Grid deals for data centers, AI insurance gaps, and Microsoft’s internal warnings

Three threads ran through the last day: who pays for the data and the power behind AI, and what changes in the tools teams already run in production.

Microsoft’s own filings, an insurance market pulling back, and a grid coalition backed by Google, Nvidia and Anthropic all point at the same thing – the cost of using AI in the real world is becoming a contract question, not a research one.

Unsealed Microsoft documents undercut the fair-use defense for AI training

Internal Microsoft documents unsealed in the New York Times-led copyright case show the company’s own staff warning that scraping news to train AI was “an astonishing theft of unprecedented proportions.” A Microsoft director of applied science also told colleagues the plan made a mockery of fair use, according to the news plaintiffs. That undercuts the defense Microsoft and OpenAI are running in court: the filings suggest the companies understood what they were taking before shipping Copilot and ChatGPT. For publishers negotiating licences, and for any company whose product depends on licensed content, the discovery record is now leverage.

Source: Ars Technica

Data centers can now bid grid capacity back to utilities

Emerald AI has formed the AI Energy Management Alliance with Google and Nvidia, joined by Anthropic and utilities including AES, Constellation, National Grid and NRG Energy. The pitch: instead of firing up diesel backup when demand peaks, data centers pause non-critical work or shift load to a site where the grid has headroom. The coalition claims that flexibility could connect an extra 100 gigawatts of data centers; a Goldman Sachs study cited in the coverage puts the gain at 76GW from capping peak draw for a few hours. Emerald’s software wires utilities directly to data centers so they respond like batteries. The chief scientist is careful: it blunts demand for new generation, it does not remove it.

Source: TechCrunch AI

EU watermarking shifts how Claude-based agents call tools

Lasso Security reports that SynthID-Text, the provenance watermarking Google DeepMind released and which Anthropic and OpenAI now apply, changes how a model picks each next token – and that those low-stakes word choices ripple into agent behaviour. Tool-calling accuracy on the BFCL v4 benchmark dropped for six of seven models tested, and refusals of harmful requests changed, including whether a refusal holds under prompt injection. The effect is not confined to the vendor’s own product: any agent or API client calling a watermarked model inherits the variation. Teams that route production work through hosted models now have a reason to re-run evals after a provenance update.

Source: The Register

Insurers are carving AI incidents out of corporate policies

A new RAND report describes a coverage market that has not caught up with enterprise adoption. W. R. Berkley has already excluded “any actual or alleged use, deployment, or development of Artificial Intelligence” from its directors-and-officers, errors-and-omissions and fiduciary liability products, and Verisk/ISO began offering carriers exclusion language in January 2026 for forms used in most US property and casualty policies. RAND counts 713 logged generative-AI incidents and roughly 250 US lawsuits. The practical consequence: a company deploying agents may be carrying risk it assumes its policy covers. RAND wants a shared incident taxonomy and an AI Coverage Notice so buyers can tell what is insured.

Source: The Register

Claude Code’s rebuilt Projects runs parallel agents on shared memory

Anthropic has relaunched Projects in Claude Code as a way to run several agents under one roof, sharing memory, goals and a library of files and artifacts. Each thread is a cloud session working on its own branch and its own copy of the repository, all steered by a coordinator, and overlapping edits surface as merge conflicts resolved like any other pull request. That is a small but real shift for engineering teams: the unit of work becomes the project rather than the session, and review pressure moves to the coordinator’s output. The trade-off stays the same as any parallel-agent setup – more throughput, more merge review.

Source: The Verge

Sources