The fall of Jensen Huang, an office suite for the agent age, and following the margin in AI

Jensen Huang resigned on Thursday as CEO, president and director of Nvidia, sending shockwaves through the AI industry. Colette Kress, CFO since 2013, takes over as interim chief executive. The events that removed him came to a head over the last few weeks, beginning on 29 October when the New York Times reported that US intelligence believed Alice Huang had been passing information to Beijing. She is an executive linked to the Singapore cloud firm Megaspeed and is no relation. Huang admitted an affair in an email to staff on 8 November. The board opened an inquiry the same day and said it had confidence in him. Three days later the Times reported a second relationship, with a senior executive who reported to him. Reuters has since reported a third, with another employee, citing two sources.

Jensen Huang at Stanford in April. Born in Taiwan and raised in the United States, he worked at AMD and LSI Logic before co-founding Nvidia in 1993. He ran the company for 33 years. Photo: Anderseidesvik, CC BY-SA 4.0.

The timing of that first report is best explained by a year of frustration in Washington over Nvidia's sales to China. According to the New York Times, the president was briefed on the intelligence and did not act. Huang had lobbied successfully against export controls, and the White House had cleared H200 chips for sale to China. Lawmakers are now trying to attach three chip-control measures to the defence bill, a week before Donald Trump meets Xi Jinping. Nobody has identified the source of the leak. The most plausible reading is that the patience of the hawks around Marco Rubio and Scott Bessent finally ran out.

Evidence that the controls were failing had been building for months before the scandal broke. Epoch AI estimates that between 290,000 and 1.6 million H100-equivalent chips were smuggled into China by the end of 2025. Its median figure equals about a third of the country's AI compute. The Information reported that Alice Huang helped facilitate at least $7.5 billion of Nvidia shipments to China. On the day Huang resigned, the researcher Will DePue published a forensic analysis of published model weights. It found that six recent models from DeepSeek, Moonshot, MiniMax and Ant Group were initialised on Blackwell systems, which cannot be exported to China without a licence.

Huang's commercial interests ran against the hawks on almost every question, which made him their most effective opponent. In July he led a letter from 25 companies defending open-weight models, having argued that wider use of them sells more Nvidia computers. Dario Amodei of Anthropic did not sign. He argues that frontier open-weight models carry greater risks than closed ones and that no advanced chips should go to China.

These initial revelations did not immediately bring Huang down, but the board had little room once relationships with employees were reported. Most large companies prohibit relationships between managers and their direct reports. McDonald's, Intel and BP all lost chief executives on those grounds. A rule that binds every other employee cannot be waived for the founder. Huang's email admitted one affair and left out at least two others. It is hard to believe that omission was accidental, which makes the email dishonest. A relationship with an executive at a customer under export scrutiny is also a conflict of interest of the most serious kind. If it affected what Nvidia sold or how closely it checked the buyer, that would be corruption. The US justice department has not concluded that she was a spy, so the inquiry still has to settle both questions.

Investors and the more accelerationist parts of the AI industry have objected that the board removed perhaps the person most responsible for America's lead in AI. The investor Gavin Baker wanted the directors to wait for the inquiry or hold a shareholder vote. Tim Sweeney of Epic Games compared the decision to Apple firing Steve Jobs. An internal staff post asking for Huang's return drew more than 1,600 comments. That loyalty was earned over three decades. He also had dozens of direct reports, no succession plan, a board of long-standing allies and a son-in-law running a credit programme for customers. The concentration of authority that made Nvidia fast also meant nobody was in a position to check him.

Kress arrives with a reputation that Huang himself tried to diminish, according to a Wall Street Journal account of a meeting in August. He called her "a deer in the headlights" and a "reporting CFO". The same account shows her asking how Nvidia valued the guarantees it was offering to cloud providers. She reworked a proposed $250 billion backstop for OpenAI's Ohio project, halving Nvidia's initial commitment. New deals were paused in late August while the board reviewed them. The retreat from Huang's dealmaking therefore began before the scandal.

Nvidia's products are unlikely to suffer in the near term, since the roadmap is set for the next two generations and the engineers who build it remain. The risk is in the financing. In September we described Nvidia as the central bank of the AI build-out, with about $300 billion of potential liabilities on top of about $100 billion of investments. Smaller cloud providers cannot borrow at scale without that support. Britain's Nscale is one of them, with Nvidia as supplier, shareholder and a $1 billion contributor to its latest raise. Huang told his finance team that the largest cloud companies were "running away" from Nvidia. He described the wafers and memory it had committed to buy as an "avalanche". Fewer guarantees would slow construction outside the largest companies and leave Nvidia with supply it cannot place. Google's TPUs and AMD would gain from both. Kress will face questions on this at Tuesday's earnings call.

Takeaways: Huang ran one of the world's largest companies on his own judgement, with few people able to challenge him. That built Nvidia, and it also produced the relationships and the conflict of interest that ended his tenure. The board had started reviewing his financing commitments in August, so his departure speeds up a change that was already under way. The chips will keep shipping. What is uncertain is whether Nvidia will keep guaranteeing the debts of smaller cloud providers, because much of the AI build-out outside the largest companies depends on it. Anyone who has reserved compute with an Nvidia-backed provider should listen closely to what Kress says on Tuesday.

Anthropic launched new features this week that continue its push into the world of knowledge work and business content creation. Claude Sheets and Claude Whiteboards arrived in beta on all paid plans, while Claude Data, previously called Dashboards, became generally available. Claude Meetings, in beta for Team and Enterprise customers, lets Claude join a Google Meet call as a visible participant that takes notes and drafts plans while people are still talking.

These releases build on the Docs and Slides features launched in September, which means Claude now covers most of the file types that make up a working week. Much of the press coverage described the launch as a direct challenge to Office and Google Workspace. Anthropic's own guidance is more modest, since it tells users to export to .xlsx or switch to Claude for Excel when they need a full spreadsheet.

Having tested the main features at ExoBrain, we found them very light but interesting. You can leave a note on a specific cell or slide, ask Claude a small question about it and get a targeted change. That's a better way to work than describing the whole file again in plain chat. Data is useful for building a quick story around figures held in systems such as Snowflake or Salesforce.

Anthropic's illustration of Claude Whiteboards: three people and Claude on one board, with Claude turning their clustered notes into a plan and asking them to correct its draft. Source: Anthropic.

What these tools do well is start a piece of collaborative work, but they cannot yet run it from beginning to end. The gap is control. There is no sophisticated way to configure what Claude should produce or to check that the result matches what you intended. You end up iterating many times instead, which in some cases takes longer than building the sheet or deck yourself.

The more promising direction is in what surrounds these features. MCP apps, richer interfaces inside chat and chat embedded in other products all suggest that the boundary between an assistant and an application is becoming less fixed. Microsoft has spent two years adding an agent to its apps, while Anthropic began with the agent and is now adding apps. Both companies are arriving at a similar product from different starting points, although one with less legacy to work around.

That similarity is the reason to be cautious about treating this as how work will be done. Spreadsheets, decks and whiteboards were designed so that people could coordinate with other people. If agents become able to work together on these tasks, they will not always rely on these formats. Anthropic is trying to compete with Office 365 and Google Workspace in a category that likely won't exist in a few years.

Takeaways: Even after an extension this large, we find Claude's products more coherent than ChatGPT's mix of chat, work and features. The concept that holds them together is the artefact: each new tool produces one, whether a document, deck, sheet, board or dashboard, that people and Claude can both open and change. That is a useful construct for anyone designing AI solutions. We hope Anthropic will in the future let developers define artefact types of their own, which would also help the models learn to create and manipulate them. This central idea could outlast the collaborative surfaces, as multi-agent systems take over work we now do in office software.

This week's chart comes from Tomasz Tunguz at Theory Ventures, who has used public filings to trace AI spending from the power station to the agent. By converting every layer to dollars per watt per year, he finds that applications take 40% of the gross profit, neoclouds 28%, inference vendors 24%, data centres 5% and power 3%.

Although the chart is a useful map, the neocloud figure is measured before GPU depreciation, which is that layer's largest cost. CoreWeave grew revenue 168% last quarter and still lost $709 million. Its new multi-year contracts are priced at $20 to $25 per watt a year, close to the chart's entire $25 end price, so either customers pay far more or the layers above absorb the difference.

Nscale's latest IPO prospectus shows what sits behind that 28%. The London company has doubled its active GPUs to 50,000 in under three months and holds $105.8 billion of contracts, yet only $6.4 billion is live. It estimates revenue of $307.5 million for the first nine months of the year against losses of $1.69 billion. Financing for its $45 billion Anthropic campus is not yet agreed. Anthropic can cancel any tranche that arrives late.

Tunguz's chart describes how AI revenue is shared once capacity is running, but it leaves out depreciation, debt and delivery risk. Neoclouds receive a large share because they borrow the money and build against deadlines, while labs such as Anthropic keep the right to cancel.