AMD Q2 2026: The Gross Margin Guide Stayed at 56% and the Stock Lost 8%
Advanced Micro Devices beat on revenue, beat on earnings, beat on operating margin, and guided the September quarter about $500 million above consensus. The stock fell as much as 8% in after-hours trading, giving back a 7.7% regular-session gain that had carried it to roughly $513. It changed hands near $472.50 in the extended session, which puts the market capitalization just under $800 billion on a diluted share count of about 1.66 billion.
The quarter itself was close to the best AMD has ever printed. Revenue of $11.54 billion was up 50% year over year and 13% sequentially, against a consensus near $11.3 billion. Non-GAAP earnings came in at $1.66 per share versus $1.62 expected. Non-GAAP operating margin reached 27%, more than double the 12% posted a year ago, which is the cleanest evidence available that the data center mix shift is doing what it was supposed to do. Data Center revenue was $6.7 billion, up 107% year over year, and now accounts for 58% of the company. EPYC and Instinct are no longer the growth segment. They are the company.
The number that mattered was not on the results line. It was in the outlook. AMD guided third-quarter non-GAAP gross margin to approximately 56%, the same level it just delivered, with non-GAAP operating expenses stepping up to roughly $3.65 billion from about $3.3 billion. Flat gross margin, in a quarter where data center revenue is guided to grow strong double digits sequentially, is a statement about what the MI450 ramp costs. Jean Hu told investors in May that MI450 carries a gross margin below the corporate average and would create puts and takes into the fourth quarter. Investors got the confirmation on schedule and did not enjoy it. A mix shift toward the highest-growth product line that does not lift the margin line is a mix shift that pays in revenue and charges in structure.
The consumer side supplied the second bruise. Client revenue of $3.1 billion was up 23% on continued Ryzen share gains against a degraded Intel. Gaming revenue of $779 million was down 31%, mostly on semi-custom, and management guided a strong double-digit sequential decline for the September quarter, which offsets modest client growth and leaves the combined segment down. The cause is not demand. It is memory. DRAM and NAND contract pricing has repriced the entire consumer bill of materials, and AMD is one of the buyers rather than one of the sellers. That is the awkward symmetry of this cycle: the same supply constraint that makes Micron and SanDisk the trade of the year makes AMD’s gaming margin a cost problem. Embedded, at $977 million and up 19%, is the quiet accretive segment nobody prices.
The competitive question is where the moat actually sits, and it is not where the multiple sits. AMD’s durable advantage is the EPYC server CPU franchise. It operates inside an x86 duopoly with a structurally impaired counterparty, the switching costs run through validated enterprise stacks rather than through software rewrites, and the share gains compound because each cloud design win locks in a refresh cycle. That is a real moat with a defensible price structure. Instinct is something else. It is a challenger position in a market where the incumbent’s advantage is a software ecosystem, and ROCm still trails CUDA in developer gravity by a wide margin. The gigawatt commitments announced at the July investor event, up to 2 GW with Anthropic, 6 GW with Meta, and 2.5 GW with Core Scientific, are the market’s evidence that hyperscalers want a credible second source. They are not evidence that AMD has closed the software gap. They are evidence that the buyers are willing to pay for optionality against Nvidia’s pricing power, and buyers who purchase optionality negotiate hard on price. The dilutive MI450 margin is what that negotiation looks like on an income statement.
The position going into the print was the real culprit. AMD had risen roughly 142% year to date against Nvidia’s 13.6% and the S&P 500’s 13%, and it added another 7.7% during Tuesday’s session before the release. At that level a beat and raise is the entry fee, not the surprise. Intel demonstrated the same mechanic in July, beating by $1.7 billion and dropping 11%. The Philadelphia Semiconductor Index lost 20.6% in July, its worst month since October 2008, and then recovered into month-end. This is a cohort trading with high beta to sentiment and low tolerance for anything that reads as a plateau.
Base case runs $460 to $540 over the next two quarters. That range assumes the MI450 ramp lands on schedule, Q4 gross margin dips a point or two on mix and recovers into 2027, and the CPU franchise continues taking share. Lisa Su’s comment that AMD now expects growth to substantially exceed the prior long-term target of more than 35% and to significantly exceed the $20 annual EPS target supports the upper half of that band if the Street starts modeling 2027 rather than the next print. Bull case is $600 to $700, roughly where Benchmark’s $685 target sits, and requires MI450 margins to converge toward corporate average faster than guided, which would mean AMD has pricing power in accelerators rather than volume alone. The consensus target near $590 is the compromise view. Bear case is $340 to $400 and does not require AMD to miss anything. It requires the AI infrastructure cohort to derate on capex-digestion fear, at which point a name trading on 2028 earnings gets marked back to 2026 earnings regardless of execution. That is what happened in July, and nothing in this quarter changes the exposure.
The single number to carry forward: third-quarter gross margin guidance of 56% against operating expenses of $3.65 billion. If the December guide holds 56% while MI450 ramps significantly, the accelerator business has better economics than management has been willing to promise. If it guides to 54% or below, the gigawatt deals were bought with margin, and the multiple has to account for it.