AMD unveiled the complete Instinct MI400 series of AI accelerators, led by the flagship MI455X that delivers up to 40 petaflops (PFLOPS) of FP4 performance and packs 432GB of HBM4 memory with 19.6 terabytes per second of bandwidth. Built on TSMC's 2nm process node with the CDNA 5 architecture, the MI400 series represents AMD's most ambitious challenge yet to Nvidia's dominance in the data center AI chip market. Analysts at S&P Global project the MI400 lineup will generate approximately $7.2 billion in revenue in its first year.
CEO Lisa Su presented the full product family at an investor event, demonstrating competitive benchmark performance against Nvidia's Blackwell B200 on key large language model inference tasks. Microsoft Azure and Oracle confirmed as launch partners, and AMD raised its 2026 data center AI revenue target. The MI400 series ships in the second half of 2026, with the MI455X targeting hyperscale customers, the MI440X for enterprise AI training, and the MI430X for sovereign AI and HPC workloads.
What happened
AMD initially previewed the MI400 architecture at CES 2026 in January, but the May investor event provided the first complete specifications, benchmarks, and commercial availability timeline. The MI455X flagship packs 320 billion transistors across 12 TSMC N2 compute chiplets and 3 advanced 3nm chiplets. AMD also introduced Helios, a rack-scale AI architecture comprising 72 MI455X accelerators with 31TB of HBM4, designed to compete directly with Nvidia's NVL72 rack-scale solution.
The benchmark data showed MI455X delivering comparable inference throughput to Nvidia's Blackwell B200 on GPT-class models while consuming similar power. AMD claimed a 2.4x improvement over its previous-generation MI300X at the same power envelope. The company also raised its 2026 data center AI revenue target, reflecting growing confidence in the MI400 pipeline based on advance customer commitments from hyperscalers and enterprise customers.