Anthropic Says It Will Be Profitable. The Math Does Not Look Friendly.
Anthropic's "Profitability" Is a Two-Month Accounting Artifact
On May 6, the Wall Street Journal ran a straightforward milestone story: Anthropic was on track to post its first operating profit, $559 million in the June quarter, on revenue that would more than double to $10.9 billion. The framing was uncritical and celebratory. A startup that spent years burning billions was finally making money. The story spread everywhere.
The framing is misleading in nearly every dimension.
The $559 million profit rests on a compute contract discount that covers exactly May and June, the two months of the projected quarter. According to SpaceX's S-1 filing, Anthropic is paying $1.25 billion per month for access to the Colossus cluster. At full rate, that is $15 billion per year. The discount during the May-June ramp-up period is what makes the profit number exist. Once the contract moves to full-rate billing, the economics change substantially. This is not a structural profit. It is a two-month window dressed up as a milestone.
The scale of the underlying infrastructure bet is also worth sitting with. Analyst Ed Zitron estimates Anthropic's total compute spend at roughly $45 billion per year across SpaceX, Amazon Web Services, Google Cloud, and Microsoft. Against that number, $559 million is 1.2 percent of annual infrastructure cost. The Journal described this as "mind-blowing growth." It is more accurately described as a rounding error on a very large burn rate.
Anthropic is not subject to public company accounting requirements. It is unclear what methodology it uses to book revenue and costs. The Journal acknowledged this, which is more than most coverage managed.
The harder number to reconcile is the one Anthropic's own CFO certified under oath.
Krishna Rao filed an affidavit in federal court on March 9 in Anthropic's lawsuit against the Department of Defense. In that filing, he stated that Anthropic's total revenue "to date" was "exceeding $5 billion." That is every dollar the company has ever recognized from its founding through March 9, 2026. The filing is on record at CourtListener.
Six weeks later, on April 6, Anthropic announced it had reached $30 billion in annualized recurring revenue.
Those two numbers cannot be reconciled through ordinary growth. To make them consistent, you would have to believe that Anthropic processed roughly 85 percent of its entire lifetime revenue in the approximately 100 days between March 9 and mid-April. That is not how software businesses operate.
The arithmetic problem has been documented in detail by Zitron and analyst Davi Ottenheimer at flyingpenguin. They compiled Anthropic's own ARR announcements from January 2025 through March 2026. The table is straightforward:
January 2025: $1 billion ARR
March 2025: $1.4 billion ARR
March 2025: $2 billion ARR
May 2025: $3 billion ARR
July 2025: $4 billion ARR
July 2025: $5 billion ARR
October 2025: $7 billion ARR
December 2025: $9 billion ARR
February 12, 2026: $14 billion ARR
March 3, 2026: $19 billion ARR
Dividing each ARR figure by twelve to estimate monthly recognized revenue, then adding across the full period with conservative gap estimates, produces a cumulative implied revenue of roughly $6.66 billion through early March 2026. Rao's court filing says the actual figure exceeds $5 billion. The gap is not a rounding error. Zitron's estimate is a floor, not a ceiling, because it uses the lower bound for each interval. If you interpolate between announcement dates rather than using the lower figure, the implied cumulative revenue climbs higher.
There is one charitable reading of the discrepancy. ARR is a forward-looking contract metric. It can represent signed agreements for future service, not money already earned. Rao's "revenue to date" likely means recognized revenue, the money Anthropic has actually earned under accounting rules. Under that reading, Anthropic has billions in contracts where service has not yet been delivered, which would make the ARR look large while recognized revenue stays lower.
If that is the explanation, Anthropic was reporting unearned contract value to the press as though it were operating revenue, while reporting actual earned revenue to a federal court. That is not an accounting distinction. It is two different narratives tuned for two different audiences.
The contrast is not subtle. ARR is presented as a growth metric to investors and press. The court filing was an opportunity to state the company's commercial position under penalty of perjury, and Rao used the more conservative figure. In a filing where Anthropic was trying to demonstrate commercial scale to a federal judge, the CFO said "exceeding $5 billion" rather than "approaching $6 billion" or "nearly $7 billion." That word choice tells you where the real number sits.
Anthropic also has a pattern of loosening rate limits immediately after announcing major infrastructure deals, which suggests the constraints were negotiable rather than physical. The company announced the SpaceX partnership and simultaneously raised usage limits. The Information reported in January 2026 that Anthropic had missed its gross margin projections, with inference costs running 23 percent above what the company had modeled. That miss predates the SpaceX deal and predates the "profitability" narrative.
The broader context matters here. The AI infrastructure build-out is real, and the compute costs are real, and the revenue growth is real. But "real" in the sense of durable, structural, and independently verifiable is different from "real" in the sense of a press release. The $559 million operating profit is a projected figure built on a specific two-month discount window. The $30 billion ARR is a forward-looking contract metric that may include revenue not yet earned. The $5 billion in court filings is a backward-looking certified fact. These are three different things being reported as though they are the same thing.
Every headline that reported annualized revenue as though it indicated actual business scale was describing a speedometer, not an odometer. Every valuation model built on ARR trajectory was fed inputs that may not mean what they appear to mean. If the reader takes one number away from this, it should be this: $559 million in profit against $45 billion in annual infrastructure spend is not a milestone. It is a discount window.