Miguel Ángel TempranoEconomics, geopolitics and investment
ES·EN
The column by Miguel Ángel Temprano

Nvidia without Huang: what is really at stake

By · November 16, 2026· Investment· 10 minute read

Jensen Huang’s resignation as chief executive of Nvidia has been reported as a personal scandal, but its consequences are economic and geopolitical. Over the past two years the company did not just sell the chips that underpin artificial intelligence. It also financed its own customers, guaranteed their obligations and bought capacity from them, which is why it had become a kind of central bank for the sector. With a cautious chief financial officer now in charge, that safety net may tighten just as heavily indebted data centres need to refinance. That could change how the chip market is shared out and the pressure on chipmakers, and embolden those in Washington who want to tighten export controls on China and the countries around it.

This column is not about a love affair or a spy novel. Much has already been written about Jensen Huang’s private life and about the suspicions of the US intelligence services, and whatever remains to be known will come from the investigation by Nvidia’s board and, if it comes to that, from the courts. I am interested in something else.

On 12 November the man who, over 33 years, turned a maker of graphics cards for video games into the company on which much of the world’s investment in artificial intelligence rests resigned with immediate effect. He has been replaced on an interim basis by his chief financial officer, Colette Kress, and for the first time the board has a chair who is not also the chief executive.

That change at the top has consequences that go far beyond a share price. Nvidia is not just a chipmaker. Over the past two years it has also become financier, guarantor and customer of its own customers, the great allocator of the world’s most advanced manufacturing capacity and a central piece in the technological tug-of-war between Washington and Beijing. When the person in charge of a company like that changes, the rules of several markets change at once. It is worth separating the noise from the signal.

The market reaction was as expected. On the day of the resignation the shares fell as much as 14% at the open, and for a few hours the company slipped below the $4 trillion mark in market capitalisation. From the early-October peak the decline is around 35%, and at least seven brokers have downgraded their recommendation in a week.

Yet the price is the symptom, not the disease. Nvidia closed its last fiscal year with revenue of $215.9 billion and net income of $120.1 billion. It holds more in cash and marketable securities, $56.6 billion, than in financial debt, about $33.4 billion after its June bond issue, and it keeps its AA rating. The roadmap for its next processors is committed to suppliers and customers, and does not depend on one person.

heavily indebted neoclouds face their maturity wall between 2026 and 2028 without the same safety net

What the market is pricing in is neither a solvency crisis nor a technological delay. It is the loss of two assets that appear on no balance sheet. The first is Huang’s ability to close gigawatt-scale deals with governments and the big AI labs, negotiating almost as an equal with heads of state. The second, more delicate, is the company’s willingness to put its own balance sheet at the service of demand for its products.

Nvidia’s risk lies not on its own balance sheet but on its customers’. And that is where the change at the top starts to have real consequences.

To understand why, you have to look at how the latest leg of the AI boom has been financed. Microsoft, Google, Amazon and Meta pay for their data centres with their own cash and with investment-grade bonds. But a growing share of chip demand comes from the so-called neoclouds, young companies that buy GPUs by the tens of thousands, fill entire warehouses with them and rent out the computing capacity. Many have low credit ratings or none at all, and fund themselves with debt secured on the chips themselves.

Nvidia solved that bottleneck with a simple and powerful idea. It committed to renting the capacity those operators failed to sell, at a fixed price and for about six years. It guaranteed leases and residual values, and signed memoranda with six large asset managers and banks to mobilise more than $500 billion, offering to cover up to a quarter of the risk. Its latest quarterly report shows a maximum exposure under guarantees of $108.5 billion, almost all of it linked to OpenAI’s campus in Ohio. Adding capacity purchase commitments and leases, its support for customers comes to around $164.5 billion, close to 72% of its shareholders’ equity.

The result was that banks stopped analysing the operator’s risk and started looking at Nvidia’s instead. A loan to a B-rated company came to be priced almost as if it had been signed by an AA-rated one.

The new interim chief executive is, according to the specialist press, the executive who put up the most resistance to that model. She opposed a $250 billion guarantee for OpenAI, which ended up split into two phases with only the first signed, and tried to spread the risk across more customers. There is no record of her having changed any contract, and the guarantees already signed are obligations she cannot withdraw. But everything still to be signed, which is far more than what has been signed, is now in the hands of a cautious manager.

a guarantee is only worth as much as the will of whoever renews it

The strain was there before the resignation. CoreWeave, the largest listed neocloud, ended June with $35 billion of debt, and in October its credit default swaps were pricing in a probability of default of more than 60% over five years. Much of the sector’s debt, extended at loan-to-value ratios of 60% to 70%, matures between 2026 and 2028. If, at that maturity wall, Nvidia’s backing becomes more expensive or shorter, refinancing will cost more, covenants will be tighter and some projects will never be finished.

The danger is not so much the failure of one operator as contagion. A single default would force a review of all GPU-backed credit, whose collateral depends on a single manufacturer that, moreover, sets the pace at which those chips become obsolete. The IMF and the Bank for International Settlements had already warned of the risks of this circular financing.

Nvidia being more cautious with credit does not mean that fewer AI chips will be bought. It means a change in who buys them and from whom. Customers that are not large platforms contributed about $40 billion in the July quarter, 45% of its data-centre revenue. If part of that demand loses access to cheap financing, the weight will shift towards the big hyperscalers, which negotiate prices harder and design their own processors.

This is where the competitors come in. Broadcom is already backing leases of its custom chips for up to $29 billion through a platform with Apollo and Blackstone, two private equity giants. Google can offer its TPUs with a balance sheet even larger than Nvidia’s, and AMD has spent years looking for a crack in CUDA’s dominance. For two years Nvidia not only sold the best chip, it also offered the best financing to buy it. If it gives up the second advantage, the first will have to defend itself on its own.

if Nvidia stops financing those who buy its chips, Broadcom, Google and AMD may gain ground, and manufacturers such as TSMC and the memory makers will have to ask themselves who pays for the new capacity

In manufacturing the effect is slower, but real. Nvidia has $366 billion in purchase commitments with suppliers, mainly TSMC and the makers of high-bandwidth memory. Some investors estimate that it absorbs around two-thirds of TSMC’s most advanced capacity, which could be reallocated relatively quickly if orders slow. For SK Hynix, Samsung and Micron, and for the whole advanced packaging chain, the question is no longer how much capacity will be needed but who will pay for it.

There is also a less visible risk: second-hand GPUs. If an indebted operator goes bust, its chips reach the secondary market and compete with the new generation. Nvidia was the implicit buyer of last resort for that installed base. Without that net, the hourly rental price of computing capacity becomes the indicator to watch.

The third front is Washington. Huang was the most effective advocate of the idea that selling chips to China, even previous-generation ones, kept its developers tied to American technology. In December 2025 he persuaded the White House to authorise the sale of the H200 to approved Chinese customers, in exchange for 25% of the revenue going to the Treasury. That voice disappears just when the company needs it most.

The timing could not be worse. The Democrats will control Congress from January, and the Senate’s defence bill package already includes three bills that Nvidia was fighting. One would give Congress the power to review exports of advanced chips and would bar the sale of Blackwell to China for two years. Another would require chips to be location-tracked to curb smuggling. The third would press allies to restrict chipmaking equipment. Senators from both parties have explicitly cited the case in calling for their approval.

Paradoxically, the direct blow in China would be small. Beijing is holding back H200 imports, some 600,000 chips are waiting in Hong Kong and Nvidia was already leaving China out of its forecasts. The damage that matters lies around it. The Megaspeed case, involving a Singapore-based company whose Malaysian subsidiary bought at least $4.8 billion of equipment with Nvidia chips and which is now under investigation by the Justice Department, casts suspicion on the South-East Asian channel. The risk is not losing China, which was already lost, but the national security label spreading to the entire international business.

Mandatory tracking raises compliance costs for every customer in the world. A congressional review of large sales could delay the sovereign deals with the Gulf states, which were the natural substitute for the Chinese market. And every step in that direction strengthens Beijing’s bet on Huawei, Cambricon and the other local suppliers, as well as the temptation to respond with tools it has already used, from competition regulation to rare earths. For Europe the warning is clear. The AI supply chain depends on decisions taken on Capitol Hill and in Beijing, not in Brussels.

without its main advocate in Washington, the laws restricting chip exports are moving forward

The most likely scenario is not a rupture but a rationalisation. Nvidia will honour what it has signed, will be much more selective with new guarantees and will demand more collateral and creditworthy end customers, while the neoclouds go through a consolidation that many already considered inevitable. It is a healthy change, because an industry that depends on its supplier lending it the money to buy from it is a fragile industry. But the transition will not come free.

The coming weeks will tell us a great deal. On 17 November Nvidia reports results, the first with Kress in charge, and it will be worth listening less to the revenue figure than to what it says about support for its customers. Then come the vote on the defence bill and the search for a permanent chief executive. In the credit market, spreads on debt linked to data centres will be the most reliable thermometer.

The Huang era was that of a founder acting as the central bank of artificial intelligence. The one now beginning will probably be that of an industry forced to finance itself. Markets would do well to prepare for the difference.

Share LinkedIn X Email

On the same subject

Get every column by email

You will be told whenever something new is published: a column, an analysis or a research note. No advertising, and your details are never passed on.

What would you like to receive?
Language

One click to unsubscribe in every email.