AMD Doubled Data Center and the Stock Fell 9%

AMD reported Q2 2026 revenue of $11.5 billion, up 50% year over year. Data center revenue hit $6.7 billion, up 107% from $3.2 billion, throwing off $2.1 billion in segment operating income. Non-GAAP EPS came in at $1.66, up 82%, against a $1.60 consensus. Q3 guidance landed at roughly $13 billion, about 41% growth and comfortably above the $12.51 billion the street had modeled.

The stock fell about 9% after hours, to roughly $472.

Beat, raise, sell. That combination is usually a story about positioning — a stock that ran too far into the print. It's partly that here. It is mostly something more specific, and it is visible in three lines of the release that have nothing to do with revenue.

Line one: the margin didn't move

Adjusted gross margin was 56%, up about two points year over year. Guidance for Q3: 56%, flat.

Sit with that. The data center segment more than doubled. Mix shifted violently toward the highest-value product AMD makes. And the blended margin went nowhere, and the company does not expect it to go anywhere next quarter either.

There are only two readings. Either accelerator gross margins are structurally lower than the rest of the business — plausible, given HBM cost, packaging constraints, and the fact that AMD is buying share against an incumbent — or AMD is pricing for placement rather than profit. Both readings say the same thing about the next four quarters: incremental revenue arrives at roughly the company average, not above it.

That is the exact opposite of the Nvidia pattern investors have spent three years extrapolating, where scale in accelerators pulled the whole margin structure upward. AMD is growing into the AI buildout without the operating leverage that made the buildout worth owning.

Line two: capex nearly tripled

Capital expenditure hit $808 million, up from $282 million a year ago.

AMD is a fabless designer. Its capex line is supposed to be small and boring. Tripling it means the company is funding test infrastructure, packaging capacity, validation labs, and the physical apparatus of shipping rack-scale systems rather than chips in trays — the cost of the Helios transition showing up where the marketing didn't put it.

Rack-scale is the right strategic call. Helios bundles processors, accelerators, and networking into the unit customers actually buy, and the early customer list is as good as it gets: Meta, OpenAI, and Oracle, plus an agreement with Anthropic covering up to 2 gigawatts of MI450 deployment. You do not get those names selling loose silicon.

But rack-scale is a different business with different economics. You take on integration, networking, thermal, and support obligations. You carry more inventory and more working capital. The revenue is larger per deal and the margin per dollar is not obviously better. AMD is buying its way into the room where the money is, and the entry fee is now on the cash flow statement.

Line three: free cash flow, 14%

Free cash flow was $1.6 billion — about 14% of revenue, against a 24% adjusted net income margin.

That ten-point gap is the whole trade. Adjusted net income is the number the narrative runs on. Free cash flow is the number that funds buybacks, acquisitions, and the next capacity commitment. When the two diverge by that much in a quarter where everything went right, the market reprices from "how fast is this growing" to "how much of it do they keep."

Meanwhile the client and gaming segment's operating margin fell to 15% from 21%. The PC business, which is supposed to be the stable cash engine that funds the AI attack, is getting less profitable at the same time the AI attack gets more expensive.

The rotation underneath

Zoom out and this is the market doing something it has been threatening to do all year: shifting from the training-cluster buildout story to inference economics.

In a buildout regime, whoever can ship units wins, and margin is a detail you sort out later because demand is infinite and supply is the constraint. In an inference regime, buyers optimize cost per token, second sources exist specifically to compress prices, and the vendor with the weaker software moat discovers what its silicon is actually worth. AMD's structural role in this market — the credible alternative that keeps Nvidia honest — is precisely the role that generates volume without pricing power.

CFO Jean Hu told investors data center sales would accelerate in the second half. That is almost certainly true. The question the stock asked on Tuesday is whether accelerating into a flat margin at rising capex is a growth story or a share-gain story that happens to be expensive.

What to watch

Whether Q4 gross margin guidance breaks 57%. One flat quarter is mix. Two is structure. Three is the answer.

Whether Helios revenue gets disclosed separately. Rack-scale bundles processors, accelerators, and networking into one line, which makes accelerator ASPs and attach rates unreadable. If AMD keeps them merged as the number grows, assume the blend is unflattering.

Whether the Anthropic 2GW commitment converts on schedule. Announced capacity and installed capacity are different quantities separated by power, packaging, and HBM supply. The conversion rate on that agreement is the single best read on whether MI450 is a product or a placeholder.

The read

AMD had a genuinely excellent quarter and the market marked it down, which is not irrational — it is the market noticing that this quarter's growth is not the kind that compounds into the multiple.

Doubling data center revenue while holding margin flat and tripling capex means AMD is converting the AI cycle into share, not into cash. Share is worth having. It's just worth less than the stock had been paying for.