Dylan PatelDwarkesh Podcast
Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028
Hosted by Dwarkesh Patel
- Artificial Intelligence
- AI Infrastructure
- Capital Expenditure
- China
- Data Centers
- Geopolitics
- Interest Rates
- Monetary Policy
- Regulation
- Semiconductors
Briefing
Executive Summary
Dylan Patel, founder of SemiAnalysis, discusses with host Dwarkesh Patel the accelerating centralization of AI compute in the hands of OpenAI and Anthropic. He projects that by 2028, these two labs will control most of the world's incremental compute, driven by their superior revenue generation per megawatt (up to $50M now, potentially $100M+). This concentration is enabled by massive capital expenditure (over $2T by 2028) and a supply chain that cannot keep up, leading to price increases and a bullwhip effect. Patel also highlights regulatory and political constraints that could slow this trajectory, and the potential for a sovereign debt crisis as AI investment crowds out other borrowing. He argues that the labs will increasingly allocate compute to R&D rather than inference, and that China, despite its manufacturing prowess, will remain far behind in compute due to export controls and financial system differences. The conversation underscores the profound economic and geopolitical implications of AI's growth.
Analysis confidence: High
Expert perspective
Claims & Outlook
Claim
OpenAI and Anthropic are taking 40-50% of incremental compute next year, and by end of 2028 they will control most of the world's usable flops.
- Evidence
- Paraphrase
- Condition
- if current trends continue
- Uncertainty
- Qualified
Claim
Anthropic's revenue per megawatt has reached as high as $50 million, and could reach $70-80 million per megawatt by end of 2027.
- Evidence
- Paraphrase
- Uncertainty
- Speculative
Claim
The labs are allocating less compute to inference and more to R&D, contrary to consensus.
- Evidence
- Paraphrase
Claim
China will have 30 gigawatts or less of AI compute by 2028, and its domestic chips are less efficient than American ones.
- Evidence
- Paraphrase
- Uncertainty
- Qualified
Claim
The AI buildout will require over $5 trillion of credit issuance by 2029, which will raise interest rates and crowd out other borrowing.
- Evidence
- Paraphrase
- Uncertainty
- Qualified
Claim
Regulation is slowing down AI labs more than open-source Chinese models, with examples like OpenAI not releasing Astra and stopping training for two weeks.
- Evidence
- Paraphrase
Claim
The value capture in AI is shifting: previously captured by hardware supply chain, now moving to model labs, but end users like Jane Street and Meta capture more value than the labs.
- Evidence
- Paraphrase
Claim
Compute prices will rise to $25-50 million per megawatt as labs outbid others for compute.
Analytical layer
System Analysis
These implications are system-generated analysis and are not statements attributed to the guest.
Industry
The concentration of compute in two labs could lead to a market structure where these labs have significant pricing power and influence over the AI value chain.
Analysis confidence: High
Macro
The massive capital expenditure and debt issuance for AI infrastructure could lead to higher interest rates and potential sovereign debt stress in vulnerable economies.
Analysis confidence: Medium
Policy
Regulatory actions aimed at slowing AI development could have unintended consequences, such as slowing innovation and giving an advantage to less regulated regions.
Analysis confidence: Medium
Competitive
The ability of labs to generate high revenue per megawatt may allow them to outbid other players for compute, potentially marginalizing other AI developers.
Analysis confidence: High
Technology
The shift of compute allocation from inference to R&D could accelerate AI capabilities but may also reduce the availability of AI services to the broader market.
Analysis confidence: Medium
Original context
About the episode
"Every force is screeching towards centralization."