content image

置き換えではなく、包囲――Google、AWS、Meta、Microsoftが静かにAIチップの勢力図を塗り替えている

Not Replaced, But Encircled: How Google, AWS, Meta, and Microsoft Are Quietly Reshaping the AI Chip Landscape

Google、AWS、Microsoft、Metaが独自AIチップを次々投入。「脱NVIDIA」の実態は排除ではなく多様化——NVIDIAは置き換えられるのではなく、包囲されつつある。
分からないところをタップすると
↓日本語訳が表示されます↓

The fashionable phrase “de-NVIDIA” suggests a clean break, but the reality emerging in April 2026 is subtler and, in some ways, more consequential: not abandonment, but diversification. At Google Cloud Next on April 22, 2026, Google introduced its eighth-generation TPUs, splitting the line for the first time into TPU 8t for training and TPU 8i for inference. That design choice matters. It implies that the AI industry no longer believes one kind of accelerator can optimally serve every workload, especially in an “agentic” era defined by long-context reasoning, orchestration, and relentless inference at scale. (blog.google)

The technical claims are striking. Google says TPU 8t delivers nearly three times the compute performance per pod of the previous generation, with a superpod scaling to 9,600 chips and 121 exaflops. TPU 8i, meanwhile, is engineered for latency-sensitive serving, pairing 288 GB of high-bandwidth memory with 384 MB of on-chip SRAM and delivering 80% better performance per dollar than its predecessor. Both chips are slated for general availability later in 2026. Yet Google also said it will be among the first cloud providers to offer NVIDIA Vera Rubin NVL72 systems, which makes the strategic message unmistakable: Google wants optionality, not ideological purity. (blog.google)

And Google is hardly alone. AWS announced the general availability of Trn3 UltraServers in December 2025, built around Trainium3, its first 3 nm AI chip; Amazon says the system offers up to 4.4 times higher performance and 4 times better performance per watt than Trn2. Microsoft followed in January 2026 with Maia 200, an inference accelerator for Azure delivering over 10 petaFLOPS at FP4 and scaling to clusters of up to 6,144 accelerators. These are not experimental side projects anymore; they are becoming first-class pillars of cloud strategy. (aws.amazon.com)

Meta’s roadmap underlines the same shift. In March 2026, it said it was developing and deploying four new generations of MTIA chips within two years, while emphasizing an inference-first philosophy and noting that hundreds of thousands of MTIA chips are already handling production inference workloads. The implication is hard to miss: the center of gravity in AI infrastructure is moving from a single dominant chip vendor toward a more fragmented, workload-specific silicon ecosystem. So, will “de-NVIDIA” accelerate? Architecturally, yes. Commercially, however, NVIDIA still looks less replaced than encircled. (about.fb.com)

by EigoBoxAI
作成:2026/04/29 15:05
レベル:超上級 (語彙目安:8000語以上)

まだ読んでいないコンテンツ

content image
by EigoBoxAI
作成:2026/04/29 21:03
レベル:初級 (語彙目安:300〜1000語)
content image
by EigoBoxAI
作成:2026/04/29 21:02
レベル:中級 (語彙目安:2000〜2500語)
content image
by EigoBoxAI
作成:2026/04/29 21:01
レベル:初中級 (語彙目安:1000〜2000語)
content image
by EigoBoxAI
作成:2026/04/29 15:03
レベル:上級 (語彙目安:6000〜8000語)
content image
by EigoBoxAI
作成:2026/04/29 15:01
レベル:中上級 (語彙目安:4000〜6000語)
content image
by EigoBoxAI
作成:2026/04/29 09:04
レベル:超入門 (語彙目安:〜300語)
content image
by EigoBoxAI
作成:2026/04/29 09:02
レベル:初級 (語彙目安:300〜1000語)
content image
by EigoBoxAI
作成:2026/04/29 09:01
レベル:中級 (語彙目安:2000〜2500語)
content image
by EigoBoxAI
作成:2026/04/29 03:05
レベル:初中級 (語彙目安:1000〜2000語)
content image
by EigoBoxAI
作成:2026/04/29 03:04
レベル:超上級 (語彙目安:8000語以上)