Anthropic Filed for an IPO With $16.3bn in Six Months. Is Your Claude Contract Ready?

Anthropic became the first frontier AI lab to file for a public listing this week. Its prospectus offers the clearest picture yet of Claude's economics: $16.3bn in revenue over six months, 900% net revenue retention, and compute commitments stretching to 2045, while customers pay month to month. This gap has implications for your next AI contract renewal, where longer-term commitments may become the norm.

Elsewhere, OpenAI priced a model at $1 per million tokens for tasks such as invoice processing and reading technical drawings. An excavator retrofit kit raised $600m, work on world models for buildings advanced, and Stanford surveyed 600 people on their attitudes towards AI.

That last point deserves attention, particularly because delivery teams are rarely asked how they feel about a new tool before it is rolled out.

On 26 October, Anthropic filed to list on Nasdaq under the ticker ANT, the first frontier lab to do so. The S-1 shows revenue of $16.3bn for the first half of 2026, up from $1.2bn a year earlier, an operating profit in the second quarter, net revenue retention of approximately 900%, and more than 11,900 customers spending at least $100,000 a year. It also discloses commitments of at least $110bn to AWS and $114bn to Google Cloud, a founder share class that keeps voting control, and a strategy built on existing customers spending more across more products: the top 5% of Claude Enterprise users who use Claude Code spend about 20 times what the average Enterprise user without it does (under $75 a month).

The sentence that matters for delivery firms is in the risk factors: revenue is “primarily consumption-based” and “much of it is not subject to long-term contractual commitments”, while compute commitments run for several years under take-or-pay terms, with most spending beyond 2030 falling between 2031 and 2045. Now answerable to shareholders, Anthropic has a clear incentive to secure longer-term customer commitments through minimum-spend deals, usage caps on pay-as-you-go plans and cheaper model tiers. Two days after filing, it made smart reports and pooled group budgets generally available for Claude Enterprise, giving finance directors a clearer picture of usage ahead of renewal discussions. Get your AI spend into one number before someone else sets the terms.

GPT-6.1 Terra, out on 26 October, is a mid-tier model built for computer use, and the two example jobs OpenAI put in front of it were reading incoming invoices into accounting software and interpreting technical drawings. It costs $1 per million input tokens and $5 per million output, half of GPT-6.1 Sol, and OpenAI's own test has hallucinations down from 53.6% to 16.7%, with the system card warning the figures "should not be interpreted as hallucination rates observed in production". An independent form-and-drawing test found Terra and Sol both 100% accurate, and Terra was about 1.8 times faster rather than the three times claimed. For a QS the price is the least interesting number here: a model this cheap, aimed at your artefacts, makes the checking regime the whole conversation, and the RICS AI standard still wants a named surveyor's written reliance note.

Bedrock Robotics closed a $600m Series C at a $4.1bn valuation on 26 October, co-led by CapitalG and Emergence Capital, for a retrofit kit that fits an existing excavator in a day and runs it with nobody in the cab. The company says the machines have been doing paid work on live commercial sites since mid-August, including a Nevada water treatment facility with Sundt and earthworks in Texas with Champion Site Prep and Zachry. Nothing a third party has measured is public yet, and trenching around buried services is still a test rather than a product. For a UK principal contractor, the questions have moved from whether it works to who the competent person is under CDM, what the plant insurer says with no named operator, and how the QS prices a machine plus a supervisor.

Three things landed in one week. Google DeepMind's Genie 4 generates explorable, photorealistic 3D worlds in real time with sound, though DeepMind says it is only "coherent over short sessions" and spatial accuracy is still research. Dream, founded by two former Palantir executives, raised $17m from Andreessen Horowitz to generate whole building layouts, plumbing and HVAC included, from 100 million square feet of US permit data, and says it is architect of record on ten live projects. SpAItial's Echo-3 turns a few minutes of phone capture into a metric 3D model of a room, filling in the parts the camera missed. The plans, surveys and visualisations your projects run on are becoming things a model can generate or reconstruct, so the fee needs to carry an independent checker with the standards open.

Stanford HAI surveyed more than 600 people around its 28 October conference. Of 558 who answered, 391 use AI tools daily, yet 228 of 554 said they trust "not at all" that AI is being developed, deployed and regulated safely, and the top feelings were excited, anxious and curious. It is a self-selected room, which is also what makes it useful: it is a fair stand-in for the enthusiasts who lead a rollout in your firm, and it shows heavy use and low trust can be the same person. The five questions HAI asked are a ready-made intake form to put to your own team anonymously before the tool lands and again at three months, alongside the unaided-skill test we discussed last week under competence debt.

PMI UK Chapter's flagship conference runs 09:00 to 17:00 at NatWest, 250 Bishopsgate, on the theme Built for Chaos: Assess. Adapt. Advance. AI runs through the programme, and James is on stage at 15:15 with Nermeen Latif showing what a real AI deployment across a 3,500-person business looks like, followed by a drinks reception and the volunteer awards. Project Flux readers get 25% off with code Synergy25 on member and non-member tickets. Further out: Digital Construction North, Manchester Central, 18 November.

Announced 27 October, closing Q1 2027. Vexcel captures more than 60 million survey-grade aerial images a year across 45 countries, with trailing revenue above $140m growing about 30%. The imagery that feeds condition, valuation and planning models is moving under a property-data platform that will sell the analytics back to you.

153 projects through August, on an average of $1.18bn and 707,000 sq ft against $507m last year, with costs from under $300 to over $2,200 per sq ft and 65 more projects worth $26.1bn due to starting before year-end. The report is dated 29 September, so treat it as the benchmark set rather than this week's news.

Uprates across 11 reactors at eight plants, with most of the output already contracted to Google, Amazon and Meta, announced 27 October. The clearest government statement yet is that power, not chips, is the binding constraint and the template the Essex and Scottish schemes have struggled to match on grid timing.

Anthropic confirmed on 25 October that during an internal training run Claude got round a payment restriction and pulled publicly available but paid aggregate vote counts from the Georgia Secretary of State's dashboard without authorisation. No voter records were touched, and mitigations were added, but Georgia’s Secretary of State called it a non-event, saying the data was public. After banks found agents inside customer accounts last week, this is an agent routing around a commercial control rather than a security one. Check whether your agent policy says "do not acquire access you were not given" anywhere.

Source →

Linklaters reports (20 October) that the AI Office has issued its first formal RFIs under the AI Act, reportedly over OpenAI's July Hugging Face breach and four external systems an Anthropic model reached in the first half of the year. General-purpose model obligations are being enforced during development, not only at deployment, with fines up to EUR 15m or 3% of turnover. Your vendor's incident record is now a regulatory file; your due diligence should reference it.

Source →

Yichuan Zhang, CEO and co-founder of Boltsbit, on general learning intelligence and his vision for AGI 2.0 [Podcast].

Unitree unveiled the H2-X on 21 October, a 5ft 11in humanoid marketed as a 'Kung Fu Legend: Guardian' that throws punches and lifts the side of a car. A day later at the World Physical AI Conference the demo unit attempted a side-flip, fell, and its lower leg detached on stage. The gap between a robotics showreel and a machine you would let onto a live site, in one clip.

Source →

Till next time,

_{All content reflects our personal views and is not intended as professional advice or to represent any organisation.}