The LLM ARR Debate (Sept 2026) — Anthropic's $65B Expectation Miss, the Gross-Up Fight, and the Capex Test

Research page: 2026-09-05. This is a dated snapshot of a fast-moving controversy, August 18 – September 5, 2026. Every figure is as reported by the cited source; treat run-rate numbers, leaked memos, and forecasts as claims rather than audited fact. For how these ARR numbers are constructed, see the companion page How LLM Companies Compute ARR.

TL;DR

On August 18, Bloomberg reported Anthropic's July run rate at ~$65B (+38% from $47B in May, >7x the ~$9B at end-2025). Q2 preliminary revenue was >$11.5B (~14x YoY) with positive adjusted operating income.
The number was a miss against buy-side expectations of $70–80B from trackers (YipitData, TickerTrends) and a ~$78B expert forecast — and it knocked global tech down for three days (SOX ≈ −5%, Nasdaq −1.33%, then STAR 50 / ChiNext −5%+).
The methodological fight resurfaced: a leaked OpenAI CRO memo from April had accused Anthropic of "grossing up" cloud-marketplace revenue with Google/Amazon, allegedly overstating the then-$30B figure by ~$8B.
Product-level tracking sharpened the debate: Claude Code's tracked ARR was $15.1B but grew only +5.2% MoM by August 10, while Codex's $8.8B grew +20.8% MoM.
The stakes: Anthropic internally projects ~$190–200B of 2028 revenue behind a potential ~$2T IPO (as early as October). The S-1, not the press release, settles who was right.

1. What happened: an expectation gap, not a bad number

The August 18 Bloomberg disclosure was, by any absolute standard, enormous: ~$65B annualized in July, up ~38% from the $47B Anthropic disclosed with its May Series H, and more than seven times the ~$9B run rate at end-2025. Q2 preliminary revenue exceeded $11.5B, ~14x the year-ago quarter, with adjusted operating profit positive for the first time.

The problem was relative. Two weeks earlier, third-party trackers YipitData and TickerTrends had pointed at $70–80B; the week before, an institutional expert survey produced a ~$78B forecast. The reported ~$65B was therefore "the leader slowing" rather than "the leader winning."

OpenAI's own Q2 numbers (WSJ) added fuel the same week: revenue of $6.7B (+18% QoQ from $5.7B) with operating losses expanding from $9.3B to $12.3B — more than $1.8 burned for every $1 earned — even though its total run rate passed $40B in August (CFO Sarah Friar told investors in mid-August that enterprise revenue had overtaken consumer for the first time, with July ARR up ~20% MoM and enterprise up ~32% MoM).

2. Why one company's ARR moved three continents

The market reaction on August 18–20 shows how central Anthropic's number has become:

Day	Move
Aug 18	Philadelphia Semiconductor Index ≈ −5%; Western Digital and Micron > −7%
Aug 19	Nasdaq −1.33%; semiconductor index continues down after TickerTrends data
Aug 20	A-shares sharp correction; STAR 50 and ChiNext indexes −5%+ in a day

The transmission mechanism is the AI capex narrative. The four hyperscalers' combined annual capex has climbed from ~$455.8B (2024/25 baseline) toward ~$750B in 2026, with 2027 consensus of $1.1–1.4T (roughly 1.6–1.9x 2026). Model-vendor revenue is the demand-side validation of that spend, and Anthropic — the first pure frontier-lab IPO candidate — is the proxy. If the fastest-growing lab decelerates, upstream capex sustainability is questioned, even if the slowdown is a single-company artifact.

3. The accounting fight: gross-up accusations

The debate over caliber is not new, but it is now central:

April 2026 leaked OpenAI memo. OpenAI CRO Denise Dresser's leaked internal memo accused Anthropic of inflating its then-disclosed ~$30B annualized revenue by ~$8B through full gross-up of revenue shares with Google and Amazon — i.e., recognizing cloud-channel sales at the full price the cloud charges, rather than only Anthropic's share. Coverage at the time (Gizmodo, BlockBeats) framed it as OpenAI's new competitive strategy: attack the rival's revenue quality directly.
Sacra's standing caution. Anthropic's cloud-channel revenue (AWS, Google, Microsoft) "may be recognized on a gross basis, making the revenue scale appear larger."
Why it resurfaces now. With the S-1 pending, the ~$8B charge — which was aimed at the $30B spring figure — is being extrapolated to the $47B and $65B disclosures. Skeptics ask whether the gap between run-rate (~$65B) and booked revenue (Q2 at ~$11.5B, or ~$46B annualized) is explained by growth or by accounting treatment.

The accounting question matters more than the multiplication question. A hypothetical comparison from Barclays' August research shows two labs with identical real economics can report gross margins ~17 points apart purely from business mix and gross-vs-net recognition — the Uber/Lyft reporting problem applied to AI.

4. Product-level evidence: Claude Code vs Codex

The sharpest new data point came from TickerTrends on August 19. Its tracked ARR (as of August 10, 2026):

Product	Tracked ARR	MoM growth	Share of parent tracked ARR
Claude Code	$15.12B	+5.2%	~21.9% of Anthropic
OpenAI Codex	$8.83B	+20.8%	~45% of OpenAI enterprise ARR (central case)

Claude Code's trajectory: $2.5B (Feb 12, disclosed) → ~$6.3B (end-Mar) → ~$10.3B (end-Apr) → >$14B (Jun) → $15.12B (Aug 10). Codex: ~$1.0B (early Feb) → $3.0B (Apr) → $5.6B (Jun) → $8.83B (Aug 10).

This is the bull/bear crux in miniature. Bulls read the absolute numbers as proof that coding agents are the largest verified AI payment scenario — Claude Code alone went from launch to >$500M ARR in three months, $1B in about six, and $2.5B to $15.1B in ~seven months. Bears read the delta: the leader's growth decelerated to ~5% MoM while the smaller rival grew four times faster, and a large share of parent growth now depends on whether that slowdown is a plateau or a pause.

5. The IPO math

Anthropic reportedly filed confidentially in June and could list as early as October (FT). The valuation debate, per Reuters-reported figures and Wall Street analysis:

Basis	Multiple vs a ~$2T target
Q2 revenue annualized (~$46B)	~43x
July run rate (~$65B)	~31x
Internal 2028 revenue forecast (~$190–200B)	~10x forward

Bear arguments beyond accounting: ~80% of Anthropic's revenue is API (concentrated and volatile if enterprise budgets tighten); Chinese open-source labs keep cutting prices; Anthropic itself cut Claude Opus 5 pricing on July 25.

6. What settles the debate

Nothing in the current data can settle it — that is the point. The deciding evidence is the S-1:

Audited revenue replaces run rates; the gap between booked and annualized becomes explicit.
Recognition policy (gross vs net, prepaid tokens, cloud-channel treatment) is disclosed, making the ~$8B gross-up accusation testable.
Gross margin and product mix (API vs subscription vs enterprise) quantify the Barclays-style "same economics, different report" problem.
Customer concentration and retention test whether usage spikes repeat.

Timing: only two to three revenue-update windows remain before the earliest possible October listing, so every subsequent monthly disclosure will be scrutinized for whether the slowdown is a fluctuation or an inflection.

Sources

Bloomberg — Anthropic July ARR ~$65B and Q2 >$11.5B revenue (2026-08-18), via BlockBeats and 36Kr/Gelonghui
CNBC — OpenAI CFO investor call: run rate >$40B; enterprise > consumer; July ARR +20% MoM, enterprise +32% MoM (2026-08-14)
The Wall Street Journal — OpenAI Q2 2026 figures: $6.7B revenue, operating loss widening to $12.3B (reported week of 2026-08-17)
Reuters — Anthropic internal ~$190–200B 2028 revenue projection (2026-08-16)
Financial Times (via 36Kr/Gelonghui) — IPO as early as October; ~$2T target valuation (2026-08-22)
TickerTrends — Claude Code Reaches $15.1B in Tracked ARR as Codex Climbs to $8.8B (2026-08-17/19)
36Kr / Gelonghui — Can Anthropic Take On the $2 Trillion Gamble? (2026-08-21/22)
Gizmodo / BlockBeats — OpenAI leaked CRO memo accusing Anthropic of ~$8B gross-up on the ~$30B figure (2026-04-12/13)
Sacra — cloud-channel gross-recognition caution (2026)
Barclays via CLS / China Star Market — AI lab unit economics and gross-vs-net recognition (2026-08-28/30)