The Great Sorting

The market is sorting AI spenders from receivers.

For three years, the AI trade rewarded breadth over precision. Nvidia, obviously, but also its suppliers, its customers, the cloud providers building data centers, the startups training models, the software companies adding “AI-powered” to their product descriptions—all of them rose together. Capital flowed to anything adjacent to the boom.

That pattern broke sometime in the fourth quarter of 2025. What has emerged in its place is a market that actively sorts winners from losers, and the sorting has been severe.

The clearest illustration involves two companies that both rode the AI wave, both made substantial bets on the infrastructure buildout, and have since diverged dramatically.

Two Tickers, Opposite Trajectories

Oracle’s decline has been swift. The stock is down 43% from its September peak, on pace for its worst quarter since the dot-com bust in 2001. Bondholders have filed a lawsuit alleging the company concealed plans for substantial additional borrowing. Credit default swaps on Oracle debt have hit their highest levels since 2009.

The catalyst was Oracle’s bet on becoming the infrastructure backbone for OpenAI. The Stargate project, announced at the White House in January with President Trump and Larry Ellison in attendance, committed Oracle to building $300 billion worth of data centers for ChatGPT’s maker. The stock soared, and Ellison briefly became the world’s wealthiest person.

Then investors examined the details. Oracle’s capital expenditure plans expanded to $50 billion for fiscal 2026, 43% higher than previously announced and double the prior year. Free cash flow went negative by more than $10 billion in a single quarter. The company disclosed $248 billion in additional lease commitments not reflected on its balance sheet, mostly for data centers. Reports emerged that some Stargate projects were being delayed by at least a year due to labor and material shortages.

Morningstar analyst Luke Yang observed that Oracle has “very little room for error” to execute its strategy.

Micron presents the opposite case. The memory chipmaker is up over 250% in the past year. Its stock reached all-time highs in January. On January 16th, the company broke ground on a facility in Clay, New York, one of the largest private investments in state history, backed by $6 billion from the CHIPS Act.

The difference is structural. Micron is not spending to chase AI demand; Micron is what the demand looks like. Every AI accelerator requires high-bandwidth memory, and there is not enough of it. The company’s entire HBM output is sold out through 2026. DRAM prices are rising 20-25% quarter over quarter. CEO Sanjay Mehrotra told CNBC that Micron can only meet “50% to two-thirds” of demand from several key customers.

Oracle is spending. Micron is receiving. The market has decided that distinction matters.

The Circular Financing Question

Part of what is driving the new scrutiny is a growing discomfort with how AI deals are structured.

The web of relationships around OpenAI illustrates the pattern. Nvidia has announced plans to invest up to $100 billion in OpenAI, which OpenAI will use to purchase Nvidia chips. AMD has offered OpenAI warrants for up to 160 million shares in exchange for a commitment to deploy 6 gigawatts of AMD hardware. Oracle is building data centers for OpenAI, funded partly by debt that bondholders now claim they were not properly warned about. Microsoft has invested $13 billion in OpenAI, which then spends most of it on Azure. Nvidia holds a 7% stake in CoreWeave, which provides cloud services to OpenAI, which has expanded its CoreWeave deals to $22.4 billion.

This is circular financing: investment flows from Company A to Company B, then returns to Company A through purchases. The practice is not illegal, and it is not hidden. Anyone can trace the deals. But it creates a self-reinforcing loop that can make demand appear stronger than underlying fundamentals would suggest.

Neuberger Berman analyst Jamie Zakalik told CoStar News that “the crux of the concern is whether Nvidia is juicing demand for their own equipment that isn’t real or sustainable.” Bernstein’s Stacy Rasgon, in a note following Nvidia’s OpenAI investment announcement, wrote that the deal “will clearly fuel ‘circular’ concerns.”

The comparisons to the dot-com era follow naturally. In the late 1990s, telecom companies extended billions in financing to internet startups purchasing their equipment. Cisco, Lucent, and Nortel all participated. When the bubble burst, many of those loans went bad and the lenders collapsed. The most problematic cases involved “revenue roundtripping,” where companies paid each other for services of exactly equal value to hit quarterly targets.

The current situation differs in important respects. The major AI infrastructure players have stronger balance sheets than the telecoms of 2000. The “Big Four” tech firms are expected to generate $203 billion in free cash flow in 2025, meaning most capital expenditure is funded by operations rather than debt. And there is genuine, measurable demand for AI services, unlike the speculative projections that underpinned the fiber-optic buildout.

But the circular deals do complicate the picture. When Nvidia is simultaneously chip supplier, investor, and cloud capacity buyer across multiple counterparties, separating real demand from financially-engineered demand becomes difficult. That uncertainty helps explain why investors are now differentiating, rewarding companies with clear, direct revenue streams while scrutinizing those whose fortunes depend on interlocking commitments being fulfilled.

Downstream Effects

This newfound caution has extended beyond public markets. Venture investors report that indiscriminate AI funding has dried up.

Lightspeed partner Guru Chahal told Quartz that “the market’s gotten disciplined,” adding that founders now “need real product moats or genuine go-to-market advantages, not just access to an API.”