How this was made
We used 172 of Huang's appearances from 2021 to November 11, 2026: interviews, talks, podcasts and other recordings. Each appearance counts once on each question.
Finding the videos
We searched public video, podcast and news archives for long appearances where Huang is the main speaker: interviews, talks, podcasts, press conferences and earnings calls. We used the original upload where we found one; some appearances come from other channels’ copies, credited to the outlet that recorded them.
We kept one copy of each event. We removed 35 duplicate uploads and 53 re-uploads posted long after the event under a later date, and held back 2 more whose words we could not check.
We found only 24 interviews from 2021 through 2023, so the charts count those three years as one period. After that, each point is one year, and 2026 runs through November 11.
Counting his words
For the first chapter we counted only Huang's own words, in 172 appearances: 65 before 2025 and 107 since January 2025, each event once. A word or theme counts for an appearance when he says it at least once, and the first chart shows the share of his appearances in each period that mention each theme.
The chart of what his talk is made of sorts his speech into subjects; the note under that chart says how well an independent second sort agreed.
What the word counts can’t tell you:
- He talks about what he is asked about. More of his 2025 and 2026 appearances were on TV news and policy stages.
- China, the US government and America spike in the second half of 2025 rather than rising steadily.
- July to November 2026 is a partial half-year.
- Shown terms held within each venue type (TV, stage talks, podcasts) and without the single longest file, per an independent check.
Who is speaking
We transcribed every recording and confirmed, in several independent ways, which voice was Huang’s. A recording counts only when those checks agree. Every quote on this page was then checked against the recording itself: the words are his, word for word, and they play from the second he says them.
Reading what he said
We picked seven questions that run through his interviews and through business coverage of Nvidia. Independent readers placed each passage on its question: the top answer, the bottom answer, or no clear side. They never saw the video’s title, outlet or date, and a quote counts only when they agree. The labels describe his words, not his beliefs or motives.
A separate check re-read a random sample of the placed quotes without seeing our answer. The table shows how often it agreed with the side we gave, and the likely range.
| Question | Side checks agreed | Range |
|---|---|---|
| Warnings that AI could go rogue or end humanity | 12 of 12 | 75–99% |
| US limits on Nvidia's chip sales to China | 23 of 24 | 79–99% |
| New government rules for AI | 7 of 7 | 64–100% |
| The US government | 20 of 20 | 83–99% |
| The money spent on AI data centers | 22 of 23 | 79–99% |
| Open-source AI models | 23 of 25 | 75–97% |
| What AI does to jobs | 29 of 30 | 83–99% |
First test: did he switch sides?
We first tested whether he switched sides on any question. He did not switch sides on any of the seven. The smallest chance we found that a change was noise, after allowing for asking seven questions, was about 0.56.
| Question | Interviews | Chance it was noise |
|---|---|---|
| New government rules for AI | 23 | about 0.56 |
| Warnings that AI could go rogue or end humanity | 22 | not tested (too few on one side) |
| US limits on Nvidia's chip sales to China | 36 | about 0.73 |
| The US government | 38 | about 0.73 |
| The money spent on AI data centers | 67 | not tested (too few on one side) |
| Open-source AI models | 43 | not tested (too few on one side) |
| What AI does to jobs | 67 | not tested (too few on one side) |
Then: how often he argued each side
We then counted, for each period, how many of all his interviews in that period had him arguing each side. We added this measure after the first test found no switch, so treat the tests below as a description of the counts, not as proof of a change in his views.
Each point on the charts stands for all of his appearances in that year. A busy year with many appearances does not look louder, because each point is a share of that year’s appearances.
| Question | Earlier | Later | Chance it was noise |
|---|---|---|---|
| Warnings that AI could go rogue or end humanity Plays them down. Earlier: 2024, January to June; 2024, July to December. Later: 2026, July to December. | 2 of 41 | 12 of 28 | under 1 in 1,000 |
| US limits on Nvidia's chip sales to China Argues against the limits. Earlier: 2021 to 2023; 2024, January to June; 2024, July to December. Later: 2025, January to June; 2025, July to December. | 2 of 65 | 21 of 47 | under 1 in 1,000 |
| New government rules for AI Says no new rules are needed. Earlier: 2021 to 2023; 2024, January to June; 2024, July to December. Later: 2026, January to June; 2026, July to December. | 1 of 65 | 8 of 60 | about 0.01 |
Limits
Huang mostly speaks on a subject when he is asked about it. He was asked about the dangers of AI and about China far more in 2025 and 2026, so more of his interviews in those years touch those questions. The counts show what he said, not what he chose to raise.
The dates are months. The posting date YouTube holds can be a day off, and a recording can be a few days older than its posting. Public policy dates appear only on the China chart, for reference.
This page is not a judgment of Huang or of Nvidia, and it makes no claim about why he said what he said. You can hear every quote in full and decide for yourself.
Changes
November 2026: An earlier private draft of this page marked a point where his answers changed. A stricter test found no switch of sides on any question, so we replaced those charts with counts per period before publishing.
Periods on the charts: 2021 to 2023, 2024, 2025, 2026 (through Nov. 11).