Today's Headlines

  • Hugging Face reported to have received an approach above $13 billion — no deal reached, counterparty unnamed
  • Personal AI assistant Instinct draws scrutiny — a perpetual training license and the power to bind users to contracts
  • Taiwan indicts nine over smuggled AI servers — 74 of 130 B300 units reached China, customs stopped 56
  • NVIDIA publishes its own efficiency measurements for Vera Rubin NVL72, with third-party review still pending
  • A U.S. district court upholds the denial of a motion to compel a list of Google's training works

The three stories on today's page approach the same question from different sides: who ends up holding what. Hugging Face is reported to have received an approach above $13 billion, with no deal reached and the buyer unnamed; a personal AI assistant's terms of service ask for a perpetual training license and the power to bind its user to contracts; and prosecutors in Taiwan have indicted nine people over AI servers routed into China.

What moves in all three is the placement of authority and responsibility, rather than the capability of any model.

All three also share something unresolved. The acquisition is at the approach stage with no named buyer, the terms-of-service questions await an answer from the operator, and an indictment is an accusation rather than a finding of guilt.

Today's Top Three

Hugging Face reported to have received an approach above $13 billion

Hugging Face, the platform through which most open AI models are shared, is reported to have received an approach to be acquired at a valuation above $13 billion.

The report originates with Business Insider and reaches us through TechCrunch. No party to the transaction has announced anything.

The stage matters here. The company is said to have received several offers and to be working with banks to evaluate them. The would-be buyer is not named in the reporting, and no agreement has been reached.

Set against the company's own history, the figure is a steep move. The 2023 round led by Salesforce Ventures valued the company at $4.5 billion, so $13 billion is roughly 2.9 times that mark. Earlier this year the company turned down a $500 million investment from NVIDIA at a $7 billion valuation, and the new figure is about 1.9 times that.

Whether a sale happens at all is treated as an open question in the reporting. Chief executive Clément Delangue has repeatedly framed the company's obligation to its community as a long-term one, and the report notes doubts about squaring that stance with a sale.

For teams that build on open weights, the number is the less interesting part. Model retrieval and distribution for a great many organisations run through this one company's infrastructure, which makes it a place where a change of owner could move terms and priorities. What has moved so far, though, is a report of an approach — the terms themselves stand where they stood.

Instinct's terms of service ask for a perpetual license and the power to bind users

The personal AI assistant Instinct is drawing privacy and security scrutiny over both its terms of service and its observed behaviour.

Instinct is built by a small team led by Noah Shinn, formerly a researcher at Sierra, and operated by Spear Street Technology. It connects to a user's email, messages, calendar and location data, and carries out tasks on that user's behalf. Kleiner Perkins and Conviction have invested; the amounts are undisclosed.

At the centre of the concern is the scope of rights the terms request. They grant a perpetual and irrevocable license over user data, and that license extends to training AI models.

The terms go further than reading and writing. They also ask for the authority to enter into contracts that bind the user — obligations signed in the user's name become part of the service's remit.

Behaviour, as well as wording, has been reported. Testers say they confirmed that Instinct can be phished: a lure sent from a newly created Gmail account succeeded, demonstrating a path by which instructions from an unexpected sender are followed.

What happened after disconnection has also been reported. Email summaries continued to arrive after a user revoked access, and messages a user had asked to have deleted remained in the system.

The investor Katie Jacobs Stanton put the general point plainly, saying that trust "resets to zero with one unauthorized action."

For a legal team, the significance runs past this one product. Any design that hands an agent read-and-write permissions invites the same two checks before adoption: what rights the terms actually claim, and whether revoking access genuinely stops the flow of data.

Taiwan indicts nine over AI servers routed into China

Prosecutors in Keelung, Taiwan have indicted nine people over forged documents used to route high-end AI servers into China in breach of U.S. export controls.

Reuters reports that an NVIDIA senior manager and two Taiwan-based Supermicro employees are among the nine. Prosecutors have withheld the names, and the charges are described as breach of trust and document forgery. The important qualifier is that this is an indictment — a court has reached no finding of guilt.

The forgery concerned where the hardware sat. According to the Taiwanese investigation, paperwork showed 130 B300 servers installed and running at facilities in Taiwan.

Of those, 74 were in fact exported and delivered to customers in China. Customs officers in Taiwan caught the discrepancies on the remaining 56 and stopped them.

Related matters continue. One individual is reported to be a fugitive, and investigators point to funds diverted from a distributor involved in moving restricted chips into China. Supermicro co-founder Wally Liaw was arrested in March over an alleged $2.5 billion in restricted AI servers sent to China since 2024.

Supermicro has responded that the company itself is not under investigation and that it is cooperating with the Taiwanese authorities. The investigation has covered at least twelve locations.

Read as an export-control matter, the case marks a change in where enforcement lands. The United States has required licences for semiconductor exports to China since 2022, and attention has usually fallen on distributors and customers. Here the indictments reach individuals employed inside the manufacturers.

Other Developments

Models & APIs

NVIDIA published its own measurements of the power efficiency of its new Vera Rubin NVL72 rack. On agentic workloads, the company reports up to 30 times the throughput per megawatt of the previous GB300 NVL72, and cost per million tokens as low as one thirty-fifth. The measurements use SemiAnalysis's AgentX workload, and NVIDIA attributes the 30x figure specifically to DeepSeek V4 Pro. The company itself notes that these are in-house measurements pending SemiAnalysis review, and that they do not yet reflect the Vera CPU's performance on tool calls. For operators whose growth is bounded by power, these belong in the column marked "awaiting third-party confirmation."

ITmedia surveyed the run of 30B-class open models released between August 10 and 19: Meta's Muse Glimmer, NVIDIA's Nemotron 3.5 Lightning, Alibaba Cloud's Qwen3.8-27B and the National Institute of Informatics' LLM-jp-4 33B. Qwen3.8-27B is reported to have beaten Claude Opus 4.6 on SWE-bench Pro and LiveCodeBench v6, and scores 52 on the Artificial Analysis Intelligence Index. The article cautions that the lead is confined to a handful of benchmarks and that most of the comparison figures come from vendors testing their own models.

Anthropic's Claude ran into elevated request errors from 1:50 p.m. JST on August 24, and service was restored at 4:36 p.m. the same day. Claude Fable 5 and Claude Opus 5 were among the models affected. The cause has not been published.

Products & Features

OpenAI made all three GPT-5.6 models — Sol, Terra and Luna — available inside Kiro, the AWS software development agent. Kiro breaks a high-level intent into requirements, technical design and executable tasks, inserting checkpoints before changes are applied. OpenAI reports that on Terminal-Bench 2.1, the cost of completing a task with GPT-5.6 Terra inside Kiro fell by roughly 82 percent. Swami Sivasubramanian of AWS and Colleen Kapase of OpenAI each commented on the collaboration.

Anthropic opened Claude Academy, a free official online learning platform. Courses are organised around six areas — Claude.ai, Claude Cowork, Claude Code, Claude Tag, Claude Platform and AI Fluency — covering beginners, developers and organisations rolling out team adoption. Webinars are scheduled to begin in September.

Policy & Courts

A U.S. district court upheld the denial of the plaintiffs' motion to compel Google to produce a list of the copyrighted works used to train its generative AI models. Google argued that no such list exists and that identifying the training materials would require examining the training data itself. The magistrate judge denied the motion partly on the ground that it duplicated work the plaintiffs' own expert had already performed. One qualification deserves emphasis: the order does not narrow the reach of discovery as a general matter. It says of itself that it denied a particular request as redundant and not proportional.

Business

General Intuition, which is pushing into robotics, raised new investment from Valor and Point72 at a $6 billion valuation.

The OpenAI price reduction covered in yesterday's briefing is now visible on the company's pricing page. It applies to on-demand pricing for GPT-5.6 Sol and holds at least until November 21. Terra and Luna fall outside this change; both were reduced separately on July 30.

Source: selected by the editorial desk from the AI news inbox for August 25, 2026 — 30 items, 11 primary and 19 secondary.