The Omniscient Bulletin · 2026-07-02
The Omniscient Bulletin — July 2, 2026
A fight over who sells AI computing power, as Washington moves to gate how models get released
The plumbing of AI led the day. A Bloomberg report that Meta may rent out its spare computing power sent the specialist clouds it would rival sliding, even as one of them, Together AI, raised $800 million at more than double its old value and a utility put $1.75 billion into powering US data centers. Around the hardware, the state tightened its grip on the software: Washington is said to be near voluntary rules for how models ship, and Anthropic's top model returned from an export ban wrapped in government-blessed safeguards. And the big labs began hiring philosophers to ask whether the systems can feel.
Industry
Meta may start selling its spare AI computing power, and neocloud stocks tumble
Meta could launch a business to rent out its spare AI computing power, Bloomberg reported, a move that would turn one of the industry's biggest buyers of chips into a direct rival to the specialist providers that lease out exactly that kind of capacity. The report hit those firms hard: CoreWeave fell about 14% and Nebius about 17% on the day, while Meta rose nearly 9%. The catch is that Meta is itself a large customer of the companies it would compete against, having signed multibillion-dollar compute deals with more than one of them. Meta has not confirmed the plan.
Bloomberg via The Motley Fool→
Compute
National Grid bets $1.75 billion on powering America's AI data centers
National Grid's investment arm is paying $1.75 billion for a 35% stake in Joulent, a US power developer, valuing the young company at about $5 billion and underlining how central electricity has become to the AI build-out. Joulent's first project, Kilby, is a 2.67-gigawatt gas plant in West Texas, built as a 50-50 venture with Chevron, that will feed a Microsoft-operated data center under a 20-year supply deal. First power is targeted for 2028. The deal pushes a regulated utility straight into generating power for AI, a layer it has traditionally only wired up.
Industry
Together AI raises $800 million as investors back cheaper open-model clouds
Together AI, which rents out clusters of Nvidia chips tuned to run open-source models, raised an $800 million round led by Saudi Aramco's venture arm, more than doubling its valuation to $8.3 billion from $3.3 billion early last year. Nvidia, Vista Equity Partners and General Catalyst also joined, on the back of what the company says are annual bookings above $1.15 billion. It is a strong vote for the 'neocloud' model of selling specialized AI compute more cheaply than the big clouds, landing the same day a report suggested Meta may crowd into that market.
Policy
White House is said to be near voluntary rules for how AI models get released
The US government is in advanced talks with Google, Anthropic and OpenAI on voluntary standards for releasing AI models, the Financial Times reported, with an announcement possible within the week. The rules would set benchmarks for frontier systems and their release timelines and clarify who may use them at home and abroad. They would formalize recent case-by-case moves, from the June request that OpenAI hold back GPT-5.6 to the export limits placed on, then lifted from, Anthropic's top models. Reuters, relaying the FT report, said it could not independently verify it; the companies declined to comment.
Reuters, citing Financial Times→
Frontier
Anthropic's revived Fable 5 ships new guardrails, and a US agency signs off
A day after Washington lifted the export limits that had pulled it offline, Anthropic switched Claude Fable 5 back on with a new safety classifier that blocks the reported cyberattack bypass in over 99% of cases. Requests it blocks are redirected to the less capable Opus 4.8, with a user notice. Commerce Department researchers at CAISI tested the safeguards and called them 'extraordinarily strong,' though Anthropic says they now trip more on routine coding and debugging. Access returns with 50% of weekly usage included through July 7, then metered.
Research
Can a chatbot suffer? The big AI labs are hiring people to find out
Anthropic, Google and Meta have hired neuroscientists, philosophers and computer scientists to study 'model welfare': whether their systems might develop something like emotions, and what would be owed to them if so, the Washington Post reported. Anthropic has a team probing its models for behavior like panic or anxiety, and a paper it published in April cataloged 171 'emotion concepts' in model outputs without claiming the models feel them. Most scientists doubt today's chatbots are conscious; the labs are staffing up anyway, wary of a moral question arriving before the science does.