CUDA and Effect: Why Nvidia Wins the AI Race Regardless of Who Comes Second

The word "bubble" is doing an enormous amount of work in financial commentary right now, and like most words that are made to work too hard, it is starting to show the strain. Capital Economics says the AI bubble has burst. Goldman Sachs says the market is fundamentally earnings-driven. Citigroup says there is a bubble but there is also enough liquidity to keep it inflated, which is roughly the financial equivalent of saying that the Titanic had excellent ballrooms.

All of them are making reasonable observations. Not all of them are asking the right question.

Where the Article Gets It Right

John Higgins of Capital Economics is correct about the valuation picture. Tech P/E ratios at pandemic-era lows, the premium over the broader market contracting, semiconductor valuations being pulled down alongside software: these are real, documented, not-up-for-much-debate facts. If you bought into the sector at peak premium and are now looking at your portfolio with the expression of someone who has just found out what a margin call is, this is accurate and I am sorry for your situation.

The more interesting observation, though, is the one about where the bubble actually lives now. Higgins suggests it has migrated from stock prices to earnings expectations. That distinction matters enormously. Price bubbles are unpleasant. Expectation bubbles, when they deflate, take entire business models with them, and they do it in the specific way that makes executives say things like "we didn't see this coming" while standing in the wreckage of something that was, with hindsight, entirely predictable.

He is right about this. He is not quite right about what it means.

Where the Article Misses the Point

The article conflates two things that should be kept separate: the valuation of AI companies, and the transformative utility of AI technology. These are not the same thing, and one can be dramatically overstated while the other remains perfectly real.

The dot-com bust of 2000 is the relevant precedent, and it is instructive in ways the article does not fully explore. When the dot-com bubble burst, it was genuinely catastrophic for investors in Pets.com, Webvan, and approximately 400 other companies that had turned the phrase "first mover advantage" into a business plan. The Nasdaq fell 78%. Fortunes were wiped out. The carnage was real and extensive.

And then the internet turned out to be the most transformative technology of the twentieth century anyway.

The infrastructure that was overbuilt during the bubble, the fibre optic cables, the server farms, the routing protocols, became the foundation for Google, Amazon, Netflix, and everything that followed. The companies that died were overvalued. The technology was not overhyped. These are different sentences.

AI is, I believe, in the same position. The technology is genuinely transformative. Many of the companies currently being valued on its promise are not. The bubble, if there is one, is in the gap between those two things.

The Collapse That Is Coming (and Deserves to Come)

Here is the part that the financial analysts are discussing politely and I will discuss rather more directly: a significant number of companies currently presenting themselves as "AI companies" are, in fact, API wrappers with a Stripe integration and a pitch deck.

They have built their entire value proposition on the assumption that having access to a large language model is, in itself, a competitive advantage. They have taken GPT-4, or Claude, or Gemini, wrapped it in a slightly different interface, applied it to a specific vertical, and raised money on the basis that they are doing something defensible. They are not. They are renting land and calling it a property empire.

A technical moat, in the meaningful sense, requires at least one of the following:

• Proprietary data that others cannot replicate (and that actually makes the model better in your domain)
• Deep workflow integration that creates genuine switching costs
• Network effects that improve the product as users accumulate
• Original model research that produces capabilities others do not have

 

"We call the API and charge a margin" is not on that list.

What happens to these companies when the underlying models get cheaper, better, and more widely accessible? When OpenAI or Anthropic decides to move slightly down-market and offer the functionality these companies have been reselling at a premium? When the enterprise customers who were paying for a vertical-specific wrapper discover that the general-purpose model now does the same thing for a fraction of the cost?

They do not gradually decline. They fall off a cliff with the surprised expression of someone who has just been informed that gravity has always applied to them personally.

This is not the AI bubble bursting. This is natural selection. The distinction matters because the technology survives natural selection. Only the weak survive it badly.

The Hyperscaler Problem

Morgan Stanley's figures are genuinely alarming. Hyperscalers spending 34% of revenue on capital expenditure this year, 37% by 2028, exceeding the dot-com peak of 32%. These are very large numbers being spent on very large quantities of hardware, in advance of demand that has not fully materialised.

The bull case is that the demand will materialise, that enterprise AI adoption is following the usual S-curve and is about to hit the steep part, and that the infrastructure being built now will be fully utilised within three years. This is possible. It may even be probable.

The bear case is that the demand materialises more slowly than projected, that enterprises are still in the "we ran a pilot" phase rather than the "we have production workloads" phase, and that the capex being deployed now represents a bet that the market is not yet large enough to justify. Higgins' warning about profit expectation bubbles lives here.

Both cases can be true simultaneously for different hyperscalers. The one thing that is true for all of them, regardless of which case plays out, is that the compute has already been ordered.

Which brings us to the only company in this entire story that does not have a downside scenario.

The House Always Wins

There is a principle in gold rush economics that the people who reliably make money are not the prospectors. They are the people selling picks, shovels, and sturdy denim trousers to the prospectors. The prospectors may strike gold. They may not. The hardware sellers get paid regardless.

The AI gold rush has a pick-and-shovel seller. Its name is Nvidia.

Consider the structure of the competitive landscape. OpenAI and Anthropic are competing with each other. Google's Gemini and Meta's Llama are competing with both of them. DeepSeek arrived from China and briefly caused significant distress in financial markets by demonstrating that capable models can be trained for considerably less than the prevailing consensus had assumed. Mistral is doing interesting things in Europe. There are dozens of other model providers, fine-tuned variants, and specialised architectures all competing for the same enterprise budgets.

Every single one of them, to a first approximation, runs on Nvidia hardware.

The training compute runs on H100s and H200s and B200s. The inference infrastructure runs on the same. When DeepSeek demonstrated training efficiency that alarmed the market, Nvidia's stock dropped briefly as investors concluded this meant less demand for GPUs. What it actually demonstrated was that more organisations can now train capable models, which means more demand for GPUs to run inference at scale. The market corrected this misreading within a few days.

The CUDA ecosystem is the mechanism that makes this lock-in structural rather than accidental. CUDA is Nvidia's proprietary parallel computing platform, and the entire AI software stack has been built on top of it over fifteen years of careful cultivation. PyTorch, TensorFlow, every major training framework, the tooling, the libraries, the institutional knowledge of how to make models run efficiently: all of it assumes CUDA. AMD makes competing hardware. AMD's ROCm platform is a genuine and improving alternative. AMD is still fighting for scraps at the margins of a market that Nvidia built and still owns.

What This Means Practically

I want to be careful not to suggest that Nvidia is invulnerable. No company is invulnerable. The history of technology is substantially a history of invulnerable companies discovering that they were not. IBM was invulnerable. Microsoft was invulnerable at the browser layer. Intel was invulnerable on x86. Invulnerability in technology has a half-life, and anyone who tells you otherwise is selling something.

What Nvidia has, right now, is the most defensible position in the AI stack. The CUDA moat is real. The ecosystem lock-in is real. The manufacturing partnership with TSMC is real. The demand from hyperscalers who have already committed their capex budgets is contracted and real. If the AI bubble deflates in the expectations-and-valuations sense, Nvidia gets a hangover. If the AI bubble bursts in the catastrophic-correction-and-recession sense, Nvidia gets a nasty few years. But it does not go away. The infrastructure it sold does not disappear. The next wave of demand, when it comes, runs on the same hardware.

What this means for the market more broadly:

• AI model providers without genuine research differentiation will consolidate or fail. There will be two or three dominant models, not forty.
• API-wrapper companies without proprietary data or deep integration will be commoditised by the very models they depend on. This is already happening in some verticals.
• Hyperscalers will see margin pressure if demand ramps slower than projected, but their infrastructure position means they capture value when it does ramp.
• Nvidia captures upside in all scenarios except a genuine halt to AI development, which would require something considerably more dramatic than a valuation correction.

 

The Honest Summary

The article is right that valuations have compressed. It is right that profit expectations may be inflated. It is right that the gap between investment and proven return is concerning. It is not quite right that this constitutes a bubble bursting, because it conflates the fate of overpriced stocks with the fate of the underlying technology, and these are genuinely different things.

The AI companies that will collapse are the ones that mistake access to a model for possession of a moat. They are currently numerous and will become less so. This is not a tragedy. This is a market doing what markets occasionally remember to do.

The AI companies that will survive are the ones with proprietary data, genuine technical differentiation, deep operational integration, or the specific and enviable position of having built the infrastructure that everyone else depends on.

The universe of AI investing is complex, contested, and largely uncertain. But within it there is one position that is about as close to a sure thing as the technology sector produces. It does not matter which model you run. It does not matter which cloud you run it on. It does not matter whether the bubble inflates further or deflates next Tuesday.

Somewhere in a data centre, an H100 is processing a token.

Jensen Huang is having a very good decade.

All Posts