If Jensen Huang were to return to the semiconductor business as a founder, he would face a daunting benchmark: the company he built. In the quarter to October 25th, Nvidia made $77.5bn in operating profit on sales of $115.8bn. That works out at around 67 cents of operating profit for every dollar of revenue. Huang resigned in November, with Nvidia announcing the transition on November 12th; on November 17th the board ruled out his return. 1 2 3 Speculation that he might raise $50bn to launch another chipmaker is just that: no such financing or venture has been established. 4 But it is an illuminating thought experiment. A business designed to undermine the economics of Nvidia might discover that those very economics are what investors want it to reproduce.
Too much of a good thing
Suppose a group of investors put $50bn into a new company and emerge with half its equity at exit. To triple their money in eight years (a 14.7% annual return without intermediate distributions), the company must be worth $300bn. At 20 times operating earnings, assuming zero net debt, it must make $15bn of annual operating profit. With a 55% gross margin and $4bn of operating expenses, it needs $34.5bn of sales. A $10bn initial cheque would bring that revenue hurdle down to $12.7bn under the same assumptions. Assumptions for the chart: investors retain 50% at exit, target three times their money after eight years, and the company is valued at 20 times annual operating profit with no net debt. Gross margin is fixed at 55% and annual operating expenses at $4bn. The numbers are scenarios, not forecasts. They show why the size of the cheque can be the problem rather than the solution.
The price of a second act | Required annual revenue ($bn)
A second problem is deciding what constitutes the market. Nvidia's $107.6bn quarterly data-centre sales imply a $430.4bn annualised run-rate, but include products other than accelerators. 2 Revenue of $34.5bn is equivalent to 8% of that benchmark, not 8% of the accelerator market. If only a quarter of the benchmark were technically substitutable and commercially available at entry, the required revenue would equal 32% of that smaller pool. Neither the quarter nor the run-rate is a market-size forecast. The point is that market size becomes a poor guide to opportunity once switching costs and existing obligations enter the calculation.
The customers are becoming competitors
In October 2025 OpenAI and Broadcom announced plans for ten gigawatts of custom accelerator deployment. AMD has since disclosed two customer warrants covering up to 160m shares each, equivalent in aggregate to nearly one-fifth of its September 2026 outstanding share count, before any dilution. None had vested at that reporting date; their exercise depends on specified conditions. AMD also recorded a $1.2bn warrant liability, a reminder that the concessions can have accounting consequences well before any shares are issued. 5 6 The implication is uncomfortable for a challenger. Large customers want alternatives, but would be unlikely to sponsor a new supplier merely to hand it the pricing power of the old one. A guaranteed order book may be available at the cost of a smaller share of the resulting profits.
Nvidia's purchase commitments tell the story from the other end of the supply chain. Its $390bn in supply and capacity commitments at October 25th include $203bn falling in fiscal 2028. 7 This is a schedule of obligations, not money already spent. Fiscal 2027 figures represent the remainder of that year. Fiscal 2032 includes all subsequent years. An incumbent can reserve memory and manufacturing capacity ahead of customers' requirements; an entrant risks paying for unavailable or unused capacity. There is a vulnerability in such scale, too: Nvidia disclosed gross maximum exposure of $118bn under contingent guarantees, not realised losses. A newcomer that obtains customer-backed manufacturing commitments without guaranteeing infrastructure financing could avoid some of that exposure.
The incumbent has booked the future | Nvidia commitments ($bn)
The price of a token
One plausible entry point is AI inference: running trained models rather than training them. Yet a cheap processor does not necessarily produce cheap output. Suppose a new system costs $80m annually against $100m for the incumbent at identical peak capacity, output quality and latency. If the old system operates at 80% effective utilisation and the new one at 60%, its cost per useful unit is 6.7% higher. To merely match the old system's economics, it needs 64% utilisation. Nvidia's software ecosystem matters because customers buy useful computing, not theoretical performance. A newcomer could target a narrower, repeatable workload where software portability and hardware efficiency can be proven.
When cheaper hardware costs more | Unit cost index (incumbent = 100)
There is a further wrinkle. If inference becomes much cheaper, demand may expand. But to preserve revenue when realised prices halve, the number of units sold must at least double. The beneficiary of the price reduction may therefore be the model provider or its customer, rather than the chipmaker. Mr Huang's hypothetical investors should ask who would retain the savings, not merely how much electricity a faster architecture might conserve.
The numbers favour a more modest ambition than another Nvidia: an independently developed processor for a well-defined workload, supported by binding customer purchases and financed in stages. Without a product, suppliers, ownership terms or binding orders, a credible probability of success cannot be assigned. The test would be whether a new business could attack the incumbent's margins without needing those same margins to survive.

