CASPA Panel: SemiAnalysis Ranks Substrates Above Power as AI's Binding Constraint, While Bernstein Flags Financing and Custom Silicon as the Pressure Points
Semiconductor Industry Outlook Panel at the CASPA 2026 CEO Summit, October 24, 2026
Dylan Patel, founder and CEO of SemiAnalysis, and Stacy Rasgon, managing director and senior semiconductor analyst at Bernstein, used the CASPA panel to challenge several consensus views on the AI buildout. Patel argued that power and data center space are a smaller constraint than the market assumes, and that the real chokepoints are substrates, memory and optics. Rasgon said financing is now near the top of investor worries, and that the demand and supply signals point to a tight market through at least 2028. Both agreed the harder question is who keeps the margin as component costs soar and as Nvidia and Broadcom face a wave of custom silicon.
Substrates, Not Power, Are the Tightest Bottleneck
Patel's most contrarian point was his ranking of constraints. He put substrates first, ahead of land, power and shell, and ahead of semiconductor tools. "Substrates, PCB materials, substrate materials, all of these components were severely underinvested in," he said, noting that the Japanese suppliers behind much of this supply chain are conservative and reluctant to add capacity. Memory comes next, followed by optics. On co-packaged and near-package optics, he said that even if one major company wanted to move fully to CPO in 2028, "they could not satisfy their volumes" in lasers, fiber attach and other components.
On power, SemiAnalysis research shows nearly 50 gigawatts of data centers being built globally for next year. About 35 gigawatts are in the US and 7 to 8 gigawatts in China, with the rest in places such as Australia, India and Malaysia. Patel said on-site turbines, reciprocating engines and fuel cells are helping developers work around grid limits. He called these fixes "significantly less complex than the semiconductor supply chain." Rasgon was more cautious, noting that Nvidia and Broadcom appear to view land, power and shell as their limiting factors. He also questioned turbine lead times, pointing out that Elon Musk has talked about making his own turbine blades.
Rasgon described a layered shortage in which each constraint has to be cleared before the next can be addressed. He said tool vendors are themselves complaining that short-term shortages of PCBs, chips, valves and mass flow controllers are limiting their ability to ship upside. His view on duration was blunt: "I think this year is fine. I think next year is going to be the same thing. 2028 will probably still be tight. 2029, I don't know. That turns into a demand question."
Memory and Wafer Fab Equipment Show How Tight the Chain Is
Rasgon said memory stocks trade at two to four times forward earnings even as pricing "has gone through the roof." He noted that SanDisk guided to quarterly EPS above its share price of 24 months ago. He attributed the low multiples to peak-cycle fears that supply arrives in 2028 and prices collapse. His counterpoint was structural: 70% to 80% of the silicon area in an AI rack is memory, mostly DRAM. High bandwidth memory also takes three to four times as many wafers per gigabyte as standard DDR5.
On equipment, Rasgon said a company reported this week and gave a preliminary 2027 wafer fab equipment outlook of more than $200 billion. He said he was surprised that many investors expected even more, because "we don't have the fabs." Clean rooms are full, and the industry has squeezed existing footprints hard. WFE estimates started this year near $120 billion and will likely end around $155 billion. New clean room capacity arrives modestly in 2027 and more meaningfully in 2028 and 2029, which is when Rasgon expects the market to find out whether supply and demand are mismatched.
Financing Is Rationing the Buildout, and Nvidia Is Backstopping It
Rasgon said the stock market's drawdown from peak reflects worries about the sustainability of the buildout, "and clearly financing is one of those concerns. It's probably up near the top." He noted that hyperscalers, once "the most cash generative businesses in the history of humankind," are burning cash and likely to raise equity. He cited Meta's new share class structure as one enabler. CoreWeave's unsecured bonds trade at about 15% yields, which he said is a market level rather than its current cost, but a signal of where new unsecured debt would price. He still concluded that "for the next year or two" the industry is probably fine. He also argued that constraints may help, because they push supply out and avoid an overbuild like the dark fiber left after the tech bubble.
Patel gave the more granular picture. Hyperscalers cannot fund $2 trillion to $2.5 trillion of annual capex, so their spending is self-limiting. Neoclouds, which account for nearly half of the buildout, can only finance clusters if they have offtake from a hyperscaler or a major AI lab. Even then the debt is expensive. He cited a recent cluster built for Anthropic financed with a 12% coupon sold at 95 cents on the dollar, for an effective yield of about 15%. Outside that tier, he said, "you just literally cannot" build capacity unless a hyperscaler is behind it. That is why Musk's strategy is to build capacity and rent it to startups on short-term contracts at very high prices.
Patel said Nvidia has recognized the problem and is acting as a backstop. With Sharon AI and Fermi, it takes revenue shares in exchange for guaranteeing financing. In that structure Nvidia rents the cluster at a guaranteed $3.50 an hour, which leaves the cloud operator with no profit. Nvidia then shares in any revenue above that price. Lambda gave Nvidia a 10% revenue share to build for Anthropic, versus a much larger 50% in the Sharon AI and Fermi deals. Broadcom, Patel said, has structured a rent-to-own deal with Anthropic that stretches chip payments over several years. He also flagged Meta's $1.4 trillion of compute and data center commitments as something capital markets are watching closely, though he said "Meta is completely fine."
Pricing Power Has Replaced Price Deflation, and the Cost of a Gigawatt Has Jumped
Rasgon argued that the industry's historical identity as a value destroyer has been reversed. He cited earlier work from his McKinsey days showing the industry destroyed economic profit over a full cycle in the 1990s and early 2000s, with value flowing to customers. COVID changed that, he said. "I didn't have a single one experience margin compression because of input costs going up. They all raised their prices to offset it." He expects the industry to reach about $1.5 trillion in revenue this year, up roughly 100%. He attributes 70% to 80% of that growth to memory ASPs, with units up only about 10%. Customers, he said, "don't really have any choice."
Patel said the cost to build a gigawatt of AI compute has risen from about $40 billion in 2024 to more than $60 billion. He drew a distinction in pricing behavior. TSMC has lifted gross margin from 45% to 60% over a decade, but "has not really jacked up pricing even though they could," given its near-monopoly and long-term partnerships. Memory makers, coming off negative gross margins about 24 months ago, are squeezing customers. Patel said that "if Apple's going to screw them in the downturn, they're going to screw Apple over in the upside of the cycle." Nvidia has held its margin roughly flat while absorbing higher memory, substrate and optics costs, so its customers' prices have risen faster than its margin suggests. Patel said it is unclear who ultimately absorbs the cost between chip vendors, clouds and the labs.
Nvidia's Margin Defense Is Performance per Dollar, and It Has to Keep Winning
Rasgon framed Nvidia's pricing around customer value. ASP roughly doubled from Ampere to Hopper while performance rose about 10x. Hopper to Blackwell saw roughly a 50% ASP increase against 3x to 30x performance gains. Blackwell to Rubin brings another doubling of rack cost against roughly 10x performance. He cited Nvidia's own reference figures of about $25 billion per gigawatt for Blackwell and $40 billion for Rubin. "If they were to roll out Feynman and it's only 20% better, then yeah, then I would start to worry a lot about pricing and margins." He recalled Bill Dally's Hot Chips remark that GPU performance improved 1,000x over the last decade, with a goal of another 1,000,000x over the next.
Custom Silicon and the Commoditization of Chip Design
Patel posed what he called the big question for the industry: whether AI models make chip design as commoditized as software. He said OpenAI's first-generation chip outperformed Blackwell on SemiAnalysis's open-source InferenceX benchmark running Chinese open-source models, with AI used heavily in design and programming. Anthropic is starting its own chip program and aims to finish in half the time. Google is selling TPUs externally, and Amazon's Trainium is now "fully usable in PyTorch."
Software bring-up times back this up. Patel said it took SemiAnalysis six months to get Nvidia working on InferenceX, six to nine months for AMD, a few months each for Google and Amazon, and about a month and a half for OpenAI. If design becomes easier, he argued, "then margins have to fall." Stripping the design houses' margins out of the stack, he said, would bring the cost of a gigawatt to roughly $20 billion. Meanwhile component vendors such as memory, foundry, substrate and optics suppliers have "a lot of room to run on pricing."
Nvidia is already losing share of the supply chain, in Patel's account. It has taken a smaller percentage of total CoWoS wafers and HBM in 2026 and will lose share again next year. Earnings are still expected to roughly double because the base is growing so fast. Both speakers tied this to valuation. Patel said Nvidia and Broadcom are at their lowest multiples "basically ever." Rasgon said that when he last measured, roughly nine to ten months ago, Nvidia had been cheaper relative to the broader semiconductor index on only 13 days in the prior decade, and it is cheaper today. Rasgon listed Nvidia's moats as CUDA, which he said is debatable, along with its installed base, annual product cadence, balance sheet and own use of AI in design. He described Jensen Huang as "paranoid," having "took the old Intel lessons to heart."
Enterprise Demand Is Real but Will Largely Be Bought as Tokens
Rasgon said Nvidia's revenue is split roughly 50/50 between hyperscale and non-hyperscale customers, with the labs about 20% directly and indirectly, a base he called "maybe even a little more diverse than maybe we had thought." Enterprise adoption, he said, is essential: "If we don't have enterprise adoption, there's no industry." He sees agentic coding as the first killer application, with revenues that "have gone vertical." As an anecdote, he said a member of his team used Claude to build a complete financial model from scratch for an already-covered company in 1 hour and 20 minutes, a task that would take an analyst a week. He added that the gains may show up as productivity rather than revenue, and that the labor implications are a separate question. Anecdotal evidence points to a positive return on AI, he said, but many companies remain in the experimentation phase.
Patel cautioned that this will not translate into much enterprise hardware demand. Scale economics mean most enterprises, including pharma companies and hedge funds, "are going to be buying tokens" and combining them with their own data rather than running open-source models on their own compute. That keeps AI spending concentrated among a small group of providers, reinforcing the financing and pricing dynamics both analysts described.