Claims

Factual · AI compute, hardware and data

“Today, NVIDIA’s token cost is the lowest in the world, not by 10%, by X factors, orders of magnitude, all because we did extreme co-design”

Jensen Huang, NVIDIA CEO · 1 June 2026 · NVIDIA GTC Taipei 2026 Keynote w/ CEO Jensen Huang (Transcript) (Singju Post)

NVIDIA's token cost is the lowest in the world, lower than others by large factors and orders of magnitude, due to extreme co-design.

Resolved false The record shows it did not happen or was not so.

The second judge would not resolve this claim; the verdict shown is the first judge's.

Said for an organisation they lead or work for; it counts on both records.

What would count as true or false

Judged as of 2026-06-01. The claim is that NVIDIA has the lowest inference cost per token in the world, and that its lead is large: 'orders of magnitude', meaning roughly 10x or more against the main competing platforms (for example Google TPU, AMD Instinct, AWS Trainium and other custom accelerators) on comparable models and workloads. TRUE: independent benchmarks or analyses (for example SemiAnalysis InferenceMAX, MLPerf Inference-based cost analyses or independent TCO studies) current around mid-2026 show NVIDIA's leading platform with the lowest cost per token, and about 10x or more below the main competitors' current platforms. MOSTLY TRUE: NVIDIA is clearly lowest, and the gap is several-fold (for example 5-10x) against the main competitors, or is 10x or more against most but not all of them. MIXED: NVIDIA is lowest on some major workloads or models but not others, or ties a competitor. MOSTLY FALSE: NVIDIA is lowest, but only by a modest margin (well under about 5x). The 'orders of magnitude' figure would then come from comparing against NVIDIA's own older generations rather than against rivals, so a real lead is overstated. Also MOSTLY FALSE if the only support is NVIDIA's own figures and independent evidence is inconclusive. FALSE: independent evidence shows a competitor with an equal or lower cost per token on major comparable workloads. The causal attribution to 'extreme co-design' cannot be tested independently and does not change the verdict on its own.

Evidence

1. GB300 NVL72 vs TPU7x: Qwen 3.5 397B-A17B | InferenceX SemiAnalysis (InferenceX) · 13 July 2026

   “TPU7x edges GB300 NVL72 at 439 tok/s/user on Qwen 3.5 397B-A17B — $0.02 per million tokens versus $0.05, a 160% cost-per-token gap.”

2. GB300 NVL72 vs TPU7x: Kimi K3 2.8T | InferenceX SemiAnalysis (InferenceX) · 17 October 2026

   “GB300 NVL72: $0.06 per million tokens. TPU7x: $0.08. Both at 123 tok/s/user on Kimi K3 2.8T, with GB300 NVL72 30% cheaper.”

3. TPU Inference Externalization Full Steam Ahead - InferenceX SemiAnalysis · 7 September 2026

   “Combined with the lower modeled hourly cost, Ironwood delivers 50.4% more tokens per dollar than B200 and 96.0% more than B300.”

4. Tokens per Watt Determines AI Factory Revenue as Power Constraints Tighten TechTimes · 15 July 2026

   “AMD's own analysis of SemiAnalysis InferenceX data shows that the MI355X — the current deployed Instinct GPU — already outperforms GB300 NVL72 on cost-per-million-tokens at operating points above 60 tokens per second per user in FP8 configurations without multi-token prediction.”

Reasoning

SemiAnalysis's independent InferenceX data shows Google's TPU7x cheaper per token than NVIDIA's GB300 NVL72 on Qwen 3.5 397B at one operating point: $0.02 versus $0.05 per million tokens (E1). SemiAnalysis also reports that Ironwood delivers 50.4% more tokens per dollar than B200 and 96.0% more than B300 (E3). GB300 is cheaper at one Kimi K3 operating point, but only by 30% (E2), far short of the 'orders of magnitude' lead claimed. Independent evidence shows a competitor with a lower cost per token on major comparable workloads: Google's TPU7x on Qwen 3.5 at one operating point [E1], and Ironwood against NVIDIA's B200 and B300 on tokens per dollar [E3]. NVIDIA's GB300 NVL72 was 30% cheaper at one Kimi K3 operating point [E2], which is far short of 'orders of magnitude'. Under the criteria, a competitor with an equal or lower cost per token on major comparable workloads resolves the claim false. E4, which relays AMD's own analysis showing the MI355X ahead at some operating points, is consistent with this but carries less weight.

Cite this claim

Jensen Huang, “NVIDIA's token cost is the lowest in the world, lower than others by large factors and orders of magnitude, due to extreme co-design.” (said 1 June 2026). Verdict: resolved false. Reliability API, claim c_d6d611e64e3d, method 0.1.0+03d23c0e. https://epistacklabs.com/claims/c_d6d611e64e3d