Newsletter · · Ashutosh Agarwal

Optics Move Inside the Chip Package and Micron Faces Its Test - Foundry & Chip Equipment Weekly - Week of September 27, 2026

A synthesis of what podcasts, operators, and analysts said about advanced chip packaging, memory, and AI infrastructure financing for the week of September 21-27, 2026, built around Ayar Labs CEO Mark Wade on co-packaged optics moving inside TSMC's advanced packages and heading into Micron's fiscal Q4 report on September 30. Operators and analysts split on whether memory pricing power holds or turns from shortage to glut, while one skeptic put a 2029 due date on the AI build-out.

Foundry & Chip Equipment Weekly

Week of September 27, 2026: Optics Move Inside the Chip Package and Micron Faces Its Test

For two years the story of AI chips has been told in three words: TSMC, CoWoS, HBM. The world's biggest chip factory, its special packaging line that glues processors and memory together, and the stacked high-speed memory that feeds them. This week a fourth word pushed its way into that sentence: light.

The chief executive of one of the best-funded startups in the field came on a podcast and explained, in more detail than we've heard all year, how optical connections are about to move inside the most advanced chip packages on earth, built on TSMC's lines, with Nvidia, AMD, Intel and TSMC all on the investor list. It's a big deal for anyone who owns the packaging and testing supply chain, because it makes the hardest step in chipmaking harder still.

At the same time, the memory trade walks into its biggest test of the quarter. Micron reports on Tuesday, September 30, and the podcasts are split right down the middle. One camp says there is "not one" sign that memory pricing power is fading. The other points at slowing price increases, a surge of Chinese supply, and phones and PCs being priced out of reach, and warns memory "could go very quickly from a shortage to a glut."

And behind both stories sits the question that has hung over this newsletter for a month: who pays for all of this, and when do the bills come due? This week one well-known skeptic put a date on it. It's 2029.

A note on sourcing, as always: we heard from one genuine operator this week (Ayar Labs), but still nobody from TSMC, Intel, Samsung, ASML or the big tool-makers themselves. Everything else comes from analysts, investors and commentators, and we weight it that way.

TL;DR

  • Optics are moving into the package, and TSMC is the chosen factory. Ayar Labs CEO Mark Wade says his optical "engine" pairs a 3-nanometer chip with a photonics chip on TSMC's COUPE process, with "lots of wafers flowing out of the fab right now." Leading GPU makers adopt in 2028–2029; the wider custom-chip crowd follows in 2029–2031. His market sizing: "tens of billion, probably approaching $100 billion."
  • The bottleneck moves to lasers and testing. Wade says the current "crunch points" are laser manufacturing and production test, with as many as "five test insertions" per part. Wall Street is already asking him about "probers, insertion one, insertion two, insertion three."
  • Micron on September 30 is the week's big event. Bear-leaning analyst Sara Awad (Tech Contrarians) says the stock lives or dies on the gross-margin guide after last quarter's 86% guide (on an 84.6% print), and cites TrendForce data showing DRAM price increases slowing to 13–18% in Q3 from roughly 50–60% in Q2. The Futurum Equities hosts counter: "I haven't heard one plausible evidence that memory pricing power is eroding, not one."
  • China's memory maker is coming fast. Awad says CXMT is expected to end the year with monthly output within about 20,000 wafer starts of Micron's own.
  • The bubble got a due date. Paul Kedrosky (SK Ventures): AI spending has driven "30% to 70% of US GDP growth" depending on the quarter, more than 60% of it paid with outside money, setting up a "maturity wall" in 2029–2030 when five-year GPU-backed leases must be refinanced. "Will it all get refinanced? No freaking way."
  • A data point for the other side. GPU-pricing specialist Steve Hou says rental data shows used AI chips hold their value "quite a bit higher" than accounting rules assume, and CME is launching futures on H100 and B200 rental prices.
  • Quantum had a real headline. IonQ announced what it called the industry's first end-to-end real-time quantum error decoder; shares jumped as much as 13% while Rigetti and D-Wave fell.

What's new

Light moves into the chip: an operator's view of the next packaging bottleneck

The best conversation of the week was on The Circuit, "Conversation with Ayar Labs CEO Mark Wade on Silicon Photonics, Co-Packaged Optics, and Next-Gen AI Infrastructure" (September 21). Hosts Ben Bajarin and Jay Goldberg, both long-time semiconductor analysts, sat down with Mark Wade, co-founder and CEO of Ayar Labs, and Jesse Leiter, its VP of capital strategy and investor relations. This is operator testimony, from a company that is a TSMC customer and partner.

The problem, in plain English. Today, the AI chips in a data-center rack talk to each other over copper wires. Copper has a hard limit: the faster you push data, the shorter the distance it can travel. Wade calls this "the bandwidth distance tradeoff," and he argues it is now one of the main things capping how much revenue an AI data center can earn: "large-scale AI data centers are limited in their revenue scaling by the performance of the interconnect in their compute systems." Bajarin added that GPUs and memory "are scaling faster than the interconnect wall is." The fix, Wade says, is to swap copper for light: "the only way that we know how to solve this... from the grounds up, is to swap over to optical connectivity."

Why it matters for the foundry and packaging chain. This is the part equipment investors should read twice. Ayar's product is not a separate cable or plug-in module. It is an "optical engine" that goes inside the same giant package as the AI processor. Wade described it this way:

"Right now we build in the TSMC COUPE process, which the EIC is a three nanometer EIC. And on that EIC are hundreds of millions of transistors."

(An EIC is the electronic chip that drives the optics. COUPE is TSMC's process for stacking that chip on top of a photonics chip.) Then that stack has to be fitted into what he called a "large substrate, multi-reticle, you know, CoWoS-like advanced package," the same kind of package that has been the industry's tightest bottleneck for two years. His summary of the manufacturing challenge was blunt: "difficult on difficult makes it more difficult."

The new choke points: lasers and testing. Asked where the real pain is today, Wade pointed past the wafer itself. Making the photonics wafer, he said, is "just not that hard to do." The hard part is attaching optical fiber to the chip ("probably the most challenging part"), making the lasers, and testing everything. He is personally building a "digital twin" model of the whole flow, and one of the questions it answers is "what's my optimal test policy across five test insertions?" In other words, each part may be tested five separate times on its way through the chain. Investors have noticed. At a Jefferies dinner this month, Wade said, he was quizzed about "test equipment vendors that I wasn't even aware Wall Street was tracking, you know, probers, insertion one, insertion two, insertion three."

On the TSMC question. Bajarin put the bear whisper to him directly: some people say Tower and GlobalFoundries are further along in photonics than TSMC. Wade's answer was that this compares the wrong things. A photonics chip "is nowhere near the final product form factor." What matters is the whole package, and for connecting directly to GPUs and switches, "you have to have optical chips come out of CMOS-like, leading-edge foundries." He said Ayar has worked in TSMC COUPE for "approaching four years," calls TSMC "a phenomenal partner," and has "lots of wafers flowing out of the fab right now." He was candid that the path is bumpy: "Are you guys having challenges on getting things working? Yes."

The timeline. Wade expects "a lot of exciting things from the leading-edge GPU guys" in 2028 and 2029, then adoption spreading to "the rest of the accelerator community" in 2029, 2030, 2031. For that second wave, customers building their own chips through design partners need the technology to be "essentially zero risk," which is why Ayar has built close ties with MediaTek, Alchip and GUC. His framing: "You've got to get to where these optical engines are thought about like HBM." Until then, he expects "a chaotic, you know, turbulent set of technologies, form factors" as customers scramble for 2027–2028 stopgaps.

Who is paying for it. Leiter said Ayar's Series D had strategic money from Nvidia, AMD, Intel and TSMC, and the Series E earlier this year was led by Neuberger Berman, with sovereign funds QIA and GIC and crossover investors including Sequoia, ARK and 1789. A recent strategic investment took the company's primary capital raised this year to $650 million. On market size, Wade said scale-up optics is "tens of billion, probably approaching $100 billion"; Bajarin said most estimates for the space are $30–50 billion by 2030. Leiter also noted MediaTek "winning some of those projects" for Google TPU inference chips and Alchip's work with AWS. That lines up with the MediaTek-inside-Google thread from earlier this month.

Why it matters. Every extra layer inside the package means more steps at TSMC's advanced packaging lines, more fiber-attach and assembly work at outsourced packaging houses, and more test insertions. That is a long-dated positive for the advanced-packaging and test supply chain. The timing is the catch: real volume is a 2028-plus story.

Micron's September 30 print: shortage or glut?

On Investing Experts, "What will Micron's gross margin guide be?" (September 27), Sara Awad of Tech Contrarians (an independent analyst, not an insider) laid out the most detailed bear-leaning memory case we've heard.

Her central point is that Micron's margins are now almost unbelievable for a cyclical business. The company "guided for 86% gross margins after printing 84.6," which means it "has the highest gross margins in the entire sector, even higher than NVIDIA's own peak margins." (Gross margin is the share of each sales dollar left after paying to make the product.) "Memory is still cyclical," she said, and "we don't think the market is going to be very forgiving if Micron's gross margins aren't guided to expand again."

Her evidence that the price surge is cooling:

  • Price increases slowing. TrendForce now sees DRAM prices up "only 13 to 18 percent in Q3 versus I think it's closer to 50 to 60 percent in Q2."
  • Monthly sales from memory makers with no AI exposure. Nanya's August sales hit a record 44.7 billion (currency not stated on the podcast) but rose "only 2% month over month versus increasing 49% month over month in July." She added that Micron's CEO had said two quarters ago that non-HBM margins were higher than HBM margins. Her argument is that ordinary, non-AI memory prices, not HBM, set where margins go, so "memory prices in the non-AI segment have peaked out."
  • Chinese supply. CXMT, "the main Chinese DRAM guy," is "expected to end the year within 20,000 wafer starts range as Micron's own monthly output."
  • Buyers priced out. She cited IDC forecasts for smartphones "contracting about 17% this year and 27% in the second half" and PC shipments falling "close to 19%," both blamed on memory and component costs. "You can no longer have, you know, a PC at the price point of $399."

Her conclusion: "we could go very quickly from a shortage to a glut." She expects a supply-chain "reset" in Q4 and sees the bigger bubble as "a 2027 conversation." She also made a useful point about the difference from the COVID shortage: back then companies missed numbers because they couldn't get parts; "today, no one's missing numbers... this is all driven by higher selling prices rather than more unit volume being sold."

The bull response came on Futurum Equities Podcast, "Meta is up big on Muse. AMD is priced for no mistakes. Micron is next. | Ep 58" (September 25). The hosts, analysts from the Futurum Group research firm, argued the opposite: "I haven't heard one plausible evidence that memory pricing power is eroding, not one." Their framing is that Micron is "selling everything they can make," and because "there's no new production fabs" that have ramped, the quarter is purely "an execution play... the highest possible price on every unit." Their checklist for the print:

  • An accounting trap. "Q4 has 14 weeks in it. Q1 only has 13 weeks," so the guide will look lighter than it is and generate "FUD."
  • Strategic customer agreements. More detail on these long-term deals, and whether they make "the next downturn look different from previous cycles," the argument for the stock trading above "six times earnings."
  • HBM4 share. Whether Micron keeps "that 20% market share" in the next generation.
  • Capex. Whether management "still sees major supply response coming in 2028 or whether capacity is starting to arrive earlier." For equipment investors, this is the line that matters most.

They also shared Futurum's own forecast revisions: its 2030 CPU market estimate went from $110 billion to $250 billion; its custom-chip estimate from roughly $150 billion to $385 billion; and its cumulative AI capex through 2030 now tops $14 trillion.

Why it matters for equipment names. Memory makers are big buyers of chip-making tools. If Micron signals that new supply is arriving earlier than 2028, that is good for near-term equipment orders but bad for memory pricing. If it holds to 2028, the shortage runs longer and tool orders stay back-loaded. Awad also noted that ASML and TSMC report soon after Micron. Last quarter "both reported great results, both sold off after print," as earnings acted "more as a test than a catalyst." She says she remains "bullish on both of those names."

The bill for the build-out has a date on it: 2029

On Between Two COO's with Michael Koenig, "The AI Bubble Has a Due Date. It's 2029 | Paul Kedrosky, SK Ventures" (September 22), Paul Kedrosky (investor at SK Ventures and MIT fellow; a commentator, not an operator) made the most specific version yet of the "who pays" argument that has run through this newsletter.

  • Scale. "Depending on the quarter, anywhere from 30% to 70% of US GDP growth was being driven by a single thing. And that single thing was AI CapEx." He compares it to canals, railroads, rural electrification and the fiber-optic bubble, and notes a single sector is now "40% of the S&P 500."
  • Funding. "More than 60% is coming from what's euphemistically called external financing," meaning "anywhere but free cash flow." Unlike cash flow, outside lenders want fixed payments: "your interest rates don't change just because... token prices have fallen."
  • The wall. "In about 2029, we'll hit a maturity wall, a five-year maturity wall that'll sort of take us out 29 to 30," hitting especially "the sort of classic Oracle-style five-year GPU-backed lease, as opposed to a pure operator lease around, say, like a 15-year powered shell." His verdict: "Will it all get refinanced? No freaking way." He named CoreWeave as his "favorite example" of a debt-heavy provider to be careful with.
  • The math problem. Quality-adjusted token prices are falling "60 to 80% a year." If prices fall 70% a year, "to stand still, I need a roughly 500% year over year growth," and to justify valuations, "something in excess of 1000% growth a year" for a decade. "It would be the first commodity in history to do it."

He also warned that coders, the heaviest early users, are a poor guide to everyone else because coding "has this unique property of being expansive in terms of its token use."

A counterpoint on the collateral: used GPUs are worth more than the books say

Last week's sharpest bear argument was that chips backing all this debt might be close to worthless if things go wrong. On The Business Brew, "Steve Hou - Data Over Narrative" (September 23), Steve Hou, whose firm builds GPU rental-price benchmarks, offered actual data on that question.

He said CME is launching two cash-settled futures contracts based on his firm's indices: an H100 neocloud rental-price index and a B200 rental-price index, settling against monthly averages. (A neocloud is a newer, GPU-focused cloud provider such as CoreWeave.) The goal is the first "institutional grade financial instruments" to hedge AI compute costs, as has been done "with oil and... with power."

On collateral: "over 50% of the cost of a data center today is... chips and memory." Lenders started out applying "very punitive" standard depreciation schedules. But after a few years of rental data, "you very quickly realize that... the residual value... is potentially quite a bit higher," even "relatively conservative." His caveat: "we haven't gone through a full business cycle." He also made the point that a working GPU installed in a data center is "a very different piece of asset" than one "for sale on eBay."

Why it matters. This doesn't settle last week's specific worry about custom chips that only one buyer can use. But it is the first data-driven pushback on the idea that AI hardware collateral collapses in value, and a futures market would give lenders and chip buyers a real price to anchor on.

Taiwan's supply chain is more European than you think

On Taiwanology, "How France and Taiwan Became Interdependent, From Industrial Gases to Chip Design【Taiwanology Ep.65】" (September 22), Frederick, the new director of the French Office in Taipei (a diplomat, not an industry insider), made a point that often gets lost in the "Taiwan concentration risk" debate: Taiwan's fabs depend on foreign suppliers too. "If you don't have Air Liquide, who is producing and providing industrial gases to the Taiwanese semiconductors industry, then you're lacking something." He also named Veolia (industrial recycling, ultra-pure water and chemicals) and Soitec (silicon substrates). He noted France's CEA research center working with PSMC, and French HPC chip designer SiPearl having its processor made at TSMC. He also cited Taiwan's GDP growth of 8.6% last year, forecast "above 11%" for 2026, and said Semicon Taiwan this year had "nearly 20 international pavilions."

Quantum: a real engineering milestone, and a sober timeline

On CNBC's "Fast Money," "U.S. 10-Year Yield Hits 19-Year High… And The Ripple Effects From Meta's Muse 9/23/26" (September 23), the panel covered IonQ jumping "as much as 13%" after the company said it "demonstrated the industry's first end-to-end real-time quantum error decoder," letting a processor "find, fix, and decode errors while running continuously in the background." Rigetti and D-Wave both fell. Panelist Steve called real-time error correction "the problem for the entire industry," but stressed "they're solving it, not the rest of the quantum space," and that the whole sector burns "a ton of cash." Host Mike's pushback: it's "still... kind of a research project... not a product project, just yet." (These are TV-market pundits, not quantum engineers.)

For a sober counterweight, InvestTalk, "Is Quantum Computing a Real Investment or Just Hype?" (September 25), host Justin Klein (a financial adviser-host) laid out the consensus timeline: earliest commercial use in chemistry and materials "by the late 2020s," broader logistics and finance uses in the "early 2030s," and the ability to break current encryption "by the mid-2030s." He compared pure plays IonQ, Rigetti and D-Wave to "your biotech version" of tech stocks, with "severe dilution," and suggested IBM, Alphabet and Microsoft as the lower-risk way in, since legacy businesses fund their research. His bottom line: "odds are good that while this timeline looks enticing, it gets pushed out."

The debate

The bull case: the AI up-cycle has years to run, and TSMC sits in the middle of it.

  • Demand keeps finding new legs. The Futurum hosts: "every time we think we've sort of figured out what the max demand that could be on compute, there's a new use." First chatbots, then enterprise agents, now consumer agents like Meta's Muse. Their raised forecasts ($385 billion custom silicon, $250 billion CPUs by 2030, $14 trillion-plus cumulative capex) say the ceiling keeps moving up.
  • Memory supply is capped until new fabs arrive, which the Futurum hosts expect around 2028, so pricing holds.
  • A new layer of manufacturing complexity is coming. Ayar's Wade calls customer demand for in-package optics "just unquestioned," and it runs through TSMC's leading-edge nodes and advanced packaging. More complexity means more tools, more test steps and more reasons to pay TSMC.
  • The collateral may be sturdier than feared. Hou's rental data suggests used GPUs hold value well above accounting schedules.

The bear case: pricing, supply and financing all turn at once.

  • Memory prices are slowing just as China adds supply and consumer demand breaks (Awad). If Micron's margin guide stalls, the hottest part of the chip trade reprices.
  • The money behind the build-out is borrowed and has a maturity date. Kedrosky's 2029–2030 wall, and his token-price math, say the returns can't keep pace with the debt.
  • AI labs are cutting prices hard. Awad noted Anthropic's cheaper Opus 5.5 ("40% less to run on typical workloads"), OpenAI cutting API pricing "by 50%," and Microsoft offering "30 to 50%" volume discounts. Her figures: OpenAI and Anthropic together hold "north of 1.2 trillion" in commitments across the chip space through 2030, and OpenAI projects burning "almost 280 billion by 2030."
  • Physical and political limits on building. Awad flagged Texas Governor Greg Abbott ordering regulators to halt a data-center permit, and Oracle's notice to Blue Owl about possible delays at its Project Jupiter (Stargate) site. She cited surveys showing "over 60% of Americans don't want a data center anywhere near them."

What wasn't said. Nobody on the podcasts this week argued that Intel 18A or Samsung Foundry are closing the gap on TSMC, and nobody discussed Chinese chip-equipment makers (SMEE, Naura, AMEC, Piotech) taking share from Western tool vendors. The only China supply point was on memory (CXMT). Intel showed up only as an investor in Ayar Labs. Those sides of the debate weren't voiced this week.

Names in play

Micron (MU) is the clear name in play, with a catalyst on Tuesday. The bull read (Futurum): pricing power intact, the stock at "five times earnings" prices in a lot of bad news already, and the 13-versus-14-week quarter will make the guide look worse than it is. The bear read (Awad): margins at 86% are a peak, not a floor, with Chinese supply and demand destruction closing in. Both camps agree on what to watch: the gross-margin guide, and what management says about capex and 2028 supply. Even the bulls note the market has turned tough on any margin slip. As one Futurum host put it, if Nvidia "had to go from 75 to 72, there will be people that will complain about that."

TSMC (TSM) gets a quiet long-term vote of confidence from an operator building on its lines (Ayar Labs), and Awad calls herself bullish on both TSMC and ASML ahead of their reports, while warning that last quarter both "sold off after print." The near-term question is whether October's reports act as a "test" or a "catalyst."

Read-throughs

  • CoWoS and advanced packaging. Co-packaged optics adds a whole new element (a photonics chip, a 3nm driver chip and a fiber connector) to packages that are already the industry's tightest bottleneck. That supports advanced-packaging demand at TSMC and outsourced packaging houses from 2028. Wade's "optical engines thought about like HBM" line suggests optics could eventually become a standard attachment, the way HBM stacks did.
  • Test and metrology. Wade's "five test insertions," and investors asking about "probers," point to test as a growing share of the value in advanced packages. This echoes the "bottleneck has moved into testing" theme from earlier this month. Our inference, not something said on the podcasts: test and inspection specialists (Advantest, Onto Innovation, Camtek on the task list) are the natural listed beneficiaries, but none were named directly this week.
  • Memory capex versus logic capex. Micron's capex commentary on September 30 is the next hard signal for memory-driven equipment orders. Earlier supply means near-term tool orders; a 2028 timeline keeps them back-loaded.
  • Fabless customers and design partners. MediaTek, Alchip and GUC are becoming the path by which hyperscaler custom chips get new packaging tech. That extends last month's "MediaTek inside Google" crack in Broadcom's custom-chip dominance.
  • China-domestic suppliers. The China share threat this week was in memory (CXMT), not equipment. No Chinese toolmaker was discussed.
  • Materials and gases. The Taiwanology conversation is a reminder that Taiwan's fabs rely on European gases, water treatment and substrates (Air Liquide, Veolia, Soitec). Concentration risk runs both ways.
  • Quantum hardware. IonQ's real-time decoder is a genuine step on error correction, the problem that has to be solved before quantum machines are useful. But the consensus timeline is still "late 2020s" at the earliest for commercial use, and pure plays keep raising money. IBM, Alphabet and Microsoft remain the lower-risk way to track the field.

What changed from prior weeks

  • The operator drought broke on the manufacturing side, sort of. Last week our best operator was a TSMC customer (Cerebras). This week it's a TSMC customer and partner building on its COUPE line, with wafers in production. Still no voice from TSMC, ASML or the big tool-makers themselves.
  • The bottleneck conversation moved again. From CoWoS, to substrates, to testing (September 13), and now to lasers, fiber attach and multi-step optical testing.
  • The memory debate flipped from "how high" to "how long." Last week the memory story was Tim Cook warning laptop prices could double. This week the question is whether prices have peaked, with TrendForce and Nanya data on one side and Futurum's "not one" sign of erosion on the other.
  • The bear case got a calendar. Last week's worry was that the collateral behind AI debt might not hold. This week Kedrosky put a 2029–2030 refinancing date on it, and Hou pushed back with rental data showing GPU residual values are higher than assumed.
  • The quantum timeline got pushed out again. Last week one guest argued "Q-Day" (the day quantum computers can break encryption) could be under five years away. This week's consensus framing put encryption-breaking in the mid-2030s, with a real error-correction milestone from IonQ in between.