ExpertsFMDylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028
Dwarkesh Podcast · August 25, 2026
5 Selected Lessons:
reasoning_pattern · 3 case · 1 heuristic · 1
Compute centralization driven by monetization asymmetry
Source
Dwarkesh Podcast — Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028 (August 25, 2026) · at 3:24 ▸ · at 13:13 ▸ · at 16:28 ▸
When this applies
when allocating a scarce input across buyers with differing abilities to monetize it
The takeaway
In any market for a scarce input, the buyer with the highest marginal value extraction will progressively capture a disproportionate and growing share of new supply.
Why it works
Because frontier labs generate far more revenue per unit of compute than anyone else, they can outbid all other buyers, pulling an ever-larger share of new supply toward themselves even as total supply grows.
One level up
Whenever one actor's ROI from a resource compounds faster than others', expect market share concentration to compound too, not stay stable.
Where it stops holding
This stops working when buyers' relative ability to monetize the input converges, or external constraints (politics, financing, regulation) cap the leading buyer's purchasing power.
Detail
approach: Compare revenue-per-unit generated by different buyers of the same scarce resource to forecast future allocation share.
signals:
- rising revenue per megawatt at the leading buyer
- stagnant or negative margins for other buyers at current pricing
- new entrants selling compute to the highest bidder rather than using it internally
rationale: Markets allocate scarce capacity toward whoever can pay the most, and that willingness-to-pay is set by who extracts the most value from the resource.
In their words
You've got them just controlling most of the usable, you know, flops in the world on their own.
Shift from cash-funded to debt-funded CapEx as scale outpaces cash flow
Source
Dwarkesh Podcast — Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028 (August 25, 2026) · at 44:49 ▸ · at 52:24 ▸ · at 54:15 ▸
When this applies
when an organization's growth ambitions exceed what its own cash generation can fund
The takeaway
When an industry's investment opportunity set grows faster than its internal cash generation, it will progressively escalate up the capital stack — first cutting returns to shareholders, then issuing debt — and this escalation has spillover effects on the cost of capital for everyone else.
Why it works
As CapEx needs grow into the trillions while cash flow grows more slowly, companies first cut shareholder returns (buybacks) and then increasingly turn to debt markets, eventually drawing capital away from other borrowers in the broader economy (crowding out).
One level up
A sector's financing stack (cash → reduced buybacks → debt → crowding out other borrowers) is itself an indicator of how much its growth has outstripped its underlying cash economics.
Where it stops holding
This stops working when revenue growth genuinely catches up to CapEx growth, restoring cash self-sufficiency, or when external capital markets have enough slack to absorb new debt without raising rates economy-wide.
Detail
approach: Track a sector's financing waterfall (internal cash → reduced shareholder returns → debt issuance → effect on broader credit markets) to diagnose how capital-constrained its growth ambitions are.
signals:
- companies halting buybacks to redirect cash to CapEx
- rising debt issuance by major players
- other sectors beginning to compete for the same credit pool
rationale: Hyperscalers... now don't generate cash. They actually spend everything on CapEx. And in addition, they raise debt and spend everything on CapEx.
In their words
you still end up with north of $5,000,000,000,000 of credit that need to be issued for this $11,000,000,000,000 plus bill.
Rising discount rates collapse non-AI equity valuations even in a booming economy
Source
Dwarkesh Podcast — Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028 (August 25, 2026) · at 57:54 ▸ · at 1:01:28 ▸
When this applies
when a macro shock changes the discount rate used to value all future cash flows across an economy
The takeaway
Index-level stability can mask severe individual-stock devaluation when a macro shock (like rate increases) hits segments differently based on how growth- versus cash-flow-dependent their valuation is.
Why it works
A rising discount rate reduces the present value of any firm's future cash flows; steady, low-growth 'stable cash flow' businesses are hit hardest because their value depends almost entirely on discounted long-duration cash flows rather than near-term growth.
One level up
Aggregate metrics can conceal large redistributions happening underneath — always decompose an index into its components' sensitivity to the underlying macro driver.
Where it stops holding
This stops working when rate increases are offset by equally fast growth in those firms' actual cash flows, preserving valuations.
Detail
approach: Separate a firm's valuation driver (growth optionality vs. steady discounted cash flow) to predict differential impact of rate changes.
signals:
- stock priced mainly on long-duration steady cash flows
- no offsetting growth acceleration from AI
rationale: even though the stock market as a whole might be doing fine... any individual stock will probably have just cratered in value, especially the Buffett, like Berkshire type... thirty year type cash flows.
In their words
if interest rates rise, the discount rate increases, which means that the discount cash flows of all equities are greater, which means that even though the stock market as a whole might be doing fine, like S and P 500 will be fine. Yeah. Any individual stock will probably have just like cratered in value,
Most value from AI is captured by end users and the hardware supply chain, not the model layer
Source
Dwarkesh Podcast — Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028 (August 25, 2026) · at 16:28 ▸ · at 19:43 ▸ · at 1:14:42 ▸
When this applies
when diagnosing which layer of a technology value chain is actually capturing the economic surplus it creates
The takeaway
In a value chain, the layer capturing the most economic surplus is not necessarily the one creating the most technical innovation — surplus tends to flow to whichever layer has the least competitive pressure or most differentiated downstream application.
Why it works
End users like Jane Street or Meta extract far more value from using a model (trading edge, ad engagement gains) than the model provider captures in subscription/API revenue, because competitive pricing and demand elasticity keep model-layer margins below the user's realized value.
One level up
Always map 'who creates value' separately from 'who captures value' in a stack — the two routinely diverge, and strategy should target the capture layer, not just the creation layer.
Where it stops holding
This stops working when model providers develop true pricing power (e.g., exclusive capability advantage) that lets them capture a larger share of downstream value, which Patel notes may be starting to happen as labs reallocate compute inward.
Detail
rule: Expect the most value to accrue to end users and scarce upstream hardware/components, not necessarily to the most visible 'AI product' layer.
threshold: Jane Street reportedly captures $300-500M/megawatt in value versus Anthropic's ~$100M/megawatt realized revenue from serving Jane Street.
rationale: Jane Street is capturing $300,000,000 per megawatt or $500,000,000 per megawatt... that's the one saving grace is that the rest of the economy maybe profits so much more from Anthropic.
In their words
But at the end of the day, the reason they've gone to a $100,000,000 per megawatt is because Jane Street is capturing $300,000,000 per megawatt or $500,000,000 per megawatt.
Export controls effectiveness on China's compute stock
Source
Dwarkesh Podcast — Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028 (August 25, 2026) · at 34:19 ▸ · at 38:36 ▸
When this applies
when evaluating whether a deliberate policy restriction has materially changed a competitor's resource trajectory
The takeaway
A sustained restriction on a scarce enabling input, even if imperfectly enforced, can compound into a multi-year structural gap between competitors.
Why it works
By restricting chip/tool exports since 2022, the US shifted the share of world AI compute deployment from a ~45/35 US/China split toward 70% US / <10% China, compounding over years into a durable compute-stock gap.
One level up
Judge policy interventions not by immediate effect but by their compounding effect on stock accumulation over years, since flows compound into large stock differences.
Where it stops holding
This stops working when the restricted party successfully builds domestic substitute capacity (as China is beginning to via SMIC/CXMT) or finds alternative supply channels (smuggling, resale via third countries).
Detail
situation: Pre-2022, China was adding 30-35% of world AI compute; the US was adding 45-50%.
decision: US imposed export controls restricting chip/tool sales to China.
outcome: By the data cited, the US deploys ~70% of world AI watts and China deploys <10%, with China projected to have only ~30 gigawatts (of lower quality) by 2028, versus a single leading lab potentially exceeding that.
lesson: Sustained supply restrictions on a critical compounding input can produce a large structural capability gap, though domestic substitution can eventually erode it.
In their words
Since 2022, we've had big regulations against China and a dramatic increase in America. So, today, 70% of Watts are being deployed in America. And China is really a very small number.