AI Industry Daily Briefing — October 2, 2026

The Executive Read

Today’s news is about the physical and human limits around AI, not the models themselves. OpenAI said it parted ways with three safety researchers after concluding they mishandled sensitive information; the Wall Street Journal, which broke the story, reported that the material went to an outside AI-safety group. The company has not named the people or the group. Elsewhere, the hardware that AI runs on is being pushed into new places and policed in old ones. Google launched a small test satellite carrying four of its AI chips, to see whether data-centre hardware can survive in orbit. Federal prosecutors charged a California man, who has not been tried, with smuggling Nvidia-powered servers worth about $300 million to China. Michigan regulators approved the power contracts for a 1-gigawatt Google data centre, with conditions meant to keep the cost off other ratepayers. And a Canadian equipment maker said it will build 450 megawatts of gas-fired generators for an unnamed data-centre developer. The common thread is that building AI is now a question of electricity, export law and trust inside the labs as much as a question of algorithms.

Top AI Headlines

OpenAI Says It Parted Ways With Three Safety Researchers Over Handling of Sensitive Information

What happened. OpenAI said it has “parted ways with three individuals for violating our policies on accessing and handling sensitive company information,” according to a statement quoted by TechCrunch. The company said an investigation “confirmed that these individuals mishandled sensitive information outside established company procedures.” The Wall Street Journal broke the story and, per TechCrunch’s summary, reported that the information went to an outside AI-safety organisation. Neither the researchers, the organisation nor the information involved has been named. TechCrunch said it could not verify identities speculated about on social media. The WSJ story is paywalled, and the reporting is theirs.

Why it matters. The departures follow a New York Times report, which TechCrunch says ran two days earlier, about OpenAI executives dismissing employee safety concerns. I could not read that report and am not relying on its details. OpenAI has also, in recent days, cancelled its planned GPT-6.1 Astra release, a story covered in earlier editions. What is established today is narrow: the company has said what it concluded and that it acted. What is not established is what was shared, with whom, or whether it was a leak, a whistleblowing disclosure or ordinary research collaboration. The company’s own account is the only account.

Business implication. Labs that sell trust to enterprises and governments are now being asked how they treat people who raise safety concerns. Connecticut, as described below, has a law that took effect yesterday protecting employees at large frontier developers who report catastrophic risks. Nothing reported suggests it applies here, and I am not suggesting it does. But a dispute over where internal safety work ends and disclosure begins is the kind that tends to end up in front of regulators. Buyers should watch whether OpenAI says more.

Sources: TechCrunch; The Wall Street Journal (paywalled; original reporting).

Google Launches a Four-Chip AI Test Satellite

What happened. Google’s Project Suncatcher satellite launched on October 1 at about 2:32 p.m. Eastern aboard a SpaceX Falcon 9, according to Scientific American. It carries Google’s Tensor Processing Units (TPUs), the custom chips Google uses to train and run AI models. Reporting on the mission puts the count at four chips powered by solar panels of about one kilowatt; I confirmed that detail in secondary coverage, not in a Google statement. The satellite is a test: it is meant to show whether the chips survive launch and the radiation of orbit, and whether a new cooling system works where there is no air to carry heat away. Google’s own research post, published in November 2025, describes the longer-term concept as compact constellations of solar-powered satellites linked by laser connections, and says Google had tested its chips against radiation up to three times the expected five-year mission dose. Two further prototype satellites, built with the satellite company Planet, are planned by early 2027.

Why it matters. One kilowatt is about what a space heater draws. The Michigan data centre above is planned at a million times that. This launch does not move any AI workload to space. It tests one narrow question, whether the hardware survives. Google’s argument, from its 2025 post, is that solar panels in the right orbit can be up to eight times as productive as on the ground and generate near-continuous power. That is Google’s own analysis. Scientific American notes that significant engineering and cost challenges remain. Google’s post says launch costs would have to fall below about $200 per kilogram by the mid-2030s for the economics to compare with ground data centres, which is a condition and not a forecast.

Business implication. The motive is easy to see: ground-based AI capacity is running into power and permitting limits, as the Michigan item below shows. For now Suncatcher is a research bet. The signals to watch are whether the chips run as designed in orbit, and whether the 2027 satellites test the laser links that any larger system would depend on.

Sources: Scientific American; Google Research, “Exploring a space-based, scalable AI infrastructure system design” (Nov. 4, 2025).

US Charges California Man Over Alleged $300 Million Nvidia Server Smuggling to China

What happened. Federal prosecutors in Southern California charged Greg Lui, 38, owner of Earthmade Computer Inc. of City of Industry, with conspiracy to violate the Export Control Reform Act, outbound smuggling and conspiracy to commit money laundering, according to Courthouse News. Prosecutors allege that from 2023 to 2024 he bought Nvidia GPU servers from US manufacturers, falsely told them the servers were going to countries that do not require export licences, and sent them through Singapore and Malaysia on to buyers in China. The alleged value is about $300 million. Courthouse News quotes First Assistant US Attorney Bill Essayli: “Protecting America’s national security means keeping our advanced Super Intelligence technology from being used to strengthen our adversaries’ military capabilities.” Bloomberg also reported the case. A charge is an allegation. Lui has not been convicted.

Why it matters. US rules restrict sales of the most advanced AI chips to China. Chips shipped to a third country without a licence requirement are one route around them. Courthouse News reports that prosecutors earlier charged two Chinese nationals in Southern California with using the same routing method. The Defense Department’s Office of Inspector General and the Commerce Department are named as investigating agencies.

Business implication. Server makers and distributors face pressure to verify end customers, not just first buyers. For Nvidia, which is not alleged to have done anything wrong in this complaint, as far as I could verify, the question regulators keep raising is how much visibility it has after the sale. Bloomberg published a feature on October 1 on that question. It is paywalled, I could not read it, and I am citing only its headline.

Sources: Courthouse News; Bloomberg (paywalled); Bloomberg feature on Nvidia and smuggling cases (paywalled; headline only).

Michigan Approves Power Contracts for a 1-Gigawatt Google Data Center

What happened. The Michigan Public Service Commission approved DTE Electric’s contracts to supply Google’s planned data centre in Van Buren Township on October 1, according to Planet Detroit. The vote was unanimous. The contract term is 20 years. Google must pay at least 80% of its contracted capacity whether it uses it or not, and pays for at least 15 years of minimum monthly charges if it ends the deal early. Google must also fund development of up to 1,600 megawatts of renewable generation and 480 megawatts of battery storage. Service is due to start in December 2027, reaching full load of about one gigawatt in December 2028. Michigan’s attorney general, Dana Nessel, had argued for a 90% minimum to protect ratepayers from “stranded assets,” meaning grid investments that cannot be recovered if a customer uses less than planned or leaves. Planet Detroit reports the Wayne County Commission voted the same day to override the county executive’s veto of a tax exemption worth more than $125 million.

Why it matters. One gigawatt is about the output of one large nuclear reactor, or the demand of hundreds of thousands of homes. Planet Detroit describes this as Michigan’s first hyperscale data-centre decision made through a contested proceeding. The 80% minimum is the number to remember: it decides who carries the risk if AI demand falls short.

Business implication. Utilities are becoming the gatekeepers of AI capacity, and the terms they extract, such as long contracts, minimum payments and self-funded generation, are becoming the template other states will be asked to match. A tax exemption fight at the county level shows the local politics are not settled.

Sources: Planet Detroit.

Model and Product Updates

Barclays says Claude is now in daily operations. Anthropic published a customer story on October 1 about Barclays, which serves more than 20 million UK retail customers. According to Anthropic’s post, Barclays’ internal Colleague Knowledge Assistant is used by more than 16,000 employees and has handled over a million searches since launch in 2025, and Claude classifies about 120,000 emails a day in its Global Markets business. Barclays expects Claude Code, Anthropic’s coding tool, to reach 50% of its developers by the end of 2026. These are figures from a vendor-published case study featuring a customer, not independent measurement. The detail worth noting is the email routing: it is a plain, high-volume job rather than a headline use. Group Co-Chief Operating Officer Anne Marie Darling is quoted saying “the true measure of any technology is the impact it has on customers, clients, and colleagues.”

Source: Anthropic.

Regulation and Policy Watch

Connecticut’s AI law starts to take effect. Provisions of Connecticut’s SB 5, the Artificial Intelligence Responsibility and Transparency Act, began applying on October 1, according to the Future of Privacy Forum’s summary. Three pieces are live now. Developers of AI tools used in hiring and firing must give employers information about them, and employers must tell workers and applicants they are in use. Layoff notices filed with the state Labor Department must say whether AI contributed to the decision. And large frontier developers must set up anonymous channels for employees to report catastrophic risks and may not retaliate against those who do. The employment and layoff provisions are enforced by the state attorney general as unfair trade practices, with a cure period through the end of 2027 and no private right of action. Chatbot companion rules start January 1, 2027, and social media provisions on January 1, 2028. I am relying on the Future of Privacy Forum’s summary and did not read the enacted text.

Source: Future of Privacy Forum.

Michigan’s data-centre order, described above, is also a regulatory story: the commission conditioned approval on protections against other customers bearing project costs.

AI Infrastructure and Market Signals

Enerflex wins a 450-megawatt order for gas power at a data-centre site. The Canadian company announced on October 1 that it will design, build and assemble about 450 megawatts of “behind-the-meter” gas-fired generators for an unnamed North American data-centre developer, meaning power produced on site without a grid connection. Deliveries start in 2027 and finish in 2028. The contract value was not disclosed. Enerflex said its 2026 plans include about $15 million of related investment and that it authorised about $85 million for the business’s facilities, mostly in 2027. CEO Paul Mahoney said, in BNN Bloomberg’s account, that demand for prime power that does not require grid connection “continues to grow.” Four hundred and fifty megawatts is under half the Michigan project’s size, and it will run on natural gas rather than the renewables Google is being required to fund in Michigan.

TSMC reportedly weighs a second US campus. Bloomberg reported on October 1, citing people familiar with the plans, that TSMC is considering a multibillion-dollar campus in Texas separate from its $265 billion Arizona commitment, with each fabrication plant costing at least $20 billion. According to the syndicated version I could read, the talks are early and depend on Congress extending an advanced-manufacturing tax credit that expires at the end of the year. That version also says Taiwan’s Economic Daily News first reported the idea earlier in the week, and quotes TSMC saying “We have no comment on market rumors.” Nothing here is confirmed by the company.

Sources: Enerflex release via Stock Titan; BNN Bloomberg; Bloomberg (paywalled); Insurance Journal syndication of Bloomberg.

Public Investment Watchlist

Informational only; this is not investment advice.

• Enerflex (TSX: EFX; NYSE: EFXT). Shares rose after the 450-megawatt announcement: BNN Bloomberg reported a gain of about 15% on the Toronto exchange, to C$37.18, while Stock Titan’s data shows a close 11.67% above the prior day. The two figures measure different markets and moments. The order value was not disclosed, so the stock move rests on an undisclosed contract.
• TSMC (NYSE: TSM). The Texas report is unconfirmed, and the company declined to comment. The stated dependence on a tax credit is the detail to follow.
• Nvidia (NASDAQ: NVDA). Named in the smuggling coverage but not, in the sources I read, alleged to have acted wrongly.

Watchlist

1. The OpenAI dismissals. Whether the company, the researchers or the outside group says anything more.
2. FTC demands to AI labs. Reporting yesterday said formal civil investigative demands were expected in coming weeks; none has been confirmed as sent.
3. Suncatcher telemetry. Whether Google reports how the four chips performed in orbit.
4. The Texas tax credit. Congress’s decision on the advanced-manufacturing credit decides whether the reported TSMC campus moves ahead.
5. Connecticut enforcement. The cure period for employment violations runs through the end of 2027, so early enforcement is unlikely; companion-chatbot rules arrive January 1, 2027.
6. Lui’s case. The first court appearances and any plea in the Earthmade matter.

Editor’s note on omissions: no startup funding story from the past 48 hours met the sourcing bar today, so there is no Emerging Startup Radar. Instinct’s $1 billion round was covered on September 29.