Jensen Huang has disposed of Nvidia shares again.
A filing Nvidia submitted to the U.S. Securities and Exchange Commission on Sept. 16 showed that Huang disposed of 45,700 Nvidia shares at roughly $212 apiece. On its face, that is the kind of disclosure that can easily be read as a signal.
The timing made it even easier to overread. Nvidia’s stock has kept climbing, while a different line of thinking has started to spread across the AI industry: spending may be moving too fast, capital expenditure is getting too large, and if models keep getting cheaper, the need for as much compute could come under closer scrutiny.
Yet Huang has not sounded cautious. On Sept. 18, he said publicly that Nvidia could sell twice as many chips next year as it does this year.
So while part of the industry is debating whether AI development should slow down, Nvidia is still preparing to sell more hardware. Even so, the 45,700-share disposal should not be taken as a straightforward bearish call.
The filing points to tax withholding, not an open-market sale
In the SEC filing, the disposal was marked with an "F," which indicates tax withholding rather than a discretionary market sale by Huang.
He made similar transactions in March and June this year, and those carried much the same character.
His stock activity this year has not been limited to disposals. The transactions disclosed in March, June, and September were mainly tied to tax payments, while he also made multiple share transfers and gifts during the year. The latest disclosure also showed that he still holds a very large Nvidia stake through different trust structures and related arrangements.
Those moves look more like long-term personal wealth management than a directional call on the stock. Huang’s wealth is heavily concentrated in Nvidia, and an asset position of that scale naturally involves ongoing tax planning, trust management, and portfolio allocation. In that context, 45,700 shares do not stand out on their own.
The more useful comparison is 2025. In June last year, Huang began a 10b5-1 trading plan worth up to about $865 million and went on to sell Nvidia stock under that framework. That was a genuine planned sale program.
The transactions disclosed so far this year are plainly different in nature. Put simply, he is handling stock, but that does not necessarily mean he is making a statement about Nvidia’s share price, and it does not support a simple bearish reading.
The AI sector has started to talk about slowing down
Since the start of the year, the discussion in AI has gradually shifted. The question is no longer only how much stronger models can become. Safety, governance, and the boundary of spending are taking up more space.
Anthropic CEO Dario Amode has publicly called for a slower pace in advancing AI capabilities, arguing that AI development is moving faster than regulation and governance can keep up with. OpenAI CEO Sam Altman and others in the industry have also publicly backed a more cautious path for frontier AI.
By mid-September, that split had moved into the open. People tied to Anthropic and OpenAI were discussing AI development speed and safety risk, while Huang and Meta CEO Mark Zuckerberg did not join the related initiative to coordinate a slowdown.
That does not mean AI companies are preparing to stop. OpenAI and Anthropic are still advancing new models, and demand for compute has not disappeared.
What has changed is the way the industry is recalculating the speed of spending. Model capability may keep improving, but data centers require land, power, network capacity, and enormous capital expenditure. As AI moves into industrial deployment, the issue is no longer only technical. Spending now has to answer a harder question: can it generate enough return?
Capital markets have already picked up that shift. For the past few years, model capability was the scarcest thing in AI. Today, compute, power, data centers, and capital have themselves become cost centers.
Huang’s answer is to keep expanding output
Huang’s response to that debate is direct.
On Sept. 18, he said Nvidia could sell twice as many chips next year as it does this year. Coming from Nvidia at its current scale, that is a consequential statement.
Nvidia’s latest quarterly revenue has already reached $96.2 billion, up 106% year over year. Talking about a doubling in chip volume on top of that suggests Huang still sees substantial room for growth in AI infrastructure demand.
And the word "chips" no longer refers only to GPUs in the narrow sense. After Blackwell comes Rubin, and beyond GPUs Nvidia is selling CPUs, networking, systems, and software. The company is pushing itself from being a chipmaker toward becoming a full AI infrastructure supplier.
The Vera Rubin platform unveiled this year makes that shift especially clear. It is no longer a standalone GPU sold by itself. It is a rack-scale system built for AI inference and agents, and Nvidia has explicitly positioned it as infrastructure for the AI agent era.
The market Huang is looking at has changed with that transition. How many GPUs Nvidia sells is only one dimension. The bigger market is how much compute, networking, storage, and power a single AI data center will need.
If models get cheaper, does demand for GPUs fall?
Over the past year, DeepSeek, Kimi, and a wide range of open-source models have kept improving efficiency. That has fed a recurring concern inside the industry: if models become more efficient and both training and inference get cheaper, will demand for GPUs fall as well?
Huang’s answer has consistently gone the other way.
In his view, better model efficiency makes AI cheaper, and cheaper AI means more tasks can be handed to models. Today a single question-and-answer interaction may require only one inference call. Once agents begin to work through tasks, a single assignment may involve dozens of consecutive calls, or more.
As AI moves into office software, search, coding, customer service, cars, robots, and phones, compute demand also spreads from training into inference.
Training is a concentrated burst of spending. Agent-based work may turn into a continuing stream of compute consumption. That is one reason Nvidia has kept stressing inference and agents.
In the past, the industry fixated on model parameters. The next metric that matters more may be token consumption. The cheaper a model gets, the more tasks it can economically handle. The more often agents are used, the more compute they require.
What Huang is really betting on is that higher efficiency and lower cost will expand AI usage rather than shrink the market for compute.
Nvidia is betting on infrastructure as a whole
That is also where Huang’s perspective diverges from that of model companies.
OpenAI and Anthropic are more focused on how far model capability can go. Huang is more focused on how much computing infrastructure those capabilities require when they are actually run at scale.
By that measure, the bill keeps getting larger. Microsoft, Google, Meta, and OpenAI are all still investing in data centers and AI compute, and AI capital expenditure has become one of the most closely watched variables across the broader technology sector.
The bigger shift is happening beyond GPUs. As compute rises, data centers need more power. Servers need more efficient cooling and power delivery. Racks need higher-bandwidth networking. Storage and optical communications expand alongside them.
This year, Nscale, an AI cloud company backed by Nvidia, disclosed 10GW of power reserves in its IPO filing and also received a $1 billion investment from Nvidia.
Power is now entering Nvidia’s business map in a more visible way. The point is straightforward: once AI computing reaches a large enough scale, chips are only one part of the infrastructure stack. What Nvidia is trying to capture is not just the life cycle of a single GPU generation, but an entire buildout of AI infrastructure.
Huang is managing personal wealth while Nvidia is buying back stock
Another detail that is easy to miss is that Huang has been dealing with his personal Nvidia holdings while Nvidia the company has kept repurchasing its own shares.
In the second fiscal quarter this year, Nvidia bought back about $19.7 billion of stock. As of July 26, it still had $99.3 billion remaining under its repurchase authorization. In May, the board also approved an additional $80 billion for buybacks.
Those are two different layers of capital activity. Huang is managing personal wealth. Nvidia is allocating corporate capital.
Viewed together, they show a simple contrast: personal assets are being diversified, while corporate capital remains concentrated in Nvidia itself.
Nvidia’s cash flow is now large enough to support both massive buybacks and investment in next-generation products. That is one of the clearest ways today’s Nvidia differs from a traditional chip company.
A traditional chipmaker mainly needs to sell products. Nvidia now also participates in building customers’ data centers, helps design servers, networks, and AI factories, and then connects the next generation of chips into that system.
What customers are increasingly buying looks more like a computing factory. What Nvidia earns is no longer just the revenue from a single GPU.
The real disagreement comes down to demand
At that point, the stock transaction itself looks less important as a signal.
The deeper issue is the split emerging across the AI industry. OpenAI and Anthropic are putting more weight on the pace of frontier AI, as well as safety and governance. Nvidia is still expanding its expectations for infrastructure demand.
In the end, that disagreement still comes back to demand: as models get cheaper, how much compute will AI actually consume?
How far the compute business can keep running after model prices fall depends on whether AI can produce more real-world demand. Huang has already placed his answer into Nvidia’s capacity plans, products, and capital spending.
Whether he is pressing the accelerator is not especially hard to judge. The harder question is whether the value created after AI speeds up can grow faster than capital expenditure.
This article was sourced from the WeChat public account Bianmian Zhiwai and written by Bianjun.