- Start
- Jul 23, 202690% CONFIDENCEfrom the source
AMD Launches Helios Rack and Instinct MI450 GPUs
AMD launched its Helios rack-scale AI system, built on MI455X GPUs and Venice EPYC CPUs, with shipments to Microsoft starting in the second half of 2026 and a 2 GW Anthropic deployment from 2027.
- AMD launched its next generation of AI data centre hardware on Thursday 23 July 2026 at an event in San Francisco, aiming to take data centre market share from Nvidia[1]
- The GPU naming is inconsistent across the coverage: the launch article's headline says “MI450 GPUs” while its body names MI455X in Helios and “Instinct MI450 Series GPUs” for the Anthropic deployment; The Next Platform explains the “Altair” MI400 series as covering the MI450, MI430X and MI455X for Helios racks of 64, 72 or 128 GPUs, and possibly an MI440X for eight-way nodes[1][4]
- AMD was also expected to formally launch the Venice data centre CPU at the same event; Venice is the codename for AMD's 6th-generation EPYC, to be made on TSMC's 2nm process[1][5]
- AMD says it expects to begin shipping Helios to customers, including Microsoft, in the second half of 2026, under an expanded partnership in which Microsoft will deploy Helios on Azure to power frontier model inference[1]
- That timing settles an earlier dispute: SemiAnalysis had claimed in February 2026 that engineering samples and low-volume production of the MI455X UALoE72 system would come in H2 2026 but that “due to manufacturing delays, the mass production ramp and first production tokens will only be generated on an MI455X UALoE72 by Q2 2027”[4]
- AMD rejected that at the time: Forrest Norrod said “I have no idea where this purported issue around thermals is coming from... We have no significant thermal issue” and that AMD was “highly confident of ramping Helios in high volume in the second half of the year”[4]
- The day before the launch AMD announced a partnership with Anthropic to deploy up to 2 gigawatts of Instinct MI450 Series GPUs in Helios rack-scale systems, with the first gigawatt beginning in the first half of 2027[1]
- AMD also committed to a strategic equity investment of up to $5 billion in Anthropic, plus a multi-year engineering collaboration using Anthropic's Claude to optimise workloads for Instinct GPUs and accelerate ROCm development, with AMD adopting Claude across its engineering and product teams[1]
- Tom Brown, Anthropic co-founder and chief compute officer: “Access to compute is central to keeping Claude at the frontier and meeting demand from our customers... By partnering with AMD across the stack, we are securing the capacity we need and optimizing it for training and serving Claude”[1]
- ZT Systems, which AMD bought for $4.9 billion in August 2024 and whose manufacturing arm it sold to Sanmina for $3 billion, underpins the rack engineering, using dummy hot plates to simulate CPUs and GPUs and retire thermal risk before silicon returns from the fabs[4][5]
- Market context: Nvidia commands upward of 95% of the data centre GPU market against AMD's roughly 4.5% on Futurum Group estimates, while AMD's data centre segment posted $5.78 billion of revenue in Q1 2026, up 57% year on year[1]
Notable features
- Helios is a rack-scale system combining Instinct MI455X GPUs, EPYC “Venice” CPUs, Pensando networking chips and ROCm software into an integrated platform for AI training and inference, positioned as AMD's first rival to Nvidia's rack-scale AI systems[1]
- AMD's launch post puts 72 MI455X GPUs in one scale-up domain over a UALoE fabric with 260 TB/s of scale-up bandwidth, alongside 6th Gen EPYC “Venice” 9006 Series CPUs and Pensando networking; one rack delivers 2.9 exaflops of dense FP4 and 1.4 exaflops of FP8 compute, 31 TB of HBM4, 1.7 PB/s of aggregate HBM bandwidth and 43 TB/s of scale-out bandwidth[3]
- Each MI455X, on AMD's CDNA 5 architecture, carries up to 432GB of HBM4 and delivers up to 40 PFLOPS of FP4; the Venice CPU uses the Zen 6 architecture with up to 256 cores and 1.6TB/s of memory bandwidth, and the rack's Pensando “Vulcano” AI NICs provide 800 Gbps each for scale-out[2]
- AMD claims Helios delivers up to 15% more AI compute, 50% more HBM capacity and 50% more scale-out bandwidth than an Nvidia Vera Rubin NVL72 rack, and that MI455X GPUs give up to 34X higher token throughput at high interactivity and up to 18X lower token cost than the MI355X on DeepSeek-V4-Flash[3]
- Helios is built to Meta's Open Rack Wide v3 double-wide rack specification, designed to operate a rack full of accelerators as if it were one single large GPU, in the manner of Nvidia's DGX GB200 NVL72; AMD first showed it in June at an event in San Jose[4][5]
References 588% CONFIDENCE
The first entry is always the pin's source. Overall confidence is a weighted average of how firmly each reference supports the start and end times used above; a reference counts half as much for every 180 days older than the newest.
- [1]90%finance.yahoo.com/technology/ai/articles/amd-launches-helios-rack-system-133630311.htmlfinance.yahoo.com· Posted Sep 20, 2026· Starts Jul 23, 2026 ✓· 28% of score
Yahoo Finance reported the launch on July 23, 2026.
- [2]88%AMD Helios Rackscale Solution – Powering Frontier AIamd.com· Added Sep 22, 2026· 29% of score
AMD's[3] Helios product page gives each MI455X "432 GB HBM4" and "up to 40 PFLOPS FP4", and the Venice CPU "up to 256 high performance cores" on Zen 6.
- [3]90%AMD Launches Helios™: The Highest Performing Rackscale AI Infrastructure Solutionamd.com· Published Jul 23, 2026· 22% of score
AMD's[2] own launch blog gives the rack as 72 MI455X GPUs with "2.9 exaflops of dense FP4 compute, 1.4 exaflops of FP8 compute, 31 TB of HBM4" and claims up to 15% more AI compute and 50% more HBM capacity than a Vera Rubin NVL72 rack.
- [4]85%AMD Says Helios Racks And MI400 Series GPUs On Track For 2H 2026nextplatform.com· Published Feb 23, 2026· 13% of score
Trade coverage of the second-half 2026 Helios schedule.[1]
- [5]75%AMD taking AI fight to Nvidia with Helios rack-scale systemtheregister.com· Published Nov 5, 2025· 8% of score
Earlier coverage of the Helios design and Nvidia rivalry.[1]
Suggest a correction
Something missing or wrong? Say it in your own words: a link that backs this pin up, a different start or end date and why, or a fact it lacks or gets wrong. The AI checks it against this pin's sources, searches for better ones, and adds any page that backs you up. The pin's own sources still count most.
