Three numbers. Same company. Same month.
Anthropic states its annualized run-rate revenue crossed $47B in May 2026, per reporting attributed to Reuters. The Wall Street Journal reportedly estimated Anthropic’s net revenue at approximately $10.9B. One analysis estimated approximately $33B, though the methodology wasn’t fully disclosed, according to The Information. The gap between the highest and lowest figures is $36B. That’s not a rounding error. It’s a measurement question.
The real story is what each figure is actually counting.
Gross annualized run-rate revenue, the ARR figure, takes recent contract and API billing activity, annualizes it, and reports the result. It’s how SaaS companies signal growth trajectory. It’s technically accurate. It’s also a ceiling, not a floor: it reflects what revenue *would be* if the current period’s billing rate held for 12 months. For a company growing as fast as Anthropic reportedly is, ARR runs ahead of realized revenue by design.
Net revenue removes the costs that don’t show up in gross ARR. For an AI lab running at Anthropic’s scale, those costs are substantial. Claude’s API delivery runs through AWS infrastructure. Revenue-sharing arrangements with hyperscaler partners apply. The compute cost structure embedded in every API call is different from a software company’s cost of goods. When the Wall Street Journal reportedly estimates $10.9B in net revenue, it’s estimating what’s left after those deductions.
The ~$33B figure from The Information likely reflects something in between, an estimate of recognized revenue that captures more than net but less than gross ARR, though without the full methodology it’s not possible to confirm exactly what it measures.
None of these figures is wrong. They’re different lenses. An investor evaluating growth trajectory uses ARR. A credit analyst or M&A team uses net revenue. A comparable-company analysis depends on which metric the comparable company reports.
This matters specifically for Anthropic because the gap is larger here than in traditional SaaS. A typical enterprise software company’s cost of goods is a small fraction of revenue. Anthropic’s compute costs, at the scale implied by the Amazon 5 GW agreement and the reported Google/Broadcom TPU arrangements, are a substantial fraction of gross revenue. The ARR-to-net-revenue spread reflects that structure.
Analysis
The $36B gap between Anthropic's stated ARR and the WSJ's net revenue estimate is larger than the entire Series H rounds raised by most frontier labs. It reflects a structural feature of compute-intensive AI businesses, not a disclosure error. Investors who use ARR without understanding the compute cost deduction are pricing a different company than the one that will show up in an S-1.
The $36B gap between $47B and $10.9B isn’t a scandal. It’s a disclosure architecture. Anthropic disclosed ARR because ARR tells the growth story it wants investors to hear before an IPO. TJS has tracked how frontier lab economics diverge dramatically depending on which revenue line you examine. Anthropic’s figures are consistent with that pattern.
The catch is that pre-IPO investors pricing a $965B valuation need to know which number to multiply. A $965B valuation on $47B ARR implies a ~20x ARR multiple, aggressive but defensible for a hypergrowth AI lab. On $10.9B net revenue, it’s ~88x. Those are different investment theses.
Watch the S-1 filing for the first standardized revenue disclosure. That’s when the $47B, $10.9B, and $33B figures all have to reconcile against a single audited line item.