In the first three months of 2026, AI startups raised $255.5 billion. That is as much money as the entire year before it, cleared in a single quarter. Two-thirds of it went to three companies. OpenAI, Anthropic, and xAI took $172 billion between them; the remaining $83.5 billion was split across 1,543 other deals. So the median AI startup in early 2026 was competing for the scraps left after three labs ate first.
The reflex read is that the three companies at the front of that line are the trade. They have the best models, the users, the revenue. The reflex is not stupid. It is just crowded, and in venture, crowded and expensive are the same word. The better question in the middle of 2026 is not which model wins. It is what every model has to pay for no matter which one wins, and where that payment is hardest to compete away. Answer that and you stop backing a horse and start owning a piece of the track it runs on. The catch, and the reason most of the picks-and-shovels advice is lazy, is that not all of the track is worth owning. Some of it gets repaved every twelve months.
Two-thirds of the money is chasing three companies
Start with the concentration, because it is the fact that reorganizes everything after it.
AI took roughly 60 percent of all US venture capital in 2025, with AI startups raising over $200 billion globally. Then the first quarter of 2026 doubled down: $255.5 billion into AI in three months, per PitchBook, matching the full prior year in a quarter. Of the 1,546 deals that made it up, three of them, OpenAI, Anthropic, and xAI, accounted for 67.3 percent of the capital. The other 1,543 deals divided $83.5 billion.
Those three labs are not cheap. Anthropic raised $65 billion in a Series H on May 28, 2026, at a $965 billion post-money valuation, the largest disclosed private AI round on record and the highest private valuation in the dataset. OpenAI announced $110 billion in new investment on February 27, 2026, at an $840 billion post-money mark, and the round closed larger, near $122 billion and $852 billion, at the end of March. Two private companies, each valued within reach of a trillion dollars, neither of them public, both funded in part by the same hyperscalers and chipmakers that sell them compute. When Amazon puts $50 billion into your customer and also books the cloud revenue, investment and vendor financing start to blur.
Here is where we grant the bull case its strongest form, because it is strong. This is not 1999 with no revenue. Anthropic's run-rate revenue went from $14 billion in February 2026 to over $30 billion in April to roughly $47 billion by late May, the fastest revenue ramp in the history of software. OpenAI's annualized revenue topped $25 billion by the end of February. The two highest-revenue categories in AI, coding and the ChatGPT franchise, are each crossing double-digit billions, and close to a dozen more startups are on their way past $100 million. The revenue is real. Anyone calling the whole thing a hallucination has not looked at the receipts.
So the question is not whether the labs make money. They do, at a rate almost nothing in business history matches. The question is whether you should pay a near-trillion-dollar private mark to own a slice of it. And the sharpest money on the other side of that trade says the thing you are buying may not stay scarce.
GMO's Chancellor and Grantham, January 2026: thanks to DeepSeek and other new entrants, AI "may even become a commodity product, like internet broadband." Grantham's line for the mega-rounds: the moats are being drained to fill the war chests.
That is the crowded trade: pay the highest private valuations ever recorded, funded partly by your customer's suppliers, for a capability the best contrarians think is on its way to becoming broadband. Maybe the winner takes most and today's price looks cheap in 2030. But that price assumes a single company holds a moat that is visibly narrowing, in the most crowded corner of the market. There is a better place to stand.
Stop asking which model wins
Every token served in 2026, from OpenAI or Anthropic or xAI or DeepSeek or whatever is on top in 2028, has to be trained and served on the same four things: power to run the chips, memory to feed them, the chips themselves, and somebody's data to ground the answer. Those are the tolls. They get paid on every token regardless of whose name is on the model. The toll-taker never has to pick the winner.
This is the model-agnostic layer, and it is where value has accrued in every prior compute build-out. The railroads went bankrupt; the towns they connected did not. The lesson venture keeps relearning is that the company burning the capital is rarely the company that keeps the return. Sequoia's David Cahn made the sharp version of this point: in a build-out, value flows to the product builders who ride falling compute costs, and away from the infrastructure owners exposed to high rates of capital incineration.
But buy infrastructure, not applications is where most of the advice stops, and it is only half an insight, because it treats all infrastructure as one thing. It is not. Some of the toll layer is a genuine, physics-bound scarcity that takes years to build and cannot be conjured by the next chip. Some of it is a shortage that the next chip erases. The whole trade is knowing which is which. So rank the model-agnostic layer not by how loud the demand is today, but by how hard the scarcity is to compete away.
The trap in the sold-out sign
Before the ranking, kill the bear case that everyone reaches for, because it is wrong right now, and being wrong about it in the correct direction is the most valuable thing in the research.
The standard short on compute is depreciation. Michael Burry made it the headline in late 2025: hyperscalers are hiding losses by pretending GPUs last five or six years when they are obsolete in two or three. It is a clean story. It is also, as of the middle of 2026, contradicted by the market it describes.
H100 one-year rental contract prices did not fall in 2026. They rose almost 40 percent, from a low of $1.70 an hour in October 2025 to $2.35 by March 2026, per SemiAnalysis, the most authoritative independent voice on compute supply. On-demand rental capacity is sold out across every GPU type, with everything coming online through August and September of 2026 already booked. H100s from 2022 contracts are being renewed at the exact rate they were signed at three years ago, some on four-year terms running through 2028. When CoreWeave's original 2022 H100 contract came up for renewal, it re-booked at 95 percent of the original price. A three-year-old chip holding 95 percent of its rental rate is not a depreciating asset. It is a scarce one.
So the obsolescence short is broken, and the crowd now piling into compute is sold out, back the truck up feels vindicated. Here is the trap. The sold-out sign is a supply shock, not a moat.
The reason H100s are scarce is not that demand will hold forever. It is that the memory that feeds them, high-bandwidth memory and the DRAM behind it, is in acute shortage, with AI absorbing something like 70 percent of the industry's memory output, on top of a Blackwell backlog running to millions of units. That is a supply-chain bottleneck rather than proof of permanent end-demand. Every GPU generation runs the same arc: premium while it is the newest thing, then decline once it is two generations back. Nvidia's B200 already delivers roughly seven times more tokens per dollar than the H100, and Rubin-class silicon is six to twelve months out from the middle of 2026. The sold-out sign is real. It is also a clock, and the clock is running.
Sold out through September is a wonderful trade. It is a terrible thing to underwrite a ten-year fund against.
Rank the scarcity by how hard it is to build
Here is the layer, ranked by durability of the scarcity rather than by loudness of the demand. This is the analysis, and the ranking is the thesis.
| Layer | What it is | Verdict | Why it holds, or doesn't |
|---|---|---|---|
| Power & grid | The megawatt, the substation, the interconnect | Buildout | A chip you can buy. A substation you wait four-plus years for. Physics-bound, model-agnostic, and it does not get lapped by the next accelerator. |
| Memory & supply | HBM, DRAM, advanced packaging | Buildout / boom | The actual cause of the 2026 GPU shortage. A tight oligopoly, years of lead time to add capacity, and every chip generation needs more of it. |
| Data infrastructure | Unstructured-data plumbing, retrieval, pipelines | Boom | Every model and every agent has to be fed and grounded. Software margins, model-agnostic, no depreciation cliff. Where a16z is putting real money. |
| GPU rental | Metered accelerator capacity, neoclouds | Boom, shot clock | Roaring and sold out today. But it is a supply shock, and the asset itself is two generations from being lapped. Great trade, dangerous moat. |
| Vertical apps | Coding, healthcare, legal, the ChatGPT franchise | Boom, unproven | The highest revenue in AI and the fastest growth. Also the least-defended, with thin wrapper margins and a brutal failure rate. |
| Frontier labs | OpenAI, Anthropic, xAI, and the rest | Bubble-risk on price | Real, fast revenue. Historic valuations. The most crowded trade in venture, funded by their own suppliers, selling a capability the contrarians think is commoditizing. |
Read the ranking top to bottom and it is a single idea: the further the scarcity is from the accelerator and the closer it is to physics, the more durable the claim.
Power sits at the top for a reason we have argued before. The build-out spent three years treating compute as the scarce thing. The real shortage was electricity and the grid behind it. More than 2,000 gigawatts of generation sat in US interconnection queues at the end of 2025, roughly twice the country's entire installed fleet, with a median wait past four years, per Lawrence Berkeley National Laboratory. In the PJM grid where the data centers cluster thickest, the 2025 to 2026 capacity auction cleared at $269.92 per megawatt-day against $28.92 the year before, a ninefold jump in one auction, with data-center load named as the driver. A new chip ships every year. A new gigawatt does not. That is the most physics-bound scarcity in the entire stack, and it is indifferent to which lab wins. We made the full case in The Megawatt Is the Moat.
Memory is the same shape, one layer below the GPU. The thing making H100s scarce is not the H100. It is the high-bandwidth memory stacked next to it, made by a handful of suppliers who cannot add capacity on an AI timeline. When a shortage is caused by a bottleneck, the durable position is the bottleneck itself rather than the thing it throttles. Our read on the customer-funded second source in silicon runs on the same logic in a different link of the chain.
Data infrastructure is the one place the smartest generalist fund is both talking and doing. a16z's Jennifer Li calls unstructured, multimodal data a generational opportunity, on the estimate that 80 percent of corporate knowledge lives in unstructured formats and that data entropy is the new limiting factor for AI companies. That is not talk-only. Li co-leads a dedicated infrastructure allocation reported around $1.25 billion, backing Fivetran, dbt, Reducto, MotherDuck, ElevenLabs, and fal. Every model needs feeding, the plumbing carries software margins, and none of it cares which model is on top. Of the layers with real venture conviction behind them, this is the one where the saying and the spending line up.
GPU rental is the trade everyone can see, and the one with the shortest clock. Own it for the boom, size it for the reversal, and do not confuse a supply shock for an annuity.
Vertical applications are the highest-reward, lowest-durability layer. a16z's Alex Immerman argues vertical AI in healthcare, legal, and housing already reached $100 million-plus in revenue within a few years, and that 2026 is when the collaboration layer becomes the moat as multi-human, multi-agent workflows raise switching costs. It is a good argument, and it is a general partner talking his own book. The figures are self-sourced and not tied to named companies, and the counter-argument is that the switching costs are interface frictions that the next model release evaporates. The revenue here is the realest in AI. Whether it is durable or a wrapper-margin illusion is genuinely unresolved, and anyone who tells you they know is selling something.
One honest gap. The question we set out to answer covered the whole open field, including robotics, AI-for-science, and defense. Capital is flowing there too. Anduril reportedly doubled its valuation to $61 billion in a May 2026 raise; Isomorphic Labs and robotics names like Skild and Physical Intelligence are drawing mega-rounds. We could not verify durable unit economics for any of them in this pass, so we are watching that shelf and holding off on a ranking. Calibrated beats confident-and-wrong.
The bear we are underwriting
Fair is fair. The ranking above leans on infrastructure, and the strongest case against infrastructure is the one we have to carry rather than wave away.
It is the return-mismatch. AI end-revenue runs in the tens of billions of dollars a year. The data-center and energy build-out to serve it runs in the trillions over five years. Cahn's arithmetic, the version that has aged best, takes Nvidia's run-rate, doubles it for total data-center cost, doubles it again for a 50 percent gross margin, and lands on the end-user revenue the build-out has to generate to pay for itself. In 2024 he put that figure at $200 billion. By December 2025 he had tripled it to $600 billion, and since then 2026 capex has run higher than the levels he modeled, so the gap is wider now than when he wrote it. Data-center capex alone is set to clear a trillion dollars in 2026. The revenue, real and fast as it is, is not within an order of magnitude of that yet.
We will call it what Cahn will not. He is careful to say this is a time lag, not a bubble, and he stays bullish on adoption; the numbers are his, the word bubble is ours. But a time lag becomes a stranding when the timeline slips, and it is slipping. The consensus on when AI pays for this hardware has walked back to the 2030s, and Cahn's own named risk is that hyperscaler capex today ends up being outdated. If the payoff is a decade out and the chips are obsolete in three years, some large fraction of what is being poured into concrete and silicon right now will never earn its return.
This is not a reason to avoid the toll layer. It is the reason to rank it the way we did. The return-mismatch is what threatens the GPU and the neocloud, the assets with a depreciation clock. It threatens the megawatt and the memory bottleneck far less, because those are scarce whether the payoff lands in 2028 or 2034, and they get re-used by whatever chip comes next. The bear case is real. We answered it by climbing down the stack to the scarcity that survives it.
If you are writing checks in the middle of 2026
The thesis, made concrete.
- Underweight the front of the line. The frontier-lab mega-rounds are the most crowded, most expensive, most vendor-financed trade in the market, and the capability may be commoditizing under them. If you own the labs, own them for distribution and revenue, not for a monopoly the contrarians already doubt.
- Buy the scarcity physics protects. Power, grid interconnection, and the memory-and-packaging bottleneck are the model-agnostic positions that a new accelerator cannot erase. They are slow, unglamorous, and years deep. That is the point.
- Own GPU rental as a trade with a shot clock. The sold-out market is real and the cash flows are excellent right now. Underwrite it to the shortage in front of you, and assume the next generation re-prices the last one.
- In applications, price the switching cost. The revenue is the realest in AI and the least defended. Pay for genuine data gravity and workflow lock-in. A wrapper that the next model release makes free is worth what it sounds like.
- Read the interconnect, not the launch. The signal that tells you who has durable capacity in 2028 is a signed power-purchase agreement, a memory-supply contract, a slot in a grid queue. Count the contracted megawatts.
Our Call
By December 31, 2027, the H100 one-year rental contract rate falls back below its pre-spike floor of $1.70 an hour, the level it sat at in October 2025 before the 2026 shortage. The whole 40 percent spike round-trips, and the sold-out market of mid-2026 is revealed as a supply shock rather than a moat. The consequence for capital: the accelerator itself is a picks-and-shovels boom on a shot clock, and the durable claim was always one rung away from it, on power, memory, and data, not on the chip.
The case: the spike was driven by a high-bandwidth-memory shortage and a Blackwell backlog rather than durable end-demand, and every input that made H100s scarce is temporary. Memory capacity is being added. The backlog clears. Rubin-class silicon ships within a year of this writing, and B200 already serves roughly seven times more tokens per dollar. By the end of 2027 the H100 is two generations old, and two-generations-old GPUs have declined in every prior cycle. The scarcity moves to the next chip and the memory behind it. It does not stay on the H100.
What proves us wrong: H100 one-year contract rates hold at or above $1.70 an hour through the end of 2027. That would mean the installed base, CUDA lock-in, and real sustained end-demand set the price, not a transient shortage, and that compute rental is a more durable moat than we are crediting. The live counter-signal is already on the board: some H100 contracts are being renewed on four-year terms through 2028 at their original rates. If that pricing holds across the market and not just in legacy renewals, the shot clock we are calling runs longer than we think.
Settles: December 31, 2027, against the SemiAnalysis GPU rental index.
Frequently asked questions
What is the best AI startup category for a VC to invest in for mid-2026?
On a risk-adjusted basis, not the frontier labs. In the first quarter of 2026, AI startups raised $255.5 billion and two-thirds of it went to just three companies: OpenAI, Anthropic, and xAI. That is the most crowded and most expensive trade in the market. The durable place to deploy is one layer down, in the model-agnostic scarcity every lab has to pay for regardless of which one wins: power and grid capacity, the memory and packaging bottleneck behind the chips, and the data infrastructure that feeds every model. Rank those by how physics-bound the scarcity is, because the harder it is to build, the longer the advantage lasts.
Is AI in a bubble in 2026?
Not as a single yes-or-no. The honest read is that AI is a real technological revolution with localized bubble dynamics. Revenue is genuinely large and fast: Anthropic's run-rate went from $14 billion in February 2026 to roughly $47 billion by late May, and OpenAI passed $25 billion annualized. The bubble risk is not the revenue. It is the price and the concentration: near-trillion-dollar private valuations for a handful of labs, funded partly by their own suppliers, against a data-center build-out running into the trillions while end-revenue is still in the tens of billions. Deploy by layer rather than making one up-or-down call on all of AI.
Why not just invest in OpenAI or Anthropic?
You can, but you are paying the highest private valuations ever recorded, Anthropic at a $965 billion post-money mark in May 2026 and OpenAI near $852 billion, in the most crowded corner of venture, for a capability the sharpest contrarians think is commoditizing. GMO's Jeremy Grantham and Edward Chancellor argue that with DeepSeek and other new entrants, frontier AI may become a commodity product like broadband, and that the moats are being drained to fill the war chests. The revenue is real. The entry price assumes a monopoly that is already in question.
Are GPUs a good AI investment if they depreciate so fast?
The depreciation short is the popular bear case, and as of mid-2026 the market contradicts it. H100 one-year rental prices rose about 40 percent, from $1.70 an hour in October 2025 to $2.35 by March 2026, on-demand capacity is sold out through September 2026, and three-year-old H100s are re-contracting near their original rates. But that strength is a supply shock driven by a memory shortage and a Blackwell backlog, not proof of a permanent moat. The next chip generation re-prices the last one. Own GPU rental as a trade with a shot clock, not as a durable position.
Where does value accrue in the AI stack?
Historically, to the product builders who ride falling compute costs and to the owners of a scarcity that a new chip cannot erase, not to the capital-intensive layer burning the money. In 2026 that points to power and grid interconnection, the high-bandwidth-memory and packaging bottleneck, and the data-infrastructure layer, all of which get paid no matter which model wins. The frontier labs absorbing most of the capital are the least defended position on a valuation basis, and raw GPU capacity is durable only while the current shortage lasts.
References and research base
This piece was built on a multi-source research pass run July 5, 2026, with claims adversarially verified before use. Figures are dated; forward-looking claims are marked as our thesis, not reported fact.
- Funding concentration. PitchBook, "Q1 2026 AI funding blows past 2025 total with three deals accounting for 67% of capital" (May 12, 2026): $255.5 billion into AI in Q1 2026; OpenAI, Anthropic, and xAI at 67.3 percent ($172 billion); $83.5 billion across the other 1,543 deals. The match to all of 2025 is against a $254.4 billion full-year figure and is close, so read it as matched, not a landslide. AI's roughly 60 percent share of 2025 US VC and the $200 billion-plus global total corroborated by Crunchbase, the OECD (61 percent global), and GMO.
- Lab valuations and rounds. Anthropic's own Series H release and the Epoch AI funding dataset: $65 billion raised, $965 billion post-money, May 28, 2026, $47 billion run-rate revenue. OpenAI: $110 billion announced February 27, 2026 at $840 billion post-money, closing near $122 billion / $852 billion at end of March (CNBC, Reuters, Epoch AI).
- Lab revenue. PitchBook and Anthropic (run-rate $14B February, $30B April, roughly $47B late May 2026); The Information via PitchBook (OpenAI roughly $25B annualized, end of February 2026). Coding and ChatGPT each crossing double-digit billions; Cursor/Anysphere and Claude Code each past $2 billion run-rate by early 2026.
- GPU rental and the obsolescence question. SemiAnalysis, "The Great GPU Shortage: Rental Capacity" (April 2, 2026): H100 one-year rental up roughly 40 percent to $2.35/hr by March 2026 from a $1.70 October 2025 low; on-demand sold out through August and September 2026; renewals at signing rates, some four-year through 2028. CoreWeave's 2022 contract re-booked at 95 percent (CNBC, November 14, 2025). The supply-shock reading, HBM/DRAM shortage and Blackwell backlog, B200 roughly 7x tokens per dollar, Rubin six to twelve months out, is our interpretation of why the strength is point-in-time.
- The return-mismatch. Sequoia, David Cahn, "AI's $600B Question" (June 20, 2024) and "AI in 2026: The Tale of Two AIs" (December 3, 2025): the run-rate doubled twice methodology; the $200 billion-to-$600 billion revenue question; tens of billions per year of end-revenue against trillions over the coming five years; the AGI window walking back to the 2030s and capex outdated as the named risk. Cahn frames this as a time lag rather than a bubble. The word bubble and the stranding interpretation are ours, not his. Data-center capex over $1 trillion in 2026 per Dell'Oro/Futurum.
- Commoditization. GMO, Edward Chancellor and Jeremy Grantham, "Valuing AI: Extreme Bubble, New Golden Era, or Both?" (January 28, 2026): the commodity-product line, on DeepSeek and new entrants. Grantham, the moats are being drained to fill the war chests (Fortune, May 19, 2026).
- The data-infrastructure and application theses. a16z, "Big Ideas 2026: Part 1": Jennifer Li on unstructured data as a generational opportunity and the roughly $1.25 billion infra allocation (Fivetran, dbt, Reducto, MotherDuck, ElevenLabs, fal); Alex Immerman on vertical AI at $100M-plus revenue and the 2026 multiplayer moat. Both are general partners describing their own investments; the revenue figures are self-sourced. The 80-percent-unstructured figure is a long-standing IDC/Gartner estimate.
- The segmented frame. Wang and Chen, "Boom, Bubble, or Buildout?" (arXiv:2606.01575, May 2026): AI as a real technological revolution with localized bubble dynamics, analyzed layer by layer. A non-peer-reviewed preprint; treat as one rigorous voice, not settled consensus.
- Power. Our own prior reporting in The Megawatt Is the Moat (June 20, 2026), sourced to the IEA "Energy and AI" report, Lawrence Berkeley National Laboratory's "Queued Up" (2,000-plus GW in queue, four-plus year median wait), and PJM/Utility Dive (the $28.92-to-$269.92/MW-day capacity-auction jump).
Source-quality note
The hard figures here, funding totals, valuations, revenue run-rates, rental prices, and capex, are drawn from primary and authoritative sources and dated to when they were reported. Three inputs are weaker and flagged in place: the a16z theses are investors describing their own book; the AGI-timeline walk-back rests on a single source; and the "Boom, Bubble, or Buildout?" framework is a preprint. Several widely repeated claims were checked and discarded, including the line that GPU compute is already commoditizing with no pricing power, which the sold-out rental data in this very piece contradicts. The robotics, bio, and defense figures are reported by their outlets and were not independently verified in this pass; we name them and decline to rank them for that reason. The ranking of the layers, the reading of the sold-out market as a supply shock rather than a moat, and Our Call are this publication's thesis, not reported fact, and should be read as such.