Debate Erupts Over Anthropic's Alleged Financials, Losses and Compute Split

Alleged Anthropic financial and compute figures triggered intense estimation and debate on X. @PythiaR (relayed by @masayoshisson) calculated ARR of $75B with 45% QoQ growth, positive adjusted operating profit, and $75B total compute spread over 8.5 years ($8.8B/year), with only 25% of compute on inference. Finance-background @menhguin argued that despite flat Q3 revenue, margin expansion is the key signal, with Opus 5.5-driven acceleration not yet reflected.

Confirmed

All figures are X users' estimates based on (alleged) financials, not official company disclosure
@AIandDesign noted $72B of the $80B loss came from non-cash accounting items (mainly stock-compensation treatment), not actual cash outflow; annual revenue grew from $1.1B to $16.3B
@EZE3D acknowledged the $72B non-cash point
@menhguin argued that if $75B goes to pre/post-training vs $25B to inference, rivals must spend comparable training capital to catch up, shortening the list of potential competitors

Unconfirmed

@EZE3D questioned whether the $1.1B-to-$16.3B revenue growth itself involves heavy accounting manipulation—unverified
@AIandDesign said he is unsure whether Anthropic survives in current form, but would not bet against it

Why it matters

The dispute is essentially whether massive paper losses or rapid revenue growth better reflect a frontier AI lab's true health
The 75% training / 25% inference split is read by some as a danger signal and by @menhguin as a moat raising the entry bar to tens of billions per year
@AIandDesign called AI the biggest thing since the steam engine, while @retr0jirachi slammed the $80B loss and backed John Carmack or Rich Sutton to reach AGI first—illustrating polarized views on frontier-lab burn rates

2026-10-21 ~ 2026-10-21 · 9 related posts

Primary sources

Anthropic math: $750B of compute ≈ $8.8B/year, $75B ARR up 45% Q/Q — PythiaR · 2026-10-21 03:55 (UTC+08:00)
Defending Anthropic's $80B loss: $72B was non-cash accounting, revenue grew $1.1B to $16.3B — AIandDesign · 2026-10-21 12:41 (UTC+08:00)
Why Anthropic spending 75% of compute on training raises rivals' bar — menhguin · 2026-10-21 04:33 (UTC+08:00)

[source] Anthropic math: $750B of compute ≈ $8.8B/year, $75B ARR up 45% Q/Q — PythiaR · 2026-10-21
Only 25% of OpenAI compute goes to inference—critic calls red flag, rebuttal does the math — masayoshisson · 2026-10-21
[source] Why Anthropic spending 75% of compute on training raises rivals' bar — menhguin · 2026-10-21
Why Anthropic's $75B training vs $25B inference split narrows the competitor field — menhguin · 2026-10-21
Why improving margins, not revenue, are the real signal in AI lab earnings — menhguin · 2026-10-21
[source] Defending Anthropic's $80B loss: $72B was non-cash accounting, revenue grew $1.1B to $16.3B — AIandDesign · 2026-10-21
Anthropic finance row: critic says the company's revenue growth numbers involve accounting tricks — EZE3D · 2026-10-21
"AI is the biggest thing since the steam engine": debater says he won't bet against Anthropic — AIandDesign · 2026-10-21
User slams Anthropic's $80B loss, bets Carmack or Sutton reach AGI first — retr0jirachi · 2026-10-21