Skip to main content
NVIDIA, GPUs & AI chips
AI
技術選定

Jensen Huang's life and technology leadership — his early years, career and the CUDA decision, checked against his own words and the record

Jensen Huang's background, career at AMD and LSI Logic, the NV1 failure, the CUDA bet and his 60-person staff, checked against his own words and filings.

Published
Reading time
30 min read
Author
友田 陽大
Share

Jensen Huang co-founded NVIDIA in 1993 and ran it as CEO for 33 years, until November 2026. He is usually described as the executive who turned the GPU into the computing base of AI, and the description often takes the form "he succeeded because he is a great man".

This article avoids that form. It checks his career and his main decisions against two things: his own words, in their original wording, and the record, such as NVIDIA's annual reports. It then tries to separate how much of the outcome came from his decisions and how much from the environment and from luck.

1. The short answer

  • Background: Born in Taiwan, he moved through Thailand and a boarding school in Kentucky and grew up in Oregon. He earned a bachelor's in electrical engineering at Oregon State University and a master's at Stanford (1992). After chip-design work at AMD (1983–85) and LSI Logic (1985–93), he founded NVIDIA with two engineers in April 1993.
  • His pattern as a technology leader: He admitted that the first product line had failed and rebuilt around the industry standard. He kept CUDA on consumer GPUs even when it cut into profit. He left markets the company could not win. These are the three decisions he keeps coming back to.
  • Organisation: Many direct reports, no one-on-ones, information shared in front of everyone. The only direct-report count confirmed in his own words is "60" (March 2026; he added "More" right after).
  • Management philosophy: He keeps returning to first principles (rethinking a problem from its premises), the idea that no task is beneath him, and the view that pain and suffering refine a company's character (section 10).
  • Limits: His recollections sometimes disagree with the record. Luck mattered a lot: the volume of the gaming market, competitors' delays and a discovery made by outside researchers. Plenty of his bets failed. And with his resignation in November 2026, the dependence on one person stopped being a hypothetical risk (my view).

Do you actually know whether self-hosting beats the API?

Self-hosted LLM technology selection — quantization, GPU budget, commercial licensing

2. How to read this article

Articles about people tend to mix facts, the subject's own story and the writer's interpretation. This one keeps three layers apart.

LayerWhat it isHow it is written
RecordSEC filings, company announcements, university and museum pages, papersStated as fact
His own wordsStatements whose original wording, date and venue I confirmed in a transcript or article published by the host or the companyQuoted in the original
AnalysisWhat I conclude from the record and his statementsMarked as my view

Quoting rule: when only a third-party transcription exists, I do not put the words in quotation marks; I summarise them as "he reportedly said". Popular "quotes" whose source I could not confirm are left out (section 15).

As-of date: this article reflects information available on November 17, 2026 (JST). NVIDIA's fiscal year ends in late January, so "FY2026" covers roughly February 2025 to January 2026.

3. Early years: from Taiwan to Oregon

Taiwan, Thailand, Kentucky

The Computer History Museum (CHM) recorded an oral history with Huang on August 9, 2024. Its catalog description summarises his path: born in Taiwan, then to Thailand, where his father was starting an oil refinery, then a boarding school in Oneida, Kentucky, then college in Oregon, and finally Silicon Valley, where he earned a master's in electrical engineering.

The school is widely reported to be Oneida Baptist Institute. The CHM description itself does not name it.

On his birth year and birthplace, here is exactly what I could confirm.

  • NVIDIA's annual report (10-K) gives his age as 63 as of February 20, 2026. That puts his birth date between February 21, 1962 and February 20, 1963.
  • Many encyclopedias give 1963 and Tainan. But the company, university and museum sources I checked say only "Taiwan". Some sources say Taipei.
  • Sources also disagree on how old he was when he left Taiwan. In his 2023 National Taiwan University speech he reportedly said he left at four (per a third-party transcription); some encyclopedias say he moved to Thailand at five and to the US at nine.

Oregon and Denny's

According to a 2022 announcement from Oregon State University (OSU), Huang and his wife graduated from Oregon high schools and met as undergraduates in OSU's College of Engineering.

According to NVIDIA's official blog (September 26, 2023), his first job was at a Denny's in Portland, Oregon, at 15. In his words:

"I was a dishwasher, I was a busboy, I waited tables."

That experience keeps returning in how he talks about management. At Stanford GSB in 2024 he cited dishwashing as the reason "no task is beneath me" (section 10).

4. The engineer years: AMD and LSI Logic

NVIDIA's 10-K describes his pre-NVIDIA career as follows.

YearsEmployerRole (as stated in the 10-K)
1983–1985AMDMicroprocessor designer
1985–1993LSI LogicA variety of positions, including Director of CoreWare, the business unit responsible for LSI's SoC (system-on-chip) work

His degrees are a BSEE from Oregon State University and an MSEE from Stanford. Stanford Engineering's "Engineering Heroes" page dates the master's to 1992, which means he earned it while working at LSI Logic.

A one-year mismatch: a third-party transcription of the National Taiwan University speech has him graduating from OSU in 1984, and NVIDIA's blog says he graduated "four decades ago". The 10-K, however, dates his AMD job from 1983, and the CHM description says AMD was his first job after OSU. None of the sources I read explains the gap.

What this period brought to his management (my view): according to the CHM description, he chose AMD for the chance to design microprocessors and moved to LSI Logic for its new generation of design-automation tools. Part of his LSI job was working with two engineers at Sun Microsystems, who later became his co-founders.

So at founding he knew both sides: designing chips, and building them into a customer's system. That combination fits NVIDIA's later direction of building not just chips but systems and software. Still, I cannot claim that his career caused that direction.

5. The founding: three people at Denny's, April 1993

NVIDIA's official timeline gives the founding date as April 5, 1993, and the founders as Jensen Huang, Chris Malachowsky and Curtis Priem.

The well-known Denny's story can be confirmed in his own words. At Stanford GSB in 2024 he recalled:

"I was an engineer at LSI Logic, and Chris and Curtis were at Sun."

The two wanted to leave Sun and asked him to work out with them where they would go, so the three met at Denny's whenever they dropped by. According to NVIDIA's blog, the restaurant is a Denny's in Silicon Valley, which installed a commemorative plaque in 2023. He explained the choice of venue:

"It had all the coffee you could drink and no one could chase you out."

He has also told the fundraising story. At GSB he said he did not know how to write a business plan, bought a book of about 450 pages on the subject, and never got through it. In the end, the CEO of LSI Logic phoned Don Valentine, the founder of Sequoia Capital, and vouched for him. His lesson: you can have a bad interview, but you cannot run away from your past, so have a good past.

One caution: his stories were organised after the success. That does not make them untrue. Every founder's memories are told with the outcome known. Later sections hold the stories up against the record.

6. The first failure: NV1, Sega and the RIVA 128 "reset"

What went wrong

NVIDIA's first product line drew 3D graphics in a way that differed from the industry mainstream. On the Acquired podcast, published in October 2023, he explained:

"NV1 and NV2 were based on forward texture mapping, no triangles but curves, and tessellated the curves."

He called the approach "completely the wrong answer". When Microsoft released DirectX, it was "fundamentally incompatible with Nvidia's architecture". And although NVIDIA had been the only consumer 3D graphics company when it was founded, "Thirty competitors have already shown up", he recalled. (The figure of 30 is his memory; I did not count them independently.)

How he handled the Sega contract

He told this story at the National Taiwan University commencement in May 2023. NVIDIA's official blog summarises it: the company had won a key contract to help Sega build a game console, rapid industry changes forced it to give up the contract, and it nearly went bankrupt until Sega's leadership helped it avoid that. The blog quotes him:

"Confronting our mistake and, with humility, asking for help saved NVIDIA."

A third-party transcription of the speech gives more detail: after a year of development the company realised its approach was wrong, he told Sega's CEO to find another partner, he still asked to be paid under the contract, and Sega agreed. Because that transcription was not published by the host or the company, I only summarise it here. I could not confirm the exact terms.

The restart with RIVA 128

He calls RIVA 128 (internally NV3) "a reset of our company". On Acquired he said that 1997, with the company's back against the wall, was "probably Nvidia's best moment", and described the first decision they made:

"DirectX is now here. We're not going to fight it."

My view: the substance of that decision was to drop the proprietary approach and ride the standard. The irony is that NVIDIA later became strong by turning its own platform, CUDA, into the industry standard. It moved from following a standard to setting one, and that shift is the counterpart to the decision in section 7.

7. CUDA: how he kept funding something nobody had asked for

CUDA is NVIDIA's programming platform for running general-purpose computation, not just graphics, on GPUs. The technology itself is covered in this series' Why NVIDIA GPUs dominate AI; this section looks at CUDA as a management decision.

About the year: NVIDIA's official timeline dates the unveiling of the CUDA architecture to 2006. NVIDIA's blog post on the National Taiwan University speech, by contrast, summarises it as "the decision in 2007 to put CUDA into all the company's GPUs". Announcing CUDA and deciding to put it in every GPU are different events, so the one-year gap is not necessarily a contradiction.

7.1 The original target was not deep learning

In hindsight CUDA looks like a far-sighted bet on AI. His own account is different. According to the transcription of the National Taiwan University speech, the original targets he named were scientific computing, physics simulation and image processing. Of 2012, he reportedly said that AI researchers found CUDA.

This matters. CUDA was not an investment that foresaw deep learning. It was an investment in a broader hypothesis: that there was demand for general-purpose parallel computing. Deep learning arrived later on top of that hypothesis and became its largest use.

7.2 The install-base argument

So why put CUDA even on consumer GeForce cards? On the Lex Fridman podcast, published in March 2026, he explained:

"we ought to put CUDA on GeForce and put it into every single PC whether customers use it or not, and use it as a starting point of cultivating our install base."

The install base is the number of devices in the field that can run a technology. Developers want their software to reach many people, so they pick the platform with the most devices. More devices bring more developers, more developers bring more software, and more software brings more devices. To start that loop you need devices first. By his account, GeForce was already selling millions of units a year, and it became the way to get those devices.

There is a record that supports this logic. The 2012 AlexNet paper (the image-recognition work widely seen as deep learning's turning point) reports training on "two NVIDIA GTX 580 3GB GPUs", which took five to six days. The GTX 580 is a gaming GeForce card. Researchers could use CUDA because it shipped on cheap GPUs a lab could buy.

7.3 What the record says the bet cost

He describes the cost of the decision this way (Lex Fridman, March 2026):

"it completely consumed all of the company's gross profit dollars."

Let us check that against the record. The code below recomputes R&D intensity (R&D expense ÷ revenue) and a few other figures from NVIDIA's 10-K filings.

  • Language and environment: Python 3.9 or later, standard library only (I ran it with Python 3.9.6 on an Apple M4 Pro)
  • Run: python3 nvidia_records.py
  • Input sources: FY2008–2010 from the FY2010 10-K, FY2024–2026 from the FY2026 10-K, share count from the FY2009 10-K
  • The printed labels are Japanese string values and are left unchanged so the script behaves identically on both language versions: 会計年度 = fiscal year, 売上高 = revenue, 研究開発費 = R&D expense, 十億ドル = USD billion, 倍 = times.
# Recompute R&D intensity and related figures from NVIDIA's annual reports (10-K)
# Python 3.9+, standard library only. Units: USD millions (share price in USD, share count in shares)

# Fiscal year: (revenue, R&D expense). FY2008-2010 from the FY2010 10-K, FY2024-2026 from the FY2026 10-K
records = {
    "FY2008": (4097.860, 691.6),
    "FY2009": (3424.859, 855.9),
    "FY2010": (3326.445, 908.9),
    "FY2024": (60922, 8675),
    "FY2025": (130497, 12914),
    "FY2026": (215938, 18497),
}

print("会計年度  売上高(百万ドル)  研究開発費  R&D比率")
for fy, (revenue, rnd) in records.items():
    print(f"{fy}  {revenue:>14,.1f}  {rnd:>10,.1f}  {rnd / revenue:>6.1%}")

# Rough floor for market value: FY2009 Q4 low price ($5.75) x shares outstanding on March 6, 2009
low_price = 5.75
shares = 542_453_547
print(f"\nFY2009 の最安値ベースの時価総額(理論値): {low_price * shares / 1e9:.2f} 十億ドル")

# Compare with his recollections (NTU 2023: 'just above a billion', Lex Fridman 2026: 'like one and a half billion')
for label, recalled in [("NTU 2023", 1.0), ("Lex 2026", 1.5)]:
    print(f"{label} の回想 {recalled:.1f} 十億ドル に対する倍率: {low_price * shares / 1e9 / recalled:.1f} 倍")

# FY2026 revenue mix
total = 215938
for name, value in [("Networking", 31376), ("Gaming", 16042), ("Automotive", 2349)]:
    print(f"FY2026 {name:<10} {value:>7,} 百万ドル  売上の {value / total:.1%}")
print(f"FY2026 研究開発に従事する従業員の比率: {31000 / 42000:.1%}")

# Time needed for a weekly 30-minute one-on-one with each of 60 direct reports
print(f"60人 × 30分 = 週 {60 * 30 / 60:.0f} 時間")

Output (verbatim):

会計年度  売上高(百万ドル)  研究開発費  R&D比率
FY2008         4,097.9       691.6   16.9%
FY2009         3,424.9       855.9   25.0%
FY2010         3,326.4       908.9   27.3%
FY2024        60,922.0     8,675.0   14.2%
FY2025       130,497.0    12,914.0    9.9%
FY2026       215,938.0    18,497.0    8.6%

FY2009 の最安値ベースの時価総額(理論値): 3.12 十億ドル
NTU 2023 の回想 1.0 十億ドル に対する倍率: 3.1 倍
Lex 2026 の回想 1.5 十億ドル に対する倍率: 2.1 倍
FY2026 Networking  31,376 百万ドル  売上の 14.5%
FY2026 Gaming      16,042 百万ドル  売上の 7.4%
FY2026 Automotive   2,349 百万ドル  売上の 1.1%
FY2026 研究開発に従事する従業員の比率: 73.8%
60人 × 30分 = 週 30 時間

Three things stand out.

  1. R&D went up while revenue went down. From FY2008 to FY2010 revenue fell from about USD 4.1 billion to about USD 3.3 billion, while R&D rose from about USD 0.69 billion to about USD 0.91 billion. R&D intensity rose from 16.9% to 27.3%. NVIDIA posted a net loss in both FY2009 and FY2010.
  2. Not all of it can be blamed on CUDA. The same period saw the 2008 global recession and a warranty charge for defective notebook chips (section 12). According to the 10-K, gross margin in FY2009 was 34.3%, of which about 6.0 points of decline came from the warranty charge. R&D also covers work other than CUDA, so the rise in R&D intensity cannot be attributed to CUDA specifically.
  3. At today's scale the ratio is lower. FY2026 R&D was about USD 18.5 billion, about 20 times the FY2010 amount, but revenue grew so much that the ratio is 8.6%. According to the 10-K, about 31,000 of roughly 42,000 employees at the end of FY2026 (73.8%) worked in R&D.

7.4 Where his memory and the record diverge

This is the point I most want to make in this section. Here is how he has described the hard years of CUDA.

StatementDate and venueThe record
"Our market cap hovered just above a billion dollars"May 2023, National Taiwan University commencement (as quoted by NVIDIA's blog)The FY2009 low ($5.75, as reported at the time) times shares outstanding (about 542 million) is about USD 3.1 billion (theoretical): about 3.1 times his recollection
"our market cap went down to like one and a half billion dollars"March 2026, Lex Fridman podcastSame calculation: about 2.1 times his recollection
"we were a 35% gross margin company"SameGross margin was 42.4% in FY2007, before CUDA was announced; 45.6% in FY2008, after it; 34.3% in FY2009 and 35.4% in FY2010. "35%" matches the FY2009–10 level after the recession and the warranty charge, not the level before CUDA

The market-value figure is theoretical: it multiplies a quarter's lowest sale price by a share count from a different date (March 2009). Because it uses the lowest price, it is a floor for the market value in that period. I expect the error from the share-count date to be a few percent at most, which does not close the gap with "just above a billion" or "one and a half billion". He may be mixing in a figure from another period (for example, before CUDA), but I did not check that. In the same Lex Fridman answer he recalled NVIDIA's pre-CUDA value as "like $8 billion or something? Like six, $7 billion" — another recollection I have not checked against the record.

My view: this is not a question of his honesty. A successful executive's memories get compressed in a dramatic direction. The lesson for readers is not to use "quotes" or "legends" as the basis for decisions as they stand. Checked against the contemporary record, the outline of the story (he kept going even at the cost of profit) holds up, while the detailed numbers do not.

7.5 The role of luck and competitors

Some of the reasons CUDA paid off were outside NVIDIA's control.

  • A discovery by outside researchers: AlexNet was the work of University of Toronto researchers using gaming GPUs. NVIDIA did not plan it. He himself reportedly said in the speech that researchers found CUDA (per the third-party transcription).
  • A rival standard came later: OpenCL 1.0, an open, royalty-free standard for parallel programming, was publicly released on December 9, 2008 (Khronos Group announcement), about two years after CUDA was announced.
  • A competitor changed course: in December 2009 Intel said that the first product from its "Larrabee" GPU project would not be launched as a standalone discrete graphics product (Intel's statement as reported by bit-tech).

None of this denies the value of his decisions. But a one-line explanation such as "he had foresight" leaves these factors out.

8. Knowing when to leave: the phone market

The third "failure" he tells is about phones. According to NVIDIA's blog, in the National Taiwan University speech he said NVIDIA entered the mobile-phone market in 2010 and retreated when the market quickly commoditized.

The record points to an earlier start. The FY2009 10-K (February 2008 to January 2009) says NVIDIA's Tegra and GoForce products "can be found primarily in multimedia cellular phones and other handheld devices", and that "During fiscal year 2009, we launched the NVIDIA Tegra APX 2500", a chip designed for mobile phones. So NVIDIA was already selling mobile-phone chips in 2008. "2010" is best read as the timing of the full push rather than the year of entry. Like section 7.4, this is a case where his memory and the record diverge.

"Our strategic retreat paid off."

"By leaving the phone market, we opened our minds to invent a new one."

The milestone of the exit in the record is the announcement of May 5, 2015: NVIDIA said it would wind down its Icera modem operations in the second quarter of fiscal 2016.

By his account, the capacity freed by the exit went into computers for robots and self-driving cars. How far has that "new market" grown? In the FY2026 10-K, automotive revenue is about USD 2.3 billion, 1.1% of total revenue. That is broadly consistent with what he reportedly said in the speech, per the transcription, that automotive and robotics had become a business of billions of dollars. But the company grew on data center, not automotive.

My view: the exit itself makes sense as a way to concentrate resources. Linking it directly to the AI success ("they left phones, so they won in AI") goes too far. No record shows a causal link between leaving phones and the growth of the data center business.

9. Organisation: how many direct reports?

The numbers disagree

His large number of direct reports is often cited as a hallmark of his organisation. But the figure differs by source.

WhenCountWho said itSource
October 202340+The hosts ("we've heard")Acquired transcript
2024 (GSB page dated April 25)50The interviewerStanford GSB transcript
March 202660Huang himselfLex Fridman transcript

The only figure confirmed in his own words is this one, from March 2026. When Fridman checked ("So you still have 60 reports"), he answered "More, yeah."

"My direct staff is 60 people. You know, I don't have one-on-ones with 'em because it's impossible."

Other numbers appear in press coverage, but I could not confirm their original sources, so they are not listed here. Any of these figures can change over time.

Why so many direct reports

At GSB in 2024 he explained:

"CEOs should have the most of the reports by definition because the people that reports to the CEO requires the least amount of management."

And on information and power:

"I don't believe in a culture, in an environment where the information that you possess is the reason why you have power."

He says he dislikes how information gets distorted at each layer, and that he explains situations in front of everyone and shows his reasoning. On Acquired he said NVIDIA is built not like a military chain of command but "much more like a computing stack", and that the company has a phrase, "mission is the boss".

The trade-offs (my view)

Arithmetic alone explains the absence of one-on-ones. A 30-minute weekly one-on-one with each of 60 people would take 30 hours a week (the last line of the code above). With that many direct reports, the only option is to drop one-on-ones and work problems as a group.

This structure has both strengths and weaknesses.

StrengthsWeaknesses
Information skips middle management, so decisions are fastThe CEO becomes the single point where information and decisions converge
Everyone shares the same contextFewer occasions for individual development and evaluation
Power is less tied to who holds what informationHard for a successor to carry on the same way

The CHM description also quotes him saying "I'm very hands-on" and, in the same sentence, "I tend not to be very operational", adding that he focuses on two areas, creating the future and solving the company's problems. For him, having many direct reports and staying out of day-to-day operations go together. Seen from outside, though, the keystone of the organisation sat largely inside the CEO's own head. That leads to section 14.

10. R&D and management philosophy

Thinking from first principles

At GSB he repeatedly stressed first principles, meaning rethinking a problem from its premises.

"You can learn how something can be done and then go back to first principles and ask yourself, 'Given the conditions today, given my motivation, given the instruments, the tools, given how things have changed, how would I redo this?'"

He applies the same idea to organisations: instead of copying other companies' org charts, design the organisation from what the company builds.

Pain and suffering

At the 2024 Economic Summit of the Stanford Institute for Economic Policy Research (SIEPR), asked for advice to students who want to start companies, he spoke about having low expectations and high resilience (SIEPR's summary). Then he said:

"For all of you Stanford students, I wish upon you ample doses of pain and suffering."

On Acquired in 2023, asked whether he would start a company again if he were 30, he answered:

"I wouldn't do it."

His reason: building NVIDIA turned out to be "a million times harder" than he expected. He added that an entrepreneur's superpower is not knowing how hard it will be.

No task is beneath me

"To me, no task is beneath me because, remember, I used to be a dishwasher"

(Stanford GSB, 2024)

My view: read on their own, these lines sound like pep talk. Set next to the decisions in sections 6 to 8, they show the same pattern: admit the mistake (NV1), put a long-term hypothesis ahead of near-term profit (CUDA), leave where you cannot win (phones). I read "pain" not as a goal but as the side effect of decisions like these.

11. Original analysis: separating what was knowable from hindsight

Successful decisions look inevitable afterwards. For six key decisions, the table below separates what was available at the time from what was learned later. The "information at the time" column is my own summary, based on the sources in each section.

DecisionInformation available at the timeAlternativeOutcomeAssessment without hindsight (my view)
1. Drop the proprietary approach and build RIVA 128 around DirectX (mid-1990s to 1997)Microsoft's standard was spreading; by his recollection there were 30 competitorsKeep developing the proprietary design for SegaThe company recovered on RIVA 128 (his account)That the standard would win was readable at the time. Whether RIVA 128 would sell was not
2. Put CUDA on consumer GeForce too (2006–07)Gaming GPUs sold millions of units a year (his account); demand for general-purpose computing was visible in areas such as scientific computingShip CUDA only in workstation and supercomputer productsShort term: pressure on profit (net losses in FY2009–10, alongside the recession and the warranty charge). Long term: the base for developersThe install-base logic made sense at the time. Its largest use, deep learning, could not be foreseen
3. The full push into phones, and the exit (2010 by his recollection, though the FY2009 10-K already records the launch of the Tegra APX 2500 for mobile phones; Icera wind-down announced 2015)Smartphones were booming; modem technology was becoming the key to competingKeep fighting for shareExit. Automotive is 1.1% of revenue in FY2026Entering was reasonable given the growth; leaving was reasonable as a way to concentrate resources. No causal link from the exit to the AI success can be shown
4. Acquire Mellanox (completed April 2020, transaction value USD 7 billion)In data centers, communication between GPUs was becoming a performance constraintLeave networking to othersNetworking revenue about USD 31.4 billion in FY2026 (14.5% of revenue)The direction could be argued at the time. But networking revenue also includes NVLink-related products, so not all of it is the acquisition's doing
5. Acquire Arm (agreed September 2020, abandoned February 2022)A semiconductor megadeal faced a clear risk of review by competition authorities in several countriesRely on a licence aloneAbandoned over "significant regulatory challenges". The USD 1.25 billion prepayment was not returned, and NVIDIA said it would record a USD 1.36 billion charge. It kept its 20-year Arm licenceThe regulatory risk was knowable in advance. The outcome should be recorded as a failed bet
6. Supply the H20 to China (export licence required from April 2025)U.S. export controls had been tightening step by stepDo not build China-specific productsA USD 4.5 billion charge in Q1 FY2026 (excess inventory and purchase obligations)The direction of tightening was readable; its timing and scope were not. An example of policy risk deciding an outcome from outside technology management

Three conclusions follow (my view).

  1. Even for the bets that paid off, what was readable at the time was the direction. CUDA read the direction (demand for general-purpose computing); its largest use came from outside.
  2. For many of the bets that failed, the risk itself was visible from the start. Arm's regulatory risk and the export-control risk could be known in advance. He presumably bet anyway because the payoff if it worked was large. Judging decisions by outcome alone is unfair, but the failures should be recorded with the same weight as the successes.
  3. The pattern is consistent. "Go for the standard and the long-term platform" and "leave where you cannot win" run through the hits and the misses alike.

12. Counterexamples and criticism: what did not work

An article about a person that contains only success stories drifts away from the real person. Here is what did not work, as recorded.

  • Defective notebook chips (2008): according to the FY2009 10-K, in July 2008 NVIDIA recorded a USD 196.0 million charge to cover warranty, repair, return, replacement and other costs arising from a weak die/packaging material set in certain versions of its previous-generation MCP and GPU products used in notebooks.
  • Leaving the phone market (2015): as above. He calls it a "strategic retreat", but the company did not win the market it entered.
  • Disclosure about crypto mining (2022): on May 6, 2022 the U.S. Securities and Exchange Commission (SEC) announced charges that NVIDIA had not adequately disclosed, in consecutive quarters of fiscal 2018, the impact of cryptomining on its gaming business. Without admitting or denying the findings, NVIDIA agreed to a cease-and-desist order and a USD 5.5 million penalty.
  • The abandoned Arm deal (2022): see section 11.
  • The export-control charge (2025): see section 11.

Criticism of founder dependence: NVIDIA's 10-K listed as a risk factor that it is "highly dependent on the services of our longstanding executive team" and that failure to ensure effective succession planning could hinder its strategy. The risk factor does not name Huang. NVIDIA also had no board chair until November 2026.

13. Avoiding "he succeeded because he is great": breaking down the factors

"NVIDIA succeeded because Huang was an excellent executive" reasons backwards from the outcome: the evidence of excellence is the success, and the reason for the success is the excellence. To avoid that loop, the table separates his decisions from everything else (my breakdown).

FactorWhat it wasHow much his decisions mattered
Market structureA huge consumer market, PC gaming, whose volume paid for GPU developmentLow to medium (the market was external; betting on it was a decision)
Developer ecosystemCUDA shipped even on consumer GPUs and was pushed into universities and researchHigh
Technical constraintsThe large matrix computations at the heart of AI suit GPU parallelism better than CPU sequential processing (see the pillar article for how)Low (the nature of the computation is external)
CompetitorsThe rival standard OpenCL came about two years after CUDA; Intel shelved Larrabee as a standalone productLow (other companies' circumstances)
Luck and timingAlexNet happened to be trained on gaming GPUs anyone could buyLow (but having the devices in the field was a decision)
Exit and focusLeaving markets it could not win and moving resourcesHigh

Survivorship bias needs attention too. By his recollection, there were about 30 3D graphics companies in the mid-1990s. Most of them are gone. Some of the founders who failed surely also believed in admitting mistakes and betting on long-term platforms. You cannot look only at the qualities of the one who succeeded and conclude that they caused the success.

He said as much himself in March 2026:

"It wasn't three people that made CUDA successful. It was 43,000 people that made CUDA successful."

14. Dependence on one person, and the resignation of November 2026

On November 11, 2026, NVIDIA's board accepted Huang's resignation as President, CEO and director, and the company announced it the next day. NVIDIA has not given a reason. CFO Colette Kress became interim CEO while remaining CFO, and Harvey Jones became board chair. NVIDIA's announcement ("NVIDIA Announces CEO Transition", November 12, 2026) calls him the company's "first-ever Chair of the Board". The sequence of events, the line between reporting and confirmed fact, and the governance questions are covered in Jensen Huang's resignation as NVIDIA CEO — separating confirmed facts, reporting and his own words.

In the context of this article, the point is that the organisational weakness described in section 9 is no longer hypothetical. In an organisation where information and decisions converge on the CEO, that CEO suddenly left. According to statements from the board and the congressional committees, the internal and congressional inquiries are continuing, and as of November 17, 2026 there are no conclusions (see the article linked above for details). Nor is it known whether the next CEO will keep the 60-direct-reports, no-one-on-ones model or rebuild the organisation.

15. How to treat the "quotes" that circulate

Search for Jensen Huang quotes and you will find many. Here are the ones whose original wording, date and venue I confirmed.

Statement (original)DateVenueSource checked
"Run. Don't walk." / "you are running for food, or you are running from becoming food"May 2023 (ceremony on May 27, Taiwan time)National Taiwan University commencementNVIDIA blog (2023-05-26)
"Confronting our mistake and, with humility, asking for help saved NVIDIA."SameSameSame
"Our strategic retreat paid off."SameSameSame
"I wouldn't do it."Published 2023-10-16AcquiredAcquired transcript
"I wish upon you ample doses of pain and suffering."Article dated 2024-03-07SIEPR Economic SummitSIEPR article
"no task is beneath me"Page dated 2024-04-25Stanford GSB View From The TopGSB transcript
"My direct staff is 60 people."Published 2026-03-23Lex Fridman Podcast #494Official transcript
"It was 43,000 people that made CUDA successful."SameSameSame

There are also lines I could not confirm. The line about the company being "thirty days from going out of business" is widely quoted, but I could not confirm its original wording and venue in a transcript published by a host or by NVIDIA, so this article does not use it.

Before quoting him, check three things.

  1. Is there an original? Lines that circulate only in translation can drift from what he said in English.
  2. Are the date and venue known? The same idea can carry a different meaning in a different setting.
  3. Who published the record? A transcript published by the host or the company is more reliable than a third party's compilation.

16. Lessons for technology leadership

Finally, here is what can be taken from his case into technology choices and business decisions, written as decision criteria (my view).

  1. Choose deliberately between following a standard and setting one. RIVA 128 survived by following a standard; CUDA won by setting one. Insisting on a proprietary approach when you are not in a position to set the standard carries the same danger as NV1.
  2. Do not count on a platform's biggest use. CUDA did not foresee deep learning. A platform's value lies in being usable for things you cannot predict. When deciding on such an investment, look for demand that does not depend on a single use.
  3. Set exit criteria in advance. The exit from phones came from judging that NVIDIA lacked what it took to win there (modem technology). Deciding at entry "what must be in place, or we leave" keeps exits from coming too late.
  4. Trust the record over the story. A founder's recollection can have the right outline and the wrong numbers. When you borrow from another company's success story, check it against contemporary filings.
  5. Pair a design that concentrates on one person with a design for handing over. An organisation with few layers is fast, but it can stall when its keystone leaves. If you choose speed, you need a succession plan to match.

Summary

  • Jensen Huang was born in Taiwan, moved through Thailand, a Kentucky boarding school and Oregon, and studied electrical engineering at Oregon State University and Stanford (master's, 1992). He designed chips at AMD and LSI Logic and co-founded NVIDIA in April 1993.
  • He admitted that the first product line had failed and rebuilt around the standard, kept CUDA on consumer GPUs even at the cost of profit, and left a phone market he could not win. Admit, persist, leave: that was his decision pattern.
  • CUDA was first aimed at scientific computing; deep learning was opened up by outside researchers using gaming GPUs. The filings support the story of investing through hard years, but the market value and gross margin he recalls for that period do not match the record.
  • An organisation with many direct reports, no one-on-ones and shared information traded speed for heavier dependence on the CEO. With his resignation in November 2026, that dependence stopped being a hypothetical risk (my view).
  • The outcome should be explained not by his qualities alone but by market structure, the developer ecosystem, competitors' delays, and luck and timing. Failed bets, such as the abandoned Arm deal and the export-control charge, are on the record with the same weight.

The company's own path, including how its revenue mix changed, is traced with figures from its filings in A history of NVIDIA.

Frequently asked questions

What is Jensen Huang's career in brief?
He was born in Taiwan and moved to the United States as a boy. He holds a bachelor's degree in electrical engineering from Oregon State University and a master's in electrical engineering from Stanford (1992). According to NVIDIA's annual report, he was a microprocessor designer at AMD from 1983 to 1985 and worked at LSI Logic from 1985 to 1993. In April 1993 he co-founded NVIDIA with Chris Malachowsky and Curtis Priem and served as CEO. He resigned as CEO and director on November 11, 2026.
Where was Jensen Huang born, and how did he come to the United States?
According to the description of the oral history the Computer History Museum recorded in 2024, he was born in Taiwan, moved to Thailand, where his father was starting an oil refinery, and then attended a boarding school in Oneida, Kentucky. He went to college in Oregon. Many encyclopedias give Tainan as his birthplace, but the company, university and museum sources I could check say only "Taiwan". Sources also disagree on his age at each move.
Is it true that he has 60 direct reports and holds no one-on-ones?
In his own words, yes: on the Lex Fridman podcast published in March 2026 he said "My direct staff is 60 people" and that he does not hold one-on-ones because it is impossible; when the host checked the figure, he added "More, yeah." On Acquired in 2023 the hosts put the figure at "40+", and at Stanford GSB in 2024 the interviewer said "50", so the number varies by date and speaker. He says he works problems in front of everyone to remove layers of information.
Are the Jensen Huang quotes that circulate online really his words?
Not all of them can be verified. This article quotes only words whose original wording, date and venue I could confirm in a transcript or article published by the host or by NVIDIA. "Run, don't walk" (National Taiwan University commencement, 2023) and "pain and suffering" (Stanford, 2024) check out. The line about the company being "thirty days from going out of business" does not: I could not confirm its original wording and venue, so it is left out.

References

友田

友田 陽大

Developer of a METI Minister's Award–winning product. With TypeScript + Python + AWS, I deliver SaaS, industry DX, and production-grade generative AI (RAG) end to end — from requirements to infrastructure and operations — single-handedly.

Do you actually know whether self-hosting beats the API?

Self-hosted LLM technology selection — quantization, GPU budget, commercial licensing

When to use AWQ / GPTQ / FP8 / GGUF, VRAM estimates that include the KV cache, the break-even point against per-token API pricing, and the license terms of Apache-2.0 / Llama / Qwen / Gemma. From hands-on experience adopting commercially usable open quantized models stage by stage against license, quality, and speed trade-offs, I help you assemble the decision criteria before you implement.

Available for both project-based (contract) and advisory engagements. Start with a free 30-minute consult.

Also worth reading