Newsletter · · Ashutosh Agarwal

The Easy Money in the Chip Shortage Is Gone as Bottlenecks Reach Testing - Foundry & Chip Equipment Weekly - Week of September 13, 2026

Foundry & Chip Equipment Weekly for the week of September 13, 2026. Podcast synthesis on chip-industry analysts declaring the bottleneck trade played out as scarcity moves into wafer testing, substrates, MLCCs and specialty glass with relief not until 2028, Broadcom's roughly $115 billion custom-chip guide and first named rival in MediaTek, and quantum becoming a foundry customer through Xanadu and a fourfold jump in private funding.

Foundry & Chip Equipment Weekly

Week of September 13, 2026: The Easy Money in the Chip Shortage Is Gone as Bottlenecks Reach Testing

For most of this year, making money in the chip build-out was almost embarrassingly simple: find the thing that was sold out, buy the company that makes it, and wait. Every few weeks a new bottleneck popped up (fancy chips, then memory, then packaging, then the circuit boards and capacitors around it) and each one minted a winner. This week a group of people who watch this industry for a living called time on that game. Not because the shortage is ending. Because everyone now knows exactly where the shortages are, so the free money that came from being early is spent.

That was the sharpest, most useful conversation of the week. It also carried a genuinely new detail worth filing away: the tightest chokepoint left isn't the glamorous stuff. It's the dull, invisible middle of the supply chain: the testing of wafers, the little ceramic building blocks called substrates and capacitors, even specialty glass. And the relief everyone is hoping for doesn't arrive next year. It arrives in 2028.

The rest of the week was quantum. Two good episodes, one with an actual company chief executive and one with the person who runs the definitive scorecard for the field, and both, in different ways, land right in this newsletter's wheelhouse: quantum's future increasingly runs through the same foundries and the same clean-room tools that make ordinary chips.

A note on whose voices these are. No one from TSMC, Intel, Samsung, ASML, or the big Western tool-makers sat down to talk. The one company boss who did was a quantum-hardware founder. Everyone else was an industry analyst or a market commentator: sharp observers, but observers. Where that changes how much weight to give a claim, it is flagged in place.

TL;DR

  • The "bottleneck trade" is over. A pair of veteran chip-industry analysts declared that the first-half-of-the-year game (buy whatever's sold out) is "fully played out." Their line: "We've turned over every rock. We've searched under every couch cushion. We know where all the bottlenecks are." From here it's careful, company-by-company stock-picking, not buy-anything.
  • The last, deepest bottleneck is the boring stuff, and it's a tool-and-test story. The scarcity has moved "past substrates into testing." Named constraints now include the machines that test wafers, the ceramic substrates chips sit on, the tiny ceramic capacitors (MLCCs), and even specialty "tea glass." As one host put it, "the industry can only scale as its deepest bottleneck."
  • Relief comes in 2028, not 2027. The same analysts think 2027 is the peak year of shortages: "it's not going to come until 2028." Every big hardware maker they tracked (Dell, HP, Broadcom, and Nvidia the prior week) said the same thing: growth is capped by supply, not demand.
  • Broadcom's custom-chip numbers are enormous, and the competition is finally real. Broadcom guided its custom-chip business to roughly $115 billion next fiscal year and about $230 billion in 2028, inside a total custom-chip market the analysts size at $400–430 billion by 2028. But for the first time Broadcom named a rival, MediaTek, that "won the compute tiles" inside Google, and its profit margins slipped.
  • Quantum is quietly becoming a foundry customer. The CEO of Xanadu, the only public quantum company built on light rather than electrons, explained that his whole strategy rests on using existing chip foundries "with the same tools, processes and materials," and running at room temperature instead of in a giant fridge. Meanwhile the editor of MIT's quantum scorecard says private money into quantum jumped roughly fourfold in a year, still "pocket change" next to AI.
  • A hard demand datapoint amid the noise: TSMC reported August sales up 53% from a year earlier.

What's new

The bottleneck trade is done, and the last one left is testing and glass

Start here, because it reframes the whole year. On The Circuit (September 7), the two hosts, semiconductor-industry analysts who follow these companies for a living (so a weight-bearing pundit voice rather than idle chatter, though not operators), declared the end of the trade that has driven this market all year.

Their framing: "The first half of the year was very much about the bottleneck trade: which company has capacity, which company is going to benefit from the bottleneck and raise prices. That trade is fully played out. We've turned over every rock. We've searched under every couch cushion. We know where all the bottlenecks are. So then what?" The answer, they said, is that the market shifts "to a much more tactical, individual-company analysis." Last year "you could have made money just buying anything. Now you're going to have to be a lot pickier and choosier."

That's not a bearish call on demand; it's the opposite. The reason the trade is "over" is that demand is so obviously supply-capped that there's nothing left to discover. Across every hardware company they tracked this week (Dell, HP, Broadcom, and Nvidia the week before), "pretty much everybody's just like: our growth is constrained by supply." That's now the baseline assumption.

Here's the genuinely new part, and it's squarely a tool-and-materials story. The scarcity has crept down into parts of the supply chain nobody was watching. As one host put it:

"The industry can only scale as its deepest bottleneck. You could get more memory or more GPUs, but if you can't get more substrates, doesn't matter. Or if you can't get enough MLCCs, or if you can't test your wafers, you're constrained by whatever your lowest bottleneck is."

Unpack that, because each item is a real chokepoint. Substrates are the engineered ceramic-and-resin bases that a chip is mounted on: mundane, and suddenly scarce. MLCCs are multi-layer ceramic capacitors, the grain-of-rice-sized components sprinkled by the thousands around every processor. And testing, the step where finished wafers and chips are checked to see if they actually work, is now called out as a binding constraint; one of the hosts said he'd written a report on testing that very week. The favorite anecdote came from Dell's earnings call: Michael Dell reportedly reeled off a constraint list that ran "all the way into tea glass," concluding that "this industry is severely underbuilt in capacity."

The timing call is the punchline. The hosts think this pain gets worse before it gets better: "relief is going to come, but it's not going to come until 2028." Their thesis, plainly stated: "2027 is the peak year of constraints." For anyone holding the equipment and testing names, that's the read: the crunch that lifts their order books has another full year to run.

Broadcom's staggering numbers, and the first real crack in the story

The concrete corporate news of the week, from the same episode, was Broadcom's custom-chip business, the designed-to-order AI chips it builds for the big cloud companies. The numbers are almost hard to hold in your head. Broadcom guided that business to roughly $115 billion in fiscal 2027 (chief executive Hock Tan's figure; the analysts noted the buy-side treats $115 billion as "probably the floor," with some arguing for $150 billion) and about $230 billion in 2028. The hosts size the entire custom-chip market (compute, logic, networking, and the bits attached around them) at $400–430 billion by 2028.

But two cracks appeared in what had been a flawless story. First, competition got named. For the first time, Broadcom acknowledged a specific rival, MediaTek, which "won the compute tiles" inside Google, one of Broadcom's marquee accounts. The hosts flagged the whiplash: two quarters ago Broadcom "was basically denying competition existed within Google," and now it's reduced to arguing "we're better than MediaTek... we deliver faster." As they put it, "certainly MediaTek is executing better than people expected." Second, profit margins slipped, and management's deflection (don't worry about gross margin, worry about operating margin) went down badly, because the entire bear worry about custom chips is exactly that customers eventually build enough of their own know-how to stop paying Broadcom's industry-best margins.

There's a modeling headache buried in here that matters for how you read every AI-chip print from now on. The business is fragmenting into pieces (one supplier wins the main compute chip, another the input-output, another the networking), so the clean old math ("Company X won Google's chip, therefore X gets all of it") no longer works. Marvell, for instance, is winning networking and attached products at Amazon and Google rather than the main chip. And Nvidia "doesn't break out networking revenue anymore." The hosts' warning: the numbers are about to get "very blurry," and "people are going to get this wrong." A quiet aside worth keeping: Broadcom's genuine edge is in networking and optical parts (its Tomahawk switches, lasers, and optical components that "nobody except maybe Lumentum can compete with"), and that strength risks getting lumped in and lost under the "ASIC" label.

Quantum is becoming a customer of the chip foundries

The most on-thesis frontier story this week came from an actual operator. On Tech Disruptors (September 8), Bloomberg Intelligence's quantum-and-semiconductor analyst Jake Silverman interviewed Dr. Christian Weedbrook, founder and chief executive of Xanadu, the only public quantum-computing company built on light (photons) rather than electrons, which went public this past March.

Why it belongs in this newsletter: Xanadu's entire strategic bet is that quantum should ride on the existing chip industry rather than reinvent it. In Weedbrook's words, the advantage of the photonic approach is "the ability to work with large-scale manufacturers, like large foundries... we can use the same tools, processes and materials. And we work with many foundries around the world. That saves us a lot of time and money." He added that they "didn't have to invent the laser" or fiber optics; they're "standing on the shoulders of the telco and datacom industries," which is precisely why existing foundries can build their chips.

The other eye-catching claim is about cooling. Most quantum computers need to be chilled to near absolute zero inside a giant fridge. Xanadu, Weedbrook said, uses "far less cooling than anyone else": the logic gates and measurements run at room temperature, with cooling needed only to start the machine. That matters for cost, footprint, and, eventually, whether these things can live in an ordinary data center; he pegged the system's size at "three to five tennis courts."

On progress: last year Xanadu published two papers in Nature about six months apart. One, called Aurora, wired four quantum computers together using light, a demonstration that, he argued, could just as well be forty or four hundred, which is how you scale. The other put a particular kind of qubit (the technical name is a GKP qubit) on a single chip for the first time. The company's roadmap targets up to 500 logical qubits by 2029 or 2030, with useful work possible at 100–200. He said Xanadu has achieved "up to 200 times loss reduction" (loss being the photonic version of errors) over the past three to four years, and flagged a near-term catalyst: fresh loss-reduction numbers on the earnings call "next Monday" (roughly September 14).

And the quiet read-through for the classical-chip world: building and running a quantum computer, Weedbrook noted, takes "a lot of classical compute: CPUs, GPUs, ASICs, FPGAs." Xanadu's named partners underline it: Rolls-Royce, Lockheed Martin (on quantum machine learning), and AMD, which supplies classical compute for the quantum system. Quantum, in other words, is turning into another customer for foundries, for advanced packaging, and for conventional silicon.

The money pouring into quantum, and who's funding it

For the wider picture, Reuters Econ World (September 9) featured the editor-in-chief of MIT's Quantum Index Report, the field's closest thing to an official scorecard, so an expert, academic voice rather than an operator or a stock-picker. Two takeaways stand out.

First, the money. The upcoming 2026 edition of the report will document "a substantial increase, a fourfold increase, in private investment in quantum computing over the last year." There are now roughly ten pure-play public quantum companies, several arrived via blank-check mergers; he named IonQ, Rigetti, D-Wave, Quantinuum, and France's newly public Pasqal. And yet, in his words, all of it is still "pocket change" next to AI: quantum funding is measured in billions, while AI this year is "a trillion." The honest framing for investors: real momentum, real capital, but a rounding error beside the AI build-out this newsletter mostly tracks.

Second, the government-money angle, which ties straight back to the foundry world. He noted that "the Trump administration recently put... two billion dollars across nine different quantum-computing companies, taking minority stakes," an unusual move for the US, and, he pointed out, part of a broader pattern of the government taking equity stakes, "Intel for being a famous example." He also flagged the US–China race: China is "outspending the US government" on some measures and is ahead on the sheer volume of patents and research papers (though US work is "a much higher standard"), and is "way ahead" specifically in quantum networking over satellites. On timing, he called the 2029–2030 breakthrough dates floated by IBM and Google "reasonable," and credited Google as being "at the frontier" of the error-correction math that makes noisy qubits behave like reliable ones: "getting the qubits to check each other's homework."

The debate

This was a one-sided week in a specific way, and it's worth being precise about it.

The bull case, that demand is real and the shortage runs deep, was made well, and mostly by the supply commentary itself. The Circuit's hosts weren't arguing demand is weakening; they were arguing it's so thoroughly supply-capped that the market has run out of hidden winners to find. "The industry is severely underbuilt," every major hardware maker is growth-constrained by supply, and the constraint reaches all the way down to test machines, substrates, capacitors, and glass. If you believe the tool, test, and materials companies sit at that chokepoint, 2027-as-peak-constraint is a bullish read for their order books. TSMC's reported August sales, up 53% from a year earlier (Stock Market Today With IBD, September 10), is the week's one hard demand datapoint pointing the same way.

The caution wasn't "demand is fake"; it was "the easy phase is over, and the cracks are individual." The same analysts who see a deep shortage also think the buy-anything trade is spent and you now have to be selective. And there were real, name-level cracks: Broadcom's first admission of genuine competition (MediaTek inside Google) and slipping margins; a custom-chip market fragmenting in ways that will trip up anyone still using last year's math. The old financing worry got a small echo too: the IBD hosts noted that Oracle, despite a strong print and cloud revenue that "doubled," has been "burning cash for a long time," the same concern that has dogged Google and Amazon.

Whose case this was: both sides were argued by industry analysts, not by the operators building the fabs, the tools, and the test gear, and the one operator on the tape runs a quantum-hardware startup, not a foundry. So treat this debate as an informed outside-in read rather than the incumbents' own testimony; the counter-case to the "shortage runs to 2028" thesis was not voiced by anyone building the equipment.

Names in play

Broadcom is the name that moved a thesis this week. The bull case is the sheer scale: roughly $115 billion in custom chips next year, about $230 billion in 2028, inside a market heading toward $400–430 billion. The bear case now has teeth it lacked a quarter ago: a named competitor (MediaTek) taking real work at a flagship customer, slipping margins, and a business fragmenting into pieces that make the whole thing harder to model. Next catalyst to watch: whether Broadcom can show it's holding the highest-value "compute tile" work rather than ceding it, and whether its genuine networking-and-optical edge gets recognized or buried.

Marvell and Lumentum ride the same wave from the side. Marvell is winning networking and attached silicon at Amazon and Google (not the main chip), and Lumentum was singled out as the one company that can rival Broadcom in optical parts, a reminder that in a supply-capped build-out, the "picks and shovels" around the marquee chip can be the cleaner exposure.

TSMC stayed off the microphone but posted the number of the week: August sales up 53% year over year. In a week where the theme was "demand is capped only by supply," that is the demand side speaking for itself.

Read-throughs

  • Testing and metrology just came back into view. The test-and-inspection theme (the likes of Tokyo Electron, Advantest, Onto, and Camtek) returned this week, secondhand from analysts rather than from the companies themselves: the bottleneck has moved "past substrates into testing," and wafer test is now named as a binding constraint. If 2027 is the peak-constraint year, the companies that test and inspect chips are pointed straight into it.
  • The humble supply chain is where the scarcity now lives. Substrates, multi-layer ceramic capacitors, and specialty glass are the newly-tight links. That widens the set of beneficiaries (and of hidden risks) well beyond the marquee chip names, and it's a reminder that a fab's output is gated by its least glamorous input.
  • Advanced packaging stays central by implication. Substrates being a top constraint is, at bottom, a packaging-and-assembly story: the step where chips are mounted, connected, and stacked. It remains structurally favored no matter which chip designer wins.
  • Quantum is a foundry-and-classical-silicon customer, not just a science project. Xanadu's pitch (build on existing foundries, existing tools, existing materials, and lean on CPUs, GPUs, ASICs, and FPGAs to run the machine) is the clearest sign yet that the value in quantum flows partly back to the same chip-manufacturing and packaging chain this newsletter tracks. The government's equity stakes ($2 billion across nine firms) and a fourfold jump in private funding give the theme staying power, even if the dollars are still small.

What changed from prior weeks

The frame flipped from "find the bottleneck" to "the bottlenecks are all found." For months the story was the hunt: which sold-out component would mint the next winner. This week that hunt was declared over. The demand read is unchanged (still supply-capped, still enormous), but the investing implication shifted from buy-the-shortage to pick-your-spots. That's a real change in posture, not in the underlying facts.

The bottleneck moved down-market, into testing and passives. In August the running theme was the shortage spreading from memory to circuit boards, power chips, capacitors, and resistors. This week it pushed further still, into wafer testing, substrates, ceramic capacitors, and specialty glass, and, notably, put a date on relief: 2028, with 2027 as the worst year. That revives the test-and-metrology thread.

Custom chips got their first real competitive wrinkle. Broadcom's story has been near-flawless for a year. This week brought the first named rival (MediaTek), the first margin slip, and a sober warning that the business is fragmenting into pieces that will make the numbers harder to trust. A maturing, more contested story.

Quantum shifted from roadmaps to plumbing and money. Recent quantum coverage was about milestones and error-correction physics. This week it was about the business of quantum: who's funding it (a fourfold jump in private money, $2 billion of US government equity), and how it plugs into existing foundries and classical chips. A more concrete, more semiconductor-adjacent framing than usual.