Nvidia Adjusts Vera Rubin HBM4 Bandwidth; AMD Helios Systems Enter Testing at Meta
The AI compute build-out continues with shifting hardware specifications and expanded hyperscaler partnerships, pushing toward multi-gigawatt infrastructure deployments.
The story
The dynamic landscape of AI hardware saw a notable adjustment this week as Nvidia lowered the published memory and interconnect bandwidth targets for its Vera Rubin AI hardware. The per-GPU HBM4 bandwidth is now set at 19.2 TB/s, a reduction from the previous 22 TB/s, while NVLink 6 bandwidth per GPU also dropped from 3.6 TB/s to 3 TB/s across all configurations. This revision reflects the challenges in pushing hardware limits and the intense pressure on the memory supply chain.
Meanwhile, Meta is advancing its AI infrastructure by deploying and testing initial AMD Helios rack-scale systems, part of a multi-year partnership that could scale up to 6 gigawatts of AI compute capacity. AMD has also committed a $1 billion investment in Meta's equity offering, deepening their collaboration across AI accelerators, server CPUs, networking, and rack-scale systems. This move signals a diversification in hyperscaler compute platforms, with Meta also planning to deploy Nvidia's Vera Rubin platform and Vera CPUs for its next-generation AI agents. The ongoing build-out underscores the industry's drive to secure and expand specialized compute resources, even as technical specifications evolve.
Silicon
RTX Spark N1X
Maker: Nvidia
What: Arm-based system-on-chip with Blackwell GPU (up to 6,144 CUDA cores) and up to 128GB shared memory.
For Whom: Windows laptops and compact desktops, with more models announced for consumers.
MI455X
Maker: AMD
What: GPU within the Helios rack system, featuring 432 GB of HBM4 memory per GPU.
For Whom: Meta Platforms, with initial Helios systems undergoing testing for AI workloads.
The build-out
Helios Systems Deployment
Who: AMD, Meta Platforms
Scale: Initial systems installed and testing, part of a partnership covering up to 6 GW of AI infrastructure.
Where: Meta's data centers
Oracle Cloud Infrastructure (OCI) Superclusters Expansion
Who: Oracle, Nvidia, Google, Cerebras Systems
Scale: Multi-gigawatt scale expansion, making Nvidia Vera Rubin NVL72 and Vera CPUs generally available, with Cerebras inference and Google TPUs planned for 2027.
Where: Oracle Cloud Infrastructure
British Data Centre Covenant
Who: UK Government, data center developers
Scale: Policy to offer priority electricity grid access for new data centers.
Where: United Kingdom
Supply & policy signals
Financial institutions (Bank of America, Citi, Morgan Stanley, Blackstone) to provide $63 billion in financing for Anthropic's compute expansion.
Implication: Secures long-term access to Broadcom XPU AI accelerators and networking products for Anthropic, starting in 2027.
Nvidia signed agreements with six suppliers (IBIDEN, Samsung Electro-Mechanics, Unimicron Technology, Amdor, KYEC, SPIL) to scale component production.
Implication: Aims to increase manufacturing capacity for substrates by 2.5 times over three years to meet high demand for AI systems.
UK government introduces British Data Centre Covenant following Microsoft's withdrawal from a planned data center due to grid capacity issues.
Implication: Highlights power grid constraints as a significant bottleneck for AI infrastructure development, prompting policy interventions to ensure energy access.
What we'll be watching
- Microsoft Ignite conference in November 2026 for further details on Maia 300 chip deployment plans.
- Volume production at TSMC's next Arizona fab, pulled ahead to Q3 2027.
- Amazon's target of late 2027 for selling its AI chips outside AWS.
- Availability of Cerebras' fast AI inference on Oracle Cloud Infrastructure starting calendar year 2027.
- Google's Tensor Processing Units (TPUs) to be offered on Oracle Cloud Infrastructure starting calendar year 2027.
Reporting + analyst voices: grounded via Google Search at publish time.