Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-25 · high
Meta plans 1 gigawatt of Helios racks by year end, one piece of a commitment for up to 6 gigawatts of AMD GPUs.
Why now: AMD detailed Helios's rack-scale design and customer list on July 23, with engineering samples due in the second half of 2026.
AMD unveiled Helios, a rack-scale AI system packing 72 GPUs per rack, taking its Nvidia fight to the rack level, per TechCrunch. Microsoft, Meta, OpenAI and Oracle are now buying at that scale, and none of them shop for single GPUs anymore. Engineering samples arrive in the second half of 2026, with mass production targeted for Q2 2027.
Meta's commitment gives the clearest sense of scale. It's planning 1 gigawatt of Helios racks by year end. That's part of a longer deal for up to 6 gigawatts of AMD GPUs, a roadmap bet more than a single order.
Beyond Helios, AMD is already talking about MI500 and Verano CPUs for 2027. MI500 marks a shift toward optical scale-up networking for AMD's chips.
Helios doesn't try to out-spec Nvidia chip for chip. Whatever wins single-GPU benchmarks isn't what buyers are ordering anymore. They're ordering gigawatts of compute, and Meta's 6-gigawatt commitment says at least one buyer is doing exactly that.
Each link below shares sources, entities, or timing with this story.
Huang used his inaugural X post on July 24 to publish "Open Weights and American AI Leadership," a three-page letter on Nvidia's own servers signed by 25 companies including Meta, Microsoft, IBM, Mistral, Mozilla, Hugging Face, a16z, Palantir and the Linux Foundation. Within a...
For about a year, "run your agent locally" meant accepting a model that couldn't reliably call a tool twice in a row. That excuse is gone. Meta Superintelligence Labs published Muse Glimmer today: a 29.6B dense causal transformer, 52 layers, 6,656 hidden dim, with a ~1.8B ViT-...
Trump's State of the Union announced tech companies must "build, bring, or buy" their own power for AI data centers. Amazon, Google, Meta, Microsoft, OpenAI, xAI, and Oracle will sign March 4. US data center energy demand expected to triple by 2028. Critics call it unenforceab...
PR #19378 landed in llama.cpp this week, and I think most people are underselling what it means. Backend-agnostic tensor parallelism via --split-mode tensor makes multi-GPU inference work across AMD, Intel, and Apple Silicon. Not just CUDA. Everything. For context: llama.cpp h...
NVIDIA's Blackwell successor is in production ahead of schedule. The NVL72 rack (72 GPUs) delivers 3.6 exaFLOPS for inference, with 288GB HBM4 per GPU. NVIDIA claims 10x lower cost-per-token versus Blackwell. The Rubin CPX variant — purpose-built for million-token inference —...
Reuters, via Tech Startups, reports capital released against deployment milestones with Anthropic deploying up to two gigawatts of Instinct MI450 starting 2027. Same structure as Nvidia/OpenAI: compute vendor capital flowing to the lab that commits to buy the silicon. A two-gi...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.