Dwarkesh PodcastDylan Patel – Two labs will soon control most of the world's workforce
CHAPTERS
- 0:00 – 3:24
Lab compute becomes a macroeconomic driver (profits flip from negative to positive)
Dwarkesh and Dylan set the frame: US growth has recently been heavily driven by AI infrastructure, and frontier labs are rapidly increasing their share of new compute. Dylan argues the labs’ unit economics have improved sharply, with inference moving from negative gross margins to strong profitability—enabling reinvestment into training.
- •AI infrastructure has been a major contributor to recent US GDP growth
- •Labs’ compute spending is ballooning from tens of billions toward hundreds of billions and potentially trillions
- •Anthropic reportedly turned profitable; OpenAI potentially approaching profitability
- •Cost basis framed as ~$10–15M per MW, while lab revenue per MW can be far higher
- •Profits from inference can be recycled into training capacity
- 3:24 – 5:52
Compute centralization accelerates: OpenAI + Anthropic approach half of new supply
They discuss how quickly frontier labs are absorbing incremental compute. Dylan forecasts OpenAI and Anthropic taking ~40–50% of new compute next year, reaching roughly half of incremental global compute by the end of next year, with the trend still accelerating.
- •OpenAI and Anthropic roughly 3–4x their compute over the year
- •~30% of incremental compute this year goes to the two labs; next year could be ~40–50%
- •New builders/lessors (e.g., SpaceX) may lease primarily to the highest bidders
- •Labs are also beginning to build/secure more of their own compute (custom chips, TPUs)
- •Because new deployments are more efficient, controlling incremental compute can mean controlling most usable FLOPs
- 5:52 – 8:01
Why “only” 70–100 new gigawatts/year? Supply-chain bottlenecks and bullwhip delays
Dwarkesh challenges whether global compute additions could be much larger given compute’s huge value. Dylan emphasizes physical and industrial constraints—especially lithography/tooling supply chains—where even massive capital can’t instantly scale output due to long lead times and coordination across many suppliers.
- •Newer chips deliver major performance-per-watt gains, compounding the impact of new deployments
- •Fab scaling constrained by specialized tooling (e.g., EUV) and upstream components (mirrors)
- •Arbitrage emerges when bottlenecks (turbines, EUV tools) become resellable scarce assets
- •Even aggressive assumptions still face slow supply-chain response (bullwhip effect)
- •“Private-equity everything” could accelerate, but requires coordinated scaling across the full stack
- 8:01 – 13:39
$6B fab capex → $1T+ end revenue? ROI logic meets financing and real-world frictions
Dwarkesh lays out a striking capital-efficiency argument: a relatively small amount of fab investment can underpin enormous downstream AI revenue over years. Dylan agrees on the direction but stresses the many additional costs (data centers, power, installation, opex) and the fact that the ecosystem is capital constrained relative to the scale of required buildout.
- •Illustrative claim: ~$6B in fab capex can enable ~1 GW/year of compute production
- •Downstream revenue per GW/year framed at ~hundreds of billions, compounding over time
- •Large “take rates” are absorbed by opex, data centers, power, and other intermediaries
- •Despite high returns, the overall system’s capex needs can still exceed labs’ cash flows
- •Capital constraints and long lead times prevent immediate scaling to theoretical demand
- 13:39 – 16:10
Compute prices rise when labs outbid everyone: who captures the markup?
They explore how compute pricing could reprice upward as labs monetize far more per MW than others. Dylan explains why most compute still transacts below extreme spot prices: it’s typically pre-contracted and financed, giving balance-sheet-rich players (Meta/SpaceX) unique optionality to build first and rent later.
- •If many actors can profit at $10–15M/MW, labs must pay far more to capture 70–80% of supply
- •Compute pricing could move toward $25–50M/MW for scarce incremental capacity
- •Spot-like deals (e.g., a high-price lease) reflect unusual conditions: already-built capacity
- •Most projects require customer commitments to raise financing; pre-contracting anchors prices
- •Meta/SpaceX can “hoard” compute using their balance sheets, then decide whether to use or lease it
- 16:10 – 18:23
Regulation and withheld model releases as an economic throttle
Dylan argues that policy and safety processes are already slowing external deployment and potentially revenue per MW growth—because top labs may not release their best models. This creates a scenario where labs’ ability to outbid others for compute depends not just on technical progress but on whether they can deploy and monetize it.
- •Claims of slowed releases/paused training and gated deployments reduce external monetization
- •If best models aren’t released, competitors catch up and revenue per MW can stall
- •Data center regulation (moratoriums, bans, new taxes/fees) can constrain supply and raise costs
- •Regulatory constraints could limit labs’ ability to pay top-of-market compute prices
- •Internal-only advantages could widen even if public progress appears slower
- 18:23 – 29:41
Where does AI surplus go? End users vs apps vs model layer vs hardware
They break down value capture across layers and how it’s shifted over time. Dylan notes that users (e.g., trading firms, large platforms) can extract more value than model providers, and that earlier in the cycle, hardware captured profits while model providers burned VC money—yet this is changing as model margins improve.
- •End users often capture outsized value relative to what they pay in tokens
- •The app layer has captured relatively little so far compared with expectations
- •Model layer shifted from negative gross margins to strong positive margins
- •Hardware and supply-chain pricing power can rise with a lag (bullwhip)
- •Value capture is dynamic and can rotate between chips, memory, hyperscalers, and labs
- 29:41 – 33:26
Inference vs R&D: why labs may allocate *less* compute to revenue over time
Dwarkesh asks whether public markets would pressure labs to maximize near-term inference revenue. Dylan presents a contrarian view: labs will increasingly divert marginal compute toward R&D/training because the strategic value of building better models dominates the short-term profits from selling tokens.
- •Non-consensus claim: over time, labs may reduce inference share and increase R&D/training share
- •Even if inference yields huge margins, training may offer higher long-run payoff (AGI race)
- •Evidence claim: recent revenue growth plateauing despite continued compute additions implies more R&D allocation
- •Internal use of best/fastest modes can be more valuable than external monetization
- •Inference becomes a means to fund and scale the training fleet rather than the end goal
- 33:26 – 34:15
Global compute growth curves and the looming capex stack (IT + data centers + power)
They estimate global additions (tens of GW per year rising toward ~100 GW/year) and stress that “AI capex” is often understated. Dylan highlights that power plants and data-center shells are long-lived assets that must be built ahead of demand, creating a much larger effective annual capital requirement.
- •Illustrative path: ~30 GW this year, ~50 next year, ~70 in 2028, ~90–100 in 2029
- •Performance improvements mean new deployments dominate effective global FLOPs
- •Total capex must include critical IT plus buildings and power generation
- •Long lead times mean future-year capacity requires earlier-year spending (turbines, plants, shells)
- •At scale, annual ecosystem capex could reach several trillion dollars
- 34:15 – 40:01
China’s compute: sub-10% today, domestic ramp later—yet quality-adjusted gaps remain
Dylan argues export controls and market structure have pushed most new AI data center watts to the US, leaving China with a small share of new deployment. He expects China to “hockey stick” later via domestic fabs and scaling, but notes domestic chips may lag leading-edge US/allied hardware, widening the quality-adjusted gap.
- •Since 2022, US share of deployed AI watts rose sharply; China is now sub-10% of new watts
- •By 2028, China might have ~30 GW or less of AI compute (headline number)
- •China’s ramp depends on fabs/equipment (SMIC/CXMT), export controls, and domestic tooling progress
- •China can scale manufacturing quickly; 2029 incremental +50 GW is framed as plausible
- •Domestic compute may be worth far less on performance terms than US/allied compute
- 40:01 – 48:27
Why Chinese labs can look close despite far less compute: research vs training bottlenecks
They discuss how much of a lab’s compute budget goes to research experimentation versus the biggest training runs. Dylan claims single large pretraining runs may only require a few hundred megawatts at peak, with larger fleets often underutilized for “one giant run” due to coordination and methodology constraints—helping explain why lower-compute actors can seem competitive for a while.
- •Compute budgets split into inference, research experimentation, and development—not just pretraining
- •Claim: major pretraining peaks may be sub-~200 MW at a time for a couple months
- •Multi-site training and very large coordinated clusters are operationally difficult
- •RL/rollouts don’t scale linearly in usefulness; more compute isn’t always better instantly
- •Over time, continual learning and automation may shift more budget toward training proper
- 48:27 – 1:07:36
Credit markets, crowding out, and the path to a sovereign debt crisis
Dwarkesh and Dylan connect AI’s high returns to rising interest rates via massive credit demand for compute buildouts. They explore how higher rates can crowd out other industries and countries, compress equity valuations through higher discount rates, and potentially trigger emerging-market defaults—well before any “singularity.”
- •Modeling claim: through 2029, ~$11T capex with ~$5T financed via debt across the ecosystem
- •Rising credit demand can lift spreads even if policy rates don’t move proportionally
- •Higher discount rates can crush valuations for “stable cash flow” equities (and reprice the whole market)
- •Developing countries with frequent refinancing needs become vulnerable to Volcker-like shocks
- •Crowding out and politics become key constraints on how fast AI capex can scale
- 1:07:36 – 1:16:52
AI labor concentration: two labs could control a large fraction of the world’s ‘workforce’
Dwarkesh frames a future where effective AI labor supply (capability-adjusted agent population) scales extremely fast, potentially surpassing human population equivalents within a decade. Dylan agrees that most forces—economies of scale, deployment learning, and recursive advantages—push toward centralization, creating a stark governance dilemma between concentrated private power and heavy-handed regulation.
- •Effective AI ‘population’ could scale ~10x/year even without full RSI, given compute and efficiency trends
- •If agents reach human-remote-worker capability, labor-equivalent concentration becomes extreme
- •Economies of scale in training amortize improvements across billions of sessions/users
- •Being slightly ahead lets labs monetize scarce compute better, reinforcing dominance
- •Society may face a choice between centralized corporate control vs government throttling—both risky
