Field Note

CoreWeave and the Rashomon Market

How one AI infrastructure company can become four different investments.

Entity: InflectAI, Inc.
Published: May 23, 2026
Status: current

Ever since CoreWeave went public at $40 in March 2025, it has been a study of two divergent realities.

By May 2026, the stock was trading above $100. Early investors had been richly rewarded. At the same time public short-interest data showed CoreWeave short interest rising from roughly 26.2 million shares at the end of October 2025 to about 50.7 million shares by February 2026. By March and April, reported short interest was above 60 million shares. Depending on the denominator used, that represented roughly 11 percent of shares outstanding or about 14 percent of public float. A skeptical pool was forming around the same company that equity investors were rewarding.

In Rashomon, Akira Kurosawa's film, the same event is retold from four different perspectives, each producing a different story. The market has its own version of Rashomon, once you know where to look. A company reports numbers. Different professional pools notice different things because their lenses are different. Those lenses are not easily swapped out. They are tied to how each group sees itself. Equity investors ask how large the company can become. Bond investors ask whether the company can pay back its debts. Short sellers look for fractures in the story. Customers and vendors ask whether the company can deliver what it promised.

The Equity Lens

The CoreWeave equity holder's lens is growth. Growth looks good.

CoreWeave revenue increased from $229 million in 2023 to $1.9 billion in 2024, then to $5.1 billion in 2025. First-quarter 2026 revenue was $2.1 billion, up 112 percent year over year. Revenue backlog was $66.8 billion at the end of 2025 and $99.4 billion by March 31, 2026. Adjusted EBITDA was $3.1 billion for 2025 and $1.2 billion for the first quarter of 2026. Meta signed multiple agreements in early 2026, including a new $21 billion commitment. Anthropic signed a multi-year agreement. Nvidia signed a $6.3 billion order to purchase unsold capacity through 2032.

If this is the first instrument you pick up, CoreWeave looks like one of the few direct ways to own scarce AI infrastructure. Revenue is growing. Backlog is enormous. Major customers are signing. Nvidia is aligned. Compute is the bottleneck. The equity story does not require mystery. It has receipts.

The Short Seller Lens

The short seller's lens is fragility.

A short seller can look at the same company and begin somewhere else. CoreWeave had $9.7 billion outstanding under delayed draw term loan facilities at the end of 2025. It issued $6.4 billion of notes in 2025, including senior notes due 2030, senior notes due 2031, and convertible senior notes due 2031. It also had $5.2 billion of OEM and software financing arrangements. First-quarter 2026 net loss was $740 million. Interest expense was $536 million for the quarter. Capital expenditures were $6.8 billion.

Those numbers do not erase the revenue growth. They make the short seller ask a different question. Not whether AI compute demand exists, but what kind of obligation CoreWeave has taken on to supply it. Whether GPU economics decay faster than the accounting schedule. How much of the backlog survives pricing pressure, hyperscaler self-build, customer concentration, refinancing cost, and Nvidia's next product cycle. Kerrisdale's public short report made that case directly: CoreWeave was not a differentiated cloud platform, in its view, but a heavily levered GPU rental business with weak pricing power and customers who may treat it as a bridge rather than a permanent strategic partner.

The Credit Lens

The credit investor's instrument is repayment.

Credit does not need the same story equity needs. A bondholder does not need CoreWeave to become the next great cloud platform. A bondholder needs to get paid. That means the important questions are collateral, contracted cash flows, counterparty quality, maturity, covenants, and refinancing access.

CoreWeave's debt financings were not floating in the air. Its delayed draw term loan facilities were collateralized by assets tied to contributed contracts and pledged contractual cash flows, generally from investment-grade counterparties. Its committed contracts often reserve capacity over two-to-five-year terms on a take-or-pay basis. In May 2025, CoreWeave priced $2.0 billion of 9.25 percent senior notes due 2030, upsized from an initially announced $1.5 billion offering. CNBC reported that CoreWeave shares rose 19 percent after the debt offering was announced.

That reaction makes sense if the market interpreted the debt raise as evidence that financing was available. The credit pool was not making the same claim as the equity pool. It did not have to say the upside was safe. It only had to say the cash-flow structure was financeable.

The Customer Lens

The customer and partner lens is capacity.

Microsoft, Meta, Anthropic, and Nvidia do not have to underwrite CoreWeave the way a public equity investor does. They have a more immediate problem. Can CoreWeave deliver AI compute capacity fast enough, in the right places, under commitments that make operational sense?

CoreWeave's platform grew from 10 data centers and about 70 megawatts of active power at the end of 2023 to 32 data centers and about 360 megawatts at the end of 2024. By the end of 2025, it operated 43 data centers with more than 850 megawatts of active power. It had about 3.1 gigawatts of contracted power capacity at the end of 2025 and said in the first quarter of 2026 that it had surpassed 1 gigawatt of active power.

For a customer trying to train or deploy frontier models, the relevant question may not be whether CoreWeave's equity is overvalued. The question may be whether waiting is more expensive than contracting. Capacity has its own logic. So does redundancy. So does speed.

When The Accounts Have To Reconcile

The equity holder can look at CoreWeave and see scarce growth. The short seller can look at CoreWeave and see levered duration risk. The credit investor can see contracted repayment. The customer can see capacity under pressure. Each group has its own first question, and each first question makes part of the company legible while making another part easier to ignore.

This is why the usual debate over whether CoreWeave is "good" or "bad" misses the more interesting structure. The same company can support multiple serious accounts because the professional pools are not using the same measurement instrument.

The divergence can last longer than feels comfortable because each pool has enough evidence to keep its own story intact. It becomes unstable when one fact begins to matter across the pools. That fact could be GPU rental pricing. It could be utilization. It could be a major customer slowing spend. It could be refinancing cost. It could be whether backlog converts into cash after interest expense, capex, and depreciation are taken seriously. It could be a new Nvidia cycle that changes the economic life of older chips. It could be a credit-market repricing that makes future capacity more expensive to finance.

Until then, each pool can keep finding confirmation in its own numbers. Once that cross-pool fact arrives, the accounts have to reconcile.

The October 2025 short interest gives us the first timestamp for the skeptical pool. The February 2026 short interest shows that the pool had grown. The stock price shows that equity belief had not collapsed. The debt raise shows that credit belief had not disappeared. The customer commitments show operational demand.

Taken separately, each fact can be explained inside its own pool. Taken together, they show a company whose market identity had not settled.

CoreWeave was not one story. It was several stories occupying the same ticker.