NVIDIA Claims 5× Speed Gain and 80% Token Cost Cut for DeepSeek V4 on Blackwell GPUs

NVIDIA has claimed that software improvements to its Blackwell GPU platform delivered a fivefold performance increase for the DeepSeek V4 AI model within a single month, cutting the cost per token to roughly one-fifth of previous levels — all without any change to the underlying hardware.

The claim appeared in a post on X from NVIDIA’s official account, which stated that its inference software “keeps driving down token costs, long after AI infrastructure is deployed.” The post went on to say that “in just one month on NVIDIA Blackwell, software optimizations improved DeepSeek V4 performance by up to 5×, reducing token costs to roughly one-fifth of previous levels.”

No independent body has yet verified those specific figures for DeepSeek V4. They should be treated as an unverified vendor self-report until third-party benchmarks confirm or challenge them.

What NVIDIA Is Actually Claiming

The headline numbers are striking, but the broader argument NVIDIA is making is arguably more consequential for the AI industry than any single figure. The company’s position is that buying Blackwell hardware is not a one-time transaction: the platform keeps getting faster and cheaper to operate as its software stack matures, through tools including TensorRT-LLM and NVIDIA Dynamo.

That matters because inference cost — typically measured as cost per million tokens generated — is one of the central economics of running AI services at scale. Lower token costs improve margins for AI providers and can, in theory, reduce prices for the businesses and individuals using those services downstream.

NVIDIA and the independent analysis firm SemiAnalysis have published benchmark material suggesting that for some large language model workloads, software updates on Blackwell B200 hardware reduced cost per million tokens for the GPT-OSS-120B model from around USD $0.11 at launch to around USD $0.02 within roughly two months — implying a similar fivefold cost improvement, though for a different model. Converted to sterling, that shift is broadly from around 9p to under 2p per million tokens, though exact exchange rates vary. SemiAnalysis’s InferenceX and InferenceMAX benchmarks are the primary source for many of these figures, and it’s worth being clear that SemiAnalysis works closely with NVIDIA on this material; these are not outputs from a neutral standards body.

Blackwell’s Architecture and the Software Angle

NVIDIA’s Blackwell generation — spanning the B200, GB200 and GB300 NVL72 system configurations — was designed with inference efficiency as a central goal. The architecture includes 4-bit tensor cores (NVFP4/FP4), higher memory bandwidth than its predecessor Hopper, and an optimised NVLink Switch interconnect fabric. These hardware features are designed to be exploited progressively by improving software, rather than delivering their full potential at launch.

DeepInfra, an AI inference provider independent of NVIDIA, has separately reported a 20× cost reduction for some Mixture of Experts models on Blackwell, taking serving costs from around USD $1.00 per million tokens for a dense 405B model down to around USD $0.05 per million tokens using NVFP4 quantisation on Blackwell — around 83p down to about 4p per million tokens at current rates. That figure is also self-reported, but it comes from a different company and points in the same direction as NVIDIA’s claims.

NVIDIA and SemiAnalysis have also reported that Blackwell Ultra GB300 NVL72 systems can deliver up to around 35× lower cost per token and up to roughly 50× higher throughput per megawatt compared with Hopper-generation H200 systems for certain low-latency workloads. These are workload- and configuration-dependent figures, not universal guarantees.

Industry Scepticism and Competition Questions

Not everyone accepts the headline numbers at face value. Some industry analysts argue that vendor-cited benchmarks tend to reflect best-case configurations that don’t always translate to the messier reality of production workloads. The SemiAnalysis benchmarks, while widely cited, are produced in close collaboration with NVIDIA and don’t carry the weight of a formal industry standard.

There are also broader competitive concerns. The UK’s Competition and Markets Authority has been monitoring the AI compute and cloud markets over worries about market concentration. NVIDIA’s continued dominance in high-end AI hardware, combined with a proprietary software stack that organisations must keep pace with to realise efficiency gains, could limit customer choice and weaken the bargaining power of public sector buyers — including those in the UK.

Dylan Patel, chief analyst at SemiAnalysis, said of the Blackwell platform’s inference economics: “The software improvements are real and they compound — the question is always how quickly they translate into lower prices for end customers rather than higher margins for providers.”

Unanswered Questions

Several things remain unclear. NVIDIA has not published a detailed methodology for the DeepSeek V4 claim specifically — the model configuration, batch sizes, latency targets and hardware setup that produced the “up to 5×” figure are not publicly specified in the tweet or in accompanying documentation reviewed at time of writing. It’s also not yet known how quickly, or whether, cloud providers will pass software-driven cost reductions on to customers rather than absorbing them as margin improvements.

DeepSeek V4 itself, as a model family, has limited independent public documentation compared with better-studied models such as DeepSeek-R1 or GPT-4o. That makes external validation of NVIDIA’s specific claim harder than it might otherwise be.

What This Means for Kent Residents

Kent residents won’t feel this directly — but if UK cloud providers adopt Blackwell-based infrastructure and pass on lower inference costs, the AI-powered tools embedded in everyday services, from NHS digital triage systems to council chatbots and retail assistants, could become faster or cheaper to run over time. Local businesses in Kent using cloud-hosted AI for tasks such as customer service, translation or analytics may see lower usage costs in the medium term if token prices fall across major platforms. Public bodies including Kent County Council and NHS Kent and Medway ICB, both operating under tight budgets, could potentially stretch existing digital spending further if the underlying cost of running AI models continues to fall — though any benefit depends entirely on adoption decisions made at a national or commercial level, not locally.