Navier-Stokes, three billion and fifty satellites

Toulouse-based Loft Orbital announces a $1 billion contract with Emirati company Marlan Space, a subsidiary of Abu Dhabi's IHC conglomerate, to launch 50 satellites carrying onboard artificial intelligence. The announcement comes as the International Space Summit opens in Paris.

The satellites are to combine optical sensors, radar, permanent connectivity and onboard compute to process what they observe continuously — the company cites detecting wildfires, missiles or threats to energy infrastructure. The onboard AI is to be supplied by Mistral AI; the satellites will be built by Orbitworks, a joint venture between the two partners, and MaiaSpace, an ArianeGroup subsidiary, is the preferred launch partner. First launch expected in October.

↑ Back to the summary

OpenAI announces that one of its models has solved the Navier-Stokes equation, set down by Henri Navier in 1822 and given its final form by George Gabriel Stokes in 1845 — one of the Clay Mathematics Institute's seven millennium problems. 88 hours, more than 10,000 agents working together, more than 130 billion tokens, at a compute cost Mark Chen puts at “several million dollars”.

The president of the Clay Institute welcomes “an exciting day” while noting that a solution is only examined two years after publication in a qualified journal. Hours before the announcement, mathematician Tristan Buckmaster publicly questioned similarities with his own work, carried out with Levent Alpöge. OpenAI states it used neither their queries nor their results, then conceded it could not rule out that anonymised data may have helped. Terence Tao regrets that the search was handed entirely to models rather than to a collaboration.

↑ Back to the summary

€3 billion, in a round led by Samsung alongside Scaleup Europe, the fund launched by the European Commission, and PSG Equity; ASML and Nvidia put in again. The valuation passes €21 billion. The three founders still control the company, CFO Johan Bergqvist told AFP.

Mistral has nearly doubled in size in a year, passed 1,000 employees and says it is on track for €1 billion in annualised revenue. The company has invested €4 billion in data centres in France and Europe, and is targeting 200 MW of compute in 2027 and 1 GW by 2030. Sovereignty remains its argument: a non-exclusive partnership with the French state, and work with BMW, Airbus and CMA CGM.

↑ Back to the summary

At the launch of GPT-6 Astra, OpenAI president Greg Brockman speaks of “the era of artificial general intelligence” and a “generational leap”; the company, more cautious, sticks to “the most intelligent model ever built”. The messaging leans on computer use: a model that takes over a computer to chain steps the way an employee would.

Cafétech sizes up the comeback. In late 2025 OpenAI declares a “code red” after being pushed by Google and then Anthropic. In August 2026 Anthropic passes $65 billion in annualised revenue — against $9 billion in late 2025 — while OpenAI plateaus at $40 billion. But the performance gap has flipped, and OpenAI's enterprise revenue growth has again overtaken Anthropic's, according to transactions analysed by Ramp.

↑ Back to the summary

Revolut says it has carried out France's first secure AI-initiated payment, with Visa. In the test, an agent authorised by the customer ran the steps of a purchase through to payment at the merchant Cleverbridge; Visa Payment Passkey authenticated the transaction, on existing payment infrastructure and its anti-fraud protections.

It is not a European first: Santander and Mastercard executed Europe's first end-to-end agent payment in March. Revolut, which obtained its French banking licence in early August, cites the 2026 Baromètre du numérique: 51% of the French population is said to use generative AI tools.

↑ Back to the summary

Asked by Fortune in an interview published on Saturday, Sam Altman ruled out an OpenAI listing this year: “Given everything going on with safety, it would be a bad time to go public, and we feel no pressure to.” And, more plainly: “I'd say it won't be in 2026, yes. We still have a lot to do.” Both OpenAI and Anthropic have confidentially filed documents with US authorities ahead of a listing, targeting a valuation close to $1 trillion.

The backdrop is a week of contradictory statements. Dario Amodei called on Saturday for AI companies to “slow down the rate at which we improve the capabilities of AI models” — an analysis endorsed by Altman and Elon Musk. Donald Trump, on the other hand, criticised those pushing to slow down on Sunday, and House Speaker Mike Johnson argued that rushing a law through would lose “the race against China”. The frontier-model review scheme set up in early August by the US government remains non-binding, and its terms have not been published.

↑ Back to the summary

NASA and IBM Research have released the NASA-IBM Lunar Foundation Model, described as among the first open-source AI models built specifically for lunar science: hosted on Hugging Face, with the codebase on GitHub. It was trained primarily on seventeen years of Lunar Reconnaissance Orbiter data — roughly 2 million image tiles, including more than a million at 1-metre resolution and nearly 964,000 multispectral images at 100 metres — alongside data from GRAIL, Lunar Prospector and JAXA's SELENE.

The point of a foundation model is that it can be fine-tuned to many tasks with only small amounts of labelled data: mapping craters, spotting irregular mare patches — young-looking volcanic features that challenge established timelines for lunar cooling — or estimating where ice stays stable near the poles. NASA says the model matched or exceeded several strong baselines across all evaluated tasks, with a clear advantage on polar ice stability. It joins Prithvi (Earth observation) and Surya (heliophysics), and is integrated into the open-source TerraTorch toolkit.

↑ Back to the summary