By Henry Wildes, Senior Business Editor
Microsoft, Alphabet, Amazon and Meta will spend more than $700 billion this year on data centers that train and run artificial intelligence, according to the companies’ own capital spending forecasts. Yet a study by Massachusetts Institute of Technology’s (MIT) Networked AI Agents in Decentralized Architecture (NANDA) initiative found that 95% of corporate AI pilot programs delivered no measurable effect on profit or loss, which is a problem for the revenue that eventually has to justify the $700 billion. Investors are no longer sure the spending will earn that money back. Their doubt hit prices this summer, when chip stocks shed more than $1 trillion of value in a single month, according to CNBC.
An AI stock can be one of three very different businesses. The companies that sell the hardware are already earning enormous profits. The companies that rent that hardware out are losing money on borrowed capital. The companies that build the models are mostly unprofitable and are priced on revenue they have not yet earned. The question of whether AI is a good investment depends largely on which of the three is being discussed.
First, hardware companies. Nvidia, now the most valuable public company in the world, designs the chips that train and operate AI models. In its most recent quarter, it reported $75.2 billion of data center revenue, 92% of its total, with revenue up 85% from a year earlier, The Motley Fool reported. Three customers account for 54% of that revenue, largely the same companies spending the $700 billion, and they pay for it with cash their older businesses already generate. Nvidia’s price is therefore a bet on that spending continuing, not on how well the models it trains actually sell. Thus, hardware sellers are the safest of the three types, since they are paid on delivery and are already profitable.
Renting that hardware is a far harder business. Companies such as CoreWeave buy chips, install them in data centers and rent the capacity out by the hour, borrowing enormous sums years before revenue arrives. According to Yahoo Finance, CoreWeave lost $740 million in the first quarter of 2026 even as revenue doubled to $2.08 billion, against $99.4 billion of contracted future revenue. The debt payments are fixed, while the revenue depends on customers who are mostly not yet profitable and may not even be solvent enough to pay in future years. Buying the stock means treating that backlog as money already earned. Renters are very exposed, because renting exists to cover short bursts of heavy training not covered by already owned infrastructure, and those bursts are the first thing cut when a lab slows down, leaving the debt behind.
The companies that build the models themselves are the hardest to value. They sell access through consumer subscriptions and through an application programming interface, the channel businesses use to run an outside AI model inside their own software. Yet neither channel has covered what it costs to build the models. OpenAI, valued at $852 billion after a $122 billion funding round in March, exited 2025 with roughly $20 billion of annual revenue against infrastructure commitments of about $600 billion through 2030, Forbes reported. Anthropic is the exception, telling investors to expect a $559 million operating profit in the second quarter, which would be the first profitable quarter any major model developer has posted. Pricing assumes revenue keeps growing and that costs per unit of output keep falling. Both have held so far, but the cost base resets every chip generation, since new chips mean training the next model from scratch, so each model has to pay for itself before it becomes obsolete. Model builders are so hard to value because they are mostly unprofitable, while subscriptions can be cancelled any month and hardware and rentals are sold under contracts that run for years.
It is worth noting that within the category of model builders, pure plays — meaning companies with no other business to fall back on — are far more exposed than diversified companies. AI failing to pan out financially would be a setback for a diversified company like Meta, but it would cost a pure play like OpenAI its existence.
For now, not all of this is information a retail investor can act on. Nvidia and CoreWeave can be bought in any brokerage account, but OpenAI, Anthropic, and many others cannot. That will change soon, however. Anthropic’s May funding round valued it at $965 billion on a $47 billion annualized revenue run rate, up from about $10 billion in annual revenue in 2025, CNBC reported, and both Anthropic and OpenAI confirmed confidential filings with the Securities and Exchange Commission in June. Yet, recent listings suggest that access may not be worth very much. Cerebras, which builds an alternative to Nvidia’s chips, went public in May at $185 a share, opened at $350 and had fallen to $226.72 by June 23, CNBC reported. SpaceX, which merged with Elon Musk’s AI venture xAI in February, fell below its offering price within a month of listing.
None of this addresses whether AI itself is overhyped, because a technology can be revolutionary while its stocks are still poor investments. Railroads, telecom fiber, and the early internet all rearranged the economy while destroying most of the capital that built them. AI investments are unlikely to be a total loss, but most of today’s valuations will probably not survive intact, and the three categories will not all land in the same place. When Anthropic and OpenAI eventually list, public investors will price a pure model developer with no diversified parent to hide behind. That will be the first real test of a business model no one has priced before. Until then, being right about a technology and being right about its stock price remain two separate bets.
Photo Credit: Unsplash