ClaimsJensen Huang

“And of course, it’s not as good as Blackwell, and I get that.”

Jensen Huang, Nvidia CEO 26 September 2025 Transcript: Jensen Huang on OpenAI, Future of Compute (BG2 Pod) (Singju Post)

Factual AI compute, hardware and data

Nvidia's H20 chip is not as good as Nvidia's Blackwell chip.

Resolved true The record shows it happened or was so.

Factual: Statements about the present or the past, checked against the record: did they hold up?

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 2025-09-26. True: published specifications or benchmarks show that the H20 (an export-restricted Hopper chip) has lower AI compute performance (such as FP8/FP16 TFLOPS or inference throughput) than Blackwell GPUs such as the B200 or B300. Mostly true: H20 is weaker on most metrics but matches or exceeds Blackwell on a minor metric (such as memory capacity versus some Blackwell variants). Mixed: the comparison is roughly even across metrics. Mostly false: H20 is better on most metrics. False: H20 outperforms Blackwell overall.

Set before the research began.

Evidence (8)

  1. E1 The H20 Export License Reversal, the AI Moratorium, and Qualcomm's Long Shot m1k.tech · 15 July 2025

    “| FP16 Tensor | ~148 TFLOPS | 989 TFLOPS |”

  2. E2 NVIDIA B200 Datasheet | GPU Smith GPU Smith LLC · 4 October 2026

    “and packaged as one SXM6 module with 180 GB of HBM3e.”

  3. E3 NVIDIA DGX B200 NVIDIA · 16 October 2026 Own record

    “FP8 Tensor Core: 72 PFLOPS**”

  4. E4 The H20 Export License Reversal, the AI Moratorium, and Qualcomm's Long Shot m1k.tech (personal blog) · 15 July 2025

    “A China-market B20, if it follows the H20 playbook, would have its FP8 fused down to maybe 15-20% of the B200 while keeping the memory fat and the NVLink intact.”

  5. E5 Nvidia's latest regulation-compliant H20 GPU for China has been delayed to early next year: Report Tom's Hardware · 24 November 2023

    “| INT8 I FP8 Tensor | 296 I 296 TFLOPS | 239 I 239 TFLOPS | 193 I 193 TFLOPS |”

  6. E6 Nvidia announces Blackwell Ultra B300 —1.5X faster than B200 with 288GB HBM3e and 15 PFLOPS dense FP4 Tom's Hardware · 18 March 2025

    “| FP6/FP8 Tensor Dense/Sparse | 7.5/15 petaflops ? | 5/10 petaflops | 3.5/7 petaflops |”

  7. E7 Nvidia announces Blackwell Ultra B300 —1.5X faster than B200 with 288GB HBM3e and 15 PFLOPS dense FP4 Tom's Hardware · 18 March 2025

    “| Memory | 288GB (8x36GB) | 192GB (8x24GB) | 192GB (8x24GB) |”

  8. E8 NVIDIA H20: Specs, Export Status and Guide (2026) Aquanode · 8 October 2026

    “The NVIDIA H20 is a Hopper-generation datacenter GPU that NVIDIA built for the China market to fit US export rules: it keeps a large 96 GB of HBM3 memory with about 4 TB/s of bandwidth, but its compute is cut to a small fraction of an H100.”

Reasoning

E10 describes the H20 as a Hopper GPU with 96 GB of HBM3 whose compute is cut to a small fraction of an H100's. E2 gives about 148 TFLOPS FP16 tensor against 989 TFLOPS in the H100 comparison table, and E7 lists FP8 tensor figures of only a few hundred TFLOPS in its 2023 H20 article. Blackwell figures are in the petaflops range: E8 gives the B200 5/10 PFLOPS FP8 dense/sparse, and E4 gives 72 PFLOPS FP8 for an 8-GPU DGX B200. Blackwell also has more memory (E9: 192–288 GB), so the H20 does not come out ahead on the metrics shown, and the claim resolves true.

Cite this claim

Jensen Huang, “Nvidia's H20 chip is not as good as Nvidia's Blackwell chip.” (said 26 September 2025). Verdict: resolved true. Reliability API, claim c_aa4bf34694f4, method 0.1.0+03d23c0e, dataset release v1.1.2. https://epistacklabs.com/claims/c_aa4bf34694f4

This claim is in dataset release v1.1.2, with the same verdict.

Reply to this claim

For Jensen Huang, or someone speaking for them: a correction or counter-evidence. Replies are read before anything is shown; an approved reply shows your name, role, date, statement and link. Your email address is never shown and is deleted 30 days after the decision (privacy).

More factual claims by Jensen Huang

  1. Mostly true Factual 4 September 2026
    “we invented a technology called a GPU, and basically a GPU does parallel processing tasks”

    NVIDIA invented the GPU, which does parallel processing tasks such as fluid dynamics, particle physics, quantum chemistry, graphics and AI.

  2. Resolved false Factual 1 June 2026
    “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”

    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.

  3. Mostly true Factual 1 June 2026
    “It’s the first one to be PCI Express Gen 6.”

    Vera is the first CPU to support PCI Express Gen 6.

  4. Mixed Factual 17 March 2026
    “We have demonstrated that NVFP4 can achieve inference without loss of precision, while delivering significant performance and energy efficiency improvements, and it is equally suitable for training.”

    NVIDIA has demonstrated that NVFP4 can achieve inference without loss of precision, with significant performance and energy-efficiency gains, and is equally suitable for training.

  5. Mixed Factual 17 March 2026
    “Over the past two years, the computational demands for inference have increased approximately 10,000 times, while usage has increased approximately 100 times.”

    Over the past two years, computational demand for inference rose about 10,000 times while usage rose about 100 times.

All of them

Verdicts are made automatically; how verdicts are made.