Open-weight models pass proprietary ones on Vercel
AI Watch, Wednesday 26 August 2026

The American company will be able to build new infrastructure dedicated to AI and cloud in the Haut-Rhin, Clubic reports. The project does not have unanimous local support, at a time when the water and electricity consumption of these facilities is already fuelling sharp tensions. Every siting of this kind plants a concrete question in France for local authorities and neighbouring businesses: which resources for what territorial benefit.

The button to report a post judged AI-generated, opened in July, has been activated more than a million times, says Hari Srinivasan, the network's head of product, quoted by Siècle Digital and Clubic. LinkedIn now adds a notice in post analytics to warn the author concerned. The figure measures a fatigue: professionals are rejecting mass-produced text, including on the network that pushed assisted writing hardest.

The specialist in cloud infrastructure dedicated to AI is aiming for this raise in an offering planned for September, ZDNet reports. The move illustrates the rise of the "neoclouds," those specialised operators that stand against the established cloud giants to rent out compute. An IPO of this size signals that investors keep betting heavily on compute demand, at a time when hardware costs are climbing.

On Vercel's AI Gateway, a widely used web development service, open models (which you download and run yourself) accounted for 54% of the token volume processed on Tuesday, against 46% for proprietary models, the South China Morning Post reports from the platform's public data. The share of open models even reached a record 62% on one given day. DeepSeek's most recent lightweight model is driving this shift. In the same move, corporate demand for Fable 5, Anthropic's most powerful model, is stalling because of its high price. The trade-off between top performance and acceptable cost is thus being replayed in the field, request by request, and it is tilting toward open. This is the subject of our analysis.

Two weeks before its autumn keynote, Apple unveiled by surprise new Mac mini and Mac Studio models and two chips, the M6 and the M5 Ultra, Numerama and ZDNet report. The M5 Ultra Mac Studio version goes up to 512 gigabytes of unified memory and supports eight external displays; Apple claims AI performance four times that of the M3 Ultra. Prices start at $2,499 for the M5 Max and $5,499 for the M5 Ultra, available on 22 September. That much memory targets a precise use: running large models locally, on a desktop, without going through the cloud.

Chris Malone, who helped lead the construction of OpenAI's data centres, has left the company, Bloomberg and TechCrunch report. Asked about the departure, the company says it "recently reorganised" its infrastructure organisation "to support the scale and pace" of its work. Malone joins a list of high-profile departures in recent months, at the very moment OpenAI is committing tens of billions to its compute capacity. These moves raise a question of continuity: building data centres at this speed requires a stable team in charge, and the company is losing its people one after another.

AI could cut the price of routine IT projects by up to 25%, the head of the Indian digital services group Hexaware told Bloomberg. The observation targets a sector built on labour-intensive tasks, where automating part of the coding and testing compresses the bill charged to the client. For providers, value shifts from the volume of hours sold toward the ability to deliver faster at equal quality. For French companies that outsource their development, it is a signal to renegotiate: yesterday's rate no longer holds if AI does part of the work.

The Chinese lab is finalising a raise that values it at about 500 billion yuan, or $74 billion, pre-money, the South China Morning Post reports. DeepSeek was seeking to raise nearly 50 billion yuan in this round, expected to close before the end of August, notably from its existing shareholders. The move sets up a listing targeted for 2027 on Shanghai's Star Market, China's technology exchange. The timing is no accident: DeepSeek's lightweight model is the one pushing open models ahead of proprietary models on Western platforms. A Chinese lab funded at this level, whose models spread free of charge internationally, is redrawing the map of value in AI.

Xiaomi is pursuing its bet on proprietary silicon to prepare for complex AI workloads on smartphones and cars, the South China Morning Post reports, as its profit declines and component costs soar. On Monday the group presented its Xring O3 AI processor, built on a 3-nanometre process, due in September aboard the Xiaomi 18 Fold, along with a Xring O100 on 6 nanometres. Designing its own chips is expensive and weighs on margins in the short term, but reduces dependence on American suppliers subject to export controls. The approach matches that of several Chinese giants: bring compute in-house to last over time.

The Chinese giant issued 710 million shares at $14.4 apiece in Hong Kong, the largest deal of its kind ever carried out by a company on the exchange, Siècle Digital reports. The full amount will fund its AI infrastructure, while its net profit fell 75% in the last quarter. In the same move, founder Jack Ma and the group's executives bought back several hundred million dollars of shares, a signal of confidence sent to the market after a week of decline in the stock.