Newsletter · 2026-08-30 · Ashutosh Agarwal

The Chip Shortage Has Spread Past Silicon to Boards and Capacitors - Foundry & Chip Equipment Weekly - Week of August 30, 2026

Foundry & Chip Equipment Weekly for the week of August 30, 2026. Podcast synthesis on memory analyst Jim Handy reporting that the AI buildout has run the supply chain short of printed circuit boards, power management chips, capacitors and resistors as well as memory, Jeff Snider turning the bear case into a credit-cycle argument about Nvidia residual-value guarantees, Bernstein expecting a good Nvidia print, and a materials scientist explaining why quantum computing is a clean-room problem.

Foundry & Chip Equipment Weekly

Week of August 30, 2026: The Chip Shortage Has Spread Past Silicon to Boards and Capacitors

For two years this newsletter has tracked a shortage that kept moving. First it was the fanciest chips. Then it was the memory that feeds them. This week the story took another step, and it is a telling one. The scarcity has spread past the silicon entirely, down into the dull, unglamorous parts nobody thinks about: the circuit boards, the capacitors, the resistors, the little power chips that sit around a processor. When even the boring stuff runs out, you are not looking at a hot product anymore. You are looking at a whole supply chain bent out of shape by one enormous spending spree.

That was the through-line from a memory-industry veteran fresh out of a big Silicon Valley conference, and it pairs neatly with the other real conversation of the week: a sharp, uncomfortable argument about whether the mountain of borrowed money paying for all this can keep flowing. Add a genuinely clarifying discussion of why quantum computers are, at bottom, a materials-and-manufacturing problem, and you have a week heavy on the kind of plumbing that decides who wins.

A note on whose voices these are. No chief executive of TSMC, Intel, Samsung, ASML, or any of the big tool-makers sat down to talk this week. The speakers below are an industry analyst, a covering Wall Street analyst, a macro commentator, and a materials scientist: sharp people, but observers rather than operators. Where that changes how much weight a claim carries, it is flagged in place.

TL;DR

The shortage has jumped from chips to the whole board. A memory-industry analyst back from the Future of Memory & Storage conference says it is no longer just high-bandwidth memory that is tight, it is ordinary memory, flash, hard drives, and now the circuit boards, power chips, capacitors, and resistors that go around AI servers. His blunt line: "everything's going into a shortage because of this great, enormous spending binge."
Memory prices are up 80 to 95 percent on contract, and high-bandwidth memory is sold out through 2027, heading for 2028. One presenter, from IBM, guessed the crunch might only start to ease around 2028.
The bear case matured again, this time into a credit-cycle argument. A closely-followed macro commentator laid out, in detail, how Nvidia's customers are being financed off balance sheet through Wall Street special-purpose vehicles, with Nvidia quietly guaranteeing the value of the chips used as collateral. His warning: if the credit cycle turns, the whole structure wobbles, and Nvidia sits in the middle of it.
A covering Wall Street analyst still expected a "good print" from Nvidia heading into its late-August earnings, and made the point that free cash flow has become scarce, and it is nearly all flowing to Nvidia.
Quantum's real bottleneck is materials science, not physics. A Northwestern professor explained that superconducting quantum chips are built in the same clean rooms as ordinary silicon, and that the chemicals fine for silicon are quietly poisoning the qubits. Fixing it is a fabrication problem, which is very good news for the tool and materials chain.

What's new

The shortage stopped being about memory and became about everything

Start here, because it is the most concrete thing anyone said all week. On Grey Beards on Systems (August 22), Jim Handy, a long-time memory-industry analyst at Objective Analysis and about as close to an insider view as this week offered, walked through what he saw at the Future of Memory & Storage conference in Santa Clara. His summary of the mood was that the AI "super cycle" has quietly broken the entire supply chain, not just the parts you would expect.

The familiar part first: high-bandwidth memory, the stacked, expensive memory that AI accelerators need, is "already sold out through 2027," and Handy thinks that in "the next couple of months" it will be sold out through 2028 as well. One conference presenter, from IBM, "thought maybe the constraints would start to release around 2028," which Handy called "pretty obscene," meaning obscenely far off. The reason the big three memory makers keep pouring capacity into it is simple margins: high-bandwidth memory "is a much higher margin product," so they keep converting ordinary memory lines over to it.

Here is the twist that makes this a supply-chain story rather than a memory story. According to Handy, the shortage would exist even if high-bandwidth memory did not: "even if HBM were not an issue, the way that the hyperscalers are spending, they would have consumed so much [ordinary memory] that it would have still caused a shortage." So the crunch cascades. Making high-bandwidth memory stole wafers from standard DRAM. The same buyers snapping up memory were also buying flash storage, so flash went short too. Then the hard-drive makers, he named Western Digital, Seagate, and Toshiba, "found out that they didn't have the capacity to support this growth," which pushed buyers toward solid-state drives, which made that shortage worse. And, tellingly for this newsletter: "the wafers that are being used for Nvidia processors are taking away from TSMC's capacity to produce wafers." Everything is drawing on the same finite pool of fab and packaging time.

All of a sudden everything is basically falling into short supply. And so you talk about the PC boards… everything's going into a shortage because of this great, enormous spending binge.

That is Jim Handy of Objective Analysis. The genuinely new detail is how far down the parts list the scarcity now reaches. A conference speaker described AI-server circuit boards needing more layers and higher-frequency connections than ordinary boards, which is straining the printed-circuit-board supply chain. And it goes further still: "power management ICs, ceramic capacitors, resistors, inductors, all those are starting to become short on supply because of the data center buildup." When a build-out is big enough to run the world short of resistors, the demand signal is not subtle.

The pricing tells the same story. Handy said DRAM contract prices are up "as much as 80 to 95%," and that long-term supply agreements, customers locking in volume years ahead, have become common. He also flagged a structural wrinkle worth filing away: ordinary DRAM chips "haven't been shrinking very much for the past 10 years," so you cannot easily get more memory per chip. Instead, the makers are turning to clever packaging, including "hybrid bonding," a way of stacking the chips inside a memory stack closer together to shorten the wiring. That is an advanced-packaging problem, which is exactly where the tool-makers live.

A few other conference notes that matter for the ecosystem:

High-bandwidth flash is becoming real. Handy said SanDisk and SK Hynix both presented high-bandwidth flash, a fast storage layer aimed at AI, and that the industry has finally "agreed upon a spec," with eye-watering performance targets (on the order of 100 million operations per second on a single module). It is being pitched partly as a way to hold the "KV cache," the working memory an AI model needs to keep a long conversation in context.
The hyperscalers are reusing old memory to plug the gap. Handy pointed to a recent Meta paper describing something he says is happening across the big cloud companies: instead of throwing away servers with older memory, they are pulling the old DDR4 memory out of retired machines, dropping it onto expansion cards (using a standard called CXL), and bolting it onto newer servers, literally scavenging to fill a supply hole that new memory cannot.
A memory type that has been a curiosity for decades is suddenly interesting. As chip processes shrink to 14 nanometers and below, an old on-chip memory (NOR flash) simply cannot be built anymore, so a different technology, MRAM, is filling the gap in the elaborate microcontrollers that need it, a small but real read-through for the specialty-memory names.

Handy's closing color captured the strangeness of the moment: the hyperscalers, he noted, no longer measure themselves in output. "You never hear them say we're going to process a kabillion tokens. They measure themselves by their spending," by billions of dollars and gigawatts of power. One conference slide, he recalled, put projected AI-related spending in 2030 at around $7 trillion, a number so large it would rank among the largest economies on earth. Treat that as a conference slide, not gospel. But the direction is the whole point.

The bear case grew a balance sheet

If the shortage is the bull case, demand so intense it is running the world short of capacitors, the bear case this week got its most detailed airing yet, and it has moved a long way from where it started. A month ago the worry was China. Then it was demand concentration. Then, the funding structure of Nvidia's $500 billion financing consortium. This week, on Eurodollar University (August 21), the macro commentator Jeff Snider turned it into a full-blown credit-cycle argument. Snider is a financial-markets analyst, not an operator or a chip insider, so weigh this as a framework rather than a scoop. But it is a coherent one.

His frame: Nvidia is the "gray swan" of the AI boom, not a hidden risk but an obvious one "hiding in plain sight" that everyone chooses to ignore because "it's bubble time." The mechanism he described is the part worth understanding, because it explains a run of headlines from the past two weeks about Nvidia's off-balance-sheet arrangements. In plain terms:

Nvidia's customers increasingly are not paying cash. Money flows from Wall Street, through big asset managers like BlackRock and Apollo, into special-purpose vehicles, financing structures set up to make loans. Those vehicles borrow from pensions and insurers, then lend to Nvidia's customers so they can buy chips, using the chips themselves as collateral. The snag: those chips are "relatively illiquid," specialized, and "depreciate relatively quickly." So Nvidia sweetens the deal with what is called a residual-value guarantee, a promise to "step in and make the [lender] whole" if the chips lose value and the customer cannot pay. As Snider put it, that guarantee "doesn't show up on a balance sheet anywhere," because in Nvidia's own modeling "there's almost no chance" it gets triggered, an echo, he argued, of the credit-default swaps of 2008.

Snider's cautionary tale was the dot-com telecom bust: Nortel and Lucent lent their own customers the money to buy their gear, were right about the internet but wrong about the timing, and went bankrupt when the customers could not pay. His worry is that AI is running the same play at far greater scale, and that the broader credit cycle is already turning, with stress showing up in private credit (he cited troubled lenders and rising non-accrual loans) and a looming refinancing wall for software borrowers. If lenders get "colder feet," the money that funds Nvidia's customers dries up, the collateral clauses trigger, and "suddenly everybody's entangled." He noted, pointedly, that Nvidia already trades "at a discount to peers on cash flow," a sign, he thinks, that the market is quietly starting to price this risk.

You do not have to buy the doom to take the point: the demand is real, but a growing share of it is being financed on structures that only work if the good times keep rolling. That is the live question hanging over every tool order in this cycle.

Quantum's problem is a clean-room problem

The frontier read this week was unusually clarifying, and it lands right in this newsletter's wheelhouse. On Scientific Sense (August 21), Mark Hersam, a materials-science professor at Northwestern, explained why the quantum computers being built by IBM and Google are, underneath the hype, a manufacturing challenge, and why that is good news for the semiconductor tool chain.

The key fact: the superconducting quantum chips that IBM and Google pursue "are made using conventional cleanroom processing, the same processing that's used for making silicon transistors." That is deliberate, it lets quantum lean on decades of semiconductor manufacturing know-how. But it creates a subtle trap. The make-or-break metric for these qubits is "coherence time," how long a qubit holds its state before it falls apart, and therefore how much computing you can do. And Hersam's research shows that ordinary clean-room chemistry, harmless to silicon, is quietly wrecking it.

Two examples he gave: the fluoride-based chemicals normally used to strip surface oxides off the niobium metal in these chips instead form "niobium hydride," which shortens coherence time; and the niobium's own surface oxide is lossy too. His lab's fixes are pure fabrication engineering: different etching chemicals, and an "encapsulation" trick where you cap the niobium with a thin layer of gold (or other metals) while it is still in the vacuum chamber, so air never gets a chance to form the damaging oxide. His bottom line was almost a slogan for this whole newsletter:

It's a material science problem right now… If you can't figure out how to make it, then it doesn't really matter.

That is Professor Mark Hersam of Northwestern. And that is the read-through. As quantum grinds forward, the value does not just sit with whoever designs the cleverest qubit, it sits with the deposition, etch, encapsulation, and materials steps that decide whether the qubit works at all. Quantum, increasingly, is a customer for the same kind of process expertise the chip-equipment industry has spent fifty years perfecting.

The debate

This was a one-sided week, and it is worth being clear about which side.

The bull case, that the shortage is structural and durable, was made well, and mostly by the demand data. Jim Handy's account is the strongest version: the scarcity now reaches so far down the parts list that it cannot be explained by a single hot product or a bit of double-ordering. Contract prices up 80 to 95 percent, capacity sold out for years, customers signing long-term agreements, hyperscalers buying up fab line capacity and openly indifferent to price. This is what a genuine, broad-based shortage looks like. If you believe the tool-makers sit at the chokepoint of all of it, this is the fact pattern you want.

The bear case was not "the demand is fake," it was "the money paying for it might stop." Jeff Snider's credit-cycle argument is the most developed version of a worry that has been maturing for weeks. Note the shape of the disagreement: the bull and the bear do not actually dispute the demand. They dispute the financing. The bull sees contracted demand years out; the bear sees that a rising share of it is funded by borrowed money on off-balance-sheet structures that assume nothing goes wrong. Both can be right for a while, which is precisely what makes the timing so hard to call.

On how much weight to give either side: the people who could most directly settle the argument, the operators actually building and selling the tools and the fabs, were not on tape this week. The bull case came from an industry analyst relaying a conference; the bear case from a macro commentator; the nearest thing to a name-covering Wall Street voice was Bernstein's Stacy Rasgon, and even he was talking about Nvidia rather than the tool chain. Treat this week's debate as an informed outside-in read, not the incumbents' own testimony.

Names in play

Nvidia was the name every thread orbited, even though it never took the stage itself. On CNBC's Fast Money (August 21), ahead of Nvidia's late-August earnings, Bernstein analyst Stacy Rasgon, a genuine covering analyst and so a more weight-bearing voice than the pundits above, said he expected "a good print," while candidly admitting he had "no idea what the stock's going to do." His more useful observation was about positioning: over the prior month the broad chip index (the SOX) was down about 5 percent while Nvidia was up 4 to 5 percent, a reversal after a long stretch of the stock treading water, and he framed it as the start of Nvidia's "next product cycle," which historically has been good for the shares. Host Mike Santoli captured the bull's one-liner: "there's a scarcity of free cash flow now in this market, and it's all going to Nvidia." The counterweight is everything in Snider's segment, that the free cash flow looks great partly because so much of the risk has been pushed off the balance sheet.

Broadcom earned a passing but notable mention on the same program: the company is "eyeing a massive debt deal." On its own that is just a headline, but set against this week's financing anxiety, another AI-chip winner reaching for the debt markets is a data point worth watching, not ignoring.

The memory makers (Micron, SK Hynix, Samsung) and the storage names (SanDisk, Western Digital, Seagate) remain the most direct beneficiaries of the shortage Handy described: sold out for years, with real pricing power, and now extending their reach into high-bandwidth flash. The bear counter is the same one that hangs over the whole complex: pricing this good, for this long, is exactly what eventually invites new supply and substitution.

Read-throughs

Advanced packaging keeps climbing the priority list, this week via hybrid bonding. With ordinary memory chips barely shrinking anymore, the way to get more performance is to stack and connect them more cleverly. Handy's description of hybrid bonding to shorten the wiring inside a memory stack is an advanced-packaging problem, and it is structurally good for the packaging and metrology tool chain no matter which memory maker wins.
The bottleneck is now the whole bill of materials, not just the wafer. Printed circuit boards with more layers and faster interconnects, plus power-management chips, capacitors, resistors, and inductors, are all going short. That widens the set of companies that benefit from, and are constrained by, the AI build-out well beyond the marquee chip names.
Fabless customers and TSMC are colliding over the same wafers. Handy's point that Nvidia's processors are eating into TSMC's overall wafer capacity is the clearest reminder that leading-edge logic, memory, and packaging are all fishing from the same pond. When that pond is this shallow, allocation, who gets wafers and at what price, becomes the whole game.
Quantum is a materials-and-fab read-through, not just a qubit race. Hersam's work makes the case plainly: the value pools in deposition, etch, encapsulation, and materials engineering, the semiconductor-equipment industry's home turf, because those steps decide whether a qubit holds its state at all.

What changed from prior weeks

The shortage story broadened from memory to everything. For weeks memory has been described as structurally tight. This week the same industry analyst extended the tightness down the entire supply chain, to flash, hard drives, circuit boards, and passive components, reframing it from a memory story into a supply-chain story. That is a genuinely new and more bullish framing of the same demand for the tool chain.

The bear case moved from who funds it to how it is funded, and whether the credit holds. Two weeks ago the concern was the existence of Nvidia's $500 billion financing consortium; last week it was Google and Amazon posting negative cash flow. This week it hardened into a mechanical, credit-cycle argument about off-balance-sheet vehicles and residual-value guarantees, the same worry, now with the wiring diagram drawn out.

Quantum shifted from timelines to fabrication. Recent weeks' quantum coverage was about roadmaps and milestones; this week's was about the clean-room chemistry that determines whether any of those roadmaps are buildable, a more concrete and more semiconductor-adjacent framing than usual.