Jensen Huang's Exit Shows Why Nvidia Is Hard to Replace

Jensen Huang's departure has exposed how deeply AI companies depend on Nvidia's chips, software and computing capacity. Businesses can increasingly switch between models for different tasks, but replacing a hardware supplier involves far more than changing a setting. That difference helps explain why Nvidia remains difficult to displace.

Nvidia confirmed on Thursday that Huang had resigned immediately, with chief financial officer Colette Kress appointed interim chief executive. The announcement arrived in a week when Google made switching between AI models easier and a British chip challenger raised hundreds of millions before shipping its first commercial product. Together, these events reveal a peculiar imbalance: the software is becoming easier to substitute, while much of the machinery underneath remains stubbornly concentrated.

The founder was also a bridge

Huang was more than the executive announcing new processors at conferences. Over three decades, he helped make Nvidia a central supplier to companies building and running advanced models, and became an influential advocate for its business in Washington and abroad. Nvidia's board says it expects customer commitments and the product roadmap to continue without interruption under Kress. That assurance matters, but a functioning engineering organisation and confidence in the political environment around its products are two different things.¹

His resignation followed reports concerning his personal conduct and alleged links to a woman suspected by US authorities of intelligence ties to China. Those allegations have political consequences even where claims have not been established in court, and the company's announcement did not adjudicate them. US lawmakers were already debating tighter restrictions on access to advanced chips, including rules addressing tracking and remote use. Huang's departure gives that debate a new urgency without proving that any proposed restriction will become law.²

The commercial stakes are unusually tangible. Reuters examined more than 150 procurement notices, contracts and company disclosures and identified over $3.2 billion in Chinese computing projects involving restricted Nvidia processors. That figure includes buildings, power equipment and other infrastructure, rather than representing sales of Nvidia chips alone. Some projects were planned, others contracted or already operating. The distinction matters because it shows a network of long-term commitments, not a simple order book that could be cancelled or reassigned overnight.³

Developers depend on more than physical silicon. A company may have trained its engineers on Nvidia's CUDA software, optimised code for particular accelerators, contracted for a specific class of cloud instance and built performance expectations around that configuration. Moving a workload can mean rewriting components, testing numerical behaviour, checking compatibility with specialist libraries and negotiating new capacity. A chip from another manufacturer may be excellent and still be an expensive replacement for the system already in use.

This is why the resignation is an infrastructure story rather than only a boardroom story. A cloud provider that has already invested in thousands of compatible servers will still need to operate them tomorrow; an AI startup running on those servers will still need predictable capacity next month. The leadership transition does not itself mean customers will lose supply or that Nvidia's roadmap will fail. It reveals how many organisations are exposed when the same supplier sits inside both their engineering decisions and a sensitive geopolitical relationship.

Goldman Sachs expects annualised US investment related to AI hardware and intellectual property to reach $1.4 trillion by the end of 2027. That is a forecast, not committed spending, and it captures a much wider market than Nvidia alone. Even so, it illustrates why questions about chip access now extend beyond the fortunes of one listed company. The more infrastructure companies build around a particular supplier, the more costly uncertainty about that supplier becomes.⁴

Fractile still has to ship

Two days before Nvidia's leadership announcement, Fractile raised $720 million at a reported $7 billion valuation. The London and Bristol startup is developing processors for inference, the stage when a trained model produces answers or performs work. Its pitch centres on reducing energy use and improving the speed of delivering those answers. The first commercial chip is planned for 2027, so the valuation is a wager on what its hardware may achieve rather than evidence of a mature alternative already operating at scale.⁵

That distinction does not diminish the engineering ambition. Most people encounter AI through inference: a support assistant answering a question, an application describing an image or a software agent repeatedly calling tools. Each interaction uses computing capacity, and the demand can multiply as products become more useful. A faster or less power-hungry inference chip could lower the cost of serving those interactions, especially when workloads are stable enough to optimise for particular hardware.

Yet an improved processor is only part of a usable computing service. A buyer still needs a supported software environment, predictable availability, networking, cooling, maintenance and a route for moving an existing application without breaking it. Large customers will want evidence about total cost over the lifetime of a deployment, not only a headline benchmark. Startups can demonstrate promising performance long before they have the production, distribution and developer support needed to challenge an incumbent.

Torus approached a neighbouring problem this week. The company announced $500 million in growth capital and introduced Foundry, modular installations that bundle computing with power, cooling, storage and networking. It says systems ranging from one to twenty megawatts can be deployed in twelve to sixteen weeks using existing power connections. Those timelines are Torus's claims and should be tested against completed projects, but the underlying proposition is revealing: putting useful compute somewhere can be as difficult as designing the processor.⁶

Fractile is pursuing new silicon; Torus is trying to bring infrastructure online more quickly. Neither announcement establishes that Nvidia's market position has suddenly weakened, and they tackle different constraints. Fractile must prove its chips in commercial use, while Torus still needs processors and customers whose applications run effectively on them. Both companies are competing for a share of demand that is expanding faster than many traditional infrastructure projects can accommodate.

For a business buying AI services, the lesson is not to become an amateur semiconductor analyst. It is to ask what happens when its chosen provider is short of capacity, increases prices or changes which workloads it supports. A contract with two AI software vendors is not necessarily meaningful diversification if both rely on the same underlying hardware provider and region. Supplier diversity has to be examined below the application layer, where the expensive dependencies sit.

Google's cheaper division of labour

Google's Gemini 4 Hydrogen launch shows how quickly the software side of this market is changing. Hydrogen is designed for faster everyday coding and agent workflows, alongside Google's more powerful Carbon and Argon models. Instead of requiring developers to choose one model for an entire task, Google introduced an Orchestrator that can move work between them. In Google's tests, a Carbon and Hydrogen combination approached Carbon's task success while using 24% less quota and finishing 19% faster.⁷

These are company-reported benchmark results, and they do not mean every application will see the same saving. They do demonstrate a more useful design goal than simply making the next model score higher. If a demanding reasoning step needs a premium model but routine execution does not, software can assign each portion appropriately. A customer asking for one result need not know which model handled the research, which wrote the code or which checked the output.

Google also announced a Decisions API with an even narrower purpose. It evaluates up to 1,000 predefined questions in a single pass across text and other media, returning scores rather than a generated essay. Google lists a price of $0.10 per million input tokens with no billed output tokens and latency of 150 to 250 milliseconds. Those figures concern this specific bounded-decision service, not the cost of all AI reasoning or an invitation to assume its scores are always correct.⁸

Consider a customer enquiry arriving at a small online shop. Before anyone drafts a reply, the software may need to classify the message as a return, delivery issue, product question or suspicious request. Those choices can be described in advance and checked against examples from the business. Sending the entire enquiry through an expensive general-purpose model to produce a paragraph explaining its classification may offer little advantage if a simpler decision service does the job accurately.

The same logic applies after classification. A routine delivery question may be resolved by looking up an order status and applying a clear policy, while an unusual complaint needs a more capable model or a person. For content work, selecting the correct product photographs and checking a caption against known facts is a different job from developing the creative idea. Better AI content tools will separate these tasks, rather than treating every click as a reason to call the largest model available.

This flexibility is real but should not be confused with automatic portability. Models differ in their instruction following, tool interfaces, reliability, licensing and treatment of sensitive data. Replacing one with another requires evaluation against the work it actually performs. Even if an application can switch among several models, all those calls must still run on physical computing infrastructure. The software layer is gaining choices faster than the hardware layer can be rebuilt.

Deployment is the economic reality check

The financial incentive to use those choices is no longer hypothetical. Capgemini surveyed 500 executives at businesses with annual revenue above $1 billion across ten countries. Among respondents, 59% said AI expenditure had significantly exceeded expectations or budgets over the previous year; for those experiencing overruns, the average was 24%. Capgemini also found growing interest in open-weight models, which companies can deploy more independently than services available only through a provider's hosted API.⁹

An open-weight model is not automatically cheaper once a company pays for the machines, engineers, security work and ongoing operation. A provider's published token price is similarly incomplete if an agent repeats tasks, makes excessive tool calls or generates work that people then have to correct. The useful comparison is the cost of finishing a task to an acceptable standard, including failures and exceptions. Otherwise a cheaper model can simply turn a visible computing bill into an invisible labour bill.

This week also offered an example of a company measuring the job rather than only the model. Northern Trust and Microsoft reported that an AI-assisted onboarding process cut preparation of institutional clients' historical investment data from roughly two and a half weeks to three days, an 83% reduction. The bank measured real onboarding cases including analyst review and validation, not only the time taken to generate a spreadsheet. Its 96.5% first-pass F1 quality score came from a separate controlled test of seven representative cases, so it should not be read as proof of flawless performance in every situation.¹⁰

There is a useful economic difference between that result and a model provider promising cheaper tokens. Northern Trust identified a particular costly process, worked with engineers and domain experts, and measured what changed for the people performing it. The time saving can be evaluated against the cost of deployment, monitoring and maintaining the system. It also has a human consequence: analysts can spend less effort reformatting historical records and more time resolving cases that require their expertise.

Smaller firms may never buy their own GPUs or build a dedicated agent platform, but they still inherit these economics through subscriptions and usage limits. A restaurant wants a dependable way to turn its real dishes and photographs into useful posts; it does not want to pay extra because an assistant repeated an elaborate reasoning process to schedule Thursday's content. A retailer needs product descriptions that match the actual stock, not an impressive demonstration that quietly invents details. In both cases, the provider's choice of models and infrastructure eventually appears in the customer experience and the price.

That makes the week's apparently disconnected announcements closely related. Nvidia's leadership change raises questions about the supply of the computing on which many services depend. Fractile and Torus are attracting money by promising to expand or improve that supply, while Google is finding ways to use it more selectively. Capgemini's survey suggests customers are already pushing back against the cost, and Northern Trust shows why a measurable business result can be more persuasive than a technical benchmark.

The replacement test

A company preparing its next AI product has a more practical question to ask than whether Nvidia will retain its market share. Which parts of the product would still work if a model price doubled, an inference region became unavailable or a preferred accelerator could not be rented? The answers will differ by workload: some tasks can move to a cheaper model, some can be handled by ordinary code, and some genuinely depend on specialised computing capacity. Mapping those differences is useful planning, not a prediction that every provider will fail.

The mistake would be to treat choice in one part of the stack as choice everywhere. Google's Orchestrator is evidence that model selection can increasingly happen inside an application, but it cannot manufacture power connections, chips or data-centre capacity. Fractile's funding is evidence that investors see room for alternatives, but a commercial chip planned for next year cannot replace deployed servers today. Nvidia's succession may prove operationally uneventful, yet the week has made its customers' dependence much harder to overlook.

There is an encouraging side to this story. Better software design can reduce waste, specialised processors may eventually improve the economics of inference, and clearly measured applications can free people from repetitive work. The strongest businesses will know which dependencies they can change quickly and which ones they have to plan around for years. Nvidia's next chief executive will be judged in part by how well the company keeps earning those long-term commitments, not merely by what happens to its share price this week.

Sources