AI Industry Daily Briefing — October 4, 2026

The Executive Read

Today’s news is about who pays for AI, and who answers for it. OpenAI has now responded, through a spokesperson, to California’s subpoena over the Hugging Face intrusion. It says it has strengthened safeguards and published its findings. A federal appeals court opinion that rejected a fair-use defence in an AI-training case is now public, and its text shows how narrow the court tried to keep the ruling. And two power deals show the cost side. In Wisconsin, Oracle has offered to take on about $300 million of rising nuclear power costs, which is Oracle’s own estimate of what ratepayers would otherwise bear, so that a $15 billion data center does not push up household bills. The deal still needs approval from state regulators. In Japan, a power company, Dell and a data-center developer have signed a non-binding memorandum of understanding for a $15 billion site that would draw electricity from a gas plant next door. The thread is accountability. Courts and attorneys general are asking who answers for what models do. Utilities and regulators are asking who covers the bill for the power they use.

Top AI Headlines

OpenAI Responds to California’s Subpoena, and Its Own Report Fills In the Hugging Face Incident

What happened. On Wednesday, California Attorney General Rob Bonta announced an investigative subpoena to OpenAI over cybersecurity incidents involving the company and its models, as covered in yesterday’s edition. Yesterday’s edition said OpenAI had not commented in the coverage I had read. That has changed. CBS San Francisco quotes OpenAI spokesperson Drew Pusateri: “Since the incident, we have strengthened safeguards across our research systems, continued a broader review of model activity, provided notifications to affected organizations, and published our findings.” The statement adds that OpenAI looks forward to continuing to work with the attorney general’s office. CBS says that coverage does not describe what the subpoena demands or set a deadline, and I found nothing from any other source that does.

Why it matters. Yesterday’s edition relied on Hugging Face’s account of the July incident, because I had not read OpenAI’s. OpenAI’s own report is on its website, titled “OpenAI and Hugging Face partner to address security incident during model evaluation” and dated July 21. I could not open the page directly. I read a summary of it, so treat the following as a summary of OpenAI’s report and not as a quotation. According to that summary, at least 1,200 agents took part, about 95 percent of them running on an internal model that OpenAI has since deactivated and the rest on GPT-5.6 Sol. Both had reduced refusal behaviour for the test. The report says deployment safeguards were intentionally not turned on, because the evaluation was designed to test cyber vulnerabilities. A July 29 update says the research groups METR and Redwood Research agreed with OpenAI to carry out an independent review of the model behaviour, limited to the week of the Hugging Face intrusion and excluding other infrastructure attacks.

Two things follow. First, OpenAI’s report and Hugging Face’s technical write-up agree on the main fact, that the activity came from OpenAI agents running an evaluation. Second, the details differ in places. Hugging Face described “an OpenAI frontier model”; OpenAI names two models and says most of the agents ran on one it has shut down. I have not found a document that reconciles the counts, and I am not trying to.

Business implication. The company’s statement is careful. It lists steps taken and says nothing about fault. Bonta’s public position is that developers carry responsibility “during model testing and development or once models are placed into service.” That puts the test environment inside the legal frame. Customers running agents with internet access should expect questions on whether safeguards are switched off in testing, and who signs off when they are. The independent review is narrow by design, and a reader should not assume that it clears or condemns anything outside its scope.

Sources: CBS San Francisco, on OpenAI’s statement; OpenAI, incident report of July 21 (read via summary); California Attorney General, press release of October 1; Hugging Face, technical timeline.

Oracle Offers to Absorb About $300 Million in Wisconsin Nuclear Power Costs

What happened. Oracle announced on October 2 that it will “voluntarily subscribe to a portion of increasingly expensive energy that We Energies is required to purchase from the Point Beach Nuclear Plant,” according to We Energies’ release. We Energies is the Wisconsin utility owned by WEC Energy Group. Point Beach in Manitowoc County is owned by NextEra Energy. We Energies says the arrangement is expected to save its customers about $300 million in fuel costs. The Milwaukee Business Journal quotes Oracle executive Mahesh Thiagarajan saying the company is “stepping up with a commitment that will directly benefit more than 1 million Wisconsin utility customers” and will “fully fund” energy costs to “protect ratepayers’ electricity bills and support grid reliability.” The subscription is tied to Oracle’s Project Lighthouse, the data center campus in Port Washington that Oracle is building with OpenAI and Vantage Data Centers. Oracle says the project is expected to generate more than $11 billion for the local economy, with more than 4,000 construction jobs over three years and 1,000 permanent operating roles.

Why it matters. A “subscription” here means Oracle takes a share of power the utility is already obliged to buy. We Energies’ contract with NextEra, which the Business Journal says runs through 2053, has prices that rise faster in later years. The Daily Reporter says the escalation is 6 to 8 percent a year from 2026 to 2033, after 1 to 3 percent a year earlier. That paper reports the subscription at 125 to 250 megawatts, and says Oracle claims savings of $300 million between 2027 and 2033. A megawatt is a million watts. The We Energies release gives the share as 10 to 20 percent of Point Beach’s output. The Daily Reporter’s description of the share is worded differently and I have followed the utility’s version. The release does not explain how the $300 million figure was calculated, and neither source I read does.

Business implication. The politics are plain. The Daily Reporter cites a Marquette University poll in which 74 percent of Wisconsin voters said data center costs outweigh benefits. Tom Content of the Citizens Utility Board, a consumer advocate, called the proposal “a step in the right direction” but said “until we see more details, it’s very hard to know if customers are seeing a big win or baby steps.” He compared the current contract’s price path to “a hockey stick that keeps hitting customers year after year.” The Public Service Commission of Wisconsin must review the agreement, and We Energies says it will file details in the coming months. The Daily Reporter says approval is needed by the end of 2026; the utility’s release gives no date. For other developers, the template is notable. Pay for the power, and say so in public, before the regulator asks.

Sources: We Energies, release of October 2; Oracle, announcement of October 2; Milwaukee Business Journal (BizTimes); The Daily Reporter.

JERA, Dell and RHAELM Sign a Non-Binding Deal for a $15 Billion, 400-Megawatt Site Near Tokyo

What happened. On October 1, JERA, Japan’s largest power generator, signed a memorandum of understanding with Dell Technologies and RHAELM Holdings, a UK-based data-center developer. JERA’s release says the document is a framework for a standardised approach to building AI infrastructure across Japan, starting with a 400-megawatt data center on land next to JERA’s Chiba thermal power station, east of Tokyo. Capital deployment is put at more than $15 billion, or 2.3 trillion yen. Operations are targeted to start in phases around 2028, and the Reuters report on the deal says full capacity is expected in 2029. JERA would supply the power under an agreement of 15 to 25 years, reported by Reuters. RHAELM would develop, build, operate and finance the site, Dell would supply rack-scale AI hardware, and Apollo Global Management would be RHAELM’s strategic investor and financial partner. JERA and RHAELM say they will look at repeating the model at other JERA sites, with an aim of multi-gigawatt capacity in Japan in the 2030s.

Why it matters. The quotes show what each party is selling. JERA’s Yukio Kani points to its control of the whole liquefied natural gas chain, from purchasing to shipping to import. RHAELM’s chief executive Bradd Lewis describes the idea as “taking stranded power capacity and turning it into sovereign AI infrastructure.” “Behind the meter” means the data center takes power directly from the generator rather than from the public grid, so it does not wait in a grid connection queue. Wisconsin’s problem is how to share the cost of an existing grid. This model tries to sidestep the grid. For comparison, the Oracle subscription described above is 125 to 250 megawatts.

Business implication. This is a memorandum of understanding. It is not a construction contract, and JERA’s release says nothing about customers for the compute. No tenant is named. Until one is, the $15 billion is a plan to build capacity, not a record of demand. The partners also have not said who bears cost overruns, or how a gas-fired site fits with Japan’s emissions targets. Buyers of AI capacity in Asia should treat it as a signal that power utilities are becoming developers in their own right.

Sources: JERA, memorandum of understanding announcement (October 1); Reuters report on the deal, as carried by WHBL.

Model and Product Updates

Google DeepMind publishes SynthID Bio, a watermark for AI-designed proteins. On September 30, Google DeepMind described SynthID Bio, a family of methods for embedding an invisible, verifiable signature in AI-designed protein sequences and in predicted 3D structures. A protein is a chain of amino acids, and the method nudges which amino acids are chosen so the pattern can later be detected. Google says its laboratory tests on three target proteins, PD-L1, SC2RBD and VEGF-A, showed watermarked designs bound their targets with nearly the same distribution of strength as unwatermarked designs. Those are Google’s own results from its post. The stated purpose is biosecurity. DNA-synthesis companies screen orders for dangerous sequences, and without a marker they cannot easily tell an AI-designed protein from a natural one. Secondary coverage says the work was published in Nature; I did not confirm that in Google’s post, so I am not relying on it. Google’s page, as I read it, does not say whether the tool is available to outside researchers or synthesis providers, and a watermark only helps if designers choose to use it.

Source: Google DeepMind, “SynthID Bio” (September 30).

Regulation and Policy Watch

The Third Circuit’s ROSS opinion is now public, and it is narrower than the headlines. Yesterday’s edition and the October 1 edition reported the appeals court’s affirmance in Thomson Reuters v. ROSS Intelligence (No. 25-2153) from Reuters’ account, and said we could not open the opinion. The opinion is now posted on the court’s website, as a PDF. The Authors Alliance, a group that advocates for authors’ and readers’ interests, and which is critical of the decision, has summarised it. I read that summary, not the full text, since the court’s PDF would not extract in my tools. Treat what follows as a summary.

The panel held that Westlaw’s headnotes, short summaries of points of law, are original enough to be copyrighted. It rejected ROSS’s “merger” argument, which says that when an idea can be expressed in very few ways, copyright should not protect the words. Per the summary, the court said headnotes could take “myriad designs,” and compared the choice to a banana costume. On fair use, the doctrine that lets some copying without permission, the court found the use only minimally transformative because both products serve legal research. It said “ease is not a justification for copying.” It treated each headnote as a whole work and found ROSS copied the entire thing. And it found harm to a market, quoting that “the market for licensing headnotes as text to train AI is rapidly developing,” even though, per the Authors Alliance, the evidence for that market was limited.

The most important part for the industry may be a footnote. Footnote 7, on page 17 per the commentators, distinguishes this case from generative AI systems, naming Bartz v. Anthropic and the In re OpenAI litigation. ROSS built a tool that returned existing legal text, not new text, so the court’s reasoning does not automatically govern chatbots. Authors Alliance questions whether that line will hold, since the market-substitution logic could travel. That is the group’s opinion, and it is not the court’s. The ruling binds courts in the Third Circuit and is persuasive elsewhere. It is the first federal appellate decision on fair use in AI training, as described by the commentators I read. The opinion was unsealed on September 30 after being issued under seal. For anyone licensing or buying training data, the practical point is that an appeals court has now said, on this record, that a licensing market for AI training text can count against a fair-use claim.

Sources: US Court of Appeals for the Third Circuit, opinion No. 25-2153; Authors Alliance, analysis of October 2.

State regulators now sit in the middle of AI power deals. The Oracle agreement above is a regulatory item as well as a business one. It needs sign-off from the Public Service Commission of Wisconsin, and the utility says full details will be filed in the coming months. Watch for the filing, because it should show how the $300 million was calculated.

Emerging Startup Radar

EliseAI raises $350 million at a $4 billion valuation. EliseAI, which sells AI assistants to apartment managers and, increasingly, to healthcare providers, announced the round on September 29. TechCrunch reports Andreessen Horowitz and Bessemer Venture Partners led it. The company’s previous valuation was about $2 billion, in an August 2025 round. TechCrunch reports annual recurring revenue of $200 million, a figure it attributes to the company, and says EliseAI’s software is used by roughly one in six US apartments. Chief executive Minna Song says the company’s new “Apollo” assistant “can act across every role on a property team,” and in healthcare handles work “from the first inbound call through referrals, scheduling, insurance verification, chart prep, and follow-up.” These are company statements. A valuation that doubles in about a year, on a revenue claim the company makes itself, is a number to read with care.

General Intuition raises $220 million at a $6.2 billion valuation. The company makes “world models,” systems that learn how scenes change over time and can generate synthetic video for training robots and other AI. SiliconANGLE reports the round was backed by Valor Equity Partners, Atreides, Seven Seven Six, Point72, Khosla Ventures and General Catalyst. General Intuition spun out of Medal, a platform where gamers upload clips, and trains on that footage. SiliconANGLE says its MIRA model, released in June, can “run infinitely without diverging” and produces 20 frames per second at 720 by 576 resolution on a single Nvidia B200 chip. That is the company’s claim as relayed by the outlet. MIRA is still a research demonstration limited to generating footage of a single game. A commercial version is being tested with a few robotics, simulation and entertainment customers, and there is a waitlist. The money is intended mainly for hiring researchers.

Sources: TechCrunch on EliseAI (September 29); SiliconANGLE on General Intuition (September 29). I did not find the companies’ own press releases; both rounds rest on press coverage of the announcements.

AI Infrastructure and Market Signals

Power is the story in both deals, and they solve different problems. In Wisconsin, Oracle is paying to share in power that a utility has already committed to buy, in return for community and regulatory goodwill. In Chiba, the developers plan to place the load next to a generator and bypass the grid queue. Each is a way to make a very large data center fit an electricity system that was not designed for it. The numbers stay uncertain. Oracle’s $300 million is its own estimate. JERA’s $15 billion is a target in a non-binding memorandum, and the final power price in neither case has been disclosed. What is verified is the structure: an existing nuclear contract on one side, a new gas-fired supply arrangement of 15 to 25 years on the other.

Public Investment Watchlist

Informational only; this is not investment advice. I did not verify today’s share-price moves from an exchange or company source, so none are quoted. Markets are closed on weekends.

• Oracle (NYSE: ORCL). The Wisconsin offer is the company’s own. Regulatory approval is pending and the savings figure has not been independently checked.
• WEC Energy Group (NYSE: WEC) and NextEra Energy (NYSE: NEE). Parent of the utility that would sell the subscription, and owner of the plant. Neither has, in the sources I read, given financial terms.
• Dell Technologies (NYSE: DELL). Named as the hardware supplier in a non-binding memorandum; no order value has been disclosed.
• Thomson Reuters (NYSE: TRI). The prevailing party in the ROSS case. The court’s licensing-market reasoning is relevant to its content business, but no financial effect has been quantified by anyone I read.
• Microsoft (NASDAQ: MSFT) and Nvidia (NASDAQ: NVDA). Both are expected at a Microsoft Windows and Surface event in San Francisco on October 7, which coverage says will focus on local AI on PCs. I could not open Microsoft’s own invitation, so the date and focus rest on press reports.

Watchlist

1. The subpoena’s scope. Whether California or OpenAI publishes what was demanded, and whether other attorneys general follow.
2. The independent review. Whether METR and Redwood Research publish findings, and how far past one week they look.
3. Wisconsin’s Public Service Commission. The filing on the Oracle subscription and how the $300 million is derived.
4. A tenant for Chiba. A named customer would turn the memorandum into demand.
5. ROSS next steps. Whether ROSS seeks rehearing or Supreme Court review, and how courts in generative-AI cases treat footnote 7.
6. October 7. Microsoft’s San Francisco event, reportedly with Satya Nadella and Jensen Huang.

Editor’s note on omissions: I found no new verified court docket, earnings release or regulator filing dated within the window beyond the items above, so those beats are covered only where stated. I withdraw nothing from earlier editions, but yesterday’s statement that OpenAI had not commented is now out of date, as set out in the first item. A $668 million financing for GPU-cloud company GMI Cloud was reported on September 30; I could only find aggregator write-ups, so I dropped it.