← Back to feed

Nvidia’s Blackwell Ultra ramp has reached a scale that industry analysts now estimate at $89 billion in production value — a figure that encompasses both enterprise deployments and cloud provider buildouts, though Nvidia does not publish official revenue breakdowns at the chip level.

The Numbers

Third-party estimates for a fully configured GB300 NVL72 rack range from $3.7 million to $6.5 million depending on configuration and source. On the cloud side, individual B300 GPU access was priced at approximately $9.08 per hour on demand as of July 2026. Those figures make GB300 significantly more expensive than its Hopper predecessor but reflect the performance uplift across training and inference workloads.

Blackwell Ultra became the primary driver of Nvidia’s Q2 data center growth. The company reported that large-scale Blackwell infrastructure deployments by cloud service providers and AI enterprises are the main revenue engine, with Blackwell entering “broader commercial deployment.”

The Cadence Shift

The more consequential development is structural. Nvidia has shipped a new flagship data center architecture or major revision once per year since 2022. That compresses what was a two-year GPU development cycle — the cadence that governed the Volta-to-Ampere-to-Hopper arc — into an annual rhythm.

The acceleration is a direct response to competitive pressure. Google’s TPU program, AWS’s Trainium and Inferentia lineup, and Microsoft’s Maia custom silicon each shortened Nvidia’s window to coast on any single generation. The result: a Blackwell-to-Blackwell-Ultra-to-Rubin sequence that forces buyers to upgrade more frequently and keeps Nvidia’s revenue cycle tighter.

Rubin, Nvidia’s next architecture, has not shipped.

Google’s Counter

Google split its eighth-generation TPU into two specialised chips rather than a single general-purpose design:

• TPU 8t (training): 121 FP4 exaFLOPS per superpod, optimised for large-scale model training
• TPU 8i (inference): 10.1 FP4 petaFLOPS per chip, optimised for low-latency serving

Google claims meaningful performance-per-dollar gains for each chip relative to its seventh-generation Ironwood TPUs. The specialisation strategy is the opposite of Nvidia’s unified-GPU approach and reflects Google’s ability to co-design hardware with its own workloads.

The Strategic Picture

The key tension is whether specialised custom silicon can erode Nvidia’s grip on training workloads — historically its most defensible position. Google’s TPU 8t is one attempt. AWS Trainium 2 and Microsoft’s Maia are others. None have published third-party benchmark results that directly compare against GB300 at equivalent system scale.

For buyers, the annual cadence creates a new problem: any multi-year infrastructure commitment now carries architecture obsolescence risk that did not exist in the two-year cycle. The $3.7M-$6.5M per-rack price range is not a purchase you can make lightly when the successor is 12 months away.