The chip selloff is picking sides, and AMD is on the wrong one

A broad selloff takes everything down together. This one is not doing that, and the pattern it is making instead is the useful information.

On Wednesday, AMD closed at $466.42, down 3.71 percent. Intel fell 4.02 percent to $92.80. Nvidia closed at $217.56, down 0.99 percent. Three companies in the same industry, exposed to the same AI capital cycle, on the same day, and one of them fell by roughly a quarter as much as the others. The divergence was visible enough to lead the sector coverage.

That is not risk-off. That is the market making a distinction.

The positions are further apart than one day suggests

A single session is noise. The structural picture is not.

Nvidia sits 5.1 percent above its 50-day average and 11.5 percent above its 200-day, about 8 percent below its 52-week high. On every conventional reading it remains in an uptrend that has not broken.

AMD is 8.6 percent below its 50-day, 20.2 percent below its high, and 42.8 percent above its 200-day. Intel is 15 percent below its 50-day, almost 35 percent off its high, and 31.2 percent above its 200-day. Both have lost their 50-day average, which is the simplest available definition of a trend that has broken.

All three had enormous years. AMD's 52-week range runs from $149.22 to $584.73. Intel's runs from $22.78 to $142.35. Those are not the ranges of stable businesses. But over the past fifty sessions, two of them rolled over and one did not.

The 200-day figures are what stop this being a doom story. AMD sitting 43 percent above its 200-day average is not a company being abandoned. It is a violent rerating giving back part of its most recent leg while the longer trend holds. Those are different things, and they get conflated constantly, usually by people with a position.

Remember how AMD got here

The context that makes the current drawdown legible is the size of the move that preceded it.

AMD traded as low as $149.22 within the past twelve months and as high as $584.73. That is close to a fourfold range, and the rerating was driven by a specific change in the story. AMD stopped being a credible-second-source narrative and started signing capacity commitments large enough to matter, including a multi-year supply agreement with OpenAI covering gigawatt-scale deployments, alongside a warrant giving OpenAI the ability to acquire a substantial equity stake as the arrangement progresses.

A stock does not quadruple on execution alone. It quadruples when the market changes its mind about what kind of company it is looking at. That is what happened, and it is why the current 20 percent drawdown reads differently from a 20 percent drawdown in a business nobody rerated.

It also means the position is crowded, and crowded positions unwind faster than they build.

What the market is actually pricing

The stated reasons for the sector's weakness are consistent across the coverage: concern that AI infrastructure spending may be peaking sooner than assumed, valuation compression as multiples adjust, and rising memory and storage costs pressing on margins. The sector has already been through one violent version of this, when more than a trillion dollars came off chip stocks in late July.

The first two reasons apply to everyone. If AI capex is peaking, Nvidia has the most to lose, because it sells the most into it. If multiples compress, the most expensive names compress hardest. Neither explains why Nvidia held while AMD and Intel did not.

The third one does.

Input costs are rising across the supply chain, and they are not landing evenly. We reported yesterday that Samsung raised advanced foundry prices by 5 to 15 percent, with the steepest increases falling on customers with the fewest alternatives, and that TSMC has its own increase queued for 2027. Memory has been the other half of the squeeze all year, a dynamic we traced through the memory pricing cycle.

Every accelerator vendor buys wafers and buys HBM. The question is not who faces higher input costs. Everyone does. The question is who can pass them on.

Pricing power is the whole variable

Nvidia holds roughly 80 to 85 percent of the AI training accelerator market and has spent two years selling every unit it can produce. A vendor in that position does not absorb a wafer price increase. It reprices, and the customer pays, because the customer's alternative is a queue.

AMD is in a structurally different seat. It is winning real business, and the wins are genuine rather than cosmetic. Data centre revenue grew 107 percent year over year to $6.7 billion in the second quarter. The Meta arrangement and the OpenAI commitment are meaningful decisions by buyers who had every reason to stay with the incumbent.

But those wins were priced to displace an incumbent. Share is bought with total cost of ownership, and a challenger that responds to rising wafer and memory costs by raising its own prices hands back the reason customers switched in the first place. So the cost lands on gross margin rather than on the invoice.

That is the asymmetry the tape is expressing. Same cost shock, different ability to refuse it.

Intel sits in a third position again, carrying foundry economics as well as product economics, which is why it is furthest below its own trend line and why its drawdown from the high is the deepest of the three.

The half of AMD that is not an accelerator

One complication to the thesis, and it cuts in AMD favour. The company is not only an accelerator vendor. EPYC server processors remain a substantial and structurally different business, and in that market the competitive position is reversed: AMD is the one taking share from an incumbent that has been losing it for years, and it has considerably more room to hold price.

That matters for a cost-pass-through argument, because a company blended across two businesses does not face a single margin exposure. Rising wafer prices squeeze the accelerator line hardest, where the incumbent sets the reference price. On the CPU side the same increase can be carried more comfortably. We looked at how tightly the two are coupled when we examined the CPU attach rate that comes with rack-scale deployments, and the answer was that they are coupled enough to matter. A blended margin that holds up better than the accelerator line alone would suggest is one of the ways this call goes wrong.

The part that argues the other way

There is a real case that this reading is too tidy, and it deserves stating properly rather than as a token paragraph.

Nvidia's relative strength may simply be positioning. It is the largest, most liquid and most widely held name in the sector, and in a nervous tape the biggest position is often the last one sold rather than the most defensible. Outperformance over a handful of sessions is not proof of pricing power. It is at least as consistent with investors trimming the second and third names first and reaching the first name later.

The margin thesis is also not yet visible in reported numbers. It is an inference from cost inputs and competitive position, and inferences of that kind have a poor record against actual gross margin lines. AMD reported a data centre business growing at triple digits. Nothing in the current quarter says the cost squeeze has bitten yet.

And AMD has genuine architectural momentum. Its claims around disaggregated inference point at exactly the workload shift that could make accelerator share look different in two years, and the head-to-head against Nvidia is closer on paper now than at any point in the last decade.

The honest position is that the divergence is real and observable, and the explanation for it is a hypothesis with a clean test attached.

The test

Gross margin, next earnings, both companies.

If the input cost story is right, AMD's gross margin comes under visible pressure while Nvidia's holds, and the gap between them widens rather than converges. That is specific, falsifiable, and answered on a scheduled date rather than by argument.

If AMD's margin holds while it keeps taking share, the current derating is a positioning artefact, the market got the cost pass-through question wrong, and a stock 20 percent below its high with a triple-digit data centre growth rate is a considerably more interesting proposition than the tape implies.

Two things to watch in the meantime. Whether AMD reclaims its 50-day average, which is the simplest signal that the momentum unwind has run its course. And whether foundry and memory price increases keep being announced, because every one of them widens the gap between vendors who can reprice and vendors who cannot.

The selloff is telling you something. It is just not telling you that AI is over.