AMD launches Helios rack-scale AI system and commits $5 billion to Anthropic partnership; NVIDIA releases Vera Rubin platform with Spectrum-6 networking
Product launch ● Confirmed 72% confidence first seen
AMD launched Helios, its first rack-scale AI system to compete with NVIDIA's Vera Rubin, with major customers including Microsoft, Meta, OpenAI, and Anthropic committing to deployments. Simultaneously, AMD announced a $5 billion investment in Anthropic with plans to deploy up to 2 gigawatts of AMD Instinct MI450 GPUs. NVIDIA unveiled the Vera Rubin platform featuring seven co-designed chips and the Spectrum-6 switch system, achieving 10x performance improvements in tokens per megawatt compared to predecessors, with early deployments from CoreWeave, Microsoft, and other major cloud providers.
Decision brief
- What changed
- AMD launched its first rack-scale AI system, Helios, with Microsoft, Meta, OpenAI, and Anthropic as customers, and committed up to $5 billion to Anthropic alongside a plan to deploy up to 2 gigawatts of MI450 GPUs (first gigawatt online H1 2027). Concurrently, NVIDIA released its Vera Rubin platform with seven co-designed chips and the Spectrum-6 switch system, claiming 10x tokens-per-megawatt gains over Grace Blackwell, with early deployments at CoreWeave, Microsoft, and Bristol Myers Squibb.
- Why it matters
- Two dominant AI infrastructure vendors are simultaneously escalating rack-scale system launches and multibillion-dollar customer commitments, signaling accelerating capex cycles and a potential shift in data-center GPU market share (AMD projects moving from 4.5% to 20-25%) away from NVIDIA's near-monopoly. Pricing differences (Helios at $5-5.5M vs. Vera Rubin at $3.5-4M) and claimed efficiency gains will directly affect total cost of ownership and vendor negotiating leverage for any enterprise planning large AI infrastructure buildouts.
- Evidence
- The core facts—AMD's Helios launch, the $5B Anthropic commitment, and NVIDIA's Vera Rubin/Spectrum-6 release—are corroborated across independent outlets (TLDR, TechCrunch, The Verge) and NVIDIA's own newsroom, with consistent customer names (Microsoft, Meta, OpenAI, Anthropic, CoreWeave) appearing in multiple reports. Performance claims like 10x tokens-per-megawatt are sourced primarily from NVIDIA and a CoreWeave benchmark cited in NVIDIA's own release, making them vendor-reported rather than independently audited.
- What remains uncertain
- It is unclear whether AMD's projected market-share gain (4.5% to 20-25%) is an internal projection or independently validated, and the relationship between the separate Meta-Anthropic $10B compute lease talks and the AMD-Anthropic deal is not clarified in the coverage. Actual delivery timelines, real-world performance parity claims, and whether announced 'commitments' from Microsoft, Meta, and OpenAI represent binding contracts or non-binding pilots remain unverified.
- Monitor next
- Watch for confirmation of Helios's actual 2025 shipment dates and the first gigawatt of AMD MI450 deployment coming online in H1 2027 as a test of whether these commitments translate into real infrastructure and market-share shifts.
Analytical support, not advice — assumptions and open questions stated above.