Thirty-six billion dollars separates the highest and lowest published estimates of Anthropic’s revenue as of May 2026. That gap is the story.
Not because one figure is wrong. Because the AI industry’s revenue disclosure norms haven’t caught up with the business model’s complexity, and investors who don’t understand what they’re reading will price Anthropic’s $965B valuation using the wrong denominator.
The Three Numbers and Their Sources
Start with what was actually said and by whom.
Anthropic states its annualized run-rate revenue crossed $47B in May 2026, per reporting attributed to Reuters. This is a vendor-stated figure. Anthropic made the disclosure. Reuters reported it. The original Anthropic source URL is inaccessible for independent confirmation. The figure should be read as “Anthropic states”, not “Anthropic confirmed independently.”
The Wall Street Journal reportedly estimated Anthropic’s net revenue at approximately $10.9B. The WSJ source is dead; this figure hasn’t been independently confirmed beyond the Wire’s reporting. It should be attributed explicitly: “The Wall Street Journal reportedly estimated.” It is not a competing figure to $47B, it’s measuring something fundamentally different.
One analysis estimated approximately $33B, according to The Information. The methodology wasn’t fully disclosed in the available excerpt. This figure lands between the other two, which is consistent with a recognized revenue estimate that captures more than net but less than gross ARR.
Three sources. Three methodologies. The discrepancy is a feature of how revenue gets measured in this business model, not evidence of misreporting.
What ARR Measures, and What It Misses
Annualized run-rate revenue is SaaS accounting’s most useful growth signal and its easiest figure to misread simultaneously.
The mechanics: take a recent period’s contract bookings and API billing activity, annualize it (multiply by 12/months in period), report the result. The figure tells you how fast revenue is growing. It doesn’t tell you how much revenue is in the bank.
For Anthropic, $47B ARR means the company’s recent billing activity, if held at the same rate for 12 months, would produce $47B in gross contract and API revenue. Two things to note. First, “gross” matters here. ARR typically captures top-line contract value before deducting the costs of delivery. Second, “annualized” means it’s a projection, not a realization. A company growing at Anthropic’s reported pace will have ARR figures that consistently run ahead of actual trailing-twelve-month revenue.
Evidence
Evidence
This isn’t unique to Anthropic. Every hypergrowth SaaS company reports ARR ahead of recognized revenue. What’s different for an AI lab is the size of the deduction between gross and net.
The Cost Structure That Creates the Gap
A conventional SaaS company’s cost of revenue, the direct costs of delivering its software, might run 20-30% of gross revenue. The margin structure is favorable by design: software scales without proportional cost increase.
Anthropic’s cost structure doesn’t follow that model. Claude runs on AWS infrastructure. Every API call consumes compute. At the scale implied by the confirmed 5 GW Amazon agreement and the reported Google and Broadcom TPU arrangements, covered in depth in the infrastructure-focused brief in as of publication, the compute cost embedded in each unit of revenue is substantial.
Revenue-sharing arrangements compound the effect. When hyperscalers are both investors and infrastructure providers, commercial terms typically include some form of revenue alignment. The $15B hyperscaler commitment in the Series H, including $5B from Amazon, didn’t come without commercial conditions. Those conditions show up in the gap between gross ARR and net revenue.
The WSJ’s reported $10.9B net estimate, if accurate, implies roughly 77% of Anthropic’s gross ARR is consumed by infrastructure costs and revenue-sharing arrangements. That sounds alarming. It’s worth calibrating. For a company training frontier models and serving API traffic at scale, early-stage gross margins in the 20-30% range are consistent with what’s been reported for comparable businesses. Anthropic’s margin structure is a scaling problem, not a business model problem, the expectation is that margins improve as compute costs decline and revenue grows faster than infrastructure spend.
The ~$33B Middle Estimate
The Information’s approximately $33B figure likely captures recognized revenue, closer to what an accountant would put on an income statement than either ARR or net margin estimates would suggest. Recognized revenue under GAAP smooths out the timing distortions in ARR (you can’t book a 12-month contract as 12 months of revenue in month one) while capturing more than the net-of-costs figure the WSJ reportedly used.
Without access to the full methodology, The Information is paywalled, this reading is inferential. The placement of the figure between ARR and net revenue is consistent with recognized revenue as a methodology. It’s informational context, not a verified figure.
Why This Matters for Pre-IPO Pricing
TJS has covered how frontier lab economics produce dramatically different pictures depending on which revenue line you examine. Anthropic’s three-figure situation is the clearest illustration of that dynamic yet.
Analysis
Frontier AI labs now have a structural incentive to disclose ARR over net revenue in pre-IPO periods: ARR tells the growth story at its most compelling, while net revenue exposes margin compression that early investors would rather not highlight. This isn't unique to Anthropic. It's the disclosure architecture of the category. The S-1 will force the reconciliation that pre-IPO coverage can't.
What to Watch
At $965B post-money valuation:
– 20x ARR (on $47B), aggressive, but in range for a hypergrowth AI lab with dominant enterprise market share
– ~29x on ~$33B, still aggressive, defensible on a high-growth SaaS comp set
– ~88x net revenue (on $10.9B), outside the range of most precedent transactions; only justifiable if the market is pricing significant margin expansion into the future
Investors are pricing some combination of all three, weighted by how much they believe the net margin will expand before an IPO. The management team is presenting $47B because it tells the growth story. The investment thesis lives in what the margin trajectory looks like by the time an S-1 hits.
The $965B valuation has been contextualized against OpenAI extensively. The more useful frame is the implied margin expansion required to justify it. If Anthropic reaches 50% gross margin on a $100B ARR base, plausible within three to four years if compute costs continue declining, the current valuation is rational. If margins stay compressed at current levels, the valuation is pricing a growth story that the unit economics don’t yet support.
The Investor Framework
Use ARR for growth trajectory assessment, it’s the right tool for measuring how fast the revenue engine is growing. Use net revenue for unit economics and margin analysis, it’s what tells you whether the growth is profitable. Use the gap between them to assess infrastructure cost exposure, the larger the spread, the more the business depends on compute cost reduction to reach attractive margins.
Anthropic’s disclosed ARR is vendor-stated and should be labeled as such in any model. The WSJ’s net revenue figure is reportedly estimated and single-source. The Information’s ~$33B figure is paywalled and methodologically opaque. None should be treated as audited fact.
The only audited figures will come from a prospectus. Watch the S-1 filing. That’s when all three numbers have to reconcile against a single income statement line, and the question of which figure actually measures Anthropic’s economics gets answered on the record.