Newsletter · 2026-10-05 · Ashutosh Agarwal

Micron Says Memory Stays Tight Through 2028 - AI Accelerators: GPUs, Custom Silicon & Optics - Week of October 5, 2026

AI Accelerators: GPUs, Custom Silicon & Optics for the week of October 5, 2026. Podcast synthesis on Micron's fiscal Q4 beat and its CEO's warning that the memory shortage gets tighter through 2027 and 2028, the strategic customer agreements and HBM price resets now running to 2031, why the cheap stock barely moved, CoreWeave and Crusoe operators on rising GPU rental margins and depreciation, Fractile on the memory-bandwidth wall, Anthropic's roughly $518 billion compute bill, and a big unexplained rally in optical networking stocks.

AI Accelerators: GPUs, Custom Silicon & Optics

Week of October 5, 2026: Micron Says Memory Stays Tight Through 2028

AI Accelerators: GPUs, Custom Silicon & Optics. Issue 030, Monday, October 5, 2026. Covers podcasts published September 28 to October 5.

Last Thursday I had to leave a hole in the middle of this newsletter. Micron ($MU) had reported after the close on September 30, but I couldn't confirm the numbers yet, so I didn't print them.

The hole is filled now. The numbers are big, and the guidance is bigger. The line that stayed with me all weekend came from Micron CEO Sanjay Mehrotra on CNBC the next morning:

"Supply is tight in '27 and '28, even tighter than '26."

Think about what that says. Memory chips have been the most boom-and-bust corner of the semiconductor industry for forty years. Now the CEO of the only US memory maker is telling investors the shortage gets worse for two more years. He's not saying it's ending.

The stock barely moved. Micron closed at $1,065.11 the day of the report and $1,074.89 on Friday, up less than 1%. Meanwhile the optical networking stocks, which almost no podcast talked about, rose 12% to 17% in two days.

That gap between what's being said and how the stocks are moving is the story of this issue.

TL;DR

• Micron beat and raised, and said the shortage lasts longer. Fiscal Q4 (June to August 2026): EPS of $33.42 against a $31.61 estimate, revenue of $54.23 billion against $51.0 billion expected, and an 87% gross margin (Closing Bell, 9/30). Next quarter's guide is about $61.5 billion in revenue (The Rundown, 10/1).

• The one soft spot was margin. Guidance of roughly 86.25% gross margin came in a touch below the Street's 86.6%. Bloomberg Intelligence pinned part of it on pay raises.

• Micron now sells memory years in advance. More than 75% of fiscal 2027 shipments are already committed. Customers are signing supply deals that run to 2031. Strategic customer agreements went from 16 to 26 since August.

• The bull case from Melius: Ben Reitzes ($2,200 target) argues the stock is priced as if margins "tank back into the 60s," when high-bandwidth memory margins are "actually going up a lot."

• The operators agree on scarcity. CoreWeave CEO Mike Intrator says margins are "expanding faster than the rates are going up" and sees no drop in orders. Crusoe CEO Chase Lochmiller says Hopper GPUs cost more in 2026 than when they were new in 2023.

• The newer debate is about chip design. Fractile CEO Walter Goodwin says the industry is hitting a memory-bandwidth wall. He's building a chip that targets "25 times more bandwidth per chip than an HBM based chip," ramping in the second half of 2027.

• Anthropic's leaked IPO filing put a number on AI's compute bill. Several podcasts cited about $518 billion in future compute commitments. Bloomberg Intelligence also raised a reported $42 billion Broadcom-linked lending deal for chip leasing, which is unverified.

• Prices (9/30 to 10/2): Coherent +17.1%, Credo +12.2%, Lumentum +11.8%, AMD +3.6%, Marvell +3.1%, Nvidia +2.4% (a record high Friday), Micron +0.9%, Astera Labs −1.6%. Nasdaq +1.2%.

1. Micron's quarter: the numbers, finally confirmed

Here is what Micron reported for fiscal Q4 2026 (the three months ended late August 2026), as read out on CNBC's Closing Bell the evening of the report (Closing Bell, "Markets Close Out September With AI in Focus 9/30/26"):

The Rundown, a podcast from Public.com, added the revenue guide for the current quarter (FQ1 FY27, September to November 2026): about $61.5 billion. That's another roughly 13% jump in a single quarter. For context, the same quarter a year earlier brought in $11.3 billion (The Rundown, "Micron Posts Monster Quarter, Apple Readies Smart Home Devices").

Mehrotra summed up the year himself on Squawk on the Street the next morning (Squawk on the Street, "9AM HOUR: … Micron CEO Exclusive … 10/1/26"):

"We delivered an exceptional quarter, exceptional year with $133 billion in revenue, more than 250% up year over year, with 87% gross margins in FQ4. And we guided to 86.25% gross margin, approximately that for our FQ1."

A quick word on gross margin: it's the share of each sales dollar left after paying to make the product. An 87% gross margin means that of every $100 in chips Micron sells, only $13 goes to making them. That's a software-company margin, earned by a company that runs factories.

Adam Parker of Trivariate Research said beforehand that a quarter like this could be "one of the best quarters ever produced by a manufacturing company in the history of humankind" (Squawk on the Street, "10AM Hour: Nvidia's $150B Buyback … 9/28/26"). He wasn't far off.

Why it matters: Micron's results are a direct read on how much AI hardware is actually being shipped. Every Nvidia GPU and every Google TPU needs stacks of high-bandwidth memory beside it. When Micron says it "cannot fulfill the demand of our customers," that's also a statement about how many accelerators its customers want to build.

2. What the operators said: shortages, contracts, and a long capex runway

The most useful material this week came from Micron's own executives, on CNBC and on the analyst call that followed the report. These are company insiders, so treat this section as management's view.

On how long the shortage lasts. Mehrotra on Squawk on the Street (10/1):

"Supply is tight in '27 and '28, even tighter than '26 … continued, strong, durable, predictable financial performance."

On Closing Bell he was quoted as having no "line of sight to when this whole supply demand imbalance will sort of peter out" (Closing Bell, 9/30).

On contracts. Micron now has 26 "strategic customer agreements," up from 16 in August. These are multi-year deals in which customers commit to volumes in exchange for guaranteed supply. Mehrotra:

"Customers are extending these agreements … even now, into 2031 timeframe."

On the analyst call, President and COO Manish Bhatia added the detail that matters for modeling. He said about three-quarters of contract revenue has a defined pricing framework, mostly with floor and ceiling bands (a minimum and a maximum price). The other quarter is renegotiated periodically as the market moves. New contracts are being set "with an eye towards the current market conditions and the future market tightness," so the price bands are moving up (Micron FQ4 FY2026 post-earnings analyst call).

On HBM pricing. Bhatia said Micron has "increased that pricing significantly for calendar year 2027." The new prices reset at the start of the year, "to narrow the profitability gap with conventional DRAM" (analyst call). Put simply, the memory used beside AI chips (HBM) had been less profitable than ordinary computer memory, because ordinary memory prices rose so fast. Micron is now pushing HBM prices up to close that gap. That matters for Nvidia, AMD and Broadcom's cost of goods next year.

On why supply can't catch up. Bhatia gave three reasons on the call:

• HBM uses up more factory capacity. Each new generation (HBM3E, then HBM4, then HBM4E late in 2027) needs more silicon wafer per bit of memory. The industry calls this the "trade ratio."

• New manufacturing nodes add fewer bits than they used to.

• Cleanrooms take years to build. A cleanroom is the ultra-clean factory floor where chips are made. Micron's Idaho fab produces its first wafers in mid-calendar 2027, but "meaningful supply growth takes a few quarters after that."

On capex. Mehrotra said first-half fiscal 2027 capex "will be around $25 billion. That's almost as much as what we invested in full fiscal year '26." He cited "$250 billion of investment in Idaho, New York, and Virginia" over time (Squawk, 10/1). On the call, Bhatia and CFO Mark Murphy explained that most of the increase is for building cleanrooms that open "in late 2028 and beyond." Equipment goes in only "when needed based on our view of the market." That's how a memory company tries to avoid the overbuild that has ended every past cycle.

On a custom Nvidia memory part. CTO Scott DeBoer said Micron has worked with Nvidia "for over a year on HBM4E, what's called NVHBM." He called it "the first major custom HBM product out in the market" (analyst call). Custom HBM ties a memory supplier more tightly to one accelerator design. That's good for pricing power and switching costs, and it's another example of Nvidia locking up its supply chain.

On CPUs. This is the demand driver few people expected a year ago. Bhatia said "agentic workloads are executed across CPU" with "a large attach rate" of memory and storage. That's why server units are growing "in the high teens." He specifically named Meta's Muse agent as an example. In his words, this makes it "clear that DRAM is the principal constraint versus logic or power to the data center." Coming from a memory maker, that claim is self-interested, but it's also the reverse of the usual view that power is the bottleneck.

On China. DeBoer said Micron is "at least two nodes ahead of the China competition." Bhatia said China sales exposure will be "in the single-digit range in fiscal 2027."

3. The debate: why didn't Micron go up?

The stock rose less than 1% from the report to Friday's close. It went up about 3% on October 1 (Stock Market Today With IBD, 10/1), then gave it back. The Rundown noted it trades at less than 7 times forward earnings, against 18.5 times for the S&P 500. Based on FactSet consensus, I get about 6.2 times this fiscal year's EPS estimate of $172.67. The average analyst price target of $1,592.92 implies 48% upside.

When a stock this cheap shrugs off a quarter this good, the market is telling you it doesn't believe the profits will last. Here are both sides, each argued at its strongest.

The bear case: this is still a cycle.

• Gross margin guidance went down a little, from 87% to about 86.25%. In memory, the first margin dip has often marked the top. Bloomberg Intelligence titled its episode "Micron Gives Bullish Forecast, Even as Pay Raises Crimp Margins" (Bloomberg Intelligence, 10/1).

• CFO Mark Murphy said price increases in fiscal 2027 will come "at a more moderate pace" (analyst call). Slower growth in prices is still growth, but stocks tend to follow the rate of change.

• Competitors are adding capacity. On Fast Money, Tim Seymour noted Samsung and SK Hynix are "building out more memory and more capacity," and Chinese maker CXMT is ramping (Fast Money, 9/30).

• Fixed-price contracts cut both ways. Before the report, Bernstein's Stacy Rasgon asked "whether long-term agreements … have capped their ability to continue to take price" (Squawk, 9/28). Contracts with ceilings limit the upside in a shortage.

The bull case: the cycle has changed.

• Ben Reitzes of Melius Research ($2,200 target) made the sharpest version on Squawk on the Street (10AM Hour, 10/1):

"[The stock trades as if] margins are going to tank back into the 60s … [but HBM margins] are actually going up a lot. So margins might be able to stay in the mid to high 80s much longer than people think."

• Reitzes also said Micron's buyback "might be 15% a year, easy" in 2027–2028. That's his projection, and Micron hasn't guaranteed it. Mehrotra said Micron will have "well over $100 billion in cash" by the end of this quarter (Squawk, 10/1). I'm repeating that as he said it on air. It isn't in the transcript I could check.

• Wedbush's Matt Bryson said the small margin dip isn't a warning sign if it comes from shipping more HBM: "All of my price checks have suggested that pricing is still moving up" (Closing Bell, 9/30).

• Bloomberg Intelligence's Jake Silverman said meaningful new fab capacity is a "2029 story," and HBM is expected to see more price increases. He said Samsung announced a "$100 increase in at least some of their SKUs" (Bloomberg Intelligence, 10/1).

• The supply deals are the structural change. On the IBD podcast, Ed Carson summarized: "We already have 75% of our sales already baked in for fiscal 2027 … We're already selling fiscal 2028" (IBD, 10/1).

My read: I think the bears are arguing about the right thing, which is the price trend, but over the wrong time frame. The new information this quarter is in the contracts. Micron has 26 customers, deals running to 2031, and price bands being reset higher. That makes 2027 revenue far more visible than in any past memory cycle. What would change my mind is the 25% of contract revenue that floats with the market. If those prices flatten, the "stays tight" story gets tested early. The next real check is Samsung's and SK Hynix's October results.

4. The neoclouds: margins are rising faster than prices

Two operators who rent out GPUs gave unusually direct answers this week.

CoreWeave ($CRWV), CEO Mike Intrator on Squawk on the Street (9AM Hour, 9/30):

• CoreWeave says its customer Cognition is the "first client in the world to enter into production on Vera Rubin architecture," Nvidia's next-generation platform.

• On profitability: "Margins that we are earning are expanding faster than the rates are going up," and "the profitability of the infrastructure [is] higher than it was a year ago."

• On the political backlash against data centers: "There has absolutely not been any reduction in orders." Regulations "are going to relocate the infrastructure," not reduce demand.

• He sees the demand trajectory running "through 2027 and into early '28." He also flagged "substantial demand for CPUs across the board," inside and outside GPU servers. That matches what Micron said about agentic AI.

Crusoe, founder and CEO Chase Lochmiller on 20VC (20VC, "The Future of Datacentres … with Chase Lochmiller," 10/3):

• Crusoe built more than 200 MW of compute in Abilene, Texas, in a year. The next-best competing bid would have taken 2.5 years. The first eight buildings are budgeted at 140 MW each.

• On the GPU-depreciation worry, he said "the GPU is actually the most valuable thing in the entire data center." He said a six-year depreciation schedule is standard, and Hopper GPUs are priced higher in 2026 than when they were new in 2023. (Depreciation is the accounting charge for an asset wearing out. Bears argue GPUs lose value faster than the books assume. Lochmiller's point is that resale and rental prices say otherwise, for now.)

• Power gear is the bottleneck. Lead times for power distribution centers are about 100 weeks in the market, and Crusoe cut that to 28 weeks by making them itself.

• His sobering estimate: about 50% of planned data centers will not go online. That's his opinion, not data, but it's notable coming from a builder. If he's right, the companies that do get built have more pricing power, and headline capex plans overstate real chip demand.

Why it matters for the chip names: Rising rental margins at the neoclouds are the clearest sign that end demand for Nvidia's hardware is real and not just financed. It doesn't settle the "circular money" question from Issue 029. But it does make the bear case harder: you can't call GPUs a stranded asset while their used prices are going up.

5. Custom silicon and the memory-bandwidth wall

The most technically interesting podcast of the week was No Priors with Walter Goodwin, founder and CEO of Fractile, a UK chip startup of about 150 people (No Priors, "The Future of Frontier Model Architectures with Walter Goodwin," 10/2). He's an operator, but a startup one, so his performance claims are aims, not shipped results.

• Fractile is moving away from its original SRAM-based design (SRAM is very fast, very small on-chip memory) toward "much, much, much higher bandwidth to DRAM memories."

• His target: "25 times more bandwidth per chip than an HBM based chip," ramping in the "second half of next year." He's aiming for "thousands of tokens per second" on models with trillions of parameters. That target is unverified.

• On Nvidia's moat: "If you crack open an Nvidia system it has anywhere between kind of six and nine custom chips all built by Nvidia." That's his explanation of why Nvidia is hard to displace. It sells a system, not a chip.

• He listed the custom-chip field as Google's TPU (10+ years old), Meta's MTIA, Microsoft's Maia and OpenAI's chip effort. He called Broadcom ($AVGO) the largest custom-chip design house.

Why it matters: Goodwin and Micron are describing the same problem from opposite ends. AI inference (running a trained model to answer requests) is limited more by how fast data moves between memory and the processor than by raw compute. Micron profits from that bottleneck. Fractile is betting it can design around it. Over the next two years, the scarcest resource in AI hardware may be memory bandwidth, not GPUs.

AMD ($AMD) made a software move rather than a chip move. On Bloomberg Tech, CEO Lisa Su confirmed an all-stock deal to buy Fei-Fei Li's World Labs for $8.2 billion, expected to close by end of 2026 (Bloomberg Tech, "AI Takes Center Stage From Washington to Wall Street," 9/29). The pundit reaction was sceptical. Ben Pouladian on Monetary Matters noted AMD "haven't even shipped a rack yet" (Monetary Matters, 9/30). We covered his Helios skepticism in Issue 029. Dan Ives took the other side on The Pomp Podcast: AMD is "just starting," and "I just think you have to own chips" (The Pomp Podcast, "The AI Boom Is Bigger Than Anyone Thinks | Dan Ives," 10/1). AMD rose 3.6% from 9/30 to 10/2 and sits 1.8% below its 52-week high.

6. Anthropic's compute bill and the Broadcom read-through

Anthropic's leaked IPO prospectus (an S-1, the filing a company makes before going public) dominated the general tech podcasts. Most of that talk was about Anthropic's losses, not hardware. But the compute numbers matter for the chip names:

• On Bloomberg Tech, a Reuters report was cited saying the prospectus shows more than $500 billion in commitments to cloud infrastructure, compute and data centers (Bloomberg Tech, 9/29). Several other podcasts used a figure of about $518 billion, including Tom Bilyeu's Impact Theory (10/3) and Tech Brew Ride Home (9/29). The filing isn't public, so treat these as reported figures.

• On Bloomberg Intelligence, Bloomberg's Sarah Frier described a $42 billion lending deal tied to Broadcom for chip leasing. She also put total circular financing deals at "over $1 trillion … largely driven by NVIDIA" (Bloomberg Intelligence, 10/1). The exact terms weren't explained on air, so treat the $42 billion as unverified.

• In Issue 029, Erica Clower put Anthropic's Nvidia hardware commitments at $180 billion. If the total is about $518 billion, a large share of the rest is going elsewhere. Pouladian believes that's mainly Google TPUs, which Broadcom co-designs, rather than AWS Trainium. That's his opinion, not something the filing confirms.

Why it matters: Broadcom is 28% below its 52-week high, with a 50% gap to the average analyst target. That's the widest of any name we track besides Micron. If Anthropic's prospectus confirms a large TPU commitment, it's the most direct evidence yet behind Broadcom's custom-chip backlog. It's also the clearest example of the financing worry: a chip supplier helping fund its customer's purchases.

7. Optics: the biggest move, the least talk

From the Micron report (9/30) to Friday (10/2), the optical networking stocks were the best performers we track:

Source: FactSet closing prices.

The only podcast discussion was a short item on Stock Market Today With IBD. Alysa Coram noted Lumentum ($LITE) rose 7.7% on October 1 and closed above $1,000. She cited "triple-digit growth on both the top and bottom lines for a couple of quarters" (IBD, 10/1).

On Bloomberg Intelligence, Jake Silverman put co-packaged optics (moving the laser connection onto the chip package itself) on Nvidia's supply-chain list, alongside HBM4 and its TSMC wafer commitments (BI, 10/1).

That's the fifth issue in a row where optics has moved the most with the least explanation from the people talking about AI hardware. A 17% two-day move in Coherent with no podcast story behind it usually means one of two things: the move is about stock positioning, or the explanation is in sell-side research that hasn't reached the podcasts yet. Credo's rebound erased last week's unexplained 7.7% drop. I'd treat that as noise in both directions until a company says something.

Names in play

FactSet closes and consensus. Forward P/E is the 10/2 close divided by consensus EPS for the current fiscal year. Fiscal years differ by company, so compare these across names with care.

• Micron ($MU): The cheapest name in the group on earnings, with the biggest gap between operator commentary and stock reaction. The next test is the 25% of contract revenue that floats with market prices.

• Nvidia ($NVDA): Hit an all-time high Friday, with market cap above $5.5 trillion (Closing Bell Overtime, 10/2). Its new custom memory part with Micron (NVHBM) and CoreWeave's first Vera Rubin production customer are both signs the next platform is on schedule.

• Broadcom ($AVGO): The Anthropic filing is the swing factor. It has the largest gap to target outside Micron.

• Marvell ($MRVL): Investor Day is tomorrow, October 6. No podcast previewed it this week.

• Coherent, Lumentum, Credo: Rallying hard without a podcast explanation (see section 7).

Read-throughs

• Micron → SK Hynix, Samsung: "Tighter in 2028 than 2026" and an HBM price reset for calendar 2027 both point to strong October results from the Korean memory makers. Samsung's reported $100 SKU price increase (BI) is consistent.

• Micron → Nvidia, AMD, Broadcom costs: Higher HBM prices in 2027 raise the bill of materials for every accelerator. Nvidia has the pricing power to pass it on. Custom-chip customers bear more of it directly.

• Micron → equipment makers: Construction capex is going up first, and tools go in only "when needed." That puts equipment orders later in 2027–2028, not now.

• CoreWeave and Crusoe → GPU residual values: Rising Hopper prices and margins growing faster than rental rates support the idea that GPU useful lives match the books. That weakens the "stranded asset" bear case for now.

• Micron CPU commentary → server CPUs: Agentic AI is driving server units up in the high teens. That helps AMD's EPYC server CPUs and Arm-based server chips, not just GPUs.

• Fractile → HBM long term: If bandwidth-first designs work, they would raise demand for DRAM bandwidth but could change which memory products win. That's a 2028+ question, not a 2027 one.

What changed vs. Issue 029

• Micron's numbers are confirmed. Last issue held them back. This issue publishes the confirmed figures: $33.42 EPS, $54.23B revenue, 87% gross margin, about $61.5B next-quarter revenue guide, $133B fiscal-year revenue.

• The preview targets held up. Susquehanna's Mehdi Hosseini had sketched a path to "$50 quarterly" EPS (Issue 029). The guide points that way, but the stock didn't follow.

• The talk moved from Nvidia's buyback to the duration of the memory shortage. The buyback and "circular money" debate from 029 continued only in passing.

• Operators replaced pundits as the main source. Micron, CoreWeave, Crusoe, Fractile and AMD all spoke directly this week.

• Anthropic's compute number got firmer. Last issue it came only from Pouladian ("over half a trillion"). This week multiple podcasts cited about $518 billion, from the leaked prospectus via press reports.

• Optics turned from worst to best. Credo went from −7.7% (9/25→9/30) to +12.2% (9/30→10/2). Coherent went from −2.7% to +17.1%.

Coming up: Marvell Investor Day (October 6). CME GPU compute futures launch, expected around October 5. TSMC Q3 results mid-October (flagged on IBD). Samsung and SK Hynix Q3 results later in October. Nvidia reports around November 18.