My Conversation with Dylan Patel and Stacy Rasgon

Summary

• Financing is becoming a supply constraint, but Rasgon thinks the next year or two are probably okay. Patel says neoclouds increasingly need strong offtake or chip-vendor backstops, while even major AI-lab contracts can carry 13–15% debt costs. Rasgon argues those brakes might create “stronger for longer” conditions by preventing overbuilding.
• Substrates and memory may constrain AI deployment more than power through 2028. Patel calls substrates the “biggest bottleneck of all the bottlenecks” and estimates nearly 50 gigawatts of global data-center construction next year, with on-site generation helping alleviate power constraints. Rasgon thinks 2028 will probably remain tight; 2029 becomes a demand question.
• Higher infrastructure prices are shifting profits across the AI supply chain. Patel puts construction above $60 billion per gigawatt, versus roughly $40 billion in 2024, while selling ready capacity can command much more than contracting ahead. Rasgon says “inflation saved the industry during COVID”; the current contest is over which suppliers retain margins and which customers absorb rising costs.
• Nvidia’s margin defense depends on performance improving faster than prices. Rasgon cites Blackwell-to-Rubin reference costs rising from about $25 billion to $40 billion per gigawatt against a claimed 10x performance gain. He would worry if hypothetical Feynman were only 20% better: “You need to keep giving your customer value.”
• Patel expects Nvidia’s share of CoWoS and HBM to fall while earnings still at least double next year. His team says OpenAI’s first chip beat Blackwell on InferenceX running open-source Chinese models, sharpening the question of whether design becomes “more commodity like software.” Rasgon emphasizes annual releases, scale and financing as defenses; both connect depressed valuations with competitive fears.
• Enterprise adoption may generate demand primarily through productivity gains and purchased tokens. Rasgon describes Claude completing a week’s modeling work in 80 minutes, although he considers the evidence for returns encouraging and early. Patel expects batching and model-scale economics to favor cloud consumption: “Most people are just going to be buying tokens.”

Deep dive

1. Capital constraints may stretch the buildout

• For Rasgon, financing is a leading concern, but “we’re probably okay” for the next year or two. CoreWeave’s roughly 15% unsecured-debt yield reflects secondary-market trading, rather than its current borrowing cost. It signals what fresh unsecured financing might cost; whether revenues catch up depends on the sustainability of today’s compute scarcity.
• Patel says neoclouds account for nearly half the buildout, yet financing increasingly requires high-grade offtake. Even Anthropic/OpenAI-backed projects face 13–15% debt pricing. Hyperscalers cannot spend north of $2 trillion next year, he argues, while smaller customers struggle to finance capacity outside those channels.
• Nvidia’s backstop can guarantee a $3.50 rental price, “the cloud making zero money,” with revenue sharing when others rent above it. Patel cites 50% shares for Sharon AI and Fermi, versus 10% for Lambda’s Anthropic cluster. Broadcom stretches Anthropic’s payments over several years, effectively “a rent to own.”
• Rasgon’s upside is “stronger for longer”: constraints could postpone supply rather than repeat the dark-fiber overbuild. Patel says Meta has $1.4 trillion of compute and data-center commitments; the broader market can finance capacity against those commitments.

2. Materials and clean rooms may bind before power

• Patel ranks substrates first, followed by memory, then significant optics constraints. Conservative Japanese materials suppliers underinvested, restricting accelerators, CPUs and other components. Even one major company could not satisfy its full 2028 co-packaged-optics volumes, he argues, given constraints in lasers, fiber attachment and related components.
• Against the emphasis on power, Patel counts nearly 50 gigawatts of data-center construction next year: about 35 in the U.S. and seven or eight in China. On-site turbines, reciprocating engines and fuel cells offer workarounds. Rasgon agrees those exist while admitting uncertainty about turbine lead times.
• With memory stocks trading at two to four times forward earnings, Rasgon asks what happens if the feared pricing collapse as supply arrives around 2028 fails to materialize. Memory occupies roughly 70–80% of an AI rack’s silicon area; HBM needs three to four times the wafers per gigabyte of DDR5. Continued demand growth could sustain scarcity.
• This year’s wafer-fabrication-equipment estimates began around $120 billion and will probably finish near $155 billion; one vendor’s preliminary next-year outlook exceeds $200 billion. Rasgon says full clean rooms limit installation. Some new space arrives in 2027, more in 2028–29. He thinks 2028 will “probably still be tight”; on 2029, “I don’t know.”

3. Pricing power is redistributing the profit pool

• Rasgon’s framing is that “inflation saved the industry during COVID.” Industry units in 2023–24 were flat against 2019, but blended ASPs were 50% higher, even excluding memory. He expects roughly $1.5 trillion of semiconductor revenue this year, maybe more; memory ASPs account for 70–80% of growth, while units rise about 10%.
• Patel contrasts TSMC’s steady gross-margin climb from 45% to 60% with memory’s negative gross margins just a few years earlier. TSMC preserves long-term partnerships; memory suppliers respond to customers’ cyclical bargaining. If Apple squeezes them in downturns, “they’re going to screw Apple over in the upside of the cycle.”
• Patel puts construction above $60 billion per gigawatt, versus $40 billion in 2024. Waiting until capacity is ready lets Musk seek $50 billion rather than $15 billion per gigawatt; Patel says Anthropic generates $70–80 billion per gigawatt of inference capacity. Nvidia’s higher prices do not imply expanding margins: those stayed roughly flat or slipped as memory and other inputs became dearer.

4. Nvidia must keep widening the performance-to-price gap

• Lyons asks when customers finally reject rising costs. Rasgon answers with performance: Ampere-to-Hopper ASPs roughly doubled for about 10x performance; Hopper-to-Blackwell ASPs rose roughly 50% for 3–30x performance, potentially more after optimization. Blackwell-to-Rubin reference costs rise from about $25 billion to $40 billion per gigawatt, but Rasgon cites 10x performance. Customers are also getting more from each gigawatt.
• Rasgon’s condition is explicit: “as long as that gap is getting bigger,” higher prices remain palatable. If hypothetical Feynman were only 20% better, he would worry about pricing and margins. Slower innovation would force suppliers to absorb costs, producing “a dogfight” over who pays.

5. Custom silicon threatens margins before it stops growth

• OpenAI’s first chip beat Blackwell on InferenceX running open-source Chinese models, according to Patel, whose team tested it in OpenAI’s lab. AI helped design and program it. Anthropic wants its chip in half OpenAI’s development time: “aggressive goals,” dependent on models accelerating the process.
• Patel says getting Nvidia running on InferenceX took six months, versus a month and a half for OpenAI. He asks whether design follows software toward commoditization. The incentive is substantial: removing chip-design-house margins brings his roughly $60 billion-per-gigawatt stack toward $20 billion. Fixed mask and EDA costs remain around $50 million, alongside data and scaling challenges.
• Lyons asks whether Nvidia’s lost share means units; Patel prefers CoWoS wafers, HBM and associated spending because chip sizes differ. He expects further share loss next year despite earnings at least doubling. Startups face a separate hurdle: Cerebras and Groq illustrate his argument that only “really crazy things,” difficult for incumbents to copy, have a shot.
• Rasgon emphasizes annual releases, ecosystem, installed infrastructure and financing as defenses; Nvidia also uses AI in design. He allows debate over CUDA’s moat. A note nine or ten months earlier found only 13 days in a decade when Nvidia traded cheaper relative to the SOX; it is cheaper now. “It’s a race.”

6. Enterprise returns need not require enterprise-owned hardware

• Nvidia’s revenue is roughly 50/50 hyperscale and non-hyperscale, Rasgon says, although lab demand cuts across that split; labs contribute about 20% directly and indirectly. “If we don’t have enterprise adoption, there’s no industry.” Coding already supplies one killer application.
• His team previously tried AI modeling and got “garbage.” More recently, Claude built a model in an hour and 20 minutes, incorporating ten years of history, resegmentations and forecasts informed by CEO commentary. A colleague would have needed a week. Rasgon sees encouraging evidence of productivity returns, while stressing that adoption remains early.
• Patel expects most enterprises to buy tokens and combine them with their own data. Batching efficiencies and model-scale economics favor that approach over owning hardware and running models themselves. He adds that even pharma companies and hedge funds are likely to direct most compute and AI spending to models rather than buying compute to run open-source models. Enterprise adoption can therefore expand without a corresponding wave of on-premise accelerator purchases.