Nvidia lost its founding chief executive this month, and its next quarterly report showed revenue up 103%. Both facts matter to anyone buying AI services, including the many buyers who will never sign a contract with Nvidia. If your models run on GPUs, a supplier this concentrated sits in your supply chain whether it shows up in your vendor inventory or not, and this month showed what a key-person event at that supplier actually looks like.
My read is that buyers should treat this as a vendor risk exercise. What follows sticks to what Nvidia filed and said, what the press attributed to named people, and what I'd do with it. I'm not offering a view on the stock or on Huang personally.
What Nvidia disclosed
On November 12, Nvidia filed an 8-K (opens in new tab) stating that "Jensen Huang has resigned as President, Chief Executive Officer, and a member of the Board, effective immediately." The filing names Colette Kress, the company's EVP and CFO, as interim CEO "until a permanent successor has been named," and says she keeps the CFO role while she does it. The board appointed Harvey C. Jones as its "first-ever Chair of the Board" and started a search for a permanent CEO that will consider internal and external candidates.
The filing gives no reason and no separation terms. Its only operational statement is Jones's expectation that "the company's operations, customer commitments and product roadmap" will "continue without interruption." That is the new chair's expectation, not a contract term or guidance.
The quarter reported five days later
Nvidia reported its fiscal third quarter five days later, and the release (opens in new tab) quotes Kress, now interim CEO and CFO. It shows revenue of $115.8 billion, up 103% from a year earlier, and Data Center revenue of $107.6 billion, up 110%. GAAP net income was about $65.3 billion. The outlook for the current quarter is $130.0 billion in revenue, plus or minus 3%. Kress said Vera Rubin is "off to an explosive start" and that the roadmap is "on track," and the release calls the Vera Rubin NVL72 ramp "the fastest product ramp in the company's history," with racks running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius.
Look at the share first. Data Center is about 93% of revenue, so the business you depend on and the business the market is pricing are the same business. Then read the forward-looking statements. The release lists, among its risks, Nvidia's ability to "identify and integrate a new chief executive officer." The company has written its own key-person risk into its standard disclosure, and that's the line I'd track each quarter until a permanent CEO is named and settled in.
Why a buyer who never buys Nvidia still has this risk
Most enterprise and fintech teams don't buy GPUs. They buy a model API, a cloud AI service, a SaaS product with a model inside, or capacity through a managed provider. Every one of those paths ends on accelerators, and the release's own customer list for the newest racks is a who's who of the clouds and specialist hosts your vendors build on.
I covered the lab side of this in concentration risk in the three-lab AI stack: most of the AI on a typical roadmap traces back to a few labs that share chips, supply chains and balance sheets. This month adds the layer underneath. The labs and the clouds are themselves concentrated on one chip supplier, and that supplier just went through a founder-CEO departure, with only an interim successor named.
Revenue, guidance and the product ramp all held in the first report after the change, which is one data point and only one. A CEO search takes time, the permanent choice will make its own calls on allocation, pricing and roadmap, and a supplier's strategy can drift slowly after a leadership change without any single disclosure marking the moment.
What I'd map first
The job is to know your dependency before you need the answer. I'd build the map from the inside out:
- Every AI-backed service you run or buy. Internal applications, vendor products with AI features, and anything an employee reaches through a browser. If it calls a model, it goes on the list.
- The provider behind each one. The model vendor, the cloud region, and any reseller or router in between. Your contracts often name only the top layer.
- The hardware behind each provider, where they'll say. Some publish it, some disclose it under NDA, and some won't answer. A "won't answer" is itself a data point for your risk register.
- The substitution path. For each service, the model or provider you would move to, whether it runs on different hardware, and how long the move takes in practice.
Most teams will find the same pattern. The dependency is indirect, it runs through two or three providers, and those providers draw on overlapping capacity. That overlap is the concentration. Consolidating vendors at the top layer can hide it, which is the point I made in one throat to choke is also one point of failure.
What I'd ask my providers
A buyer's questions should be ones only the provider can answer. These are the ones I'd send to every AI vendor carrying regulated or customer-facing work:
- Allocation. How is your accelerator capacity secured: owned, reserved under contract, or bought on demand? If capacity tightens, where does our workload sit in your priority order?
- Substitution. Can you serve our workload on a different chip family or a different cloud, and have you actually done it in production? What changes for us in latency, cost or model availability if you do?
- Notice. Will you tell us when a change in your own supply affects our service, and through which contract clause?
- Exit. If we need to move, what do we take with us: prompts, evaluation sets, fine-tuned weights, logs?
None of these questions mention Nvidia's leadership, and they shouldn't. What you want to learn is whether your providers have options if the supplier's priorities change, for any reason.
Design so the answer matters less
Questions buy information. Architecture buys tolerance. The pattern I argued for in designing AI inference for model disappearance applies here unchanged: route model calls through one place you control, keep an evaluation set that proves a second model can do the job, and practice the switch before you need it.
That's the shape we used when we built AgentOS, a single governed egress through AWS Bedrock with an audit trail on every call, and the reason I like the shape here is simple. A single egress gives you one place to see which provider served each request and one place to change it. It doesn't remove the dependency on the chips underneath, and nothing at the buyer's layer can. It does make substitution a configuration change instead of a project.
What I'd watch in the disclosures
Nvidia's filings will tell you more than the coverage will. I'd read, each quarter:
- The 8-Ks. A permanent CEO appointment, any change to the chair or interim structure, and any separation agreement would each arrive there first.
- The forward-looking statements. Whether the CEO-search risk language stays, changes, or drops out is the plainest signal of how the company itself sees the transition.
- Revenue mix and guidance. A Data Center share near 93% means any wobble in that one line is the whole story for your supply chain.
- Your own providers' disclosures. The clouds and model vendors that depend on this supplier have their own filings and trust pages. A change in how they describe capacity tells you more about your exposure than Nvidia's will.
For a bank or credit union, this belongs in third-party risk management even though the supplier is two or three tiers down. The same discipline I'd apply in technical due diligence, where key-person risk is one of the first things a data room tends to hide, applies to your AI stack. A single founder leaving a single supplier is exactly the scenario a fourth-party review is supposed to have already priced.
Map which of your AI services depend on which providers, send the allocation and substitution questions, and put Nvidia's next 8-K and next earnings release on someone's calendar.