Everyone in the Compute Stack Is Buying at the Top at the Same Time

Modal goes long

Modal, the serverless GPU cloud, has reached a $1 billion annualized run rate, CEO Erik Bernhardsson told the Latent Space podcast — and it raised a $750 million Series D at a $16 billion valuation along the way. Two-thirds of revenue comes from Modal Functions; Sandboxes, the product for running LLM-generated code, have grown to roughly a third of revenue, "a result he didn't predict at launch." and the company has spun up nearly three billion sandboxes to date. m:…8579

The business model is duration transformation: pooling demand across thousands of customers to sell elastic compute against spiky AI workloads. "A typical capacity contract gives a customer a fixed number of GPUs for a fixed period. Plot capacity against time and you get a rectangle. Real workloads behave differently." To back the promise, Modal has moved from hyperscaler spot instances to three-plus-year capacity deals requiring significant upfront payments, working with roughly 25 cloud providers — and is now, in Bernhardsson's framing, effectively long on GPU prices. m:…8579

Bernhardsson says he does not expect an imminent collapse in GPU prices but acknowledges the financial risk of holding long-term capacity contracts — the classic buying-at-the-top problem for anyone who locked in large supply at current high prices. He also rejects the idea of cornering the compute market, citing the Hunt Brothers' failed attempt to corner silver. m:…8579

The physics of the shortage

The Dwarkesh Patel episode's Intel tour gives the supply side. AI is driving computing demand 40 to 50% above maximum industry capacity, according to the episode's Intel speaker, and only three foundries remain at the advanced node. The speaker's estimate: the industry can likely increase leading-edge wafer output by roughly 2x, but not 10x — and meaningful increases in semiconductor production capacity are unlikely before 2030, given tool and fab lead times. ASML builds fewer than 100 EUV tools a year. m:…9273

Intel's own lesson from the 10nm debacle is schedule discipline: the delay came from aggressive 'hyperscaling' and EUV tools that were not ready, and the company now limits innovation per node, using 'phase gate' maturity metrics to decide when a process moves from pathfinding to development. "I think what we've learned is that you don't want to do too much innovation in one node. Because the more innovation that you did, the more complex it becomes." the Intel speaker said. Foundries also run on brutal operating leverage — roughly 80% of costs are fixed once a fab is built — which is why scarcity rent flows to whoever has the tools: TSMC is on track for about $50 billion in CapEx this year against Intel's $20 billion. m:…9273

The peak call

Then the peak call. On Bloomberg Surveillance, Dawson, chief investment officer at Edge Wealth, said of semiconductor earnings: "The 27 is peak earnings." The semiconductor index is expected to grow earnings by more than 100% in 2026, and AI infrastructure stocks saw EBIT margins expand from 16% to 42% this year — the setup, in Dawson's telling, for counter-cyclical multiples: "You put the lowest multiple on the highest earnings." m:…7314

His broader warning: tech stocks tend to peak before the adoption cycle peaks — "The stocks are likely going to peak before the adoption peaks, which is something that we see in every technology adoption cycle." Credit markets are already pricing the risk: spreads on AI-related assets and data center bonds have widened, only 25% to 50% of planned data centers may get built given power constraints, and Micron capacity arriving in mid-2027 and 2028 raises questions about sustaining 85% gross margins. m:…7314

Who holds the duration risk

Put the three together and the margin question sharpens. Modal is long GPUs on three-year paper. The foundry layer cannot print its way out before 2030, so scarcity rent persists — which is exactly the condition that produces 42% EBIT margins and 100% earnings growth, the numbers Dawson says mark the top. The tightening is not that demand is falling; it is that everyone in the stack is simultaneously locked into the high-price regime, and the first one to blink on utilization or pricing takes the loss first. m:…8579 m:…9273 m:…7314

For a tech executive buying or selling compute, the actionable read is about contract shape, not price direction: Modal's rectangle problem is everyone's rectangle problem. Fixed rectangles against spiky workloads are fine while utilization climbs; the risk sits entirely in the duration mismatch. m:…8579