Sub-headline: UBS semiconductor analyst discusses AI infrastructure limits, memory contracts and ASIC versus merchant silicon on The Compute 100 podcast, November 4, 2026
Semiconductor Expectations Now Outrun the Grid
UBS Managing Director Timothy Arcuri argued that the debate in semiconductors is shifting from chip supply to whether the physical infrastructure exists to absorb what the market already assumes. Backing out the gigawatts implied by consensus numbers for Nvidia, AMD, Broadcom, MediaTek and Amazon, he concluded that what has been permitted on the grid is "nowhere close" to what is needed. Investors, he said, "have to believe that there's this conversion of what's not permitted currently, and you have to believe that stuff that's off-grid, a significant amount of off-grid capacity, also has to come online." He framed this as a 2028 and 2030 problem rather than a 2027 one, but said it already shapes how capital is allocated.
Arcuri said UBS published a figure on the day of the discussion for 380 gigawatts of capacity added over the next five years. That is a power number that includes a power usage effectiveness factor, not an IT compute number. He said the more important question is now how much revenue must be generated to justify it. He cited a framework of $30 billion to $50 billion of cost per gigawatt, requiring roughly $80 billion of gross profit over four to five years per gigawatt. He added that lab revenue-per-gigawatt figures, though large, imply only "a couple percentage points of GDP incrementally" over five years. "Does that sound crazy? I don't think so," he said.
He was candid about the limits of his own expertise. As a self-described "P times Q guy," he said translating token economics into demand is outside his comfort zone, and he is drawing on UBS's software, internet and utilities teams. He can say how much capacity will be added, he said, but is less able to say which "killer app" will fill it, though he suspects demand will be there within five years.
Investors Have Only Half Rebuilt Positions After the Spring Correction
Arcuri described the May and June correction as a high-water mark that the group has not yet grown back into. Most large funds he speaks with have rebuilt only about half of their pre-correction exposure, and the questions being asked are more probing than during the "performance chasing" phase. He sees grid limits, safety concerns and permitting as risks to the conversion needed to reach 2028 expectations. Slowing the frontier could push the power problem out, a point he said Nvidia also made to him that day. He noted a counterargument: if open-source models that are "pretty competitive" become the alternative to a slower frontier, demand for power could arrive sooner, and a lack of power could become "a natural pacer."
Micron: Four Times Earnings and a Different Cycle
Arcuri's most forceful argument concerned Micron, where he said the market trades the stock "as if nothing has changed." The shares trade at roughly four times earnings, and the earnings multiple sits unusually close to the tangible book multiple, because book value has not yet caught up with the speed of earnings growth. He attributed the low multiple to a market that does not know what to do with earnings rising this fast at a historically cyclical company. Long-tenured investors, he said, see 85% gross margins and conclude the momentum is gone, so the debate has become one of duration rather than direction.
He identified several structural changes. HBM runs on an annual cadence, and the prior generation is generally not usable once a new chip ships, an obsolescence that DDR, good for seven or eight years, never had. An HBM wafer also consumes roughly four times the capacity of a non-HBM wafer, so demand rises while effective supply falls. And hyperscalers are signing long-term agreements with prepayments, something the memory industry has never seen. "The best it ever was was, hey, we sit down January 1st and we talk about what I'll need this year," he said, with volume commitments but no price protection.
Arcuri did not claim the cycle is dead. "There is going to be a cycle on the other side," he said, after 30 years of hearing that this time is different. His point is that contract prices signed months ago will sit at about half the open-market price by year-end, and the contracts carry floor prices far below current levels. Even a 70% to 80% collapse in spot pricing would leave Micron earning substantial profit, generating cash and adding to book value, which "brute forces the stock higher." He said any downturn would likely be demand-led rather than supply-led, since the contracts discipline capacity additions as well as guarantee them. He credited CEO Sanjay Mehrotra with turning Micron from technology laggard to leader, and noted the geopolitical tailwind for a US-based producer. He does not cover the Korean memory makers.
ASICs Close the Unit Gap, but Google's Broadcom Volume Flattens
On units, Arcuri estimates Nvidia will ship 6 million to 7 million GPU units this year, counting two dies per unit. He expects Broadcom to ship 3.5 million to 4 million custom units this year and about 6 million next year, so ASIC units are approaching merchant GPU units. Nvidia's dollar share will remain higher because of its higher average selling prices.
He said Broadcom's own commentary implies Google's business flattens in 2028, with incremental volume shifting to MediaTek under a customer-owned tooling model. In that model Google buys HBM and wafers itself, rather than paying Broadcom, which he said charges a 60% margin for a full-service chip. Google, he said, is likely to stay reliant on Broadcom "because Broadcom can do things that nobody else can," and it took Google seven generations to build the resources to go it alone. He read Nvidia's recent MediaTek deal as part of a push to spread NVLink Fusion to ASIC players outside Broadcom's orbit.
Nvidia's Installed Base and a Contrarian Call on What Gets Cut First
Arcuri said the market underestimates Nvidia's ability to monetize its installed base. Every Blackwell and Rubin shipping today is greenfield and replaces nothing. Eventually Ampere and other older parts will be ripped out, and he sees Nvidia upgrading existing fleets with additions such as LPUs from its Groq acquisition or new networking. Cloud customers tell him that practical GPU utilization is only 30% to 50%, which makes accelerating the existing fleet a large opportunity. "Even when capex declines, there's going to still be a very strong propensity to go in and upgrade the installed base," he said.
He also staked out what he called a hot take. When capex flattens and declines, he believes ASICs will be cut before merchant GPUs, because ASICs are built for a particular purpose and buyers retrenching want flexibility. A "very smart" CIO he debated last week argued the opposite. Arcuri also said he remains more bullish on AMD's estimates than on its competitive edge. Software is no longer an impediment for hyperscalers, but he sees no compelling reason in most applications to choose AMD over Nvidia, which is also ahead on rack engineering and an annual cadence. He prefers Intel's foundry and packaging prospects to its product roadmap.
Equipment: $4 Billion of Front-End Spend per Gigawatt
Arcuri said UBS estimates roughly $4 billion of front-end wafer fab equipment spend per gigawatt of compute. Applied to 380 gigawatts through 2030 or 2031, the math implies a year with 100 gigawatts or more of additions and over $400 billion of equipment spending, versus about $155 billion to $158 billion this year. He stressed that those out-year numbers are beyond his formal estimates. He argued equipment is earlier in its capacity cycle than compute, since there has been "nowhere to put new tools," even though equipment stocks have correlated at about 0.9 with Micron and get lumped into the memory duration debate. Back-end stacking, new materials and fab layout redesign are the main innovation areas.
Cerebras and the Fast-Inference Market
Arcuri said the OpenAI agreement transformed Cerebras's fortunes. It is a framework deal of three tranches of 750 megawatts each, initially structured as a cloud deal because OpenAI lacks its own infrastructure. He assumes Cerebras eventually splits 50% cloud and 50% hardware, which he said is unusual for a company selling both and driven more by necessity than preference. He is sympathetic to the argument that wafer-scale advantages weaken for very large models but does not think it is a major issue. He said Jensen Huang puts fast inference at 10% to 20% of the market, while Cerebras CEO Andrew Feldman argues it is the entire market, and that his own view on price per token is "evolving."
Political Backlash Is the Risk That Keeps Him Up
Asked about black swans, Arcuri said demand and supply do not worry him. What does is a backlash. "I fear it's like the French Revolution where people come with pitchforks," he said, citing job losses and anti-data center sentiment as threats to political support for AI regardless of who holds office. He suggested better public relations and a unified approach between Silicon Valley and government, but acknowledged limits unless China participates. He also flagged customer concentration, noting that if Nvidia's customer base depends on a handful of buyers, so do Applied Materials, Micron and the rest of the supply chain. Even so, he said he has a buy rating on everything that touches AI.