Aug 25, 2026 · 1h 17m
Frontier AI labs poised to centralize global compute by 2028
Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028
The sheer scale of AI infrastructure spending is shifting from a tech sector trend to a macroeconomic force capable of altering global interest rates and sovereign debt stability.
- 1Leading AI labs like OpenAI and Anthropic are rapidly centralizing the vast majority of global computing power.
Don't miss
Dylan Patel explains how the massive demand for debt to fund AI data centers could drive up global interest rates and trigger a second Volcker shock.
The brief
Dylan Patel, founder of SemiAnalysis, joins Dwarkesh Patel to outline the staggering capital expenditure driving AI infrastructure and how the global compute supply chain is being rapidly monopolized.
Frontier labs like OpenAI and Anthropic are projected to centralize the vast majority of global compute by 2028, prioritizing internal research and development over external inference.
This massive demand for data center debt could trigger macroeconomic shocks, driving up global interest rates and crowding out capital for other critical industries.
As physical bottlenecks like ASML tools and wafer fabrication equipment tighten, the extreme centralization of AI capabilities risks creating an economy with unprecedented barriers to entry.
What was said on this episode
34 statements · 18 positive · 9 negative · 2 mixed · 5 neutral
Global compute capital expenditure will exceed $2 trillion in 2028.
“As we go out into 28, it's going to be more than $2 trillion.”
Listen at 1:07
Frontier AI labs will increase annual spending from tens of billions to trillions by decade’s end.
“the labs are going from companies that spend, you know, tens of billions of dollars a year, to hundreds of billions of dollars a year, to forecasting to spend trillions of dollars a year even towards the end of the decade.”
Listen at 1:18
Anthropic began turning a profit in the second quarter.
“Anthropic started turning a profit in Q2”
Listen at 1:45
Anthropic and OpenAI will take 40–50% of newly added compute next year.
“Anthropic. OpenAI are taking as much as 40 to 50% of compute next year.”
Listen at 4:17
Anthropic and OpenAI will control most usable global FLOPs by late 2028 if current trends continue.
“by the time you're in like towards the end of 2028, if this trend continues, which I see nothing that's stopping it. You've got them just controlling most of the usable flops in the world on their own.”
Listen at 6:47
An ASML EUV tool could be bought for $400 million and resold for over $1 billion.
“if anyone had like $400 million and the ability to convince ASML to sell them an EUV tool, they should totally just go buy one and wait, wait, wait, and then sell it for north of a billion dollars.”
Listen at 10:17
The supply chain will produce roughly 100 ASML tools in 2030.
“the hundred is roughly still the right number for 2030, 100 ASML tools for 2030.”
Listen at 11:47
Supply-chain expansion will not accelerate substantially over the next three years because capital is constrained.
“I don't think it'll happen this year, I don't think it'll happen next year. I don't think it'll happen the year after. Because the world is capital constrained.”
Listen at 12:03
AI labs’ cash flows cannot yet fund the required compute infrastructure investment.
“The labs have not yet gotten to the point where their cash flows can fund this stuff.”
Listen at 12:55
Anthropic and OpenAI will acquire more compute, but not at current prices.
“that they will continue to gobble up more of the compute, but ultimately they can't do it at current pricing or anywhere close to it.”
Listen at 15:55
Capturing 70% of global compute in 2028 requires labs to pay $25–50 million per megawatt.
“They do have to start paying 25, 30, $50 million a megawatt to really gobble up 70% of the world's compute in 2028”
Listen at 16:03
Most compute capacity is contracted before construction begins.
“most compute is contracted well before it's built.”
Listen at 22:55
Meta and SpaceX can choose between internal use and renting compute to Anthropic or OpenAI.
“now Meta and SpaceX have this optionality of looking around and being like, is my internal Use case going to make me more money or should I go out there and sell it to Anthropic OpenAI at crazy margins?”
Listen at 24:08
Anthropic or OpenAI could exceed $70–80 million revenue per megawatt by late 2027.
“I think, I think it could get to, you know, higher than that, like 70, $80 million a megawatt blended across a company, if not higher.”
Listen at 25:56
Improving AI models increase the value generated by those models.
“as long as the model gets better, the value generated out of it gets better.”
Listen at 28:32
In a takeoff scenario, Anthropic could keep its best model six months ahead of public models.
“in a takeoff scenario, why would anthropic not have their best model six months ahead of what is externally available?”
Listen at 29:16
Frontier AI labs will allocate a declining share of compute to inference.
“the labs are going to allocate less and less compute to inference over time”
Listen at 30:33
Anthropic and OpenAI will prioritize building AGI over distributing profits to shareholders.
“the obvious answer from Anthropic and OpenAI and not just at the executive level, but also their board, is go build AGI”
Listen at 31:12
Anthropic is currently increasing the share of compute devoted to research and development.
“they are factually increasing their compute towards R and D today.”
Listen at 33:16
Global AI compute additions should reach roughly 90–100 gigawatts in 2029.
“29 should be on the order of 90 to 100”
Listen at 33:49
China will account for under 10% of incremental new AI compute.
“China domestically still continues to have sub 10% of incremental new computer.”
Listen at 35:29
China will have at most roughly 30 gigawatts of AI compute by 2028.
“China will have like 30 gigawatts of AI compute or less by 2028.”
Listen at 35:38
China will add 5–10 gigawatts of domestically produced AI chips in 2028.
“now they're incrementally adding, you know, 5, 10 gigawatts in just 2028 of domestically produced chips.”
Listen at 36:17
China’s AI compute deployment will accelerate sharply.
“China, China is definitely going to hockey stick.”
Listen at 37:19
China could add 50 gigawatts of AI compute in 2029.
“China in 2029 can do 50 gigs.”
Listen at 37:57
Fifty Chinese gigawatts could provide performance equivalent to 20 American gigawatts.
“that 50 gigawatts is really worth as much as 20 gigawatts in America or from American chips.”
Listen at 38:03
China provides substantially greater subsidies once its financial system targets an industry.
“once Chinese financial systems choose an industry to focus on, they'll subsidize it a hell of a lot more.”
Listen at 39:41
The described interest-rate and financing effects will occur before the singularity.
“I think this all happens before singularity.”
Listen at 59:16
Rising interest rates will severely damage equity markets.
“as interest rates go up, equity markets get pummeled.”
Listen at 1:01:29
Memory will perform well, but memory stocks should not increase tenfold.
“memory is going to do great. But, you know, memory, memory stocks shouldn't, you know, 10x or whatever.”
Listen at 1:01:56
If recursive self-improvement occurs, AI compute will centralize because labs use it most effectively.
“if you believe in rsi, you believe in the labs are the most effective user of compute and can generate the most value from the compute, then the only thing that's going to happen is centralization of compute.”
Listen at 1:10:01
Deployment, scale economies, and recursive improvement all push AI toward centralization.
“All of these things point to centralization.”
Listen at 1:12:22
Without major slowdown or regulation, AI development will produce extreme resource concentration.
“Unless AI progress slows down, unless governments Regulate the fuck out of it. This is all that happens.”
Listen at 1:14:03
Anthropic can generate hundreds of millions per megawatt by using compute internally.
“what if Anthropic can just generate hundreds of millions of dollars per megawatt by using that compute internally? And that's what's happening.”
Listen at 1:16:28
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.