Where Nvidia Goes Without Jensen Huang
Jensen Huang, the co-founder and chief executive of Nvidia, stepped down on Nov. 12, after 33 years leading the company. The board had accepted his resignation the day before, and chief financial officer Colette Kress took over as interim chief executive. Harvey C. Jones became the board’s first chair. Now there is one question. Where does Nvidia go without Huang?
The background to the resignation is already known. The Wall Street Journal reported that the board was preparing to dismiss Huang, and The New York Times named Kimberly Powell, a vice president for health care, as a second affair partner. President Donald Trump said the Federal Bureau of Investigation was looking into Huang’s relationship with Alice Huang. A shareholder class action was also filed. The board opened an independent review on Nov. 8 and on Nov. 10 retained Joon H. Kim, a former acting United States attorney in Manhattan, and the law firm Wachtell, Lipton, Rosen and Katz. The Nov. 12 statement gave no reason for the resignation and did not say whether the review had concluded.
The market reacted first. Nvidia shares closed Nov. 12 at $166.19, down 4.34 percent. They had fallen as low as $151.01 during the session before recovering. By around midday on Nov. 13 they had fallen further, to $160.36. Market value, about $5.8 trillion in early October, had shrunk to about $3.9 trillion intraday on Nov. 13. Apple was again the world’s most valuable company, and Nvidia had slipped to second. On the same day AMD rose 5.86 percent and Broadcom rose 5.54 percent. Money did not leave artificial intelligence. It moved.
The moat is a stack, not a chip
To answer the question, start with the machine that makes Nvidia strong. The core is CUDA. CUDA is a software platform Nvidia introduced in 2006 that turned the graphics processing unit, a chip built for games, into a general-purpose computer. Training artificial intelligence is the repeated multiplication of huge matrices, and the thousands of cores on a GPU fit that work exactly. CUDA lets programmers run code written in ordinary programming languages across those thousands of cores in parallel.
CUDA’s strength is not chip performance but the ecosystem. Libraries like cuDNN and cuBLAS, development tools, and artificial intelligence frameworks like PyTorch and TensorFlow all run on CUDA. By Nvidia’s own count, more than 6 million developers use CUDA. Twenty years of code is tied to Nvidia chips. That is why developers do not move even when a rival matches the hardware.
The moat is not CUDA alone. Nvidia sells the chip, the software, the network and the whole data center as one package. At the center of that package are the NVL72 systems, which bind 72 GPUs into a single rack, and the Vera Rubin platform, whose shipments began in August 2026. The company expects Vera Rubin to account for about 20 percent of third-quarter data-center revenue. There is the network, too. The 2019 acquisition of Mellanox for about $6.9 billion brought InfiniBand and high-speed Ethernet in house, and networking revenue reached $31.4 billion in fiscal 2026. There is inference software as well. TensorRT-LLM runs only on Nvidia GPUs, and the fastest path through the open-source Dynamo framework runs on top of it.
The supply chain is another layer of the moat. Nvidia chips use packaging from Taiwan Semiconductor Manufacturing Co. and high-bandwidth memory from SK Hynix, Samsung and Micron. Both were reported sold out for 2026. One estimate puts Nvidia at about 60 percent of TSMC’s CoWoS packaging capacity, against about 11 percent for AMD. The exact split has never been disclosed, but the direction is clear. The company that secures the most of a scarce supply is Nvidia.
The team stays, the chair is empty
Huang’s departure does not mean the technology left. Ian Buck, who created CUDA, is still there. Buck joined Nvidia in 2004 and now leads its hyperscale and high-performance computing business. Co-founder Chris Malachowsky is still there. So are Jonah Alben, who leads GPU engineering, Debora Shoquist, who runs operations, and Kress, who has been chief financial officer since 2013. Reuters put it this way. Nvidia has lost the man who helped create the artificial intelligence revolution, but it has not lost the technology driving it.
The problem is the next chief executive. Nvidia has given no timeline for the search and no shortlist. It has not been confirmed whether Kress is a candidate for the permanent job. Reuters named Buck and Kress as the prominent internal candidates. A co-chief-executive arrangement has also been discussed, but that is analysts’ speculation, not a board plan. Outside names are circulating. Rene Haas, chief executive of Arm, and Lisa Su, chief executive of AMD, have been mentioned, and President Trump publicly suggested Tim Cook of Apple. None has been confirmed by Nvidia as a candidate. Prediction markets put Buck in front, but that is betting, not a board decision.
What depended on Huang is separate. First, the long-horizon bets. The CUDA bet of 2006 and the roadmap that puts out a new chip every year were like that. Shay Boloor, chief market strategist at Futurum Equities, said the worry is not the next quarter but the next decade, and that Huang was not simply a chief executive but the person who consistently identified where computing was going. Second, capital allocation. Huang signaled an intent to invest up to $100 billion in OpenAI, but Kress and her team clashed with Huang over such large financing deals, Reuters reported. A $250 billion plan to support an OpenAI data-center project in Ohio was cut to $120 billion. Third, customers and Washington. Reuters described Huang as the gatekeeper to scarce chip supplies. He also stayed close to Washington to preserve access to the Chinese market.
So the verdict splits two ways. Stacy Rasgon, a senior analyst at Bernstein, cut the price target in half, to $200 from $400, saying Jensen is Nvidia. Barron’s, by contrast, argued for buying the stock, saying that when the bad news is all out, when the investigations are all done, and when an inevitable new leader is named, Nvidia will go back to making chips. Chris Caso of Wolfe Research said the effect would not matter in the long term because the product lineup was already ready.
The challengers are closing in
The rivals are not waiting for the empty chair. AMD is the closest. Its data-center revenue reached $8.55 billion in the third quarter, up 97 percent from a year earlier. With its new MI455X chip and its Helios rack system, it has signed gigawatt-scale supply deals with OpenAI, Meta and Anthropic. Only the first gigawatt in each deal is binding, and the rest is intent. AMD does not break out Instinct GPU revenue, so the exact size is unknown. In MLPerf benchmarks, the MI355X matched or beat Nvidia’s B200 on the workloads AMD entered. Those were inference workloads only, and AMD did not enter reasoning or vision workloads.
Anthropic’s claim needs care. Anthropic said it got its model running on AMD’s MI355X in three days, but that covered prompt processing for a single model, and there is no independent verification. Anthropic’s securities filing does not repeat the figure.
Broadcom comes by a different road. Broadcom does not sell graphics processors. It co-designs custom accelerators for each customer. Google’s TPU, OpenAI’s Jalapeño and Meta’s MTIA are examples. Broadcom’s artificial intelligence semiconductor revenue reached $16.7 billion in the third quarter of fiscal 2026, up 221 percent from a year earlier. The company’s forecast puts fiscal 2027 artificial intelligence revenue at about $115 billion. Its Tomahawk networking chips compete with Nvidia’s Spectrum-X.
The big technology companies’ own chips are growing, too. Google agreed to supply Anthropic with up to one million Ironwood TPUs, and Amazon said Anthropic uses more than one million of its Trainium2 chips. Microsoft’s Maia 200 runs in two Azure regions, and Meta uses hundreds of thousands of its MTIA chips for inference. But there is no evidence that any company has replaced Nvidia at training scale. No case has been confirmed of a flagship model trained mainly on non-Nvidia hardware.
The other challengers are still far behind. Intel has effectively withdrawn. It canceled Falcon Shores, and Gaudi missed its targets. Cerebras is public and has a 750-megawatt deal with OpenAI, but its quarterly net loss reached $450 million and 86 percent of its projected revenue comes from two customers. Groq licensed its technology to Nvidia in December 2025, and its founder and other key people moved to Nvidia. Nvidia’s annual filing puts the deal at about $17 billion. Huawei’s Ascend works only inside China and is constrained by high-bandwidth memory shortages. A single chip’s performance is estimated at one-third to 60 percent of Nvidia’s latest.
Cracks in the moat have begun to show. First, customer diversification. OpenAI, Meta and Anthropic have signed gigawatt-scale deals with AMD, and those deals carry warrants to buy AMD stock. Second, open-source portability. Moving code inside frameworks like PyTorch and vLLM is getting easier. Some assessments put ROCm at 90 to 95 percent of CUDA performance on standard work. Third, inference. The inference moat is thinner than the training moat, and Nvidia’s licensing of Groq’s technology is itself evidence of that. But no layer has collapsed yet. Shipping silicon for the open UALink standard lags the specification by 12 to 24 months, and AMD’s Helios volume is scheduled for the first half of 2027. Nvidia’s Vera Rubin has been shipping since August 2026.
What comes next
The next test is earnings on Nov. 17, the first under Kress as interim chief executive. The company expects third-quarter revenue of about $108 billion and Vera Rubin to account for about 20 percent of data-center revenue. The actual figures arrive on Nov. 17. Options markets are pricing in a move of about 10.6 percent in the shares around the report. On Nov. 18, Kress is scheduled to meet employees to discuss next steps.
The political and legal calendar remains. House committees have asked Huang to sit for a closed-door interview on Dec. 1 and said they would issue a subpoena if he did not agree by 8 p.m. on Friday. Tim Teter, Nvidia’s general counsel, is called for Dec. 2, and Huang for a public hearing of the Senate Banking Committee on Dec. 3. The chairman of the House Foreign Affairs Committee said Huang had agreed to brief lawmakers before Thanksgiving. Whether Huang actually appears has not been confirmed. The shareholder class action has not been tested in court.
In the end the question narrows to one. Who designs Nvidia’s next decade? The next quarter and the one after are already set. Analysts see the roadmap as fixed for 12 to 18 months. Vera Rubin is in production, and earnings arrive on Nov. 17. The problem is what follows. Whether Buck, who created CUDA, leads the company, whether Kress steadies the finances and buys time, or whether a new face arrives from outside, the board has not said. What Huang built was not a one-man company but a moat built layer by layer. Whether that moat holds for another decade is now the job of the people who inherited it.
