Anthropic’s Sept-ARR number last week (~$70B vs expects of $78-$80B) added skepticism fuel to an AI trade that was already muddied, particularly on the topic of the trajectory of Lab ARRs through 2028.
We’ve spent the last week doing our best to try to model Anthropic off the leaked draft S-1 reports, press reports, investor color and usage data. The whole exercise was elucidating in helping us think through several issues and drivers in more detail. We were going to wait for our weekly, but wanted to get this out before the S-1 dropped to help frame and give context to a bunch of issues that likely will be talked about more in the coming days.
Here is where we come out on the numbers:
The 6 things modeling Anthropic made us think about:
1. What does “unlimited demand” mean?
In H1 every GW Anthropic added got sold, and ARR went from ~$9B last December to ~$65B by July. Then Q3 freed real capacity as distillation got shut off, META moved work to its own models and MSFT cut internal use. On our numbers those one-offs fell from ~$10.6B of ARR in July to ~$4.5B in September.
If assume unlimited demand, paying customers would have soaked that up at list price within weeks (with the caveat enterprise deals take time to sign). They didn’t. Anthropic cut prices (Opus 5.5 −20%, cache reads −60%, Haiku 5.5 ~−75%) and gave Pro and Max subscribers more usage: 3p measurements show a $200 Max plan’s weekly allowance rising from ~$1,400 to ~$1,750 of API-equivalent usage between late September and mid-October, even though Ramp data shows Anthropic back to 60%+ of Anthropic spend, similar to where it was before Astra. Why is Anthropic giving more allowance to subscribers instead of shifting it to ARR generating API spend if compute is really constrained?
2. How fast does volume catch up with price cuts?
The September-October price cuts (Opus 5.5 −20%, cache reads −60%, Haiku 5.5 ~−75%) take ~$10B off October ARR at constant volume according to our estimates. Ramp’s data says volume is coming back faster already. OpenAI + Anthropic token spend rose 23% in the four weeks to 10/11 on volume up 52%. We know OAI saw Luna usage 10-14x on an 80% price cut in Sept. A 20% obviously will drive less usage inflection. We model ~$8B back by December, now phased 40/40/20 across October-December, for a net Q4 price charge of ~$2B of ARR. Would love to get more detail from mgmt on this one.
3. How much of the fleet actually earns money?
We estimate about 38% of Anthropic’s fleet does inference today with the rest training, post-training and RL. We model ~70% of the inference serving paying customers, and the rest is the free tier, internal use and evals. So about a quarter of the fleet earns the revenue.
SA has said subscriptions use over 40% of inference compute for under 20% of revenue, and a GW sold as API earns ~$80B of ARR vs ~$27-30B in subscriptions.
We estimate Anthropic earns ~$56-58B of ARR per inference GW and ~$18-22B per GW of total compute.
4. If a lab trains less, does everyone else catch up?
The biggest margin lever isn’t buying more GW. It’s using more of the GW Anthropic already has to serve customers instead of training models. Our inference bull case has the same GW as our base but runs 55-62% of the fleet on inference in 2027-28, vs 44-50% in base. The FY28 compute bill is the same ~$149B, but more of it earns revenue, so FY28 adjusted operating margin is ~24% instead of ~5%.
Filling that extra capacity means ~$9.9B a month of net new ARR in 2027, vs $6.5B in our base, and probably lower prices to get there.
A 5% increase in inference share has a big impact: ~$10B more in ARR by end of 2027:
However, we saw with the OpenAI’s Astra release how fast a rival can play catch up at the frontier. On our estimate it took ~$4-6B of ARR from Anthropic in a few weeks in September, and Opus 5.5 won ~$4.5B of it back just as fast. Whoever slows training first to show margins risks the lead and the pricing that comes with it, so I don’t expect anyone to blink. We think the market is in line with this thinking and would punish any slowing of training right now. Training still takes half or more of the fleet through 2028 in our base case
5. How fast can pricing fall?
We model frontier token prices falling ~45-55% a year (Opus input price went $15 → $5 (Opus 4.5, Nov 2025) → $4 (Opus 5.5, 9/22) per million tokens. That’s down ~73% in a year…in Ramp’s data what Anthropic customers pay per token fell ~40% from late July to early October). Each chip generation roughly doubles tokens per GW and work shifts to longer agentic jobs, which offsets most of it, so we have ARR per inference GW down only ~10-15% a year. Would be interested to see what others model here.
If price falls faster than tokens per GW rise, the labs need far more GW and far more demand just to grow. Our bear case (ARR per inference GW −35% in 2027, fewer GW, weaker demand) has end-2027 ARR at ~$103B vs $164B in base. Big swing factor and should be a major focus for investors.
6. If August was that far off, what about 2027-28?
Over the summer, investors told the press Anthropic would exit 2026 well above $100B of ARR. In mid-September the NYT reported it was on pace to top $100B. We’re now at ~$86B for December, $14-34B short on a number a couple months out of those estimates.
This is a key point of why I think the ARR number last week threw a big wrench in the AI semi trade. For most of the first 9 months of the year as ARR numbers were beating, the labs could throw out whatever revenue numbers they wanted in the future, and the market just ate it up. With this miss, there’s a big credibility issue in believing whatever Anthropic’s mgmt tries to spin on the roadshow and to investors (and we expect it to be good spin). The labs are the marginal buyer of AI compute and its why Lab ARRs has been a key debate point for investors. Along with raising some of the questions above, the miss more importantly raised the question of all questions: whether end demand is there to pay for the build.
More detail + charts in how we get to our numbers, what they mean, and what we’re watching next:
You are reading TMTB Pro. For a trial or more detail, please e-mail sales@tmtbreakout.com