Tuesday AI Update: 27 October 2026 — Anthropic files for an IPO and shows the receipts

Anthropic filed publicly for a Nasdaq listing on Monday under the ticker ANT, and the prospectus ended a year of guesswork about the AI boom’s finances: $16.286bn of revenue in the first half of 2026, up from $4.585bn for all of 2025, against an $80bn net loss over the same six months (Guardian).

Two open-weight releases landed the same week: Moonshot’s Kimi K3.1 Pro on 23 October and Mistral’s trillion-parameter Large 4 today.

The week in AI

Western: the filing, in numbers

• Anthropic counted more than 14,000 enterprise customers at 30 September, 11,900 of them spending $100,000 a year or more and about 250 above $10m. It had raised $141.1bn by June and warns that a handful of customers carry much of the revenue. Analysts put it above $1.5tn (Guardian, Axios).
• OpenAI put its models on Google Cloud on 21 October, the first time developers reach them through a rival hyperscaler (CNBC).

China: an open release that trims its own cyber edge

• Moonshot released Kimi K3.1 Pro on 23 October, a post-trained K3 that keeps the 1-million-token context and adds a Highspeed variant running two to three times faster. API pricing is $2 per million input tokens on a cache miss, $10 per million output, with weights promised by 4 November (DataStudios, Bloomberg).
• The interesting part is what Moonshot removed. It stripped cyber content from mid-training and from the reinforcement-learning environments behind the open release, and reports that exploit-development success fell from 90.5% to 81.0%. That is a capability sacrifice stated as a number.

Europe: Mistral ships a trillion parameters

• Large 4 is a 1-trillion-parameter natively multimodal mixture of experts with 52bn active parameters, nicknamed Le Chonk, on preview through Mistral Studio since early October, released publicly today (Mistral, VentureBeat).
• CEO Arthur Mensch argues that renting intelligence is a strategic error, since whoever controls the model sets the price and can withdraw a version. Ecosia undercut the pitch by dropping Mistral for open-weight Chinese models, calling it a year behind (Cybernews).

MENA: sovereign compute on tour

• G42 signed MoUs with Greece’s Copelouzos Group in Athens on 22 October to explore an AI factory there, pairing Khazna data centres with Core42 cloud and compute, during a UAE trade mission that produced 18 deals worth €12.75bn. Location, capacity and financing remain open (Zawya, Reuters).
• Core42 sells reserved sovereign capacity, hardware set aside for one customer inside the country, for UAE enterprises that cannot send prompts abroad (The Circuit).

Africa: a one-stop AI agency for the region

• Rwanda is standing up a government AI agency through 2026, with an AI Factory unit that builds and tests applications with companies. Minister of State for ICT Yves Iradukunda told The Korea Herald that Kigali wants to be the entry point for Korean firms before they expand across Africa, and that the agency would coordinate data, compute, training and safety testing (Taarifa).
• The money attached is modest but real. The Rwanda AI Scaling Hub secured about RWF 25bn, roughly $17m, and the agency wants to host shared model testing for neighbours that cannot evaluate frontier systems alone (AU-Startups). The same analysis notes the constraint: 4G reaches about 96% of Rwandans, but only 21% were unique mobile internet subscribers in 2025.

Asia: money for the lab bench

• Singapore put S$150m into Marvel-AI, an ASTAR-led materials platform announced at SWITCH on 27 October that runs prediction, experiment and validation in a closed robotics loop. ASTAR cites 75 polymer samples in 24 hours, about 25 times a human team, and up to 470 metal alloys screened per day (Business Times).
• Enterprise Singapore opened its Champions of AI cohort with 24 firms, among them OCBC, UOB, Singtel, Keppel and Singapore Airlines.

South America: the deepfake count after the runoff

• Lula won Sunday’s runoff, overturning a first-round deficit to beat Flávio Bolsonaro (Guardian).
• The campaign’s AI ledger is now public. Lupa’s VigIA tracker counted 1,423 AI-made or AI-altered election posts between 16 August and 22 October, 716 of them deepfakes, and found 768 unlabeled, so 54% broke the TSE’s labelling rule. Official candidate accounts produced 548 of the posts (Rio Times).

Russia: supercomputers on the move

• Yandex relocated all three of its supercomputers after drones struck its data centres, and says none were damaged. The digital development ministry wants strategic data centres moved underground, and Kommersant puts replacing the damaged equipment at up to six months (Meduza).

Spotlight: what an AI balance sheet looks like

The S-1 is the first audited view of a frontier lab, and the shape is strange. Revenue went from $386m in 2024 to $4.585bn in 2025 to $16.286bn in six months of 2026, while the same half produced an $80bn net loss. The filing runs past 90 pages of risk factors, opens with the phrase “a country of geniuses in a datacenter,” and concedes that a few customers carry much of the revenue.

That combination, hypergrowth funded by concentrated enterprise contracts, is what investors are now asked to price, and it explains the week’s other theme. OpenAI renting Google’s cloud and Mistral giving away weights are both answers to the same question about who owns distribution.

Why this matters for Africa

Model pricing is now set by companies answerable to public shareholders, which makes it more volatile, not less. The hedge is the open-weight tier: Kimi K3.1 Pro’s weights by 4 November and Mistral’s Large 4 both land on downloadable infrastructure, the only frontier capability an African team can audit and self-host.

Rwanda’s agency model answers the capacity gap most concretely. Most regulators in the region cannot evaluate a frontier model, and one shared testing facility is cheaper than 54.

The Brazilian numbers are the warning. Labelling rules existed and a specialised court enforced them, and still 54% of tracked AI election posts carried no label. Regulation without platform-side enforcement produces statistics, not compliance.

References

• Anthropic S-1 coverage — Guardian — Axios
• Kimi K3.1 Pro specifications — DataStudios
• Mistral Large 4 — Mistral
• Lupa VigIA tracker — Rio Times

Related posts

• Tuesday AI Update: 20 October 2026 — the DeepSeek procurement ban and Kenya’s Senate vote
• Contamination checks before you trust a score — how benchmark claims get audited