AI Watch, Wednesday 7 October 2026AI pricing at work shifts from seat to meter

In three weeks, the three vendors reworked the shape of their enterprise offers, the Journal du Net reports: the per-seat subscription remains, but the work that costs, heavy requests and tasks handed to an agent, shifts to usage-based billing. The volume risk thus changes sides, from the vendor to the customer, who now pays for what its teams actually consume. For a management team, AI spending stops being predictable in advance and depends on the behaviour of each user.

French startups gathered €3.78 billion across 38 deals in September alone, a record driven by Mistral AI, the Journal du Net reports. Artificial intelligence, medtech and fintech were the most active sectors. The figure confirms that money keeps flowing into the French ecosystem despite debate about a possible market overheating.

The French group unveiled the IP9100, a network encryptor that multiplies by ten the throughput of its sovereign encrypted links, Clubic reports. Aimed at defence and critical infrastructure, the device is meant to secure flows between data centres, which have become the place where most of AI's value concentrates. First deliveries are announced for the first quarter of 2027.

The company unveiled a batch of 722 proofs covering 372 families of results on Tuesday 6 October, including solutions to long-standing open problems, produced by an unreleased frontier model, The Verge and Wired report. OpenAI accompanies the release with formalised proofs in Lean, a language that allows a demonstration to be checked mechanically, made available on GitHub. The manner grates: "There is a perception of thuggish behaviour" from the big labs, a mathematician tells Wired, faulting OpenAI for announcing breakthroughs faster than the community can verify them. The debate extends an already sharp controversy over the line between an established proof and an announcement, and over research ethics when a machine gets ahead of peer review. It also overlaps with the French case reported by Next.ink, where a supposed Anthropic biological discovery "comparable to CRISPR" turns out, according to a CNRS biologist, to be a work site where "all the work remains to be done".

From 9 October, users of Gemini's free tier will be limited to the smallest model, Flash Lite, The Verge and Siècle Digital report. Access to the Flash and Pro models, until now open without paying, will require a Google AI Plus subscription at $4.99 a month, which also gains other benefits in the process. The switch ends free advanced use of Gemini for the general public and aligns Google with an underlying trend: reserving the real power for paying accounts. For a user accustomed to testing the models at no cost, the change will be immediate.

The company is launching a programme that gives startups a year of its Claude Team enterprise offer and a thousand dollars in token credits, TechCrunch reports, the token being the unit of text billed by usage. "We created this programme because we think the benefits of AI will reach most people through the companies that build on the models, rather than through the models alone," Anthropic explains. The logic is clear: install Claude at the heart of startups' tools while they grow, before their consumption starts to count. Today's free offer prepares tomorrow's billing.

The warnings multiplied on Tuesday 6 October, Bloomberg reports. Billionaire Ray Dalio judges AI "a classic bubble" close to the breaking point, driven by rising interest rates and the need to turn wealth into cash. Essayist Nassim Taleb, author of "The Black Swan", calls the stock market a "trap" and describes a category of "naive" investors ready to bet on anything AI-related. The chief investment officer of Singapore's sovereign fund Temasek places the reversal of the AI bet and inflation among the biggest market risks for 2027. These voices converge as SpaceX seeks to borrow $40 billion to buy Nvidia chips, a sign of an increasingly strained debt-financing of AI.

On 22 September, the global AI landscape shifted in a matter of hours, the South China Morning Post recounts: in Beijing, Xiaomi engineers were livestreaming the training of their MiMo-V2.6 model; in San Francisco, Anthropic was presenting Opus 5.5; an hour later, OpenAI released its GPT-6 Sol and Luna models by surprise. For executives and developers, it was one more day in a flow that had become unmanageable. On the Chinese side, DeepSeek and its peers launched sixteen models in a single month, Nikkei notes, despite Anthropic's warning about the risks of a headlong race. Abundance creates its own problem: too many models, too fast, for teams that no longer have time to assess which one serves what purpose. Scarcity moves from access to the model towards the ability to sort.

South Korea's semiconductor exports, estimated at $60 billion, are driving the recovery of the KOSPI index, The Diplomat reports. The country's economy is now tied to the AI cycle: Samsung and SK Hynix supply a large share of the high-performance memory that feeds data centres worldwide. This dependence is both the strength of the moment and the risk of tomorrow, because a downturn in chip demand would hit hard a country that has staked everything on this segment. Seoul illustrates a broader shift in Asia, where national growth is increasingly decided in semiconductor plants.