NVIDIA Claims Blackwell Ultra Delivers 50x Higher AI Throughput Per Megawatt Than Hopper

NVIDIA has posted a claim that its latest Blackwell Ultra architecture delivers up to 50 times more AI throughput per megawatt of power compared with its Hopper generation of chips — a figure the company is using to argue that large-scale AI infrastructure is becoming heavily more energy-efficient.

The post, shared on NVIDIA’s official social media account, frames the comparison around what the company calls “AI factories” — a term NVIDIA uses for large-scale data centres built to maximise tokens, revenue, or intelligence output per kilowatt-hour of electricity consumed.

The claim centres on NVIDIA’s GB300 NVL72 system, which uses the Blackwell Ultra platform. According to NVIDIA, this system represents a step beyond the earlier Blackwell generation, itself already a successor to Hopper. The company says the GB300 NVL72 is designed specifically for AI inference workloads and the kind of always-on “AI factory” operations that large cloud providers and enterprise customers are building out.

What NVIDIA Is Claiming

The numbers NVIDIA is putting forward are striking. According to the company’s own technical and marketing materials, the GB300 NVL72 delivers up to 50 times higher throughput per megawatt than Hopper. The company also claims the system achieves 35 times lower cost per million tokens versus the same baseline.

For context, NVIDIA says the earlier Blackwell system — the GB200 NVL72 — already delivered around 10 times the throughput per megawatt compared with Hopper. Blackwell Ultra, then, is being positioned as a further major jump on top of that.

NVIDIA goes further still. Across six architecture generations, the company says inference throughput per megawatt has improved by one million times. That figure covers a longer span of chip history and is presented as evidence of a sustained, compounding efficiency curve rather than a single-generation leap.

The company’s public blog cites data from SemiAnalysis InferenceX in support of the 50x throughput-per-megawatt claim for the GB300 NVL72 versus Hopper, and describes the gains as a product of both hardware improvements and software co-design optimisations targeting latency and cost efficiency for agentic AI workloads.

How the Claim Should Be Read

These are vendor-generated figures. The 50x number does not appear to have been independently verified by a third party outside of NVIDIA’s own materials, and the tweet itself does not set out the methodology behind it.

Performance claims of this kind typically depend on workload type, system configuration, latency targets, and the specific benchmarking assumptions used. The comparison is framed around inference workloads — the process of running a trained AI model to generate outputs — rather than general-purpose computing or AI training. That distinction matters, because inference and training place very different demands on hardware.

Jensen Huang, NVIDIA’s chief executive, said at the company’s GTC conference earlier this year: “Every token of intelligence requires energy. AI factories convert energy into continuous intelligence.”

That framing is deliberate. NVIDIA is pitching Blackwell Ultra not just on raw compute power but on the economics of running AI at scale — specifically, how much useful output a data centre can squeeze from each megawatt it draws from the grid.

Why Energy Efficiency Is Now Central to the AI Debate

The energy appetite of AI infrastructure has become one of the most discussed issues in the technology industry. Large language models and AI inference services consume big amounts of electricity, and data-centre operators are under growing pressure from both regulators and investors to demonstrate that their power consumption is justified and manageable.

If NVIDIA’s efficiency claims hold up under real-world conditions, the effect on data-centre operators could be meaningful — lower electricity bills, reduced cooling requirements, and more output from the same physical footprint. But the real-world benefit depends heavily on how the systems are deployed, what workloads they run, and how the electricity powering them is sourced.

The 50x figure, even as a vendor claim, represents the kind of efficiency argument that data-centre procurement teams and cloud providers will scrutinise closely before committing to hardware refresh cycles that can run to hundreds of millions of pounds.

Blackwell Ultra in the Wider NVIDIA Lineup

Blackwell Ultra sits above the standard Blackwell generation in NVIDIA’s current roadmap. The GB300 NVL72 is the specific system configuration NVIDIA is highlighting for inference and AI factory use cases. Hopper — the platform used as the baseline in these comparisons — remains widely deployed across cloud and enterprise data centres globally, which makes it a relevant reference point for customers considering an upgrade.

NVIDIA has not announced a general availability date for GB300 NVL72 systems in the supplied materials, and the timeline for broad commercial deployment remains to be confirmed.

What This Means for Kent Residents

There’s no direct Kent connection to this announcement, but the broader picture is relevant to anyone in the county who uses AI-powered services or works in a sector that’s starting to adopt them. If Blackwell Ultra does deliver substantially lower costs per AI inference — as NVIDIA claims — that could eventually feed through to cheaper, faster AI tools used by businesses, public bodies, and consumers across the UK. Kent organisations procuring cloud-based AI services, from the NHS to local councils to small businesses, would stand to benefit if lower hardware costs translate into lower service prices over time.