Anthropic's ¥6.6 Trillion Loss Is Actually Good News

The Loss That Isn’t a Loss

Anthropic reported a $42 billion loss for 2025. On paper, it ranks among the worst annual losses ever recorded by a U.S. company—roughly matching Citigroup’s $29.1 billion write-down during the 2008 financial crisis. The headline number alone would sink any other firm.

But Anthropic didn’t spend $42 billion. It spent roughly $8 billion on operations, according to the prospectus that leaked recently. The remaining $34 billion is an accounting fiction born from the mechanics of convertible debt—and it points to a far more interesting structural story about who really controls the future of AI.

The money comes from two sources: Google, which began investing in Anthropic through convertible bonds in mid-2023 with an initial $2 billion commitment, and Amazon, which entered in late 2024 at a $46 billion valuation with an $8 billion check. Both investments are structured as convertible debt, meaning they can later be swapped for equity in Anthropic.

Under U.S. GAAP accounting rules, when the value of the shares these bonds can convert into rises, the issuer must record the increase as a loss on its income statement. That’s it. No cash left the building. No equipment was destroyed. The loss simply reflects that Anthropic is now worth dramatically more than when Google and Amazon wrote their checks.

Joshua Ronen, an accounting professor at NYU’s Stern School of Business, put it plainly: “This loss is the flip side of good news. The valuation surged so sharply that the loss ballooned to this magnitude.”

Anthropic is reportedly targeting a $2 trillion valuation for an upcoming IPO. If that materializes, the accounting losses will keep mounting—not because the company is failing, but because its investors are winning too much, too fast, and the books have to reflect it.

The Cloud Provider’s Hidden Leverage

The real pattern here isn’t accounting. It’s power.

Google disclosed $124.3 billion in startup investments across its latest earnings, noting that most of that is concentrated in a single position. In July, the company reported a $99 billion gain from investments in SpaceX and “another non-public startup”—widely understood to be Anthropic—during its April to June 2026 quarter. The New York Times reported in March 2025, citing court documents, that Google holds approximately 14 percent of Anthropic’s equity.

This is not a passive investment. Google is simultaneously Anthropic’s largest backer, its largest customer for cloud compute, and its most direct competitor in generative AI. The same infrastructure that powers Claude’s training runs also powers Gemini’s. The convertible bond structure gives Google asymmetric upside: if Anthropic succeeds, Google captures enormous returns; if it struggles, Google’s position as Anthropic’s cloud provider creates dependency that limits what Anthropic can do.

Amazon has played the same hand with a $8 billion convertible investment at a $46 billion valuation in late 2024. The parallel is striking—both cloud giants are using debt instruments to build strategic positions in frontier AI labs while extracting preferential access to compute capacity.

The Dependency Trap

The dependency goes both ways. Anthropic cannot train frontier models without massive GPU clusters, and those clusters are controlled almost entirely by the same companies investing in it. Every additional cycle Anthropic buys from AWS or Google Cloud deepens its reliance on players who could, at any point, raise prices, deprioritize its workloads, or accelerate their own competing products.

The convertible bond structure amplifies this dynamic. Each time Anthropic’s valuation climbs toward that rumored $2 trillion IPO target, the accounting loss on Google and Amazon’s convertible positions grows larger. The company’s balance sheet looks increasingly damaged even as its market position strengthens. Any misstep—delayed product launches, competitive pressure from Gemini or Amazon’s models, regulatory headwinds—could reverse the valuation trajectory and leave Anthropic with suddenly沉重的 debt obligations at unfavorable terms.

Revenue tells part of the story. At $4.6 billion in 2025, Anthropic is generating meaningful income but nowhere near enough to fund its compute ambitions independently. The gap between what it earns and what it needs to stay competitive at the frontier is being bridged entirely by external capital from the very firms that stand to benefit most if those labs falter.

What Happens Next

Several scenarios emerge from this structure.

If Anthropic reaches its $2 trillion IPO target, Google and Amazon’s convertible bonds convert into equity at terms that likely favor the cloud giants—giving them board seats, data access, and preferred compute pricing that competitors cannot match. The IPO itself would be the moment of greatest vulnerability, as the company transitions from private investor support to public market scrutiny while still burning cash on compute.

If the IPO target falters, the convertible bonds remain debt. Anthropic would face repayment obligations despite operating at an $8 billion annual loss, and its cloud-provider relationships could shift from partnership to extraction. Google and Amazon have every incentive to keep Anthropic dependent—dependency ensures revenue from cloud services and strategic influence over frontier AI development.

The broader implication extends beyond Anthropic. Any independent AI lab that cannot build its own compute infrastructure faces the same trap: capital from cloud providers creates upside capture for those providers and structural dependency for the lab. The $42 billion loss is not Anthropic’s crisis. It is the accounting signature of a new era in which the firms that control chips and clouds also control who gets to define what AI does next.