This episode features Dylan Patel (founder of SemiAnalysis) and Dwarkesh Patel discussing how AI compute economics are driving extreme centralization: OpenAI and Anthropic are on track to control the majority of the world’s compute within a few years, reshaping global capital allocation, interest rates, and the distribution of economic power.
Two labs will soon control most of the world’s compute
OpenAI and Anthropic are absorbing a rapidly growing share of marginal compute: ~30% of new compute in 2024, projected 40–50% in 2025, and over half by end of 2025.
At the start of 2024, OpenAI had ~2 GW and Anthropic <2 GW; by year-end both exceed 5 GW (3–4× growth). If the 3×/year trend continues, combined lab compute reaches ~18 GW by end of 2027 and ~54 GW by end of 2028.
Newer chips (GB300, TPU v7, Trainium3) deliver 3–5× performance per watt vs. prior generations, so the labs’ share of usable FLOP grows even faster than their share of watts.
By late 2028, if trends hold, the two labs could control most of the world’s usable compute.
SpaceX is emerging as a major new compute builder leasing to the labs; OpenAI is designing its own chips, Anthropic is deploying Google TPUs via Fluidstack.
$6B in fab capex enables $1T+ of end revenue
Producing 1 GW of compute (e.g., Nvidia Vera Rubin) requires ~55K N3 wafers, 6K N5 wafers, 170K DRAM wafers per year.
Fab tooling for 1 GW/year costs ~$3–4B; with cleanrooms/shell, ~$6B fab capex yields 1 GW/year.
That 1 GW generates $100B/year in AI revenue at current lab revenue-per-megawatt ($50M/MW for Anthropic).
Over 5 years, $6B fab capex → >$1T cumulative end revenue (100× leverage), even after subtracting opEx, data center capex, R&D, and middlemen.
Supply chain (mirrors for ASML EUV, etc.) cannot react instantly — “whip effect” means years of lag before capacity catches up to price signals.
Even if labs generate hundreds of billions in revenue, total ecosystem capex (fabs, data centers, power, semiconductors) is ~$2T+/year, far exceeding lab cash flows; hyperscalers and debt markets still fund the bulk.
Compute prices will rise if the labs outbid everyone
Today anyone can profit at $10–15M/MW (download open weights, run vLLM/SGLang, serve on OpenRouter).
Labs already generate $50–60M/MW (Anthropic) and rising; to capture 70%+ of 2028 compute (~100 GW combined), they must bid $25–50M/MW.
SpaceX and Meta, with balance-sheet-funded compute and no external customer lock-in, can auction capacity to the highest bidder (Anthropic/OpenAI), driving prices up.
Regulation (safety holds on releases, data center bans in NY/TX/OH) could stall lab revenue-per-MW growth, capping their ability to outbid.
If labs internally possess far-better models (e.g., “Mythos 2”) but cannot deploy them, revenue-per-MW plateaus and centralization slows.
Which layer will capture most of the surplus?
Most value flows to end users (Jane Street, Meta) who extract far more per token than labs capture.
App layer has captured little so far; model layer flipped from negative to massive positive gross margins (~$50M/MW now, heading toward $100M+/MW).
Hardware/fab layer captured early surplus (2023); memory vendors now earn more than TSMC.
As lab revenue-per-MW rises, they will bid up compute prices, pulling surplus back up the stack; memory and substrate prices react fastest.
Datacenter regulation and safety holds could slow AI
New York banning data centers, Texas moratoriums, Ohio property-tax schemes reduce supply and raise costs.
Safety-driven non-release of best models (OpenAI pausing training, Anthropic withholding “Mythos 2”) limits external deployment and revenue growth.
If labs cannot release best models, revenue-per-MW growth slows → they cannot sustain $50M/MW bids → centralization curve bends down.
Internal deployment of withheld models for R&D (inference optimization, architecture search) may partially offset, but external value capture is capped.
Labs are shifting compute from inference to R&D
Non-consensus view: labs will allocate less compute to inference over time, more to training/research.
Current split: ~50% research (architecture/data/hyperparam search), 10% development (final training runs), 40% inference.
Pre-training runs use <200 MW for ~2 months; RL uses less peak capacity. Most fleet sits in research.
As revenue-per-MW rises (e.g., $30M → $70M), the opportunity cost of selling inference grows; boards/executives will redirect marginal MW to internal R&D to chase AGI/RSI.
Evidence: Anthropic’s monthly compute additions keep rising while ARR growth plateaus → marginal MW increasingly goes to R&D.
China gets <10% of new compute, but its labs need less
2022: US ~45–50% of new watts, China ~30–35%. Now: US ~70%, China <10% of new AI data center watts.
China ~30 GW total AI compute by 2028 (mostly smuggled/legacy chips); domestic fabs (SMIC, CXMT) ramp in 2027–28 to add 5–10 GW/year of lower-quality chips.
2029: China could add 50 GW (some foreign-purchased), but quality-adjusted may equal ~20 GW of US chips.
Leading US lab in 2028 may have more effective compute than all of China in 2029–30.
Chinese models (Kimi, ByteDance Seed) remain competitive despite 10–50× less compute; gap matters less pre-RSI but widens post-RSI.
Export controls + US financial depth (YOLO startup funding) vs. Chinese state subsidies create divergent trajectories.
$5–10T/year capex by 2030 → sovereign debt crisis risk
100 GW/year at current prices = ~$5T IT capex + $1–2T data center/power + supply chain = ~$7–10T/year incremental capex by 2030 (~1/10 world GDP, ~1/3–1/4 US GDP).
Hyperscalers (Google, Microsoft, Amazon, Meta) fund ~half via cash + debt; all now raising hundreds of billions in debt.
Marginal borrower (Anthropic/OpenAI) willing to pay 20%+ for debt because revenue-per-MW justifies it → spreads widen for everyone.
Meta issuing at 5–6% today; could pay 8%+ → 250–300 bps spread increase across economy.
Banks suffer (liabilities reprice faster than assets); equity discount rates rise → non-AI stocks (J&J, railways) crater; developing countries (Pakistan, Nigeria) face Volcker-style defaults.
US can tax data centers; others cannot. Corporate income tax <10% of federal revenue; payroll/income taxes shrink with automation → fiscal crunch.
Interest rates could reach tens of percent in 2030s
If economy doubles yearly (fully automated labor + capital), real interest rates approach growth rate → 10–100%+.
All non-AI equities → near zero (DCF collapse); mortgages unavailable; government debt service > tax revenue unless AI taxed.
Opportunity cost of capital shifts: pension spending vs. robot-factory-that-builds-robot-factories.
Even AI stocks (Micron, Hynix) trade at 2–3× earnings because everything should trade at low multiples in high-growth/high-rate world.
World’s future workforce concentrated in two labs
Effective AI labor at frontier: compute growing 4–5×/year × algorithmic efficiency 3×/year = ~10×/year effective population growth.
OpenAI/Anthropic: 10M AI workers → 100M → 1B → >8B (human population) within a few years.
Without RSI, 10×/year continues; with RSI, 100–1000×/year.
Two labs consuming majority of compute = majority of “minds” concentrated in two private entities.
Economies of scale in training (amortize across billions of sessions) + scarce compute markup + deployment data flywheels all reinforce centralization.
No clear decentralized equilibrium unless: AI progress stalls, heavy government regulation, or government nationalization (which Patel distrusts).
Saving grace so far: users capture most surplus (Jane Street, podcasters), but labs increasingly internalize compute as internal R&D returns exceed external prices.
