AMD's 97% Data Center Jump vs Nvidia's 30x Claim: Where AI Buyers Stand

AMD booked $8.5 billion in Data Center revenue in Q3 2026, up 97% year over year — money already banked. Nvidia claims up to 30x throughput per megawatt for Vera Rubin NVL72 over GB300 NVL72 — a vendor-published target measured against its own previous generation. Treating those two numbers as equals is the most common mistake in current AMD-vs-Nvidia coverage.

As of November 12, 2026, the buying decision is no longer about which chip is faster. It's about which rack ships, which cloud rents it, and which software stack your team can run on day one.

What actually happened, and when

Nvidia and Oracle announced on October 27 that Vera Rubin NVL72 systems and Vera CPUs (Nvidia's server processor, with 88 Olympus cores per the company) are generally available on Oracle Cloud Infrastructure. AMD's November 3 results said the first Helios racks, built on Instinct MI450 Series GPUs (AMD's data-center accelerator family), were already running workloads for Anthropic, Meta, Microsoft, OpenAI and Oracle.

AMD's chair and CEO Lisa Su said she expects further Q4 growth as AMD scales Helios and the Venice EPYC platform — the sixth-generation server CPU family, generally available since November 10 on the Zen 6 architecture. Nvidia, meanwhile, says Vera Rubin is "ramping into full production," with Taiwanese server makers shipping Rubin-based systems.

What does the evidence actually prove?

Not a head-to-head comparison. No independent benchmark of MI450 against Rubin exists. Every performance figure for the newest generation comes from a vendor, and nearly all compare a vendor's new product against that same vendor's old one:

• Nvidia claims up to 30x throughput per megawatt for Vera Rubin NVL72 vs GB300 NVL72 — Nvidia versus Nvidia, under conditions Nvidia chose.
• Nvidia and Oracle state a one-tenth cost-per-token comparison, and a separate claim of up to 30x throughput per dollar; nobody outside Nvidia has reproduced either.
• An NCCL all-reduce result of 744.71 GB/s across two compute trays shows the rack fabric works in a reference cluster, not how your model scales.
• Vera Rubin NVL72 debuted in the MLPerf Inference v6.1 round on September 16 — with no AMD Helios submission to compare against.
• Nvidia's Vera CPU claim of 1.8x single-thread performance uses a baseline Nvidia picked; AMD pushed back with its own EPYC Venice numbers.

Memory, software and price

On capacity, the published numbers favor AMD. AMD describes MI455X as carrying 432 GB of HBM4 (high-bandwidth memory) with 3D chip-stack packaging; Nvidia's co-design post frames Vera Rubin NVL72 around 288 GB per GPU. Nvidia's counter is system-level: the Vera CPU carries up to 1.5 TB of LPDDR5X, coherently linked at 1.8 TB/s over NVLink-C2C, to park idle key-value cache during agentic workloads with long pauses. Dense batch inference gains less from that design.

Software splits the field differently. Nvidia's CUDA ecosystem is the incumbent — the company says its technologies power 81% of the TOP500 supercomputing list. But Nvidia's own Kubernetes recipe for Vera Rubin NVL72 is still in Preview, validated with GPU Operator v26.7.0 and DRA driver v0.5.0, with production support targeted for the R620 driver branch. AMD's ROCm stack sits at version 10.1, with MI450 identified as a CDNA 5 part targeting gfx1250.

Neither vendor publishes prices. One analyst estimate, from FlexiCloud, puts a Vera Rubin NVL72 rack at US$7.8 million to US$9.1 million and a Rubin GPU near US$55,000 — estimates, not confirmed figures.

Scale commitments and deployments

Both vendors sell the same shape of promise: gigawatts over years. AMD and Anthropic announced a partnership on July 22 for up to 2 gigawatts of MI450 Series GPUs — a capacity plan, not an installed base. Oracle planned 50,000 AMD Instinct GPUs for 2026. Nvidia says OCI Vera Rubin infrastructure will expand to multi-gigawatt scale, and a July announcement describes a Japan project with Noetra architected around 13,750 Vera CPUs and 27,500 Rubin GPUs. Meta, an early MI450 definition customer, said October 27 that initial Helios racks were installed and being tested at its sites.

For smaller buyers, the entry points are the eight-GPU and server-level products: Nvidia's HGX Rubin NVL8 (planned on OCI) and AMD's air-cooled MI350P, both suited to existing facilities.

What about the governance question?

On November 8, Nvidia's independent directors opened an investigation, with outside counsel, into reports that CEO Jensen Huang was the target of a foreign-intelligence effort. The board said it has no reason to believe national-security information or Nvidia secrets were compromised, and expressed confidence in Huang. Nothing announced changes the Rubin production plan. Buyers with long contracts should still get delivery commitments in writing.

The verdict

Nvidia is the lower-risk choice for teams that need to rent capacity and run software this quarter; AMD is the credible second source that serious buyers should qualify now. Frame that with the missing evidence: no independent MI450-vs-Rubin benchmark, no list prices from either vendor. Buyers should run their own workloads on both platforms, measure cost per million tokens at target latency, and treat every vendor multiple as an opening position in a negotiation.