Anthropic told sell-side analysts that training its Claude 5.5 model family cost more than $10 billion, and that the AI market is “winner take all”, according to posts on X on 11 November by the account Jukan (@jukan05), relayed within the hour by Liquid (@liquidtrading). Liquid also posted that Anthropic gave Q4 ARR guidance of $77 billion, lower than its October figure of $81 billion.
None of this is confirmed. Jukan named Barclays as the source for the Q4 guidance remark; we could not read any Barclays note. We did not find these claims in Anthropic’s amended S-1 of 10 November or in our saved captions of its IPO roadshow video, and we found no public comment from Anthropic on the posts as of 13:15 UTC. If the $77 billion figure is accurate, it is 4.9% below the filed $81 billion, by our arithmetic.
Business Pill · Run-Rate Revenue Explained
An evergreen Business Engineer explainer on what an AI company’s run-rate revenue number means. It teaches the general idea only and does not discuss the posts in this piece.
The key insight: As we read it, the posts put three pressures on the filed picture at once, all relayed rather than filed: a Q4 guide below the October run-rate, a training bill of more than $10 billion for one model family, and a SpaceX contract that ends around mid-2027.
At 07:31 UTC, Jukan posted that Anthropic “gave conservative Q4 guidance because engineers tend to take time off in December”, and replied a minute later: “Source: Barclays”. Jukan’s posts give no guidance figure.
At 11:46 UTC, Jukan listed “Interesting remarks Anthropic made in a meeting with the sell side”: that Anthropic has “the stability and the better talent that open source lacks”; that “The AI market is winner take all. As long as we execute properly, we will take all of it”; that “The contract with SpaceX will not be extended beyond the middle of next year”; and that “Developing Opus 5.5 cost more than $10 billion”.
At 12:27 UTC, Jukan corrected the last point: “The $10 billion was reportedly spent on training the entire Claude 5.5 model family, including Opus 5.5”. Liquid’s posts between 12:28 and 12:31 UTC repeated three of the remarks without naming a source, still describing the $10 billion as the cost of Opus 5.5, and added the $77 billion Q4 figure, which does not appear in Jukan’s posts.
Claimed vs filed: Anthropic’s run-rate revenue of about $81 billion at 31 October and the implied $71 billion at 30 September come from its S-1/A; the $77 billion Q4 ARR guidance is a claim posted on X by Liquid and is not verified. The 30 September figure and the 4.9% gap are our arithmetic.
What Was Posted on X
Interesting remarks Anthropic made in a meeting with the sell side:
– Anthropic has the stability and the better talent that open source lacks.
– The AI market is winner take all. As long as we execute properly, we will take all of it.
– The contract with SpaceX will not be extended beyond the middle of next year.
Jukan (@jukan05) lists four remarks it says Anthropic made in a meeting with the sell side, including the SpaceX contract and a cost of more than $10 billion for Opus 5.5. The post names no source; we could not confirm any of the remarks.
Correction: The $10 billion was reportedly spent on training the entire Claude 5.5 model family, including Opus 5.5.
Jukan’s correction 41 minutes later: the $10 billion was reportedly spent on training the entire Claude 5.5 model family, including Opus 5.5. This piece uses the corrected wording.
JUST IN: Anthropic gives Q4 ARR guidance of $77 Billion, a figure that is lower than Octoberโs ARR of $81 Billion
Liquid (@liquidtrading) posts a Q4 ARR guidance figure of $77 billion against October’s $81 billion, without naming a source. The $81 billion matches Anthropic’s S-1/A run-rate at 31 October; the $77 billion appears in no filing we found.
The $77 Billion Claim Against the Filing
Anthropic’s S-1/A says its run-rate revenue was approximately $81 billion as of 31 October, up $10 billion from 30 September; we covered the figure when the amendment was filed. The filing defines run-rate as 28-day consumption fees multiplied by 13, plus monthly subscription fees multiplied by 12, and says it “is not a substitute for revenue prepared in accordance with GAAP or a forecast of revenue”.
The filing gives no Q4 guidance that we found. On the relayed figure, $77 billion would be $4 billion, or 4.9%, below the 31 October run-rate, by our arithmetic. Liquid did not say what date or measure the guidance refers to, so it may not be like-for-like with the filed metric.
The SpaceX Claim Against the Filing
The S-1/A says Anthropic agreed to spend up to $120.5 billion under its SpaceXAI arrangements “over terms extending into 2029”. It adds that after non-cancelable periods ending in January 2027 and March 2027, each agreement “may be cancelled by either party upon 90 days’ advance written notice”; we set out those terms in our SpaceXAI coverage.
A contract that ends around the middle of 2027 would therefore fit inside the filed exit terms, but it would mean one side using that cancellation right. The filing does not say either party plans to, so the relayed remark is consistent with the filing without being confirmed by it.
The $10 Billion and ‘Winner Take All’
We found no training-cost figure for the Claude 5.5 family in the S-1/A. Jukan’s correction widens the claim from one model, Opus 5.5, to the whole Claude 5.5 family and adds the word “reportedly”. The “winner take all” line is a quoted remark with no figure attached, and Jukan did not name the meeting’s date, host or attendees.
As we read the timing, the guidance remark came first, at 07:31 UTC, with Barclays named as its source; the four meeting remarks followed at 11:46 UTC with no named source, and Liquid’s $77 billion figure arrived at 12:31 UTC, also without one.
As we read the filing, none of the relayed remarks contradicts the S-1/A: the run-rate is a point-in-time measure that the company says is not a forecast, and the SpaceXAI contracts carry exit rights from 2027. Confirmation would need the Barclays note, an outlet report or a statement from Anthropic.
Anthropic S-1/A, 10 November 2026
“Run-Rate Revenue is not a substitute for revenue prepared in accordance with GAAP or a forecast of revenue.”
Three Implications
FOR THE RUN-RATE The filed $81 billion is a 31 October figure built from 28 days of consumption fees plus subscriptions; the relayed $77 billion Q4 guide, if accurate, is 4.9% lower, by our arithmetic.
FOR COMPUTE The S-1/A puts the SpaceXAI arrangements at up to $120.5 billion over terms extending into 2029, cancellable on 90 days’ notice after windows ending in January and March 2027; the relayed remark says the contract will not be extended beyond the middle of next year.
FOR MODEL COSTS After Jukan’s correction, the relayed figure of more than $10 billion covers training for the whole Claude 5.5 family, including Opus 5.5; we found no training-cost figure for the family in the S-1/A.
The framework opens: “Every AI paradigm produces a different kind of moat.” It names four, starting with the Corpus Moat built by pretraining and the Verifier Moat built by reinforcement learning on domain-specific reward signals.
As we read it, a relayed “winner take all” remark is a claim that one of those moats holds, and a training bill of more than $10 billion, if accurate, is the cost of building the first two. The framework predates the posts and does not discuss these claims.
What Is Not Established
Q4 guidance of $77 billion: unverified. The figure appears only in Liquid’s post; Jukan attributed a conservative-guidance remark to Barclays without a number, and we could not read a Barclays note. Training cost of $10 billion or more: unverified, attributed by Jukan to a sell-side meeting and corrected to cover the whole Claude 5.5 family. SpaceX contract not extended past mid-2027: unverified, consistent with the filed cancellation terms but not stated in them. ‘Winner take all’: an unverified quote, relayed without a named speaker or date.
We did not find an outlet report of the meeting in Google News or web searches at about 13:00 to 13:15 UTC. We did not contact Anthropic or Barclays. The 4.9% and $4 billion figures are our arithmetic on the relayed number.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Posts on X attribute four remarks to Anthropic’s meeting with sell-side analysts, including a training bill of more than $10 billion for the Claude 5.5 family. A separate post by Liquid gives a $77 billion Q4 guide. We did not find the guidance, cost, SpaceX or winner-take-all claims in the S-1/A of 10 November, and the only named source, Barclays, could not be read.
A note on sourcing. This piece reports claims posted on X on 11 November 2026 by Jukan (@jukan05) and Liquid (@liquidtrading), which we read in full; Jukan named Barclays as the source of the Q4 guidance remark, and we could not read that note. We checked each claim against Anthropic’s S-1/A of 10 November 2026 and our saved captions of its roadshow video. We did not contact Anthropic or Barclays.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
