AMD Q2 2026: data center revenue doubled, gross margin did not move

AMD beat on revenue. It beat on earnings. It guided the next quarter roughly $480 million above the LSEG consensus CNBC reported (our arithmetic: $13.0 billion midpoint less $12.52 billion).

That combination is worth sitting with, because the explanation is not sentiment, and it is not “the bar was too high.” It is a single line item that management told you would not improve.

	Q2 2026	Consensus
Revenue	$11.54B	$11.28B (LSEG, via CNBC); $11.31B (Shacknews)	beat
Non-GAAP EPS	$1.66	$1.62 (LSEG, via CNBC; Shacknews)	beat
Non-GAAP gross margin	56%	~56% (AMD’s own guide)	in line
Q3 revenue guide	$12.7–13.3B	$12.52B (LSEG, via CNBC)	above
Q3 non-GAAP gross margin guide	~56%	—	flat

Revenue grew 50% year over year and 13% sequentially to a company record. Data center revenue grew 107%, to $6.7 billion, and is now 58% of the company against 42% a year ago (our arithmetic: $3,240 million of $7,685 million). Non-GAAP operating income was $3.1 billion, a 27% operating margin. On a GAAP basis: 54% gross margin, $2.0 billion of operating income, $2.3 billion of net income, $1.38 of diluted EPS.

Every one of those numbers is good. The last row of that table is the one that needs context.

Did AMD beat its own guidance in Q2 2026?

Yes. Revenue of $11,536 million came in above the top of the $10.9–11.5 billion range AMD guided to in its May 5, 2026 8-K (“approximately $11.2 billion, plus or minus $300 million”; the range is our arithmetic), and was up 50% from $7,685 million a year earlier. Non-GAAP gross margin of 56% matched the approximately 56% the company had guided to. AMD’s guidance covered only revenue and non-GAAP gross margin, not EPS; non-GAAP diluted EPS was $1.66 against $0.48 in a year-ago quarter that carried $800 million of MI308 inventory and related charges.

The one number that needs context

AMD guided third-quarter revenue up about 13% sequentially at the midpoint — and third-quarter non-GAAP gross margin at about 56%, the same as the quarter just reported.

Read that as arithmetic rather than as guidance. If revenue rises 13% and gross margin stays flat, the incremental dollar of revenue is arriving at roughly the company average — around 56 cents of gross profit per dollar. For a fabless semiconductor company, that is unusual and it is informative. Chip businesses normally show operating leverage on the way up: fixed costs (mask sets, design, test) spread over more units, so gross margin expands with volume. AMD’s is not expanding.

The reason is what AMD is now selling. The Instinct MI450 generation ships inside Helios, a rack-scale system. A rack is not a chip. It carries high-bandwidth memory, networking, power delivery and integration — content AMD largely buys and resells. Bought-in content passes through at low markup, so it adds revenue and gross profit dollars while pulling the gross margin percentage toward the blend. Add early-generation yields on a new accelerator, and you get exactly the picture in the release: revenue compounding, margin percentage standing still.

This is not a defect. It is the business model changing shape, and it is the correct thing for AMD to be doing — you cannot sell a gigawatt of AI capacity as loose silicon. But it does mean the old mental model, more volume equals more margin, has stopped applying.

What the 107% actually contains

The headline comparison in most coverage is that data center swung from a $155 million operating loss a year ago to $2.1 billion of operating income, at a 31% segment operating margin. That is a $2.26 billion swing (our arithmetic: $2,103 million less a $155 million loss), and it is being read as a doubling of profitability.

Part of it is not.

AMD’s Q2 2025 data center segment posted that $155 million loss because of roughly $800 million of inventory and related charges tied to the US export control on the Instinct MI308 for China. Strip the one-off out and the prior-year segment was profitable — the year-ago comparison is depressed by a write-off, not by weak operations.

So the honest split is:

• Revenue growth of 107% is clean. $3.2 billion to $6.7 billion, no charge involved. This is real demand for EPYC server CPUs and Instinct accelerators.
• The operating-income swing is roughly one-third optical. Around $800 million of the $2.26 billion improvement is the absence of last year’s charge, not this year’s earnings power.

Investors who anchor on “data center went from a loss to $2.1 billion” are overstating the change. Investors who anchor on “data center revenue doubled” are not. It is worth knowing which number you are holding.

The same effect runs through the company line: prior-year GAAP gross margin was 40% precisely because of that charge, so the 54%-versus-40% comparison overstates the improvement badly. On the cleaner basis — AMD’s own comparable non-GAAP gross margin, which excludes the charge and was 54% a year ago against 56% now — the company puts the year-over-year gain at more than 200 basis points. Real, but two points, not fourteen.

Segments: one engine, one drag, one quiet compounder

Segment	Q2 2026 revenue	YoY	Note
Data center	$6,718M	+107%	58% of company; $2,103M operating income, 31% margin
Client and gaming	$3,841M	+6%	two businesses moving opposite directions; $582M operating income
— Client	$3,062M	+23%	Ryzen demand
— Gaming	$779M	−31%	lower semi-custom (console) revenue
Embedded	$977M	+19%	$386M operating income, 40% margin

Three things stand out.

Client is doing better than the segment header suggests. Reported together, client and gaming grew 6%. Reported separately, client grew 23% and gaming fell 31%. A console cycle winding down is masking a genuinely strong PC processor quarter. The drag is finite — semi-custom revenue can only fall to zero — but it will keep suppressing the combined line until it laps.

Embedded is the best-margin business AMD owns. $386 million of operating income on $977 million of revenue is a 40% operating margin, up from 33% a year ago, on 19% revenue growth. It is 8% of revenue and nobody discusses it. In a quarter where the market is worried about margin mix, a segment earning 40% deserves more attention than it gets.

The segments reconcile. $6,718M + $3,062M + $779M + $977M = $11,536 million (our arithmetic), exactly the reported total. There is no unexplained revenue in this release, which is more than can be said for some AI-era reports.

Quality of earnings: $2.8 billion of profit, $1.6 billion of cash

Non-GAAP net income was $2.76 billion. Free cash flow was $1,558 million — a 14% free cash flow margin, from $2,366 million of operating cash flow from continuing operations less $808 million of capital expenditure.

A $1.2 billion gap (our arithmetic: $2,760 million less $1,558 million) between adjusted profit and free cash flow is not alarming for a company growing revenue 13% sequentially — inventory and receivables have to be built ahead of a Helios ramp, and that consumes working capital by construction. But it is the number to watch, because it is where a systems business differs from a chip business in the accounts as well as the margin line. Racks carry far more inventory value per unit than chips do. If revenue growth ever pauses while that inventory is on the balance sheet, the cash conversion gets worse before it gets better.

Cash, equivalents and short-term investments finished at a record $13.1 billion, up 6% sequentially from $12.3 billion. The balance sheet is not the constraint here.

For scale: 14% free cash flow conversion is what a company looks like while it is buying its way into a new market. It is not a criticism — but anyone comparing AMD to a software business on cash generation is comparing two different physics.

The targets management put on the record

This is the part worth clipping and checking later, because these are specific and dated commitments, not atmosphere.

Target	What management said
Q3 2026 revenue	$12.7–13.3 billion (about $13.0B ± $300M), roughly +41% YoY and +13% QoQ at the midpoint (8-K)
Q3 2026 non-GAAP gross margin	approximately 56% (8-K)
2027 data center revenue	expected to more than double versus 2026 (earnings call, per GuruFocus; CNBC reported it as “double”)
Anthropic	up to 2 gigawatts of MI450 Series GPUs in Helios racks (8-K), with the first gigawatt beginning deployment in the first half of 2027 (earnings call, per GuruFocus)
Microsoft	expanded collaboration to deploy Helios racks at scale on Azure (8-K)
Supply	”we have supply to more than meet our guidance” (earnings call, per GuruFocus)

The 2027 data center commitment is the load-bearing one. Doubling a $6.7-billion-a-quarter segment is a far larger absolute undertaking than the doubling just delivered, and it is the number the current multiple is discounting.

Note also what the Anthropic commitment does to the timeline: the first gigawatt starts deploying in the first half of 2027. That is three or four quarters away. Between now and then, the revenue has to come from EPYC, from MI355-class parts and from the customers already ramping — not from the headline announcement.

The concentration question nobody wants to ask

AMD’s AI growth rests on a small number of very large buyers: OpenAI, Meta and Anthropic are the “strategic anchor customers” Lisa Su named on the call (per GuruFocus); the 10-Q describes the OpenAI and Meta agreements as up to 6 gigawatts each of AMD data center GPUs. Microsoft is the newest deployment partner.

That is a superb list. It is also four names.

Customer concentration is neutral when demand is expanding and dangerous when it turns, because there is no diversification to cushion a single decision. A hyperscaler that re-phases a build by two quarters — for power availability, for a delayed data centre shell, for a change in model roadmap — moves AMD’s segment revenue by an amount no number of enterprise customers can offset. The practical point is that a handful of procurement decisions, made in rooms AMD does not control, set the growth rate.

It also frames the competitive picture honestly. AMD’s accelerators are winning share against a dominant incumbent, and much of the AI infrastructure spending of 2026 still runs on the incumbent’s roadmap. Winning a share of a market growing this fast is a real achievement. It is not the same thing as winning the market.

The bear case, stated fairly

Not “the numbers are bad.” They are not. The narrower version:

1. Margin is guided flat into a 13% revenue increase. Until rack economics mature, revenue growth converts to gross profit dollars but not to gross margin percentage — and both would be needed for margin and profit to scale together.
2. The year-over-year optics flatter. Roughly $800 million of the data center profit swing is the absence of last year’s MI308 charge.
3. Concentration. Three or four anchor customers set the AI trajectory.
4. Gaming is still shrinking 31%, and the client strength that is masking it depends on a PC cycle that has already been strong for several quarters.
5. Free cash flow is 14% of revenue while inventory builds ahead of Helios.

None of that is refuted by this quarter, and none of it contradicts the double beat and the guidance raise.

What to check at the Q3 report

Specific, falsifiable, and in rough order of importance:

• Non-GAAP gross margin against the ~56% guide. Above it means rack economics are improving faster than expected and the whole thesis strengthens. Below it means the system mix is diluting faster than management modelled. This is the single most important line in the next release.
• Revenue against $12.7–13.3 billion, and specifically where in the range. Guidance midpoints are guidance; the range tells you the confidence.
• Data center sequential growth, not the year-over-year figure. The annual comparison stays flattering on base effects; quarter-on-quarter is where a re-phased hyperscaler build shows up first.
• Whether the 2027 “more than double” language survives contact with a second quarter. Softening from “more than double” to “grow substantially” would be the most important sentence in the release, and the easiest to miss.
• Free cash flow conversion. Whether the gap between non-GAAP net income and free cash flow narrows, or widens as Helios inventory builds.
• Inventory on the balance sheet. The leading indicator for a systems business, ahead of both margin and cash.
• Client and gaming split. Whether client growth holds above 20% once the semi-custom drag laps.
• Any new anchor customer name. Diversification of the AI customer base matters more to the risk profile than another gigawatt from an existing one.

The read

AMD delivered the best quarter in its history and then guided next-quarter profitability to look the same as this one. Both statements are true and they are not in tension: the company is converting from selling chips into selling AI systems, and the transition raises revenue faster than it raises margin.

The useful way to hold this is to separate the two questions the report answers. Is AMD winning AI share? Yes — 107% data center growth against a real base, a 2027 doubling commitment, and gigawatt-scale customers with dated deployment schedules. Is that share arriving at the corporate-average profitability? Not yet, and management guided that it will not next quarter either.

For anyone holding AMD, that reframes what to track. The beat is not the signal. Gross margin against the ~56% guide is the signal, quarter after quarter, until rack economics either converge on the corporate average or fail to.

---

Sources and method. Company figures in this piece — revenue, GAAP and non-GAAP gross margin, operating income, net income, EPS, all segment revenue and operating income, operating cash flow, capital expenditure, free cash flow, the $13.1 billion cash position, the third-quarter outlook, the Anthropic and Microsoft agreements, and the prior-year comparatives including the $800 million MI308 inventory and related charges and the resulting $155 million data center operating loss — are from AMD’s Form 8-K filed 2026-08-04 (Exhibit 99.1 press release and Exhibit 99.2 results presentation, accession 0000002488-26-000121) for the quarter ended 2026-06-27, read directly from SEC EDGAR; the comparable 54% prior-year non-GAAP gross margin is AMD’s own adjusted figure from Exhibit 99.2. The OpenAI and Meta agreement sizes are from AMD’s Form 10-Q filed 2026-08-05 (accession 0000002488-26-000123). The Q2 2026 guidance is from Exhibit 99.1 to the Form 8-K filed 2026-05-05 (accession 0000002488-26-000072).

Not company filings, and attributed accordingly: consensus estimates — LSEG figures reported by CNBC (revenue $11.28 billion, non-GAAP EPS $1.62, third-quarter revenue $12.52 billion) and Shacknews (revenue $11.31 billion, EPS $1.62); and statements made on the earnings call that are not in the filed documents — the 2027 data center outlook, the first-half-2027 timing of the first Anthropic gigawatt, the supply remark and the “strategic anchor customers” — which are taken from GuruFocus’s call highlights (and, for the 2027 outlook, CNBC). Nothing else in this piece rests on market data. This piece makes no claim about the share-price reaction. AMD’s GAAP gross margin was 54% and its non-GAAP gross margin 56%; the margin concern in this article rests on the company’s own flat ~56% guide for Q3, not on a comparison between the two.

Whether the quarter beat or missed AMD’s own guidance is judged only against that previously issued guidance (Form 8-K, filed 2026-05-05) and against the year-ago quarter; the consensus comparison in the opening table is a separate, attributed outlet figure. Our arithmetic: the $10.9–11.5 billion guided range, computed from “approximately $11.2 billion, plus or minus $300 million”; the 42% prior-year data center share; the $2.26 billion data center operating-income swing; the segment sum; the $1.2 billion gap between non-GAAP net income and free cash flow; and the roughly $480 million gap between the Q3 guidance midpoint and the LSEG consensus.

Internal consistency was checked where possible: the four segment revenue lines sum to the reported total within rounding, and free cash flow reconciles to operating cash flow less capital expenditure. The interpretation — that rack-scale system content explains flat gross margin, and that roughly a third of the data center profit swing is the absence of a prior-year charge — is our analysis, not a company statement, and is not a forecast.

Related: SpaceX Q2 2026, the buyer side of the same AI buildout, and Microsoft Q4 FY2026, the other large AI-spending reporter that week. For the identical gross-margin mix problem, see Shopify Q2 2026.

Tradune shows what you actually hold across every account, and whether a position has beaten a plain index fund since you bought it. Educational analysis, not investment advice.