Within two years, the token generation rate of a 1-gigawatt AI factory will increase from 22 million to 700 million tokens per second, a 350-fold increase.

This is a forward-looking claim made on 2026-03-17: within two years (by 2028-03-17), the token generation rate of a 1-gigawatt AI factory will rise from 22 million to 700 million tokens per second, a stated 350-fold increase. Judge at the horizon against Nvidia-published or independent benchmark figures for tokens per second per gigawatt on comparable workloads, using Nvidia platforms shipping by 2028-03-17. Note that 700/22 is about 32x, not 350x, so the numbers and the multiple are internally inconsistent. If the source transcript shows the multiple was misquoted (for example 35x), use the transcript's figures. True: by 2028-03-17 the documented rate for a 1 GW Nvidia-based AI factory reaches about 700 million tokens/s (within about 10%) from a baseline of about 22 million, and the stated multiple matches the transcript. Mostly true: the rate reaches about 700 million tokens/s on time, but the '350-fold' multiple is an arithmetic or transcription error not borne out by the figures. Mostly false: the rate improves substantially but stays well short of 700 million tokens/s (for example 300–600 million), or reaches it only after 2028-03-17, or only on a cherry-picked workload not comparable to the baseline. False: by 2028-03-17 the rate is nowhere near 700 million tokens/s (for example below 300 million) or the platforms needed have not shipped. Nvidia's own projections or slides alone do not establish delivered performance.

Set before the research began.

1. E1 transcript.md NVIDIA · 16 March 2026 Own record

   “But with this architecture, we're going to take our token generation speed token generation rate from 2 million to 700,000,350 times increase.”

2. E2 Issue #97: NVIDIA GTC 2026: $1 Trillion in Orders, Seven New Chips, Groq's Place in the Ecosystem, & NemoClaw thegpu.ai · 21 August 2026

   “The maths: a one-gigawatt AI factory running Vera Rubin generates 700 million tokens per second. On x86 plus Hopper-era systems, that figure was 0.2 million. A 3,500x increase in two years.”

3. E3 Vera Rubin Inference Results Leave Nvidia’s 75% Training Claim Untested | TokenPost TokenPost · 26 October 2026

   “When the systems were measured at their respective throughput peaks, Vera Rubin’s total throughput per megawatt was 1.14 times GB300’s.”

4. E4 Vera Rubin Inference Results Leave Nvidia’s 75% Training Claim Untested | TokenPost TokenPost · 26 October 2026

   “Oct. 15 AA-AgentPerf results put Vera Rubin NVL72 at 5.7 times GB300’s total throughput per megawatt at a specific speed”

5. E5 NVIDIA GTC 2026 Keynote with Jensen Huang (transcript, section 'NVIDIA Vera Rubin: 7 Chips - 5 Rack Systems') NVIDIA · 16 March 2026 Own record

   “we're going to take our token generation speed token generation rate from 2 million to 700,000,350 times increase”

6. E6 Issue #97: NVIDIA GTC 2026: $1 Trillion in Orders, Seven New Chips, Groq's Place in the Ecosystem, & NemoClaw thegpu.ai (newsletter by Ben) · 21 August 2026

   “The maths: a one-gigawatt AI factory running Vera Rubin generates 700 million tokens per second. On x86 plus Hopper-era systems, that figure was 0.2 million.”

7. E7 NVIDIA GTC 2026 Keynote with Jensen Huang (transcript, section 'NVIDIA Extreme Co-Design' / Groq integration) NVIDIA · 16 March 2026 Own record

   “We run Dynamo, this incredible operating system for AI factories on top of it and you get 35 times increase.”

8. E8 NVIDIA Vera Rubin NVL72 NVIDIA · 21 October 2026 Own record

   “When paired with LPX, Vera Rubin NVL72 delivers up to 35x higher throughput per megawatt for trillion-parameter models.”

9. E9 NVIDIA Vera Rubin debuts in MLPerf Inference v6.1 — Data Phoenix Data Phoenix · 18 September 2026

   “On DeepSeek-R1, the Vera Rubin submission recorded 652,750.14 tokens per second in Interactive, 1,183,326.85 in Offline and 1,175,890.22 in Server.”

10. E10 NVIDIA Vera Rubin debuts in MLPerf Inference v6.1 — Data Phoenix Data Phoenix · 18 September 2026

    “The six Vera Rubin result rows do not include full-system power measurements, however, and MLCommons validates power only when those measurements accompany a submission.”

11. E11 Nvidia Vera Rubin Debuts on MLPerf 6.1, Up to 2.5x Faster Than GB300 Pasquale Pillitteri · 23 September 2026

    “The result carries a preview label, though, and redoing the math per GPU on the published numbers brings the measured margin down to between 1.6 and 2 times.”

The keynote transcript (E1, E5) gives the figures as 2 million to 700 [million] tokens per second, a 350-fold increase. That suggests the claim's '22 million' baseline is a transcription error. E2 and E6 give a baseline of 0.2 million, but they are newsletter restatements and do not show measured performance. The independent measurements give per-rack throughput without power data (E9, E10, E11) or only per-megawatt ratios against GB300 (E3, E4), so none of them documents an absolute rate of about 700 million tokens per second for a 1 GW Nvidia-based AI factory. Nvidia's own 'up to 35x' marketing figure (E8) does not establish delivered performance either, so the evidence does not settle the claim.