Hardware

The Market Just Remembered AI Runs on Memory, Not Magic

NVIDIA fell hard this week, the chip sector fell harder, and the headlines blamed a CEO scandal and an election. Underneath is one engineering fact the market keeps re-learning: AI inference is bound by memory bandwidth, so the whole trade is a bet on memory, power, and buildout.

a silicon wafer with a grid of microchip dies
Photo by Laura Ockel / Unsplash

I own thirty dollars of semiconductors. Not thirty dollars a share — thirty dollars total, spread across eleven positions I built over roughly a single day of trading, a dollar here, five dollars there, with a standing habit of adding one more dollar each Friday whenever something has dropped. NVIDIA, AMD, Broadcom, TSMC, Intel, Micro­chip, Super Micro, a D-Wave quantum-computing lottery ticket, an IT index fund, a telehealth name that has nothing to do with chips, and a managed "bold" bucket that is currently down more than forty percent. It's a toy. I built it to watch how the machine behaves, not to make rent.

This week the machine behaved like it had seen a ghost. My toy portfolio dropped a couple of percent in a few days, which on thirty dollars is a rounding error I had to squint to see — but the thing it was tracking, the actual semiconductor complex, had one of its worst days in six years. NVIDIA alone shed hundreds of billions in market value, fell again the next sessions, and by the time the dust settled on 10 November it had given back its gains for the year. The headlines blamed a CEO scandal and an election.

Both of those are real and I'll get to them. But neither is the story, because neither explains why memory companies fell twice as hard as NVIDIA did. To understand that, you have to stop reading the finance page and start reading the hardware, because what the market did this week was briefly remember a fact it spends most of its time forgetting: an AI accelerator is only as fast as the memory feeding it, and the whole trillion-dollar edifice rests on a supply of memory chips that physics and three companies control.

A note before any of this: I'm not a financial advisor, nothing here is investment advice, and the fact that I own these tickers should be read as a disclosure, not a recommendation. This is an article about technology that happens to have a stock chart attached.

Two Different Panics Wearing One Headline

The first thing to untangle is that "the chip selloff" this week was actually two separate events with two separate causes, and conflating them is how you end up with a wrong model of what happened.

Wednesday 5 November was the sector event. The Philadelphia Semiconductor Index — the SOX, thirty of the largest chip companies — fell more than ten percent intraday, on track for its worst single day since the COVID crash of March 2020. The damage was not evenly distributed, and the distribution is the clue:

Company Ticker 5 Nov drop (intraday) What they make
SanDisk SNDK ~18% Flash memory
AMD AMD ~14.5% GPUs, CPUs
Marvell MRVL ~13.5% Data-center silicon
Micron MU ~13.5% DRAM and HBM memory
Qualcomm QCOM ~10.5% Mobile SoCs (also missed earnings)
TSMC TSM ~9.5% Contract fabrication
NVIDIA NVDA ~8% AI accelerators
Broadcom AVGO ~8% Networking, custom AI silicon
Intel INTC ~6.5% CPUs, foundry

Look at the top of that list. The hardest-hit names are the memory makers — SanDisk, Micron — not NVIDIA. Hold that thought; it's the whole article.

The trigger was political, not technical. The 4 November US midterms handed Congress to the Democrats, and within a day the market's attention swung to a specific local-politics problem: communities voting to slow or stop data-center construction. Ohio's governor-elect backed a "conditional moratorium" on new data centers; Texas's governor-elect campaigned on an immediate one. Data centers are where AI chips go to live, so a credible threat to the buildout is a credible threat to every revenue line that depends on it. Nothing was enacted — these are campaign positions and early signals — but the market priced the risk immediately, because the AI trade is a bet on an uninterrupted buildout, and "uninterrupted" is exactly what a moratorium threatens.

Monday 10 November was the NVIDIA event, and it was company-specific in a way the numbers make obvious: NVIDIA fell 5.52% to close at $168.80, on volume more than triple its three-month average, while AMD rose 1.30% to $515.21 and Broadcom climbed 3.91% to $331.98 the same day. It did that, moreover, on a session when hotter-than-expected inflation data was already dragging the broad Nasdaq down 0.74%. When the macro tape is red, the sector peers are green, and only one megacap chip is bleeding, the cause is inside that one name. When the peers go up and one name goes down, the cause is inside that one name. Reports had surfaced of an independent board investigation into Jensen Huang connected to an alleged foreign-intelligence matter, with the specifics thin and unconfirmed and a congressional subpoena threat attached. I'm going to report that and move on, because the allegation is unproven, the details are not public, and speculating about a named individual's conduct is neither my job nor the point. What matters for this article is structural: it's a governance overhang landing nine days before NVIDIA's 17 November earnings report, and it is entirely separable from the question of whether the technology is overbuilt. The market treated them separately. So will I.

The Fact Under the Ticker: Inference Is Memory-Bound

Here's the engineering that explains why memory stocks lead the sector in both directions, and it is the single most useful thing to understand about the economics of AI hardware.

When a large language model generates text, it does it one token at a time, and each token requires reading the model's entire set of weights out of memory. A 70-billion-parameter model in 16-bit precision is about 140 gigabytes of weights. To produce one token, the accelerator streams all 140 GB from memory through the compute units, does a comparatively tiny amount of arithmetic on them, and emits a single word-piece. Then it does it again for the next token. And again.

This is the part people miss: during that process, the expensive matrix-multiply units on the GPU are mostly idle, waiting. The operation has what engineers call low arithmetic intensity — very few floating-point operations performed per byte loaded from memory. The classic way to see this is the "roofline" model, which says achievable performance is the lesser of two ceilings: the chip's peak compute, and its memory bandwidth multiplied by how much arithmetic you do per byte. Token generation lives far out on the memory-bandwidth side of that roofline. The compute ceiling is irrelevant; you hit the bandwidth ceiling long before you get near it.

Which means the number that determines how fast an AI accelerator generates text is not its advertised TFLOPS. It's its memory bandwidth. An NVIDIA H100's headline is its compute, but its real governor is the roughly 3.35 terabytes per second of bandwidth from its High Bandwidth Memory. Double the bandwidth and you roughly double token throughput; double the FLOPS while holding bandwidth constant and, for this workload, little happens. The accelerator is a very expensive straw, and HBM is the diameter of the straw.

That's why the picks-and-shovels of the AI boom are not really the shovels. They are the memory. And memory, it turns out, is the most supply-constrained and most violently cyclical link in the entire chain.

We recommend reading The RAMpocalypse! How AI Companies Are Destroying the Hosting Industry and Your Wallet to continue reading our selection of content. The same memory shortage that moved these stocks this week is the one that has been quietly doubling your VPS and dedicated-server bills.

The RAMpocalypse! How AI Companies Are Destroying the Hosting Industry and Your Wallet How AI's appetite for memory chips drove a shortage that reaches all the way down to ordinary hosting invoices. SudoSecurity

HBM: Three Companies, One Bottleneck, 121% Price Moves

High Bandwidth Memory is not ordinary RAM. It is a stack of DRAM dies — twelve high, in the current generation — mounted directly beside the accelerator die and wired to it through thousands of connections, so it can feed data far faster than the DDR5 sticks in your laptop ever could. The current standard, HBM4, which JEDEC finalized in April 2025, doubles the per-stack interface from 1,024 to 2,048 I/O lines and pushes past three terabytes per second per stack. It is one of the hardest things in the industry to manufacture, and essentially three companies on earth can make it: SK Hynix, Samsung, and Micron.

That concentration is the fragility. Industry estimates put SK Hynix at 60–70% of the HBM4 going into NVIDIA's next-generation Vera Rubin platform, Samsung at 25–30%, and Micron at the remainder. When demand from hyperscalers and AI labs outruns what those three can fabricate — which it has — the price does not drift, it spikes. Contract pricing for HBM4 is projected to rise around 121% year over year into 2027. A single twelve-high, 36 GB HBM4 stack is estimated to go from roughly $600 this year to roughly $1,300 next. SK Hynix has warned the shortage could persist past 2030. AMD's Lisa Su, who needs this memory for her own accelerators, put it plainly: "Memory supply is very tight."

Now the cyclicality makes sense. Memory is a commodity with brutal operating economics: when it is scarce, margins explode, and when the fear of oversupply appears, the same force runs in reverse and the stocks fall off a cliff. This week the market got scared that the buildout feeding that demand might slow — the moratorium story — and it sold the memory names hardest because they are the most exposed to the demand curve bending. Micron falling 13.5% while NVIDIA fell 8% is not noise. It is the market correctly identifying that the memory makers are the most amplified bet on exactly the thing that got threatened. The hardware tells you where the risk concentrates; the stock chart is just the echo.

The Depreciation Argument That Will Not Go Away

There is a second technical fight underneath the selloff, and it is about accounting for hardware rather than hardware itself — which makes it sound boring and makes it, in fact, the most interesting structural question in the whole trade.

Michael Burry — the "Big Short" investor — has been pressing it since late 2025, and this week's drop dragged it back into the conversation, so let me date it plainly: this is a resurfacing argument, not a new November 2026 revelation. The claim is this. Hyperscalers depreciate their GPU fleets over roughly five to six years on their books. But the real useful life of an AI accelerator, Burry argues, is closer to two to three years, because each NVIDIA generation is two to three times more efficient than the last, which makes the previous generation economically obsolete long before it physically dies. If you depreciate a two-to-three-year asset over five to six years, you understate your costs, which overstates your profit. Burry's estimate is that this inflates industry profits by something like $176 billion between 2026 and 2028, overstating Oracle's earnings by as much as 27% and Meta's by around 21% by 2028. He is short NVIDIA and Palantir, so weigh the source accordingly.

The counterargument is also technical and also real, and it is worth stating because the truth is far from settled. A chip losing its performance crown does not mean it loses its earning power. CoreWeave reported re-booking H100 capacity coming off expired contracts at about 95% of the original price. Jensen Huang has pointed to H100 rental rates rising 22% in a single month, and to a contract renting decade-old A100 chips out through 2029. A GPU can be two generations behind and still rent profitably, because demand so exceeds supply that even "obsolete" silicon finds a tenant. So the useful-life question does not have a settled answer, even though it is the one that determines whether the hyperscalers' reported earnings are truthful. Anyone who tells you it obviously does is selling something, long or short.

For our purposes the point is not who wins. It is that the valuation of the entire AI complex rests on an accounting assumption about how fast specific pieces of hardware wear out economically, and reasonable, technically literate people disagree about that number by a factor of two. That is a lot of trillion-dollar weight resting on a depreciation schedule.

The Circular Part

The last structural thread is the one that turns a hardware story into a systemic-risk story. NVIDIA does not merely sell accelerators; it invests in the companies that buy them. It has committed enormous sums to OpenAI, taken stakes in "neoclouds" like CoreWeave and Nebius, and generally helped finance the demand for its own product. Analysts have mapped something like a three-quarters-of-a-trillion-dollar web of financing arrangements radiating out from NVIDIA into its own customer base.

The engineering-adjacent reason this matters: the neoclouds borrow money to buy GPUs, and increasingly they borrow against the GPUs — the accelerators are the collateral. That is fine as long as the collateral holds its value and the rental demand stays hot, which loops you straight back to the depreciation question above. If GPUs really do depreciate on a two-to-three-year curve, then a lot of debt is secured by assets melting faster than the loans amortize. If they hold value like Huang's rental figures suggest, the structure is sound. The circularity is not inherently fraudulent — vendor financing is ancient — but it concentrates risk, because the same asset is simultaneously the product, the collateral, and the thing propping up the customer that generates the revenue. When confidence wobbles, as it did this week, every one of those roles wobbles at once.

We recommend reading What Is Supply-Chain Security? to continue reading our selection of content. The AI hardware boom is a supply chain with a single dominant vendor financing its own demand, which is a concentration risk whatever you think of the technology.

What Is Supply-Chain Security? Why the structure of how value flows through a supply chain matters as much as any single link in it. SudoSecurity

What My Own Eleven Positions Actually Did

Here is the part my toy portfolio taught me that a textbook wouldn't, and it is a correction to my own thesis as much as a confirmation of it. If the memory-bound story explained everything, my eleven positions would have fallen together, in rough proportion to how close each one sits to the HBM bottleneck. They didn't. They fell for five different reasons, and the spread is the lesson.

  • Super Micro — the AI-server builder — dropped about 10% on 5 November with no company-specific news at all. That is the pure risk-off move: a box-maker whose entire order book is AI buildout, sold because the buildout got questioned. This one is the memory story at one remove.
  • Microchip fell almost exactly as hard — about 10%, to $58.35 — on 6 November, and for a completely unrelated reason. It had just beaten on earnings, $0.97 in adjusted EPS against a $0.936 estimate, and still cratered, because its December-quarter guidance was soft and its margin outlook slipped. That is not the AI-memory trade. That is an analog- and embedded-chip company with its own humdrum automotive-and-industrial cycle, and it would have fallen this week with or without NVIDIA.
  • The D-Wave lottery ticket fell about 18% intraday on 5 November, worse than any actual chipmaker I hold. Quantum-computing stocks have essentially nothing to do with memory bandwidth, HBM, or AI inference; they are pre-revenue bets that trade as pure high-beta proxies for the market's appetite for speculative tech. When the mood sours they fall hardest, because they are the most mood. No news required.
  • NVIDIA fell on its own scandal days later, while AMD and Broadcom rose — covered above.
  • And the telehealth name I threw in as a control — a company with no silicon exposure whatsoever — spent the week trading on its own earnings calendar, indifferent to the entire drama.

Five holdings, five mechanisms: sector risk-off, company-specific guidance, high-beta sentiment, idiosyncratic governance, and total irrelevance. The memory-bound thesis is the right explanation for the center of the move — why Micron fell 13.5% while NVIDIA fell 8%, why the memory names always lead in both directions. It is not a theory of everything. My own portfolio, which cost me less than a sandwich, is a standing reminder that "chip stocks fell" is four or five different stories wearing one headline, and that the engineering only explains the ones actually bolted to the bottleneck.

What This Doesn't Cover

I am a technologist writing about the engineering under a market move, and that is the only lane I am claiming. I have not valued any of these companies, I am not predicting where any stock goes next, and the numbers here — the intraday percentages, the HBM price projections, the depreciation estimates — are reported from financial press and industry analysis as of this week, not computed by me. Intraday figures in particular vary by source and by the minute; treat them as the shape of the move, not a ledger.

The Jensen Huang investigation is deliberately handled in two sentences because the allegations are unproven and the details are not public. If you came here for that story, this is the wrong article, and any article claiming to know more than "there is an investigation" right now is ahead of the facts.

The memory-bound explanation is a description of the dominant inference workload — autoregressive token generation — and it simplifies. Prompt processing (prefill) is compute-bound rather than memory-bound; training has a different profile again; batching, quantization, speculative decoding and mixture-of-experts architectures all shift the arithmetic intensity around. The roofline picture is the right first-order intuition for why bandwidth rules decode, not a complete account of every AI workload. And I have not touched the power and grid-capacity constraint, which is arguably as binding as memory and deserves its own piece.

What I'm Watching Next

This is a living piece, and the next hard data point lands on 17 November, when NVIDIA reports earnings into the middle of all this. I'll update this section as the numbers come in. The things I'll be reading for are not the headline EPS — they are the lines that either confirm or break the memory-bound thesis:

  • Data-center revenue and, more importantly, forward guidance. The moratorium fear is a bet that the buildout slows. Guidance is where a slowdown would show up first, long before any town actually blocks a data center.
  • Gross margin. HBM4 contract prices climbing around 121% into 2027 is a cost NVIDIA has to either absorb or pass on. A margin dip is the memory squeeze reaching the income statement in black and white.
  • Any language about supply. If Huang echoes Lisa Su's "memory supply is very tight," that is the bottleneck talking directly, from the one company with the clearest view of it.
  • Whether the governance overhang gets addressed or deferred. Separable from the hardware, but it will set the tone of the call regardless.

If the buildout really is bending, the first place it becomes a fact rather than a fear is a hyperscaler or NVIDIA trimming forward numbers. Until then, everything priced this week is anticipation — and anticipation is exactly the thing that reverses fastest. Check back here after the 17th.

The Actual Point

My thirty-dollar experiment taught me one actually useful thing this week, and it had nothing to do with money, because thirty dollars can't teach you anything about money. It taught me to watch which things fall and by how much, because the spread between them is where the information is. NVIDIA is the name on the headlines, but the memory makers moved twice as far, and that spread is the market doing, clumsily and in public, the same analysis an engineer would do quietly: find the bottleneck, and price the risk where the bottleneck is.

The bottleneck is memory bandwidth. It's been memory bandwidth the whole time. Every layer of this story traces back to one physical fact: an AI accelerator is a machine for reading weights out of memory very fast, and the memory is the scarce part. Why HBM commands the margins, why three companies hold the whip hand, why a buildout slowdown terrifies the sector, why the depreciation of specific chips can move a trillion dollars of valuation, why so much debt is secured by silicon — all of it is downstream of that single constraint.

The market spends most of its time pricing AI as if it were magic: unbounded, frictionless, limited only by imagination. Weeks like this one are when it remembers that the magic is made of DRAM stacks fabricated by three companies, installed in data centers that towns can vote to block, powered by grids that are already strained, financed by debt secured against chips whose useful life nobody can agree on. None of that is a reason to buy or sell anything, and I am not suggesting you do either. It is a reason to understand that when you read "NVIDIA tanked," the interesting question is never the ticker. It's always the hardware under it.

My toy portfolio is down about a dollar. The thing it is pointed at is down a great deal more, for reasons you can read straight off a spec sheet if you know which line to look at. That, and not the gains or the losses, is what I bought it to see.

Market figures and company statements in this article are drawn from financial press and industry reporting from the week of 3–10 November 2026 and reflect intraday or as-reported values that change constantly; they are illustrative of the moves, not a record. The memory-bandwidth and roofline explanation is the author's own, from established computer-architecture principles. Nothing here is investment advice.

$ tail -f /forums

Discussion

Comments live on the community forums — sign in there to reply.