AINEWS
AMD printed $11.5 billion in Q2 revenue, up 50% year over year. Shares still dropped because the market is now demanding roughly 70 times next year's earnings to justify the multiple.
AMD reported record Q2 2026 revenue of $11.5 billion on August 6, up 50% from a year earlier, with data-center sales of $6.7 billion more than doubling at 107% year-over-year. The stock fell. The print is no longer the test. The roughly 70-times forward-earnings multiple the market is now demanding is.
The number that matters more is the quarter-over-quarter trend: about 16%. The year-over-year comparison is flattered by the April 2025 US ban on chip sales to China, which depressed the prior-year base. Strip out the easy comp, and the data-center business is still growing fast, but it is not the triple-digit story the headlines imply.
Behind the print sits a credibility test on a second straight year of data-center doubling. Adjusted operating margin reached 17% in the quarter, up from a 2% loss a year earlier. Earnings per share landed at $1.66, up 246% year-over-year. CEO Lisa Su told investors AMD has lined up roughly 6 gigawatts of committed capacity apiece from OpenAI and Meta Platforms for its new MI450 chips and Helios rack systems. Helios is AMD's rack-scale AI platform; it pairs MI450 accelerators with EPYC server CPUs and the ROCm open software stack. Six gigawatts is enough contracted power to run several major cities. The number is a CEO commitment, not a shipped order book.
AMD's stock trades at nearly 70 times forward earnings, a multiple that demands another year of roughly 100% data-center growth to defend. Su has pegged the AI data-center chip market at $1.4 trillion annually by 2030 and forecasts AMD's data-center sales will roughly double once more in 2027. That forecast is the price of admission for the multiple. Every incremental data point now has to clear a higher bar than the print alone did.
The customer wins are real, even if the market is no longer rewarding them at the prior pace. The OpenAI and Meta capacity commitments sit on top of that base. None of that changes the math: the 6-gigawatt capacity letters only matter if AMD ships MI450 racks on time and earns the second doubling Su is underwriting.
That is the context for the deal AMD announced the same week. AMD confirmed it would acquire Taalas, a Toronto-based inference-silicon startup founded in 2023 that has raised about $169 million in venture funding. Inference is the work of running already-trained AI models, not training them. Taalas's approach, originally reported by The Next Platform, optimizes how data moves through a chip during inference, cutting the compute and memory bottlenecks that throttle general-purpose accelerators. AMD plans to fold the technology into its accelerator roadmap and pair it with the Instinct GPU line, sitting on top of a stack that already includes Helios rack systems, EPYC CPUs, and the ROCm software stack. The deal, also covered by The Register, is still subject to regulatory approval and extends AMD's long-standing Canadian engineering footprint, where the company says it intends to keep hiring.
The acquisition is not the story. The story is what the market is now asking the print to deliver. Taalas is one signal in a broader inference push, alongside the MI450 ramps, the OpenAI and Meta capacity commitments, and a software stack that has to scale with the silicon. If AMD ships the second doubling in 2027, the 70-times multiple gets its defense. If it does not, the easy-comp cushion the China ban provided is gone, the QoQ trend is the only growth signal left, and the post-earnings drop looks less like a quarter-end wobble and more like the start of a longer repricing.