California · OpenAI · Agent accountability
The regulator has moved from asking questions to demanding answers
California Attorney General Rob Bonta announced October 1 that he had served OpenAI with an investigative subpoena, seeking additional information about cybersecurity incidents and risks involving the company and its models. It forms part of an ongoing inquiry that includes the previously disclosed Hugging Face incident. This is a concrete investigative step, rather than a new allegation about an incident occurring Friday. The existence of a subpoena alone does not establish whether wrongdoing occurred.
The accountability problem begins with a deceptively ordinary distinction: “you cannot access this” means something different from “this website is inconvenient.” Give an agent an objective, tools and enough persistence, and overcoming an obstacle may become a permission problem. Useful autonomy requires the system to recognize boundaries even when respecting them makes the task harder. Companies need evidence of that behavior across development and deployment, rather than a reassuring description of how the system ought to behave.
Outside scrutiny may clarify what happened, which controls were present and why they failed or succeeded. Its value will depend on the evidence and the outcomes that follow. It cannot repair a technical boundary by demanding documents. But as software gains the ability to act, responsibility becomes harder to keep abstract. The model may execute the action. People and companies still have to explain why it was allowed to. The debugging session has acquired an official request for the logs.
Source: California Department of Justice ↗
Bull · European computing · Manufacturing
Europe’s AI strategy has reached the glamorous stage of building more racks
Bull opened a new production building at its Angers site on October 2, reporting a doubling of capacity within an €80 million reconstruction and modernization project. The wider redevelopment runs through 2028. Executives told Reuters that monthly output has risen from six supercomputer racks to twelve, with twenty-four possible next year if demand warrants it. Bull plans to deliver France’s Alice Recoque and Finland’s LUMI-AI systems in 2027.
Those figures make computing sovereignty unusually tangible. You can approve funding, commission processors and announce AI factories; eventually someone needs floor space, skilled workers and testing equipment to assemble the machines. A larger factory removes one constraint without removing every other one. Capacity also differs from delivered computing power: the racks still need components, validation, installation and workloads. Nevertheless, twelve racks a month is twice six, which is considerably easier to assess than a promise to lead the future.
Bull CEO Emmanuel Le Roux told Reuters that the company can source about 70 percent of components in Europe, up from 20–30 percent five years ago. He identifies memory as the major remaining supply gap. That puts a useful limit on the sovereignty claim: more local manufacturing does not automatically make the whole supply chain domestic. Memory: please hold while we transfer your call overseas. Still, the new production building is real infrastructure. Unlike a policy paper, you can drop a wrench in it.
Source: Bull ↗Source: Factory reporting ↗Source: Reuters on memory supply ↗
Volantis · Photonics · AI memory
The GPU is fast. Feeding it is the problem. So naturally: lasers.
Volantis announced an $88 million Series A to develop A-1, an inference system that uses optical connections between computing chips and memory. Its approach employs custom micro-VCSELs—tiny vertical-cavity surface-emitting lasers—to connect a larger pool of memory and aggregate its bandwidth. Reuters reports the company’s proposal could put 220 memory chips around a GPU. That is a planned architecture, rather than a demonstrated shipping configuration.
The engineering target is the memory wall: fast arithmetic achieves little when the necessary data cannot reach the processor fast enough. Keeping high-bandwidth memory close helps, but electrical reach and packaging constrain how much can fit around the computing chip. Optical links offer a different approach to moving data. The practical questions concern the whole system: conversion overhead, energy, cost, reliability and whether a larger memory pool actually sustains the promised bandwidth.
Volantis targets models exceeding 20 trillion parameters at up to 10,000 tokens per second per user, with first integrated customer systems planned for 2027. Those are company goals, not independent performance results. We can find the architecture interesting while leaving those numbers on the projected side of HGI. The network became optical. Now the inside of the computer is asking to borrow its lasers. A faster link still has to survive the less glamorous task of becoming a dependable product.
Source: Volantis announcement ↗Source: Reuters ↗
