The two largest US model providers just staked out opposite positions on the question every regulated buyer asks first: what happens to my prompts? On 19 August, OpenAI reaffirmed Zero Data Retention for eligible API customers — prompts and responses are not stored after a request is processed — and previewed a system it calls Private Safety Processing, which is meant to spot abuse patterns across related requests without any OpenAI employee seeing the underlying content. The rollout starts with early customers, with a technical white paper promised in September, per Axios.
The timing is pointed. Anthropic recently began requiring 30-day retention on its most capable “covered” models — Fable 5 and its Mythos-class tier — so that automated safety classifiers can run over flagged content, while leaving Zero Data Retention available on its other, non-covered models (Anthropic’s own docs). For an EU compliance officer, the split is the story: two credible vendors now disagree on whether safety requires holding data at all. It is a useful reminder that “no retention” is a contractual and architectural choice, not a default — and one worth checking on every frontier model you route to.
Europe’s AI data centres are chasing the power, not the cities
Where the compute physically sits is shifting fast. Hyperscale campuses planned across EMEA for 2026–2028 will sit an average of 175km from major hub cities, up from 46km for sites delivered in 2022–2025, Reuters reported on 19 August. The share of pipeline capacity in inner-city locations is set to fall to 5% from 13%. The driver is electricity: grid-connection queues in the established FLAP-D markets now run 7 to 10 years and up to 13 in the most congested, according to JLL data, so developers are following cheap, available power to the edges — of nine proposed gigawatt-plus European sites, only one sits near a major city (Paris), the rest scattered from rural Spain to northern Sweden. For anyone weighing where regulated European workloads should run, the map is being redrawn around megawatts, and it increasingly points to wherever the power is.
Hugging Face’s open-model census: small models win, Chinese labs set the ceiling
Hugging Face published its State of Open Models: Summer 2026 report, and the numbers cut against the frontier hype. Public model repositories grew to 2.96 million over the first seven months of the year, but usage is brutally concentrated: 1.5% of repositories account for 99.2% of all downloads, and models under 1B parameters take 83% of lifetime downloads while everything above 100B takes just 1%. In other words, the models people actually deploy are small enough to self-host.
The report also documents where the frontier of open weights now lives. In five of seven months, the largest, best-performing open model from a Chinese lab was bigger than anything a US lab released for itself — the Chinese high-water mark ran from 754B up to 2.78T parameters — and several of the year’s largest US open releases were adaptations built on top of Chinese models. The most prolific repository publishers were AMD and Nvidia, whose output is mostly hardware-optimised conversions rather than new architectures. For European teams, the practical takeaway is that a strong open-weight catalogue is now overwhelmingly a catalogue of models you can run yourself.
Quick Hits
- Nvidia’s H200 chips are flowing to China again — throttled. — Beijing is allowing limited imports at roughly 13% of the per-customer licence ceiling, with ByteDance and Tencent each said to have received about 10,000 units, per Benzinga.
- Warp launches “software factories.” — On 18 August the terminal maker introduced Warp Factories, a closed-beta cloud layer for orchestrating coding agents from intake through a reviewed pull request, with hooks into GitHub, GitLab, Linear, Jira, Slack and Teams.
- IBM and Together AI commit $240M to an open-model inference cluster. — The multi-year deal puts a large cluster of Nvidia HGX B300 systems on IBM Cloud, aimed at serving open-source models, with availability from Q1 2027.
- EU transparency code passes ~190 signatories. — The Commission’s voluntary Code of Practice on marking AI-generated content, which backs the Article 50 obligations now in force, will spin up two implementation task forces in September for signatories to compare notes.