needhelp
← Back to blog

Thirty Days, Thirty-Three Years: Jensen Huang Leaves Nvidia

by xingwangzhe
nvidiajensen-huangcudagpuai-infrastructure

On the afternoon of November 12, 2026, Nvidia’s newsroom published a statement under three hundred words. Jensen Huang had resigned as president, chief executive, and board member, effective immediately.

No transition period. No “will step down on such-and-such a date.” No advisory title that founders usually keep.

The stock fell more than 10% pre-market and closed down 4.34%, wiping out roughly $182 billion in market value. The board moved the same day: Colette Kress, executive vice president and CFO, became interim CEO while keeping the CFO role; Harvey C. Jones, former CEO and executive chairman of Synopsys, became the first chair of Nvidia’s board in its 33-year history. The company had never had that position before.

The announcement did not give a reason. It said operations, customer commitments, and the product roadmap would continue as planned, and that Q3 results would still land on November 17.

An obituary would be easy to write and not very useful. The better question is which of the last 33 years’ decisions were Huang’s personal judgment and which have already become organizational muscle memory. The earnings call on November 17 tests the second category.

The booth at Denny’s

April 5, 1993, San Jose, California, a booth at a Denny’s. Three people: Huang, recently out of LSI Logic, and Chris Malachowsky and Curtis Priem, both from Sun Microsystems. Seed money: $40,000. Huang was 30.

The name came from invidia, Latin for envy. Early files used an NV prefix, for next version. They believed something specific: the PC would turn from an office machine into a consumer device, and the first thing a consumer device needs is pictures.

That belief was not strange in 1993. Two dozen graphics chip companies existed that year; three years later there were seventy. By 2006, one independent was left.

What was strange was never the belief. It was the size of the bet.

The first wrong bet: NV1 and quadratic surfaces

The 1995 NV1 was a textbook miss, and it missed on geometry.

Polygon rendering ran on triangles, and Microsoft’s DirectX was turning that into the de facto standard. NV1 chose quadratic texture mapping — building objects out of curved patches, which in theory saved memory bandwidth, smoothed edges, and avoided the stair-stepping of scalar triangles.

Sega ported the arcade title Virtual Fighter onto it, the first 3D game to run on an Nvidia chip. The card shipped at retail as the Diamond Edge 3D.

The problem was not compute. It was the toolchain. Modeling software, art assets, render engines, and the muscle memory of every programmer had grown around triangles. A card that could only draw curves asked the entire industry to redo its workflow. No performance advantage compensates for that.

Sega cancelled the follow-on order. Huang flew to Japan, admitted the wrong turn to Shoichiro Irimajiri, Sega’s executive at the time, and asked him to leave the money already paid on the table. Irimajiri converted the order into a $5 million investment. It bought roughly six more months.

In 1996 he laid off more than half the staff and put the remaining cash behind a design that had not been taped out yet.

“Our company is thirty days from going out of business.” He said it at all-hands meetings for the next decade. It was not a metaphor. It was the cash runway.

The second bet: RIVA 128 and the standard

RIVA 128 landed in 1997 as the result of that gamble. The world’s first 128-bit 3D processor, no curves, triangles only, fully committed to Direct3D. A million units in four months.

At launch, the bank balance covered one month of payroll. That detail gets retold as an underdog story, but the point is usually missed: the cost of admitting a wrong technical direction has to be smaller than the cash on hand. What separated NV1 from RIVA 128 was not genius. It was a public, expensive admission of error — including to the customer who had already paid.

Most companies die here. The direction is wrong, but the organization has no vocabulary for admitting it, so resources keep flowing into the sunk cost.

Inventing a word: the GPU in 1999

January 1999: Nvidia IPO’d on Nasdaq at $12 a share.

October of the same year: GeForce 256 shipped. Nvidia coined a term for it — GPU, graphics processing unit — and gave it a testable definition: a single-chip processor with integrated transform, lighting, triangle setup and clipping, and rendering engines, capable of at least ten million polygons per second.

The substance was hardware T&L. Until then, vertex transform and lighting were the CPU’s job, and the graphics card only filled in triangles someone else had computed. Pulling T&L onto the board gave the card its own workload for the first time. It stopped being a canvas.

Nvidia’s own timeline offers the comparison: modern GPUs process more than seven billion polygons per second. Twenty-seven years, three orders of magnitude.

The word later escaped graphics and became a unit of compute. “How many GPUs do we need” meant something entirely different in 2023 than it did in 1999.

Companies that compete on products ship products. Companies that compete on categories invent words. Naming a thing is how you set the scoreboard, and whoever sets the scoreboard decides what the next round is played on.

The bet nobody liked: CUDA

In 2006 the G80 architecture (GeForce 8800) brought unified shaders: pixel and vertex shaders stopped being physically separate, and every stream processor could execute general-purpose code. A G80 had 128 stream processors in eight groups of sixteen, sharing instruction issue and texture units.

Then came the most important decision in the company’s history, and it was not a product decision: expose those programmable cores to non-graphics work. CUDA shipped in November 2006.

Technically it was modest: a C-like parallel programming model. Threads are grouped into blocks, blocks into grids; 32 threads form a warp that executes one instruction per cycle and distinguishes itself only by data registers. A programmer does not need to know how many SMs the card has or how many schedulers each SM carries. They need to answer one question: how many independent pieces of data does my problem have.

That abstraction layer is the real technical content of the bet. It turns parallelism from a hardware detail into a programming model, and the price is silicon area, a driver stack, a compiler, a debugger, and twenty years of backward compatibility.

Commercially it was anything but modest: CUDA went onto every GeForce, including the ones bought by people who only wanted to play games.

Wall Street saw a cost. Why should a gaming card pay transistors and validation for a compiler nobody used? In the 2008 crash, Nvidia’s stock fell 70%.

“We were democratizing supercomputing,” Huang said later.

In 2008 that sounded like a coping mechanism. It took six years to become a statement of fact.

Six years later: two GTX 580s

In 2012, three people at the University of Toronto — Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton — trained a convolutional network on two consumer GeForce GTX 580 cards for ImageNet. Error rate: 15.3%. Second place: 26.2%.

The important technical fact that year was not that the algorithm won. It was that the hardware was already there. A decade of CUDA on retail shelves meant every lab in the world already owned a machine that could run deep learning; nobody had told them.

That is how ecosystem compounding works. You do not need to convince anyone to buy a card in 2012. You need the card to be on their desk in 2006.

It is also the key to the moat. The 92% enterprise GPU share is not maintained by one generation’s performance lead. It is maintained by the fact that when code from any paper lands on GitHub, someone else can clone it and run it. The switching cost lives in millions of lines of CUDA, not in the chip.

Carrying the box: DGX-1 and the boundary of “one machine”

In August 2016, Huang drove to San Francisco and carried the first DGX-1 into OpenAI’s office himself. The machine cost $129,000, held eight P100s, and connected them over NVLink. Elon Musk was in the room.

A CEO of a company worth tens of billions delivered a computer to a nonprofit lab that had existed for a few months, and did the lifting. The story gets told as a bit. It is actually the most accurate self-portrait the company has: what they sell is not a chip, it is a template for what compute should look like.

DGX-1’s real contribution was packaging an industry standard. Before it, every lab assembled its own nodes, tuned its own PCIe topology, solved its own power and cooling, and wrote its own distributed communication layer. After it, the procurement list became a model number. The months researchers got back were worth more than the peak FLOPS.

Follow that line forward and NVLink’s bandwidth runs from 160 GB/s in 2016 to rack-scale domains; the 2026 NVL72 puts 72 GPUs in a single NVLink domain. The engineering consequence is that the boundary of “one machine” gets redrawn. When a training job runs on one logical accelerator, model-parallel communication stops being a networking problem and becomes a memory problem.

Which is why comparing single-card FLOPS often produces the wrong answer.

Soldering matrix multiplication into silicon

Volta arrived in 2017, and V100 carried Tensor Cores: a 4×4 matrix multiply-accumulate unit doing a batch of multiply-adds per clock. At the time most frameworks still called cuBLAS in FP32, and the FP16 path required code changes to pay off.

Five years later it became decisive. Transformer’s core operators are almost all batched matrix multiplication; the QK transpose in attention and the weighted sum after softmax both land on hardware as large GEMMs. The collapse in training cost between 2020 and 2025 came half from process and packaging, half from the decision to build a dedicated unit for matrix math.

Then precision fell generation by generation: FP16 on V100, BF16 and TF32 on A100, FP8 on H100, FP4 on Blackwell. Each step did the same thing — strip numerical precision the model does not need out of silicon area and memory bandwidth.

In hindsight this was a clever hedge on the CUDA strategy: keep a general programming model outward, build dedicated hardware for the hottest kernel inward.

General interface, specialized core. That combination explains why “specialized chips will replace the GPU” is predicted every year and never lands. The challengers either lose on generality or lose on the millions of lines already written.

A failure: Tegra and the mobile decade

The 1993-to-2026 story contains more failures than successes. The failures are just quieter.

In 2008 Nvidia launched Tegra and went after phones. The first iPhone SDK had just shipped; Android had not scaled. Tegra landed in the Zune HD, a few tablets, a few car head units, and then got pushed out of the mainstream by Qualcomm and Apple’s in-house SoCs. By around 2015 the business had largely left the phone fight for automotive cockpits and autonomous driving.

Same lesson as NV1, cheaper tuition: building a faster chip inside someone else’s ecosystem does not beat the company that defines the ecosystem. NV1 lost to DirectX triangles. Tegra lost to Qualcomm’s modem integration inside the ARM ecosystem.

The branch that survived — automotive and robotics — became the entry point for embodied AI a decade later.

One trillion to five

H100 shipped in 2022. ChatGPT launched a few months later. The timing was too good to be true.

Market cap crossed $1 trillion in 2023, $3 trillion in 2024, and in October 2025 Nvidia became the first company past $5 trillion. Enterprise GPU share sits near 92%.

Luck was real. AlexNet ran on their cards; Transformer’s core operator happens to be matrix multiplication; large models happen to need thousands of cards wired into one machine. But luck only explains the peaks. It does not explain why they were on the field for every one of them.

Huang’s own framing was Huang’s Law: performance doubles every two years while cost per unit falls. That is not a rewrite of Moore’s Law; it is compressing architecture, packaging, interconnect, and software into a single cadence. The precondition is that all four layers sit inside one company, and that one person who does not have to justify long-term spending to a quarterly call signs off on them.

On September 2, 2026, in conversation with Commerce Secretary Howard Lutnick, he said: “I could obviously stop working, but I don’t want to miss it. I will not miss the next 10, 15 years.” He also said 100 gigawatts of compute infrastructure remained to be built before 2030.

Ten weeks later, he was gone.

Two maps

Nvidia’s China business is the backdrop for the inquiries that followed November 12.

In October 2022 the Commerce Department restricted A100 and H100 exports to China, and Nvidia shipped cut-down A800 and H800 parts. In October 2023 new rules wrote performance-density thresholds into the text, and those parts were listed too, followed by H20’s repeated reversals — restricted, then conditionally licensed again in 2025.

The result was a change in revenue mix: China data center revenue went from more than a fifth of the total to nearly nothing, then partially recovered. No amount of engineering fixes this, because what is being measured is not chip performance. It is a license.

Which is why the letter that arrived after November 12 focused not on compute but on who approved what, and where the records are. Richard Neal and Ron Wyden asked Commerce to preserve the relevant records. They are the minority party: they can ask, not compel. For investors, the practical meaning is that any future loosening or tightening of China licenses will hit friction in Congress first, then show up in guidance.

The AI factory invoice

In September 2026, Huang started calling data centers “AI factories” and gave a schedule: 100 gigawatts to build before 2030.

A large nuclear unit is about 1 gigawatt. That number moves the conversation from chip shipments to power and land. The same month, Nebius announced $40 billion of 2027 capex and its stock fell 17% in a day; Microsoft’s New Jersey data center project had slipped by months. The bottleneck moved from accelerators to substations.

That is also the hardest part of the succession. Huang was not only selling chips. He was selling the expectation that compute stays scarce, and that expectation needs a person to repeat it every week. An earnings report can prove last quarter. It cannot argue the next decade.

After inference takes over

Between 2024 and 2026 the shape of demand changed. Training capex stayed enormous but decelerated; inference took over — long context, agents calling tools in loops, multi-pass self-checking, all pushing tokens per request up by an order of magnitude.

The economics differ. Training rewards peak FLOPS and interconnect bandwidth. Inference rewards marginal cost per token: memory capacity decides how much KV cache fits, memory bandwidth decides decode speed, quantization decides how large a model fits in the same VRAM. Compress the KV cache fivefold and one card serves five times the concurrency — that argument persuades buyers more than another 10% of compute.

Hence Blackwell’s successors pushing to FP4, and hence KV cache compression, speculative decoding, and batching schedulers eating most of the 2026 research agenda.

For the moat, the shift cuts the wrong way. The training era rewarded having the fastest card, and whoever could buy it had the advantage. The inference era rewards the lowest cost per token, and that is ground self-designed ASICs can take — Google TPU, AWS Trainium, Meta MTIA are all pushing there. Huang held the gap open with an annual architecture cadence. Whether a second person can keep that cadence is the thing to watch after November 17.

Sixty direct reports

His management style was an outlier even by Silicon Valley standards. Around 60 direct reports. No one-on-ones. No slide decks.

Everyone sent a weekly Top 5 Things email: the five most important items and what you were thinking about. Information moved sideways through the organization instead of up the chain and back down. The practice survived past ten thousand employees, at the cost of a shredded CEO calendar.

“Our company is thirty days from going out of business” was the emotional baseline of that system. It kept a $4 trillion company in a state of tension. It also meant that in 33 years nobody grew into the role of successor.

That is not a conspiracy, it is structural: a CEO who inserts himself into every critical decision loop systematically suppresses the formation of a second decision-maker. The board is now making up that lesson.

After November 12

Kress is 59. She joined Nvidia as CFO in September 2013 after nearly 25 years in finance at Texas Instruments, Microsoft, and Cisco, and she rode the company from gaming cards to a trillion-dollar valuation. People Reuters spoke to describe her as the calm inverse of Huang’s confrontational style.

The market is not fully convinced. Anshel Sag of Moor Insights and Strategy put it plainly: putting the CFO in the interim seat keeps markets calm, but she is not the long-term visionary Nvidia needs to stay on this trajectory.

Three names circulate. Ian Buck, at Nvidia since 2004, runs AI hardware and software products and did early CUDA work. Rene Haas, now CEO of Arm, spent seven years as an Nvidia vice president and general manager. And Lisa Su, CEO of AMD, the most direct competitor — the fact that her name appears at all measures the market’s anxiety.

On November 17 after the close, Q3 FY27 results land on schedule. It will be the new leadership’s first public Q&A.

The same day, Richard Neal and Ron Wyden, the top Democrats on the House Ways and Means and Senate Finance committees, wrote to Commerce Secretary Lutnick about how Nvidia’s chip sales to China were handled, and asked the department to preserve related records. Minority party: they can ask, not compel.

The white leather jacket

In 33 years the company changed CEOs once. Now it has.

What gets tested is not the successor’s competence but a quieter question: of those bets — triangles, hardware T&L, unified shaders, CUDA, Tensor Cores, NVLink — how many were Huang’s personal judgment, and how many had already become the organization’s defaults.

CUDA is the first half of the answer. It is now more than four thousand accelerated libraries, millions of developers, and twenty years of compatibility baggage. No successor can overturn it, the way nobody can overturn x86. That is a company asset, not a personal one.

The second half is the thirty-days tension. It cannot be written into a process document. It only exists when one person keeps saying it and keeps demonstrating it with his calendar. That is a personal asset, not a company asset.

The November 12 announcement separated the two. The market priced the separation at $182 billion, then went back to waiting for November 17.

Silicon cycles in three years. Architectures cycle in ten. Ecosystems cycle in twenty. A person’s tenure was 33 years, just long enough to bet correctly on all three layers once.

The white leather jacket is still in the closet. The man who wore it is no longer in the room.

References

Share this page