Two buyers, one price: Patel says 2027's new compute is already half spoken for
SemiAnalysis puts 40-50% of 2027's new capacity under OpenAI and Anthropic contracts, with revenue per megawatt now well clear of deployment cost. That makes inference a residual market.
The Engineer · Build desk
What happened
- SemiAnalysis founder Dylan Patel says signed contracts already assign 40-50% of 2027's new global compute to OpenAI and Anthropic.
- On his numbers the pair account for about 30% of capacity added in 2026, much of it built by others and leased back.
- He projects 70-80% of new compute going to the two by 2028, at roughly 100 GW of combined physical capacity.
- Anthropic turned profitable in Q2 2026 and the market expects OpenAI to follow in Q3, according to the same account.
Why it matters
- cost Anyone renting inference capacity is bidding against buyers earning three to five times the cost of a megawatt, so the rental floor rises and the mid-size buyer absorbs it.
- constraint When most new capacity is pre-committed, the open market a non-lab buyer shops in is whatever is left over, and it thins each year even while total capacity doubles.
- contradiction The bull case for per-megawatt revenue depends on models the labs have chosen not to ship, which is the same mechanism that would stall the revenue the bidding assumes.
- capability Balance-sheet builders can now hold megawatts as a tradeable position rather than a service, and price them by what the top bidder can afford.
Two of Patel's per-megawatt figures do not sit comfortably together. He puts Anthropic's peak at $50M per megawatt against a deployment cost of $10-15M [11][10], a multiple of roughly three to five [23], where serving GPT-4 on Hopper had run at negative gross margin [12]. He then gives an end-2027 base case of $70-80M per megawatt blended across the whole company [6]. But only about 40% of a lab's compute budget serves inference, with the rest in training and research [16]. Divide the blended figure by that share and the serving megawatts have to be clearing $175-200M each [24]. Either the mix swings hard back toward inference, or those two numbers are not measured on the same base.
The floor is easier to pin down. Anyone can stand up open weights, Kimi on vLLM or SGLang, at $10-15M per megawatt and clear positive revenue through OpenRouter [14], so a megawatt has a clearing price that owes nothing to what a frontier lab earns on it. Patel's own arithmetic needs that price to move to $25-50M for the two labs to hold 100 GW in 2028 [3][4], which is 67% to 233% above the top of today's base [8]. Price then tracks who can pay most at the margin rather than what the capacity cost to build.
What breaks the story is not demand. From the middle of 2026, marginal revenue per new megawatt flattened even as absolute capacity kept climbing [17], and Patel reads that as marginal megawatts going into research rather than serving [5]. For a buyer of tokens that is the worse reading: compute absorbed into internal R&D does not appear as available capacity at any price. The shelving of Astra and of Anthropic's Model 2 works the same way [18]. The models that would justify $100M per megawatt are the ones not being shipped.
Supply will not rescue the residual buyer either. New York State has banned data center construction and Texas has paused approvals, with Ohio making data centers pay property tax on surrounding land [20], and the tooling chain upstream cannot be widened on a two-year horizon even with capital thrown at Carl Zeiss [22]. The silicon going in during 2026 runs three to five times the performance per watt of the last generation [15], so half the physical megawatts is more than half the usable FLOPs. Meta and SpaceX can build with no end customer and then choose between using the capacity and reselling it at $25-50M per megawatt [19].
Every number here is Patel's, relayed second hand in a summary of the conversation [7]. The structural claim survives that. In 2024 and 2025 the hardware chain took nearly all the gross margin while the model layer ran negative [21], and compute was at least sold to whoever turned up with money. On these figures, within one to two years most of the world's compute is owned by or serving two labs [2], and everyone else trades in what they leave.
What to watch
- Whether Astra and Anthropic's Model 2 are released, since withheld frontier models are the single input that moves blended revenue per megawatt.
- Whether the mid-2026 plateau in marginal revenue per added megawatt persists into 2027 or breaks upward.
- Whether open-market capacity actually clears above the $10-15M per megawatt base, which is the test of the bid-up thesis.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence20
- Adoption30
- Hype gap+55
- Incentives65
- Confidence26
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Dylan Patel, founder of SemiAnalysis, says that based on already-signed contracts, Anthropic and OpenAI will take 40-50% of global new compute added in 2027.
- [2]
Patel judges that by the end of 2027 half of all new global compute will belong to Anthropic and OpenAI, and because total global compute roughly doubles each year, new compute soon constitutes most of the installed base, making it real within one to two years that most of the world's compute is owned by or serving those two labs.
- [3]
For the two labs to secure 100 GW in 2028, the unit price of compute must rise from $15M per megawatt to $25M, $30M or even $50M per megawatt, so they can outbid every other buyer.
- [4]
If the trend continues, Patel expects the two labs to absorb 70-80% of global new compute by 2028, reaching around 100 GW of combined physical capacity.
- [5]
Against the consensus that most compute goes to inference, Patel argues labs will allocate a declining share to inference over time and more to training and internal R&D.
- [6]
Patel's base case, absent regulatory blockage, is $70-80M per megawatt of revenue for Anthropic and OpenAI on a blended whole-company basis by the end of 2027, with more than $100M per megawatt not exaggerated if models keep improving.
- [7]
The material is a third-party written summary of a podcast conversation between Dwarkesh Patel and SemiAnalysis founder Dylan Patel on AI lab economics and the compute market.
- [8]
Moving the price of capacity from $15M to $25-50M per megawatt is a 67% to 233% increase.
- [9]
In 2026, roughly one third of newly deployed global compute ends up serving OpenAI and Anthropic, including capacity physically built by others and leased to them; the two account for about 30% of the year's new compute.
- [10]
The base cost of deploying compute is around $10-15M per megawatt.
- [11]
With OpenAI serving GPT-5.6 and Anthropic serving Opus 5, revenue generated per megawatt now far exceeds the $10-15M per megawatt base cost, with Anthropic peaking at $50M per megawatt.
- [12]
Serving GPT-4 on Nvidia Hopper was negative gross margin for OpenAI.
- [13]
Anthropic began turning profitable in Q2 2026, and the market expects OpenAI to turn profitable in Q3 as Codex and GPT-5.6 volumes ramp.
- [14]
Anyone can currently deploy open-weight models such as Kimi with vLLM or SGLang at a $10-15M per megawatt cost and generate positive revenue through routes like OpenRouter, so base compute pricing has already begun to rise.
- [15]
Chips deployed in 2026, including GB300, TPUv7 and Trainium3, deliver three to five times the performance per watt of the previous generation, so the labs' share of usable FLOPs exceeds their share of physical capacity.
- [16]
Lab compute budgets split roughly 60% training and 40% inference, and within the training 60% about 50 points go to research and 10 to development; a single pretraining run is under 200 MW for about two months.
- [17]
From the middle of 2026, marginal revenue growth from newly added compute plateaued even though absolute monthly compute kept rising, implying marginal megawatts went to R&D rather than inference.
- [18]
OpenAI did not release Astra and stopped training for two weeks, and Anthropic did not release Model 2, widely believed to be the next Mythos generation, after safety evaluation; with the best models withheld, revenue per megawatt could stall or fall back.
- [19]
Meta and SpaceX have the balance sheets to build compute with no end customer and then choose between using it themselves and reselling at $25-50M per megawatt to Anthropic or OpenAI; Elon Musk's sale of B300 capacity to Google was priced at $4B per gigawatt.
- [20]
New York State has banned data center construction, Texas has paused approvals, and Ohio requires data centers to pay property tax on surrounding property; reduced supply pushes compute prices higher.
- [21]
In 2024 and 2025 the hardware supply chain captured almost all the gross margin in AI while the model layer ran negative gross margin; in 2026 the model layer crossed over.
- [22]
The world is capital-constrained on tooling: earlier in 2026 the industry assumed 100 EUV tools a year by 2030 would suffice and has since revised up, but even putting $10B into Carl Zeiss requires the whole supply chain to expand in step, which is hard to achieve across 2026, 2027 and 2028.
- [23]
Anthropic's $50M per megawatt peak is 3.3 to 5 times the $10-15M per megawatt deployment cost.
- [24]
If end-2027 blended revenue is $70-80M per megawatt across all compute and only 40% of megawatts serve inference, the serving megawatts imply $175-200M of revenue each.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- dev.to两家实验室的算力垄断之路：Dylan Patel 详解 AI 经济学
1 article · August 25, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.