AMD said at Citi's 2026 Global TMT Conference on September 8, 2026 that its AI push is lifting the company's outlook, with production shipments of its MI450 platform beginning in the third quarter of 2026. Management described the start of the MI450 ramp as the beginning of a larger cycle, with a significant step-up in the fourth quarter of 2026, another increase in the first quarter of 2027, and continued growth through 2027. The MI400 program is planned for the second half of 2027 and is expected to become the main product in 2028.
AMD said enterprise server revenue rose by more than 70% in the second quarter of 2026, while its embedded business posted double-digit year-over-year growth. Gross margin for the third quarter is guided at 56%, with margins expected to edge lower in the fourth quarter of 2026 and in 2027 as the MI450 ramp scales. AMD's gross margin over the last twelve months was 55.7%, which management attributed to pricing power in high-demand AI products.
AMD updated its view of the AI market opportunity to $2 trillion by 2030, a step up from figures presented during its Advancing AI Day. The data center business is expected to double in 2027 to about $70 billion in sales, with AI GPUs contributing in the low $40 billion range and server CPUs making up the remainder. The server CPU market is now seen at $220 billion by 2030, up from a prior estimate of $60 billion, with agentic AI expected to account for more than half of that opportunity. Server CPU revenue is expected to grow more than 80% year-over-year in the second half of 2026 and more than 70% into 2027. AMD is targeting a $100 billion server business opportunity, about half of the total server CPU dollar market, and has secured $29 billion to $30 billion in purchase commitments.
CFO Jean Hu said the company's three major anchor customers, Meta, OpenAI and Anthropic, are tied to multi-gigawatt deployments and multi-generational engagements. Matt Ramsay, corporate vice president of financial strategy, said technical engagement with the leading model companies is influencing AMD's roadmap over a multigenerational period and that the current generation of products offers tokens-per-dollar advantages for large-scale inference.
AMD said it faces tight supply in advanced process wafers, HBM memory, advanced packaging capacity and substrates. The company received the largest allocation increase from TSMC among its customers. The Helios rack-scale system is ramping at the rack level, requiring coordination across GPUs, CPUs, HBM memory and smaller components with ODM partners. AMD runs a weekly execution process with customers to validate components, mechanical design and software stack performance.
On the technology roadmap, the MI400 program is expected to support scale-up domains larger than 72, using copper-based and near-package optics-based connectivity that are expected to coexist for several generations. Infinity Fabric traffic is expected to continue over both Ethernet and optics, with future references to Instinct MI500 and MI600 programs. AMD also said its partnership with Cerebras will integrate AMD Helios systems with Cerebras wafer-scale engine technology in Cerebras' cloud, while the Taalas acquisition adds talent and chiplet architecture technology for internal ultra-low latency inference silicon. The earlier ZT Systems acquisition is described as building toward software capabilities and system-level expertise.
