このイベントの根拠となる記録
4-
企業
AMD
The credible second source: Instinct GPUs, EPYC CPUs and the ROCm stack.
Under Lisa Su, AMD went from near-bankruptcy to the industry's main alternative to NVIDIA: Instinct accelerators won Microsoft, Meta, Oracle and OpenAI as customers, EPYC took large x86 server share from Intel, and the Xilinx and ZT Systems acquisitions added FPGAs and rack-scale engineering. Gigawatt-scale agreements followed: OpenAI (up to 6 GW, October…
レコードを開く → -
企業
Meta
Open-weights pioneer turned closed-model superintelligence spender.
Meta's Llama releases made open weights a mainstream strategy and seeded a global developer ecosystem. In 2025 Mark Zuckerberg reorganized the effort into Meta Superintelligence Labs, hiring Scale AI's Alexandr Wang via a $14.3B stake. Its first model, Muse Spark (April 2026), shipped closed inside Meta's apps with only a private API preview, a break from…
レコードを開く → -
チップ
AMD Instinct MI400 / MI450
AMD's 2026 rack-scale GPU, and the chip behind its gigawatt deals with OpenAI, Meta and Anthropic.
The MI450 Series is AMD's 2026 rack-scale GPU family on the CDNA 5 architecture. The flagship MI455X carries 432 GB of HBM4 at 23.3 TB/s and up to 40 PFLOPS of FP4, and ships 72 to a Helios rack; a lower-tier MI450X also fits Helios, while MI430X targets HPC. It anchors AMD's largest agreements: OpenAI and Meta for up to 6 GW each and Anthropic for up to 2…
レコードを開く → -
システム
AMD Helios
AMD's first rack-scale system: 72 MI450-series GPUs on open standards (UALink + OCP).
Helios is AMD's first rack-scale system and its answer to NVIDIA's NVL72 racks: 72 MI455X or MI450X GPUs, 18 EPYC 'Venice' CPUs and Pensando networking in a liquid-cooled OCP Open Rack Wide frame, with UALink scale-up and Ultra Ethernet scale-out. AMD rates it at up to 2.9 exaflops of FP4 with 31 TB of HBM4. Production began in 2026, and in Q3 2026…
レコードを開く →

