The AI Industry Is Becoming an Infrastructure Industry

OpenAI discontinued its own video-generation software last spring. Sora was a working product — the kind of thing a software company ships to make money. OpenAI killed it anyway, because the GPUs running Sora were needed for something else: training GPT-5.4.

OpenAI’s decision to exit the video generation business and to shift its priorities elsewhere. — The Walt Disney Company

A software company chose hardware over software. It was not a close call. Sam Altman made the logic explicit not long after.

We are taking a forward bet that it will continue to grow, and that not only will ChatGPT keep growing, but we will be able to become one of the important AI clouds, that our consumer device business will be a significant and important thing, that AI that can automate science will create huge value. — Sam Altman

That is not a slogan. It is an admission. The CEO of the company that launched the AI boom is telling you his firm is no longer in the business you thought it was in. The same retreat is happening across the stack, in different words. Meta signed a 20-year contract for the entire output of a nuclear plant in Clinton, Illinois ^[1]. Microsoft is restoring the reactor at Three Mile Island ^[2]. Jensen Huang now describes Nvidia in terms no software CEO would use.

Nvidia is an "AI factory" that converts electricity to tokens, the core unit of AI output. — Jensen Huang

Oracle is "transitioning into an AI infrastructure provider" ^[3]. Meta is "pivoting toward AI and neocloud services" ^[3]. Every major player is saying the same thing in a different register: the business is no longer selling intelligence. It is selling the physical capacity to produce it. The financial system is being rebuilt to match. Nvidia has backstopped up to $125 billion in residual GPU values to convince institutional investors that AI compute is durable collateral. Larry Fink compared the move to the emergence of mortgage-backed securities in the 1970s.

We ourselves can backstop up to $125 billion in deals, 25% of what the deal could be. — Nvidia

Nvidia has signed agreements with six Wall Street firms — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — to mobilize $500 billion in third-party capital for AI infrastructure.

NVIDIA has reached an important milestone. We began by building chips; today, we are helping create a new class of productive, investable infrastructure: AI factories. — Nvidia

The chipmaker is securitizing GPUs as infrastructure. The model is not the asset. The hardware that runs it is. The numbers make the pattern impossible to miss. The companies that own the physical layer are capturing the value: Micron's revenue grew 196% year over year, Sandisk's 251%, TSMC's 39% ^[4]. CoreWeave, which rents GPU compute, has a $100 billion revenue backlog; Nebius reported 684% revenue growth ^[5]. The companies building the models are losing money. OpenAI expects $14 billion in losses this year on $20 billion in revenue ^[6]. Alphabet's AI spending is burning $490 million a day, producing its first negative quarter since going public ^[7]. The mechanism is straightforward: OpenAI's CFO has warned the company may struggle to fund $600 billion in compute commitments without accelerated growth ^[8].

In hindsight, that caution looks less like discipline and more like underestimating how fast demand would arrive. — OpenAI

A study found that 95% of businesses investing in AI have failed to make money on a combined $40 billion in spending ^[9].

Despite high-profile investment, industry-level transformation remains limited. — Massachusetts Institute of Technology

The software revenue is not covering the infrastructure commitments. So the companies are becoming infrastructure companies instead. Microsoft's Copilot generates $37 billion in annual software revenue, and Indeed used AI to more than double its stock price ^[10][11]. But Microsoft's real growth is in Azure cloud — infrastructure revenue — and Indeed's AI is a cost-saving tool, not a model-selling business. The exceptions are real, but they do not reverse the direction of travel. What is emerging is a reorganization of the entire industry around a single insight: the AI model is becoming a loss leader. It is sold at a loss to drive demand for the compute, the power, and the floor space underneath it. The companies winning the AI race are not AI companies. They are power companies, chip foundries, and landlords of compute. The intelligence is the thing that gets customers in the door; the physical asset is what they pay for. Investors are beginning to discriminate between AI spenders on precisely this basis — rewarding those with infrastructure positions and punishing those without ^[12][13] — and the financial system is beginning to price in the difference.