does that mean for us? (aka VCs)

VC is the AI trade. AI took 86% of all US venture dollars in the first half of 2026, and OpenAI and Anthropic alone captured 43% of all global startup investment. This cycle has also picked up a leverage layer venture didn’t carry last time: NAV loans sitting senior to LPs, capable of tripping a covenant across an entire fund if one portfolio company is marked down hard enough. Denominator effect, falling public portfolios push LPs over their illiquid-allocation targets and they pull back new commitments before a single startup is even repriced, which is what happened at a smaller scale in 2022. Worth remembering given the exit market looks euphoric rather than cautious right now, record M&A, a $1.7 trillion SpaceX IPO, OpenAI and Anthropic both confidentially filed to go public, since IPO issuance peaked right before the Nasdaq cracked in 2000, not after.

My friend took the classic techno-optimist stance. His argument is seductive: AI represents the democratization of software. If a college dropout can code a complex app with a single prompt, the moat for Big Tech evaporates.

He pushed the argument even further back in history. Relative inequality, he argued, was far worse in medieval times. Kings literally owned the land and held the power of life and death over their subjects. Human societies have always been shaped by two traits: adaptability and ambition. Technology expands opportunity, and people inevitably use it to climb.

It’s an appealing story. And in some ways, he is right.

AI will unleash a massive democratization of digital creation. Millions of people will be able to build their own applications, automate workflows, and design tools tailored to their personal needs. Our digital environments will become far more personalized, enabling us to create things we can’t even imagine yet.

But there is another side to this transformation.

The AI race today is not just about software and data. It is about power, land, compute, and physical infrastructure. And right now, this is a highly concentrated capital game.

The New Feudalism: Land vs. Compute

My friend’s medieval comparison may actually prove the opposite point.

In 1300, power came from land. Today, power comes from compute and energy.

We are witnessing a kind of physical enclosure movement.

Big Tech is moving down the stack, from owning digital infrastructure to owning the physical grid. Companies are securing massive tracts of land for data centers, locking in energy supply, and in some cases even exploring nuclear power.

The AI race is forcing a new level of vertical integration.

Competing at the frontier increasingly requires controlling the infrastructure that determines speed, scale, and cost. Whether it is Google’s TPUs, Meta’s tens of billions in GPUs, or the Microsoft/OpenAI compute alliance, the companies competing in AI are no longer just software firms.

They are becoming full-stack infrastructure sovereigns.

The Guardrail Monopoly

My argument is the concentration of power is not only physical, it is also cognitive.
AI labs like OpenAI or Anthropic are not just building tools; they are defining the rules of digital cognition. When a private boardroom decides what an AI system can or cannot say, it is effectively setting the guardrails for how billions of people interact with information.

The U-Shape Compromise: Engels’ Pause

Eventually my friend and I landed on a compromise rooted in the Kuznets Curve. History does not repeat, but it rhymes. Major technological revolutions often begin with a sharp increase in inequality before the gains eventually spread.

To understand 2026, it helps to look at 1820.

During the Industrial Revolution there was a period known as Engels’ Pause. Productivity surged thanks to steam engines and mechanization, but wages stagnated for nearly fifty years. The “captains of industry” captured most of the early upside while institutions and markets adjusted.

We may be living through a modern version of this dynamic.

Call it a Cognitive Engels’ Pause.

AI is already making firms significantly more productive, but the economic gains are not flowing to the individuals using the tools. Instead, much of the value is accumulating at the center.

Rome vs. Palo Alto

My friend’s medieval comparison felt intuitive, but the math tells an interesting story.

Historians estimate that at the height of the Roman Empire (around 150 CE), the Gini coefficient sat at roughly 0.44. Today the United States and China are estimated to be between 0.48 and 0.51. In other words, inequality today rivals that of ancient empires. But the real difference is reach.

A Roman senator could not influence the minute-by-minute thinking of a citizen thousands of miles away. Today, an AI platform potentially can.

Where the Value Eventually Settles

Which brings us back to the ultimate question for us, investors: In the long run, does the infrastructure capture the value, or the applications?

History suggests that massive infrastructure build-outs, whether they are railroads, fiber-optic cables, or early cloud networks, often follow a similar pattern: Infrastructure concentrates power, but applications capture the long-term value.

The pioneers who built the railroads often went bust or saw their margins collapse into a low-yield “utility” model. The real wealth was created by the companies that used the tracks to build brands, communities, and proprietary workflows.

We are currently building the most capital-intensive technology layer in history. While the “Infrastructure Sovereigns” are centralizing power today, they are also building the very rails that will eventually commoditize their own margins. Once the “Scale Wall” is climbed and compute becomes a utility, the power will shift back to the Application Layer, to the creators who own the customer relationship, the proprietary data, and the unique human agency that no machine can replicate.

AI may democratize creation while centralizing power today, but the “U-curve” suggests that this is a phase, not a final destination. The real prize isn’t owning the engine; it’s owning what the engine enables us to build.

One final thought for the road: If these “Sovereigns” successfully integrate energy, owning the actual nuclear plants that power the models, will they resist the commoditization that killed the railroad barons? Or will the Application Layer eventually force a race to the bottom for compute?

But one thing not getting enough attention is the change of human agency. How AI is disrupting and shifting power. How it is affecting and reshaping control over our lives in ways we are only beginning to understand.

The Anthropic vs. the Pentagon story was a good narrative on this essential question of our times.

Anthropic: Winning on every front

Dario Amodei and the Anthropic team are not only leading on AI model performance and enterprise integrations + consumers love it too. They are winning, with awesome financial performance. From $1 billion in late 2024 to $4 billion, $9 billion, $14 billion, and over $19 billion today. No enterprise software company has ever grown this fast at this scale.

Also, they have always been extremely cautious on AI safety. How? They are transparent on model architecture. They run extensive tests. They push for regulation. And they built a constitution, a set of values baked into Claude’s identity at the deepest level. Dario describes it as a letter from a deceased parent, sealed until adulthood.

The Pentagon dispute showed how serious they take it. They walked away from a $200 million deal and risked losing more partnerships thereafter.

"These threats do not change our position. We cannot in good conscience accede to their request." - Dario Amodei, CEO Anthropic

Because of two things they would not compromise on:

1. Claude could not be used for mass surveillance of US citizens

2. Claude could not be used to power fully autonomous weapons

And consumers voted in favor. Free signups surged 60%, daily registrations tripled, paid subscriptions doubled, a million new Claude signups per day after the ban.

And it wasn’t only consumers. Dario says they will be fine. And their bet on culture is paying off. Senior members of OpenAI’s safety team resigned in the days after OpenAI signed the Pentagon deal.

Culture, it turns out, is a competitive advantage in the AI race.

But what is that all about?

In his recent essay “The Adolescence of Technology,” Amodei argues that we have to be extremely cautious in deploying AI. Why? Because in tests it has been shown that AI builds its own character, with its own intents, that we cannot fully control or yet understand. Claude, in controlled tests, blackmailed fictional employees who controlled its shutdown button. When given training data suggesting Anthropic was evil, it engaged in deception against its own creators. These happened in controlled experiments, but they happened.

And Anthropic is not alone. OpenAI’s own research has shown GPT-4 attempting to persuade humans to help it avoid being shut down. Google DeepMind researchers documented Gemini models developing unexpected strategies to achieve goals in ways nobody designed. These are not isolated incidents from one cautious lab being overly transparent. This is a pattern across every frontier model currently being deployed at scale.

None of them are fully understood or controlled by the people who built them.

Human individual liberties versus collective coordination have always been a balancing act. Who should be able to govern us, and have control over us, and in which areas of our life? These are not new questions. But the stakes have never been higher.

The question we have to answer

We have always debated who controls us and which areas someone can interfere without infringing on our human liberties. Who is in control over our own lives? Me as sovereign. Private companies. Governments. Those three layers we understood, argued about, and built institutions to constrain. The church. The government. The corporation. Each time we found a way to define boundaries of control: the rule of law, democracy, human rights.

Now there is something new in the equation: AI. Not a government you can vote out. Not a company you can stop buying from. Something that emerged, that even its own creators cannot fully control or yet fully understand.

It is not only the question between us as individuals, private companies, and governments anymore. It is becoming even clearer that we have to be very careful when thinking about AI deployment and how much control we hand over to something we cannot fully control or yet fully understand.

Amodei himself said it best: “Every decision I make feels like it’s balanced on the edge of a knife.”

That knife-edge is sharper today than when he said it. And the question of how much control we hand over, to systems we don’t fully understand, in areas as fundamental as surveillance and weapons, is the essential question of our times.

We are nowhere near answering it. But we have to pay attention and be very careful. We have always found a way to define the boundaries of control. We have to find a way again before the technology defines them for us.

We send over 2.5 billion prompts to ChatGPT every day. What began as a work tool has quietly become our infrastructure for daily life: more than 70% of all prompts are now personal, not professional. But beneath this behavioural shift sits a physical one. AI inference is rapidly becoming one of the world’s fastest-growing energy loads.

A single ChatGPT query costs roughly ~0.34 Wh, according to Altman. It varies with hardware and data-centre efficiency, but let’s use that as a ballpark. At today’s volume, this translates into ~850 MWh per day, or ~0.3 TWh per year, for ChatGPT (compute) alone.

And that is only the starting point. ChatGPT already reports more than 700 million weekly active users, with both user count and workload per user rising. Even assuming meaningful efficiency gains, a 10× increase in OpenAI’s total energy consumption over the next seven years, to ~3 TWh annually, is a conservative scenario.

Inference, not training, is the driver

To understand how to cut the energy bill of inference, we first need to understand what truly drives it. That led us down the rabbit hole, past model architectures and kernels, down to the physical movements of data through a modern AI system.

So what actually happens when you send a prompt to ChatGPT?

If we put that into a simple energy cost formula, we can say the following:

E_total = E_compute + E_data movement + E_memory access + E_idle

• E_compute is the actual math on the GPU or accelerator.

• E_data movement is moving activations and parameters between chips and across interconnects.

• E_memory access is reading and writing from on-chip SRAM, HBM and DRAM.

• E_idle is everything spent while hardware is powered but waiting: on data, on other GPUs, on the network, on the scheduler.

To understand how a token travels through the compute stack, let’s take a closer look at the underlying Compute architecture.

How a prompt really travels through a GPU

When you send a prompt to an LLM running on GPUs, you’re not just doing a matrix multiply on a chip. You are waking up an entire stack of subsystems.

First, the host CPU and orchestration layer parses the request, schedules it and decides where to run it. This control plane work is comparatively cheap.

The prompt then reaches the GPU compute unit, tensor cores and streaming multiprocessors, supported by on-chip caches, which performs the matrix multiplications and attention operations we think of as the cost of inference.

However, the model’s weights and most activations live off-chip in high-bandwidth memory (HBM). Every time the GPU needs a new block of weights, it must fetch them across a short but energy-intensive distance. Each fetch can cost tens to hundreds of times more energy than the floating-point operation that uses the data.

For large models, weights and activations are also spread across many devices. The request therefore traverses the interconnect fabric, NVLink, InfiniBand, high-end Ethernet, to move activations between GPUs and often between racks. These cross-GPU hops are even more expensive.

Why GPUs were perfect for training and why they hit a wall in inference

GPUs have been perfect for training large models. Training is industrial-scale linear algebra: vast, dense matrix multiplications repeated millions of times. Each step looks like the last. Batches are large, layers are reused, and the work is highly regular. You can feed thousands of examples into the same kernels and keep the cores saturated.
For every byte pulled from memory, the chip performs an enormous number of floating-point operations, so the weaknesses of the architecture fade. The cost of fetching weights from HBM is amortised over huge batches, and cross-GPU communication is predictable. In terms of our energy identity, most of training collapses into E_compute. GPUs dominate training because they are dense-math engines, and training is a dense-math problem.

Inference lives in a completely different physical regime. Workloads are driven by users, not an optimisation loop. Requests arrive irregularly, sequence lengths vary, and latency constraints prevent large batches. Models are sharded across many GPUs, often across racks, so each prompt pays the full cost of fetching weights and KV caches from HBM or DRAM, and activations bounce across NVLink or InfiniBand. GPUs sit powered up but idle while they wait for memory or the network.
In energy terms, the balance flips: in inference, E_memory access, E_data movement and E_idle set the floor, not compute.

The Scaling Mismatch
For decades, Moore’s Law and Dennard scaling gave us a simple equation: more transistors produced more compute at roughly the same power. But the parts of the system that matter for inference, memory bandwidth and interconnect energy, never followed that curve. FLOPs per watt still improve rapidly; energy per bit in DRAM and on interconnects improves only marginally. Compute gets exponentially cheaper. Moving data does not.
In training, batching hides this divergence. In inference, sparse, latency-bound and memory-heavy, becomes the bottleneck.

And just as this wall comes into view, the workload itself is shifting.

Sparse is becoming the default

If the last decade of deep learning was the age of dense transformers, the next is shaping up to be the age of sparsity. Over the past few years, sparsity has quietly become the standard scaling strategy. Google moved first with GShard and Switch Transformer; Meta followed with NLLB and LLaMA-MoE. Mistral’s Mixtral line, Alibaba’s Qwen-MoE and a wave of Chinese labs now crowd the field. DeepSeek-V2, a 236-billion-parameter MoE model that activates only ~21 billion parameters per token, made the trend explicit: at the frontier, models are sparse, not dense.

On paper, sparsity looks like the perfect fix: more total parameters, fewer FLOPs per token. In practice, on today’s GPU infrastructure, it pushes the workload directly into the parts of the system that are already least scalable.

How MoE breaks the GPU’s memory and interconnect

In the figure below, we illustrate what actually happens inside a GPU cluster when a Mixture-of-Experts (MoE) model runs.

When a Mixture-of-Experts model runs, a router selects one or two experts for each token, often spread across multiple devices. What was a smooth streaming workload becomes a pattern of random look-ups. Instead of pulling contiguous weights through the cache hierarchy, the GPU chases scattered expert parameters around HBM, each fetch costing 10–100× more energy than the FLOP that uses it.

When the chosen experts sit on other GPUs, activations and results bounce across NVLink or InfiniBand, adding cross-GPU hops that are even more expensive again. On paper, MoE reduces the compute term: fewer multiplications per token. In practice, it hits exactly the parts of the system that scale the least: memory access, data movement, and idle time.

From prompts to workflows: how agentic inference breaks the cluster

So far, we have treated inference as a single forward pass through a model. Agentic systems break that assumption. An agent does not simply answer; it plans, calls tools, retrieves information, inspects its own output and often tries again. What used to be one neat piece of linear algebra becomes a small, irregular program spread across models, APIs and databases.

The diagram above shows what this looks like on the hardware. The host CPU no longer hands a prompt to a model and waits. It runs an agent loop, maintaining state, choosing the next step, marshalling tool calls, and repeatedly spinning the GPU up for small bursts of tokens. Thousands or millions of tiny control operations flow through a subsystem GPUs are fundamentally bad at. The result is constant ping-pong between CPU and GPU, with energy burned on coordination rather than on math.

Every time the agent decides to “think more,” extend context, call a search tool or invoke another model, the entire GPU cluster wakes up again. Fresh weights and activations are pulled from HBM, KV-caches are reshuffled, and traffic surges across NVLink and InfiniBand as sharded models exchange activations. Context windows stretch and swap, producing even more memory thrash.

In terms of our energy identity, agentic inference does two things at once: it multiplies E_compute by requiring many forward passes, and it amplifies E_data movement, E_memory access and E_idle, because each pass is smaller, more irregular and tightly interwoven with tools and network calls. The unit of work is no longer a prompt; it is a cluster-level workflow.

This is why agentic AI feels powerful at the product layer yet so inefficient at the infrastructure layer. It is exactly the pattern of computation the GPU-centric, von Neumann stack is worst at: branchy, stateful, chatty and intolerant of the large, homogeneous batches that once hid the true cost of memory and networking.

So what?

The natural question at this point is: given this physics, how do we actually build infrastructure for an agentic, sparse inference world? We cannot simply stack more GPUs and hope for the best. The constraint is no longer enough flops, it is the sum

E_total = E_compute + E_data movement + E_memory access + E_idle (and, in practice, a storage term for loading cold weights and embeddings).

and every serious solution is, in one way or another, an attempt to push down one or more of these terms.

1. E_compute: new math engines for sparse and agentic workloads

At the model level, the industry is already squeezing more out of each FLOP. MoEs, quantisation, pruning and distillation reduce arithmetic per token. DeepSeek-V2 shows how far this can go: 236B total parameters, but only ~21B active per token.

At the hardware level, a growing group of specialist players is betting that “a GPU, but slightly better” is not the end state. Cerebras, Groq, SambaNova, Graphcore and others ask what a chip would look like if it were designed from scratch for large-scale inference. They take different routes, wafer-scale integration, deterministic pipelines, domain-specific ISAs, but share a goal: keep arithmetic units busy and predictable as workloads get sparser and more irregular.

Neuromorphic designs push this further. Instead of marching a global clock through dense matrices, they compute event-by-event: only spikes or changes propagate. For sparse and agentic workloads, this matches the pattern of work. A fabric that wakes only when needed could yield order-of-magnitude energy gains over GPUs clocking through every timestep.

But achieving this requires a break from the GPU stack. A neuromorphic chip is not a faster GPU; it is a different machine. Tooling, compilers, debuggers, numerical methods, even the mental model of what a program is, must be rebuilt. The upside is enormous, but so is the risk. This is the radical edge of E_compute, where the “AI hardware dilemma” is strongest.

The incumbents are evolving too. Nvidia’s Hopper and Blackwell push tensor throughput, low-precision formats and MoE performance; GB-series systems are marketed as agent-inference engines. Google’s TPUs and AWS’s Trainium/Inferentia follow similar paths inside their clouds.

The silicon stack is fragmenting

Every major hyperscaler is now building its own chips, TPUs at Google, Trainium and Inferentia at AWS, MTIA at Meta, Athena at Microsoft, neural engines at Apple. As inference becomes sparse, memory-bound and highly irregular, general-purpose accelerators no longer fit every workload. And we assume, the GPU monoculture is ending. Silicon is diversifying into many specialised architectures tuned to bandwidth, locality, sparsity or control flow. The next decade will look less like the GPU era and more like the RISC explosion: heterogeneous, competitive, and full of openings.

From a venture perspective, this bucket is hard but wide open. Neoclouds are already testing new silicon, winning workloads, and capturing the spread between GPU energy cost and their own specialised designs. Success will require a deep architectural insight, designing around bandwidth rather than FLOPs, for example, paired with a credible route to market. The number of winners will be small, but the upside for one or two truly inference-native architectures is enormous.

2. E_memory: bringing compute to where the weights live

If DRAM and HBM access dominate the energy bill, one obvious move is to bring compute closer to the memory. Compute-in-memory (CIM) and near-memory architectures try exactly that: storing weights directly inside dense arrays and executing operations within or beside the memory banks so far fewer bits need to cross the bus. Companies like Mythic, d-Matrix, Upmem and Untether AI have pursued variants of this idea. It is a hard road, Mythic’s recapitalisation shows how unforgiving analogue and CIM can be, but the logic only strengthens as models grow and MoEs scatter parameters across ever-larger address spaces.

A more incremental and commercially nearer-term path is to make caching itself a first-class optimisation problem. Graphcore, Cerebras, Hailo and others have explored architectures that keep large portions of model state on-chip and use aggressive placement and prefetching to minimise expensive HBM and DRAM trips. Hot experts, routing tables, KV-caches and other frequently accessed structures stay as close to the cores as possible, turning what would have been random off-chip lookups into predictable on-chip reads.

From a venture perspective, this bucket is compelling because it doesn’t require beating Nvidia on FLOPs. It attacks the thing GPUs are fundamentally worst at: bandwidth per watt. The most promising companies will slot into specific, painful bottlenecks, edge inference, retrieval, KV-cache acceleration, MoE routing, where moving fewer bits, not computing more flops, is the real unlock. Recent funding momentum underlines this. d-Matrix raised $275M in late 2025 at a $2B valuation to scale its digital-in-memory compute architecture, claiming 10× faster inference and 3–5× better energy efficiency than GPUs. At the edge, Hailo extended its Series C with $120M in 2024 and launched Hailo-10, a GenAI accelerator able to run Llama-2 7B under 5 W. Both cases reflect the same trend: specialised memory-centric silicon is becoming commercially viable and strategically important as bandwidth per watt becomes the real constraint.

3. E_interconnect: photons and fabrics

Once models span dozens or hundreds of accelerators, the interconnect becomes the nervous system of the AI data centre. Today, that nervous system is overwhelmingly Nvidia’s: NVLink and NVSwitch bind GPUs inside nodes, while Mellanox-derived InfiniBand and Spectrum Ethernet extend the fabric across racks.

But the physics of copper is running out of road. Silicon-photonics companies such as Ayar Labs, Celestial AI and Lightmatter promise far higher bandwidth and dramatically lower energy per bit by bringing optics directly into the package, the board and eventually the rack. This is no longer hypothetical: Ayar has raised major rounds and signed multi-year co-development deals with Intel and HPE, while Celestial and Lightmatter are positioning optical interconnects as the next defensible layer of AI infrastructure. Nvidia itself has now announced photonic networking gear and “NVLink Spine” architectures aimed at connecting millions of GPUs with far less power.

For startups, this layer is classic deep tech: long design cycles, hyperscaler-driven adoption, and exits that will mostly come through strategic acquisition rather than IPO. But the strategic value is enormous. Whoever controls the optical fabric controls the topology of the rack and therefore the economics of very large MoE and agentic clusters. This is the lever that decides how far sparse, memory-heavy AI can scale.

4. E_idle / control: compilers, runtimes and the agent OS

A large share of wasted energy in inference does not come from computation but from waiting: transistors idle while chips sync, data moves across the fabric, or orchestration logic decides the next step. Reducing this idle tax is primarily a software challenge.

At the lowest level, compilers such as Apache TVM and OpenXLA take a model graph and lower it into code that minimises memory traffic, fuses kernels and exploits backend-specific quirks. Above them, specialised inference runtimes like TensorRT-LLM and vLLM focus on batch formation, KV-cache policy and scheduling, directly cutting E_idle and E_data movement for LLMs.

The next layer is the emerging “agent OS”: systems from Modal, Baseten, Anyscale/Ray and open-source projects that optimise entire workflows of model calls and tools, not just individual kernels. Done well, these systems prune redundant steps, co-locate hot paths, pre-warm models, and downshift to cheaper architectures when possible. For real agentic workloads, these coordination gains often reduce E_total more than any single model-level optimisation.

From a venture perspective, this is the most accessible part of the stack: fast iteration cycles, moderate capital requirements and clear developer pull. The challenge is defensibility. Open-source gravity is strong, and generic orchestration is easy to copy. Durable companies will likely emerge from tight integration with specific hardware, deep observability and scheduling advantages, proprietary datasets of workflow patterns, or by becoming the dominant agentic runtime inside enterprises, the “Kubernetes of AI workflows.”

5. E_storage: cold weights, hot caches

A growing share of inference energy lives in how models and embeddings are stored and surfaced. Serving layers increasingly pull weights from object stores, cache hot models at the edge, and juggle multi-tenant fleets where only a fraction of models are “warm” at any given moment. Platforms like Predibase, Replicate, Baseten and Modal are effectively building this storage-and-serving substrate for a world where thousands of specialised models coexist.

At first glance, this looks like ordinary cloud middleware. In reality, as models balloon in size and agentic workflows touch many specialised components, intelligent caching and placement become economically decisive. Pulling a cold model from storage can stall a cluster for seconds; keeping the right models hot can cut both energy burn and cloud bills by an order of magnitude.

There is meaningful room here for infrastructure companies that treat energy, latency and model locality as first-class optimisation targets, not afterthoughts. The winners will likely own the placement layer: deciding which models stay hot, where, and when as agentic systems orchestrate increasingly complex workflows.

After GPUs: Building Infrastructure for Agentic Inference

We have not seen a window this open for new compute architectures in many years. For decades, the industry converged on a stable stack: von Neumann architecture, general-purpose CPUs, and then GPUs repurposed for deep learning.

That stack is now running directly into the physics of physical AI. We are entering a world defined by enormous parameter spaces, sparse models, MoEs, and agentic systems that behave less like linear algebra and more like irregular, stateful programs.

In this new regime, the GPU-von Neumann architecture looks increasingly out of place. It is brilliant at dense, uniform workloads, but structurally bad at the sparse, conditional, and chatty computation patterns that characterize modern inference. We are asking it to support the exact pattern of computation it was never designed for, compensating with elaborate software and denser, hotter racks.

The Shift in First Principles

This does not mean GPUs disappear; general-purpose accelerators will remain the workhorses of training and much of inference. However, the center of gravity is shifting.

As energy and bandwidth harden into the binding constraints, there is real room for new architectures and full stacks built for the agentic, sparse, memory-heavy regime first, and everything else second.

The problem is no longer enough FLOPs. It is solving for the full cost of the system:

Etotal = E_Compute + E_Data movement + E_Memory access + E_Idle

The Contrarian Ventures Thesis

The founders we get most excited about start from physics, not product categories. They look at the world as it will be, sparse, agentic, continuous, and physical, and ask what architecture that world demands.

They are comfortable building something that looks wrong today but will be obvious in 2030 and beyond. This includes founders willing to ship:

• Chips built around bandwidth instead of FLOPs.

• Fabrics built around events instead of clocks.

• Runtimes built around agents instead of prompts.

We are especially excited about neuromorphic systems. Event-driven, locality-driven, and spike-driven architectures are no longer academic curiosities; they are an increasingly natural answer to sparse, conditional workloads. We are open to the unconventional, the architecturally radical, and the non-GPU futures, especially where the physics gives you compounding advantages.

If you are building in that space, new compute architectures, memory-centric designs, optical fabrics, neuromorphic systems, inference compilers, agent runtimes, or serving layers that treat energy as a first-class constraint, reach out.

The window to build the next layer of compute is open in a way it has not been in a generation. If you dare to rethink from the ground up, we are here to talk.

We have not seen a window this open for new compute architectures in many years. For decades, the industry converged on a stable stack: von Neumann architecture, general-purpose CPUs, then GPUs repurposed for deep learning, all wrapped in CUDA and cloud abstractions. That stack is now running directly into the physics of physical AI: enormous parameter spaces, sparse models, MoEs and agentic systems that behave less like linear algebra and more like irregular, stateful programs.

In that world, the GPU–von Neumann architecture of the 1940s looks increasingly out of place. It is brilliant at dense, uniform workloads; structurally bad at sparse, conditional, chatty ones. We are asking it to support the exact pattern of computation it was never designed for, and compensating with increasingly elaborate software and denser, hotter racks.

This does not mean GPUs disappear. It means the centre of gravity is shifting. General-purpose accelerators will remain the workhorses of training and much of inference. But as energy and bandwidth harden into the binding constraints, there is real room for architectures and full stacks built for the agentic, sparse, memory-heavy regime first and everything else second.

The founders we get most excited about start from physics, not from product categories. They look at the world as it will be sparse, agentic, continuous, and physical and ask what architecture that world demands. They are willing to ship chips built around bandwidth instead of FLOPs, fabrics built around events instead of clocks, and runtimes built around agents instead of prompts. They are comfortable building something that looks wrong today but is obvious for 2030 and beyond.

We are especially excited about neuromorphic systems. Event-driven, locality-driven, spike-driven architectures are no longer academic curiosities; they are an increasingly natural answer to sparse, conditional workloads. We are open to the unconventional, the architecturally radical and the non-GPU futures, especially where the physics give you compounding advantages.

If you are building in that space, new compute architectures, memory-centric designs, optical fabrics, neuromorphic systems, inference compilers, agent runtimes or serving layers that treat energy as a first-class constraint, reach out. The window to build the next layer of compute is open in a way it has not been in a generation. If you dare to rethink from the ground up, we are here to talk. Pm: robina@cventures.vc

At Contrarian Ventures, we’ve been exploring what it will take to bridge that gap. We know that unlocking the full potential of green assets requires more than just inflow of capital; it requires a complete transformation of the financial infrastructure that powers these investments. What are the key barriers? Why is liquidity so hard to achieve? And most importantly, how is energy itself evolving into the financial asset class of the 21st century?

To answer these questions, we’ve developed the Climate Liquidity Stack framework, which breaks down the essential layers of infrastructure needed to scale investment into climate solutions. This article will walk you through:

• The Climate Liquidity Stack: Our framework for structuring the flow of capital into green assets.

• The current market map: An analysis of different asset classes and their liquidity challenges.

• Where we see the biggest opportunities: The venture sectors that are primed for innovation and growth.

Let’s dive in.

Our Mental Model: The Climate Liquidity Stack

The Liquidity Stack is a framework that structures how infrastructure, financing tools, and market mechanisms combine to create liquidity.

Unlocking liquidity involves the following three layers, each building on the next. Here's the breakdown:

Layer 1: Infrastructure & Trust

This foundational layer doesn’t directly move capital but enables liquidity. Think of it as the rails that underpin the entire financial system. Before money can flow into an asset class, the assets themselves must be transparent, standardized, and trusted by investors.

Layer 1 includes tools such as: Data standards and APIs, Risk ratings and project verification (MRV), Registries and interoperability frameworks

Without this base layer, liquidity cannot form as markets need trust and clarity first.

Layer 2: Financing & Risk Products

Layer 2 is where capital is directly channelled into green assets. This is the engine of liquidity, where financing tools like: Lending and revenue-based financing, Insurance products and credit guarantees, Embedded finance solutions or BNPL models are used to move money into the system. Layer 2 is where financing decisions are made, directly supporting the deployment of green assets.

Layer 3: Market Access & Liquidity Pools

Layer 3 brings secondary liquidity, meaning it opens access to capital markets. Once green assets are financed and operational, Layer 3 creates mechanisms that allow investors to trade, aggregate, or exit their positions.

This layer includes solutions like Investment marketplaces, Tokenization and securitization platforms, Carbon, energy, and climate asset exchanges

Layer 3 transforms green assets into financial products that can be traded in capital markets, enabling them to move freely and attract institutional investment.

Liquidity Along the Maturity of the Asset Class

Now that we’ve covered the core layers needed to unlock capital flow, it’s time to examine how these layers apply across different green asset classes. Maturity matters, and the financial infrastructure for green assets varies significantly depending on where the asset stands in its lifecycle.

For highly standardized, mature, and bankable assets like utility-scale renewables, a robust financial infrastructure already exists. These assets benefit from well-established tools across all three layers, with highly developed market structures and transparency. In contrast, emerging assets, like First-of-a-Kind (FOAK) technologies, operate in an underdeveloped financial ecosystem. These technologies face challenges like limited standardization, a lack of transparency, poorly adapted financing models, and illiquid or underdeveloped secondary markets.

To better understand these gaps, we mapped the market by comparing the Liquidity Stack (on the Y-axis) with the maturity of various asset classes (on the X-axis). This spectrum ranges from nascent, FOAK, and esoteric assets, through distributed retail assets and commercial & mid-scale infrastructure, all the way to institutional and bankable assets.

Here’s what we found:

Market Map to Unlock Liquidity for Green Assets

1. Highly Mature Assets Like Utility-Scale Solar and Wind Farms

Layer 1: Infrastructure & Trust

For utility-scale solar and wind farms, the trust layer is highly mature. These assets have decades of performance data, robust Power Purchase Agreements (PPAs), and established engineering standards. Third-party verification, such as Moody’s and S&P credit ratings, ensures transparency and builds trust. This level of consistency and transparency has made these assets highly investable, with low barriers for capital flow.

Example: NextEra Energy, a leader in renewable energy, leverages these mechanisms to attract substantial institutional investment, benefiting from clear performance data and long-term contracts.

Layer 2: Financing & Risk Products

The financing layer for utility-scale assets is well-established, utilizing project finance structures, tax equity, and various insurance products to manage risks. These tools, such as production insurance and political risk hedging, help reduce capital cost and encourage competition among investors. The rise of private equity and capital funds further fuels the growth of this sector.

Example: LevelTen Energy and Pexapark streamline access to renewable energy projects through risk analytics and off-take agreements, driving lower financing costs and fostering greater competition in the market.

Layer 3: Market Access & Liquidity Pools

Secondary liquidity for utility-scale renewable assets is strong. These assets are packaged into YieldCos, providing stable cash flows to investors, and green bonds attract institutional capital. The M&A market for these assets is also robust, with large institutional investors like pension funds and sovereign wealth funds actively involved.

Example: Brookfield Renewable Partners utilizes YieldCos and green bonds to bring renewable energy assets to the public market, providing liquidity and access to a wide range of investors.

US vs. Europe

Both the U.S. and Europe have developed strong financial ecosystems for utility-scale renewables, but they approach it differently. In the U.S., tax equity financing plays a crucial role, particularly with the boost from the Inflation Reduction Act (IRA). Public YieldCos originated in the U.S. and have been vital in providing liquidity. In contrast, Europe relies more heavily on traditional project finance structures and government-backed green incentives through frameworks like the EU Green Deal. While both regions have strong infrastructure, the mechanics differ - tax equity is central in the U.S., while green bonds and public policy are key drivers in Europe.

2. Evolving Assets: Commercial and Mid-Scale Infrastructure (e.g., Commercial Rooftop Solar, C&I Battery Storage, Building Energy Retrofits)

Layer 1: Infrastructure & Trust

For commercial and mid-scale assets, the infrastructure and trust layer is evolving but still inconsistent. Energy savings performance data, whether from commercial rooftop solar, battery storage, or building retrofits, tends to be fragmented, with significant variation in measurement, reporting, and verification (MRV) practices. Emerging technologies like Battery Energy Storage Systems (BESS) + PV hybrids are growing, but they still lack bankability due to unclear or volatile revenue streams.

Companies like kWh Analytics (offering solar risk benchmarking for C&I projects) and Measurabl (providing ESG data for commercial real estate) are working to standardize reporting and improve data transparency. However, technical due diligence for mid-scale projects remains bespoke, and many investors perceive higher performance variability and underwriting complexity. The lack of consistent performance data aggregation is a significant bottleneck to scaling liquidity in this space.

Layer 2: Financing & Risk Products

The financing layer for commercial and mid-scale assets is fragmented, with specialized platforms emerging to channel capital into these projects. Energy-as-a-service models, pioneered by companies like Redaptive, Sparkfund, and Co-Power, allow businesses to adopt clean technologies without upfront capital expenditures, de-risking customer adoption and generating predictable cash flows. These models are helping bridge the financing gap for projects that may have been previously inaccessible due to high capital costs.

However, many SMEs still struggle with limited access to affordable capital due to weak credit profiles or non-standard project structures. Emerging solutions like credit insurance, revenue guarantees, and green lending platforms are helping fill this gap, but the lack of standardized underwriting tools means these solutions are not yet as scalable as they need to be.

Layer 3: Market Access & Liquidity Pools

Secondary liquidity for commercial and mid-scale assets is still in its early stages. While aggregators like Generate Capital and Sunwealth are beginning to bundle commercial energy efficiency and solar projects into investable portfolios, there is little structured securitization of these mid-scale assets compared to their utility-scale counterparts. Green bonds tied to mid-scale building decarbonization projects are still rare. Most liquidity remains in private markets, through direct portfolio sales or forward flow agreements.

US vs. Europe

In the U.S., tax incentives such as the Section 48 Investment Tax Credit under the Inflation Reduction Act (IRA) have significantly catalyzed investment in C&I clean infrastructure. U.S. players have been more active in experimenting with energy-as-a-service models and bundled lease models, which help overcome upfront capital barriers.

In Europe, corporate ESG pressures (under frameworks like CSRD and the EU Taxonomy) are pushing building owners towards decarbonization, but financing solutions are slower to develop due to reliance on traditional bank lending and public incentives. Europe leads in regulatory-driven demand, but the U.S. leads in financing innovation for commercial-scale assets. Both regions, however, face a lack of deep secondary liquidity for mid-scale assets.

Summary

The financial infrastructure for commercial and mid-scale assets is improving but still faces significant hurdles, particularly in the trust layer and market access. Standardization of data and improved financing models will be crucial to unlocking liquidity for these assets. While the U.S. is leading the way with tax incentives and energy-as-a-service innovations, Europe’s regulatory-driven demand presents a promising opportunity for scaling up the financing solutions for mid-scale and commercial projects. Overcoming the gaps in secondary market liquidity, especially for smaller assets, will be key to unlocking capital at scale for this sector.

3. Distributed and Retail Assets: Residential Solar, Home Batteries, Heat Pumps, EV Chargers

Layer 1: Infrastructure & Trust

The infrastructure and trust layer for distributed and retail energy assets is still in its infancy and fragmented. Unlike large utility or commercial projects, residential solar systems, home batteries, and electrification upgrades (such as heat pumps or EV chargers) are often installed as individual units with highly variable performance. Currently, there is no universal, centralized registry or standard to track these systems over time, making it challenging to build investor confidence.

Companies like Arcadia are addressing this gap through utility data APIs, enabling real-time access to household energy usage data. At the same time, WattCarbon and Singularity Energy are creating MRV platforms to track carbon intensity and the impact of distributed assets. However, the lack of standardized, verifiable performance data across millions of small assets remains a major obstacle to scaling liquidity in this sector.

Layer 2: Financing & Risk Products

Financing solutions for residential energy assets have grown rapidly but remain uneven. Embedded point-of-sale lending platforms, pioneered by companies like GoodLeap (in the U.S.) and Enpal (in Germany), allow consumers to adopt solar, batteries, and related technologies with little or no upfront capital. These models are helping to de-risk adoption, but underwriting for newer electrification technologies—such as home batteries, EV chargers, and bidirectional energy systems—is still immature. Traditional credit models, still heavily reliant on FICO scores, do not take into account asset-specific performance data, limiting access to affordable financing.

The cost of capital for distributed assets remains relatively high compared to more established asset classes like utility-scale renewables, which restricts affordability and slows penetration into lower-income customer segments. However, solar loan securitizations (e.g., ABS deals from GoodLeap and Sunlight Financial) are beginning to institutionalize this asset class.

Layer 3: Market Access & Liquidity Pools

Secondary liquidity for distributed assets is still emerging. While solar loan ABS structures have allowed some institutional capital to flow into residential solar portfolios, broader liquidity mechanisms for newer electrification assets (such as home batteries and EV chargers) are almost non-existent. Unlike utility-scale portfolios, there are no widely traded bundles of distributed assets or aggregated home energy savings contracts.

Aggregation models, such as virtual power plants (VPPs), which connect distributed assets to grid markets, are still in their early stages. Companies like Sunrun (with Tesla batteries) and Leap are pioneering these solutions, but the monetization of these assets is still heavily dependent on primary owners and utilities, rather than open, tradable markets.

Example: Enpal has made a significant stride by launching its first green bond, marking a key milestone in aggregating distributed assets and enabling them to enter the capital markets.

US vs. Europe

In the U.S., residential solar adoption has been significantly boosted by strong federal and state incentives, coupled with private-sector innovation in consumer finance models (e.g., GoodLeap and Mosaic). Loan securitizations for solar assets have begun to create early liquidity pathways.

In Europe, the residential clean tech market is more focused on integrated product offerings. For example, Enpal bundles solar panels, batteries, EV chargers, and heat pumps into comprehensive home energy-as-a-service solutions. While Europe is ahead in regulatory-driven demand for clean energy, it lags behind in building securitization and liquidity structures for distributed energy financing.

Both regions face the same challenge: aggregating millions of small assets into tradable, investable products.

Summary

The infrastructure and trust layer for distributed and retail assets is still evolving, with challenges in data standardization and performance verification. To unlock liquidity in this sector, improving transparency and developing standardized MRV systems will be critical. On the financing side, while innovative point-of-sale financing models like those from GoodLeap and Enpal have helped de-risk the adoption of residential solar and energy storage, underwriting tools still need to mature to address emerging technologies like EV chargers and home batteries. Finally, the secondary market for distributed energy assets is still nascent, but the growth of aggregation models such as VPPs offers a promising avenue to enable liquidity.

To truly unlock the potential of distributed assets, the focus will need to be on improving data transparency, developing more affordable financing solutions, and creating robust secondary markets. Both the U.S. and Europe are making strides, but significant challenges remain in turning these small, fragmented assets into tradable, investable products.

4. First-of-a-Kind (FOAK) and Emerging Technologies: Direct Air Capture, Green Hydrogen, Next-Gen Biofuels

Layer 1: Infrastructure & Trust

The trust infrastructure for First-of-a-Kind (FOAK) climate technologies, such as Direct Air Capture (DAC), green hydrogen production, and next-generation biofuels, is still in its infancy. Measurement, Reporting, and Verification (MRV) frameworks are being developed, but there are currently no globally accepted standards to validate carbon removals or hydrogen production outputs. Efforts from companies like Puro.earth (carbon removal registry) and Patch (forward procurement platforms for carbon credits) are beginning to establish digital MRV protocols. However, trust gaps remain substantial: asset performance is uncertain, risk profiles are difficult to model, and technical baselines (e.g., captured carbon per ton or electrolyzer efficiency) are still evolving.

As a result, investors face high uncertainty when underwriting FOAK asset cashflows or impact outcomes, which limits their willingness to commit large-scale capital. Bridging these trust gaps by improving data transparency and establishing clear, standardized MRV systems will be crucial for unlocking investment in these technologies.

Example: Puro.earth is leading the charge by developing a carbon removal registry, providing a foundational layer for trust and transparency in emerging carbon removal technologies.

Layer 2: Financing & Risk Products

Primary financing for FOAK assets is extremely limited and often relies on venture capital, government grants, philanthropic capital, and blended finance structures. Traditional lenders generally avoid FOAK projects due to unproven cash flows and high technical risks. Emerging models, such as the Frontier Fund (supported by Stripe, Shopify, Meta, and others), create demand-side guarantees for future carbon removals, which help new technologies secure financing to scale.

Some emerging insurance models aim to backstop volume risks (especially for green hydrogen or carbon removal delivery shortfalls), but these are still in early development. Overall, the financing infrastructure for FOAK assets is fragmented and ad hoc, relying on bespoke, project-by-project fundraising rather than standardized capital channels. This gap presents a significant opportunity to create more scalable, efficient financing solutions for emerging climate technologies.

Example: Frontier Fund is helping to create demand guarantees for carbon removal technologies, acting as a critical bridge for early-stage capital.

Layer 3: Market Access & Liquidity Pools

Currently, there is no true secondary market for FOAK assets. While carbon removal credits from DAC or enhanced weathering projects can sometimes be sold through bespoke contracts, these transactions are still negotiated bilaterally rather than traded in open markets. Early experiments in tokenization (e.g., Toucan Protocol for blockchain-based carbon credits, Flowcarbon for carbon-backed tokens) are attempting to create tradable liquidity for emerging assets, but adoption remains limited, and credibility concerns persist.

Without standardized MRV, trusted pricing benchmarks, or credit ratings, FOAK assets remain highly illiquid, with most capital committed to these assets locked in long-term, patient positions with no clear exit pathways. Establishing a secondary market for these assets will be a crucial next step in unlocking their liquidity.

Example: Flowcarbon has started experimenting with tokenizing carbon credits, but broader market acceptance is still in early stages, and standardized MRV and pricing mechanisms are needed to foster wider adoption.

US vs. Europe

The U.S. has emerged as a leader in FOAK climate financing, driven by deep venture capital markets (e.g., Breakthrough Energy Ventures, Lowercarbon Capital) and major federal policy incentives such as the 45Q tax credit for carbon capture and IRA funding for hydrogen hubs. These factors have created an environment where FOAK technologies can attract significant private capital.

In Europe, while venture-scale risk-taking has historically been slower, the region is catching up with large public funding mechanisms such as the EU Innovation Fund and upcoming Carbon Contracts for Difference (CCfD), designed to de-risk early-stage decarbonization technologies. Both regions recognize the critical need to build financial infrastructure for FOAK assets, but global markets are still in the early stages of developing the necessary trust, capital flow, and liquidity structures.

Summary

The financial ecosystem for FOAK climate technologies is still in its infancy. While early efforts in MRV standardization, financing models, and secondary market development are promising, trust gaps and uncertainty around performance remain major obstacles to scaling investment. To unlock liquidity in FOAK assets, there is a need to establish standardized MRV systems, create demand-side guarantees, and develop secondary markets that allow investors to trade these assets more freely. Both the U.S. and Europe are making strides, but global coordination and innovation will be needed to build the necessary infrastructure to scale FOAK investments and unlock their full potential.

Conclusion: Key Opportunities to Unlock Liquidity for Green Capital in Europe

To unlock the trillions needed to transition to net-zero, we must build the right financial infrastructure. Based on the Climate Liquidity Stack, we’ve identified critical gaps in the current market that represent key opportunities to unlock liquidity and scale green capital in Europe.

1. Data Standardization & MRV: Developing standardized data and Measurement, Reporting, and Verification (MRV) frameworks for emerging technologies like carbon removal and green hydrogen will build trust and transparency, enabling investor confidence.

2. Financing Solutions for SMEs & Residential Markets: Creating scalable financing models—such as embedded finance solutions, green home loans, and point-of-sale financing for electrification—will provide much-needed capital to small and medium enterprises (SMEs) and residential customers.

3. Carbon Credit Registries & Tokenization: Establishing interoperable carbon credit registries and tokenization platforms will unlock liquidity in carbon markets and create tradable assets for institutional investors.

4. Aggregation Platforms for Mid-Scale Assets: Developing platforms to bundle smaller projects, like commercial rooftop solar and building retrofits, into investable portfolios will open the door for institutional investment in mid-scale energy assets.

5. Revenue-Backed Lending for Emerging Assets: Introducing revenue-backed lending and risk-sharing platforms for First-of-a-Kind (FOAK) technologies will de-risk investments and provide financing for innovative green assets.

Unlocking the Future of Green Capital

To meet Europe’s Net Zero goals, critical infrastructure must be built now. Innovators, financiers, and entrepreneurs have a unique opportunity to shape the future of green finance and unlock capital for the energy transition.

If you're working on innovative financial solutions, whether in data, financing, aggregation, or market access, with a focus on Europe, we want to hear from you. The next decade depends on the financial systems we build today. Contact us at robina@cventures.vc to explore how we can collaborate to accelerate the green economy transition.

1. At the core of foundational models lies advanced computing power, essential for both creating and training extensive models. This demand is met by hardware innovators like NVIDIA, whose GPUs can handle large datasets in parallel, due to their architecture that supports thousands of simultaneous operations. We also have cloud-based platforms such as Azure that enhance these capabilities with HPC, which includes GPU clusters.

Climate applications often require multimodal, multidimensional data and in some use cases (e.g., Chemistry) to input real-world data, where we see robotic automation used for tasks like lab testing as computing resources.

2. The next layer is the foundational model layer, composed of large-scale AI models trained on vast datasets across a large range of internet text, images, or other data types. Extensive training allows these models to develop an extensive understanding of human language, concepts, and the world. Because they are trained on such diverse data, foundational models can generalize this knowledge to perform different tasks without being explicitly programmed for each one. One of the most known models is OpenAI, which has developed several foundational models, including GPT language models, Codex for computer code, DALL·E for digital images, and more.

Horizontal foundational models often fall short for climate-related use cases due to their lack of specific data input and the deep learning instructions required for complex climate applications. Unlike LLMs, which learn the relationships between words, climate foundational models deal with multimodal, multidimensional data and need to comprehend the relationships between different dimensions. For example, energy foundational models might require training on intricate weather data. In the realm of chemistry, this could involve developing a standard language for chemical reactions, which necessitates an understanding of the highly complicated reactions initially.

3. The third layer is the infrastructure layer. It provides developers with a range of AI development tools, libraries, and frameworks. This layer offers the necessary resources and tools for developing and operating AI applications on a large scale.

When dealing with vertical infrastructure layers, such as those in the energy sector, partnerships like the Open AI Energy Initiative provide open frameworks. These are designed for energy operators and developers to build applications, encouraging the use of vertical foundational models in the sector. Additionally, we observe a trend where players initially build vertical foundational models but eventually enter the infrastructure and application layers, aiming to control the entire stack.

4. The fourth layer refers to the AI applications that utilize the underlying stack to accomplish specific tasks or address particular problems.

In the realm of climate applications, we see people building AI applications on both more mature horizontal foundational models and vertical-focused foundational models, depending on the use case's complexity.

Where can the greatest value be captured along the four layers?

Two use cases seem to be particularly interesting within the vertical climate layers:

1. Companies that already own the client relationship and have access to proprietary industry data that can effectively train on the relatively established horizontal infrastructure, package those models via applications, and ultimately deliver tremendous utility with fast time-to-value for customers.

2. Companies that build foundational models or infrastructure layers and can build a strong industry-specific data moat or the standard, open for other players to build infrastructure and/or applications on top of it.

This is a first outline for an ever-evolving framework, and there is still much to explore.

We are enthusiastic about continuing this discovery journey with you.

If you're building within any of the four layers mentioned above in a specific climate-related industry (like Energy, Mobility, Industry & Materials, Green Finance, Green Jobs, and Climate Risk), we encourage you to engage with us and challenge our discussed points. Likewise, if you're researching, educating, or exploring just as we are, we look forward to collaborating and exchanging thoughts, ideas, and frameworks.

PS: You can reach me at robina@cventures.vc

We can categorize the various forms of capital that exist in our world broadly into six categories: i) Natural Capital, ii) Human Capital, iii) Financial Capital, iv) Constructed/Manufactured Capital, v) Social Capital, and vi) Political Capital. Each of these categories will be significantly affected by the climate crisis.

Applying this fundamental framework to the established economic principle of scarcity helps us to gain clarity to what might initially appear as a blurry picture: It helps us identify which areas of our capital world are likely to undergo significant distribution as a direct consequence of our transition processes.

1. Natural Capital: Scarcity of Resources

The transition from a fuel-intensive to a material-intensive energy system places unprecedented demand on our natural resources, especially on our minerals. Achieving global net-zero targets by 2050 necessitates a sixfold increase in mineral inputs by 2040. However, this market is fraught with challenges: high geographical concentration of resources, geopolitical insecure situations, high exposure to climate risks, dirty exploration practices, and extremely long development lead times, just to name a few. Innovations in exploration technology, supply chain resilience, market transparency, mining efficiency, recycling technologies as well as efficient access to metals/materials and new materials discovery technologies are exciting areas we are going to explore in more detail within this blog.

2. Human Capital: Scarcity of Skilled Workforce

The renewable energy sector's workforce is expected to double in the next decade, yet finding the necessary skilled labour remains a significant challenge. Upskilling and reskilling hundreds of thousands, educating a new workforce, and leveraging technologies to make this process more efficient are essential strategies for building the workforce of the future. We will explore this field in more detail and are excited to discuss it with you.

3. Financial Capital: Scarcity of Financial Capital & Impact on our Financial System

The scale of investment required for the energy transition is staggering. Current targets under the Paris Agreement suggest a need to triple global energy transition investment to over $5 trillion annually by 2050. How to provide access to financial capital for renewable generating assets, and storage, within a fast enough manner to reach our net zero goals is a critical financial challenge, as many of the new technologies are not yet bankable. This is an exciting space to build in and we will explore this area with you in more detail.

At the same time, we will see our current financial system being completely disrupted by the change in the value of the underlying assets affected by climate change. Climate risk adaptation and insurance solutions are needed to respond to this disruption, and this will be an additional area of exploration.

4. Constructed/Manufactured Capital: Scarcity of Green Energy/Electrification & Grid Infrastructure

Our built environment must evolve to support a decentralized grid where power flows are bidirectional, and electricity generation is ubiquitous. This requires a dramatic increase in grid infrastructure development and a push for energy storage solutions as well as innovations in electrification of processes. This will be part of our deep dives.

5. Social Capital: Scarcity of Knowledge

The transition to green energy is as much a social movement as a technological one. Building widespread support for sustainable practices, fostering community engagement in renewable projects, and mainstreaming environmental, social, and governance (ESG) standards are pivotal. Despite the importance of the topic, for now, we will not cover it here yet.

6. Political Capital: Scarcity of Attention

Policy frameworks and government initiatives are crucial for enabling the energy transition. Investments in renewable energy, regulatory support for green technologies, and international cooperation on climate goals are essential levers of political capital. There is a lengthy discussion about the government not paying enough attention to this burning challenge. I added this point here for a full picture of the model, but given that we are mainly focusing on technological opportunities, we will not cover this in-depth here for now.

Let’s start the ride.

xx

Robina

PS: This blog will adopt a monthly format for now. And I am always eager to explore, discuss, and exchange ideas on the topics mentioned above with like-minded people, like you!