OpenAI and Anthropic are on track to absorb roughly one third of all newly added AI compute in 2026, about half in 2027, and most of the world's effective capacity by the end of 2028. SemiAnalysis founder Dylan Patel laid out the full math on the Dwarkesh Podcast on August 25, and the takeaway is blunt: the AI race no longer runs on parameter counts or benchmark leaderboards. It runs on revenue per megawatt.
Two Labs Are Taking a Third of All New AI Compute
At the start of 2026 OpenAI held roughly 2 gigawatts of compute and Anthropic less than 2 GW. Both are on track to cross 5 GW by the end of the year, a three-to-four-fold jump in twelve months. One gigawatt is roughly the output of a large nuclear power plant, so the two labs are effectively adding three nuclear plants of new capacity a year.
Globally, newly deployed compute is about 30 GW this year, about 50 GW next year, and about 70 GW by 2028. The one-third share consumed by the two labs comes straight out of that pipeline. There is also a hidden multiplier: a watt of compute bought today is not equal to a watt installed two years ago. GB300, TPUv7 and Trainium3 deliver roughly three to five times the performance per watt of the previous generation. That is why Patel frames the 2027 milestone as half of new compute but far more than half of effective compute, and why the concentration looks worse than the raw numbers suggest.
The New Winning Metric: Revenue per Megawatt
A year ago the token business lost money. OpenAI running GPT-4 on Hopper chips earned negative gross margin on every token, and Anthropic was burning investor capital. That has reversed. Anthropic turned profitable in the second quarter of 2026, and OpenAI is reportedly set to follow in the third. With the GPT-5.6, Opus 5 and Fable 5 generation of models, both labs now clear the $10-15 million per-megawatt cost line by a wide margin. Anthropic leads at roughly $50 million per megawatt.
The competition is no longer about hoarding GPUs. It is about how much revenue a megawatt can generate. Earn more, and you can outbid everyone for scarce capacity, which buys better models, which earns even more. The flywheel is the whole game.
How the Flywheel Compounds Concentration
Ownership and usage have split. According to Epoch AI, about 71% of global compute is still owned by the five big clouds, but usage is concentrating in two labs. The clouds are landlords who buy land and build data centers; OpenAI and Anthropic are the two tenants renting most of the buildings.
A third player type has emerged: build-first operators such as Meta and SpaceX, who construct capacity before signing tenants. Musk is reportedly selling compute at $40 billion per gigawatt against a fair value of around $15 billion, and Anthropic is renting SpaceX data-center capacity at $1.25 billion per month, per WSJ. Both labs are also starting to build their own infrastructure, OpenAI with self-designed chips and Anthropic with TPUs deployed through Fluidstack, while Stargate hit its 10 GW target ahead of schedule and Anthropic locked in a multi-gigawatt TPU order with Google and Broadcom starting in 2027.
Most Compute Goes to Research, Not Training
Lab workloads split roughly 50% research, 10% development and 40% inference. The largest bucket is research, not training. Anthropic trained Mythos with under 200 megawatts at a single site for about two months, and reinforcement learning used even less. Most of the capacity sits in distributed sites where data movement, not compute, is the binding constraint. Training more does not automatically mean training better.
The counterintuitive part is that the inference share will keep shrinking. At $30-40 million per megawatt, devoting 40% of capacity to selling tokens makes sense. At $60-70 million, a board faces a real choice: keep selling tokens and pay dividends, or pull the megawatts back into R&D and chase AGI.
The $11 Trillion Question: AGI or Dividends
SemiAnalysis estimates cumulative AI capital expenditure of about $11 trillion between 2024 and 2029, of which roughly $6 trillion can be funded from earnings and $5 trillion must be borrowed. All of it draws on the same credit pool, which pushes rates up; Meta is issuing debt at 5-6% and would reportedly pay up to 8%. The tension is structural because both labs are heading toward IPOs. Investors want dividend-grade returns; the labs want to pour every new megawatt into research. The signal is already visible: Anthropic's monthly new compute keeps rising while annualized revenue growth has flattened, meaning new megawatts are going to R&D, not revenue. This is the same consolidation pressure reshaping the compute market.
Regulation can break the loop. New York is limiting data centers, Texas has a moratorium, and Ohio paused data-center tax exemptions on May 27, 2026. Withholding frontier models also caps the flywheel: OpenAI has not released Astra, and Anthropic is holding back a next-generation model in safety review. A model that cannot ship cannot lift revenue per megawatt, which means it cannot justify outbidding rivals for compute.
Dwarkesh closed with a sobering projection: frontier compute grows four to five times a year while the compute needed to reach a given capability falls to about one third per year, so effective AI labor grows roughly tenfold annually. A lab holding ten million effective AI workers this year would hold a hundred million next year and a billion the year after. A single company's effective workforce could soon exceed the human population, the same kind of always-on agent labor that persistent agents are starting to deliver.
What This Means for You
Three practical takeaways. For builders: GPU supply, availability and price now track two balance sheets, so do not build a stack that depends on single-lab access. For enterprises: API pricing follows revenue per megawatt, so benchmark your own unit economics per token rather than trusting list prices. For investors: watch the research-versus-inference split and revenue per megawatt, not flops or leaderboard positions. Three years ago the race was about parameters; two years ago it was about benchmarks. Today it is about how much each megawatt earns, and the real question is no longer who builds AGI first but who gets to call it.
Frequently Asked Questions
Which two companies are consuming the most new AI compute?
OpenAI and Anthropic. They are on track to absorb about one third of newly added global compute in 2026, roughly half in 2027, and most effective capacity by the end of 2028, according to SemiAnalysis founder Dylan Patel.
Why is revenue per megawatt the decisive metric in AI?
Because compute now costs about $10-15 million per megawatt while leading models earn $30-50 million per megawatt. Whoever earns more can outbid rivals for scarce capacity, train stronger models, and widen the gap.
Will AI compute keep concentrating in a few companies?
The trend is accelerating, but regulators are pushing back with data-center limits, tax pauses and withheld frontier models. The flywheel only spins if the best models keep shipping.
