Today's Headlines

  • A June fire at the $3.2 billion Lake Mariner AI data center — firefighters found no working alarm, no suppression system and three dead hydrants
  • Payouts from Anthropic's $1.5 billion settlement begin, and publishers and literary agents file claims on authors' shares
  • Microsoft rolls out GPT-6 Astra in Copilot Cowork and Copilot Studio
  • Artificial Analysis updates its benchmark and doubles the share of private test data to 40 percent
  • Design-industry bankruptcies in Japan hit 40 in the first half of 2026, up 166.6 percent year on year and the highest in 14 years

Two of today's stories ask who carries the duty and who takes the money. In both, the answer sits across several companies bound by contract.

Deployment moved as well. OpenAI's flagship GPT-6 Astra arrived inside Microsoft's Copilot for work.

Today's Top Three

A June fire at a $3.2 billion AI data center — ownership, operation and funding split across separate companies

A fire broke out in early June at Lake Mariner, the $3.2 billion AI data center in Somerset, New York, and reporting since then has left the question of who owns safety at the site unanswered, Ars Technica writes.

The fire started in a still-unfinished building. Firefighters found no working alarm, no suppression system and three dead hydrants, and the safety data sheets they are legally entitled to see reportedly burned in the blaze. Steve Matisz, chief of the Barker Fire Department, said his crew went in "kind of blind" through heavy black smoke from chemicals they could not identify. No one was hurt, and construction has barely slowed.

The site is a former coal mine on Lake Ontario and one of the largest AI data center buildouts in New York. TeraWulf owns and operates it on land leased from a company owned by its own CEO. Fluidstack, a UK company, will run the center. Google holds warrants for a future 14 percent equity stake and guarantees Fluidstack's lease payments. Anthropic is among the AI companies whose compute demand the facility exists to serve.

TeraWulf told Ars Technica that it is responsible for operational safety and emergency preparedness at the Lake Mariner Data Campus, including required safety systems and coordination with local first responders. Kerri Langlais, the company's chief strategy officer, said it had put in Knox boxes, additional hydrants and safety data sheet "go-bags" after an after-action review with the county. Matisz described a different picture in mid-August: a Knox box representative had visited the site with a meeting scheduled, while on hydrants he said, "As far as I know, [they] are still dry… I haven't seen work done on them."

The town is still waiting for answers of its own. TeraWulf's communications team has stayed unresponsive to community requests for a meeting, including one from Somerset town supervisor Jeffrey Dewart. Residents have asked what the town gets in return. A 2024 planning board presentation put employment at the fully built site, drawing up to 500 megawatts, at 35 to 40 people. The figures attached to the 2019 application for the site's power discount were 165 permanent jobs and $85 million in capital investment, filed by the landlord company owned by TeraWulf CEO Paul Prager.

State rules stop short of the facility. The New York State Legislature passed the Responsible Data Center Development Act in early June, and three months later it remains unsigned. The executive order Governor Kathy Hochul signed in July pauses new permits for facilities above 50 megawatts for a year, and Lake Mariner's 500 megawatts were permitted before it took effect. Hochul told reporters in July that despite the scale of these projects and the utility demands, data centers do not deliver significant, long-term jobs, and that she is challenging the industry to do it in a smarter way.

The reach of Anthropic's own commitment is part of the same picture. The company pledged in February to cover the electricity price increases consumers face from its data centers, and Lake Mariner falls under that covenant. What the pledge covers is electricity pricing, and verification of noise and clean energy commitments runs on a separate track. A source familiar with Anthropic's thinking said they had no details on what the company would need to see from Fluidstack or TeraWulf to confirm clean energy commitments are being met at the site, or whether Anthropic had ever asked.

Matisz closed the interview with a line that describes where the site stands. "If there's no hydrant for me, that fire is going to burn under their asses."

Anthropic's $1.5 billion settlement — publishers and literary agents claim shares of author payouts

Payouts from Anthropic's $1.5 billion copyright settlement have begun to flow, and publishers and literary agents are filing claims on authors' payments, TechCrunch reports.

The terms set the shape of the dispute. Anthropic settled a copyright class action last year, and the deal received final approval in July, allowing payments to move. The authors of nearly 500,000 titles will be paid $3,000 for each pirated work. Where a book is still in print with a traditional publisher, the money splits 50-50 between author and publisher. Where a book was self-published, or where the publisher let it go out of print and the rights reverted, the whole payment goes to the author.

What authors received this week were emails telling them someone else had claimed part of that money. Mystery and thriller author April Henry posted that HarperCollins claimed one of her books on the settlement seventeen years or more after the rights reverted, and that on the same day a credit alert told her the publisher had been added as her employer, a job she never held.

Victoria Strauss sorted the complaints at the blog Writers Beware into two broad categories. In one, publishers seek payment for works whose rights have reverted and on which their claim has lapsed. In the other, publishers seek the full 100 percent where they are entitled to 50. She wrote that she is "reluctant to attribute to malice what can be plausibly explained by poor recordkeeping," and noted that some publishers have already called it a mistake and asked Anthropic to fix it.

On scale, Strauss allowed that what she sees is "a peek through a small crack in a massive wall," then wrote: "But the unusually large number of reports I've received over the last two days, as well as the fact that authors are reporting the exact same errors over and over, suggest to me that these aren't the kind of routine glitches you might expect from such a large operation, but something much more wide[s]pread and systemic."

Authors Guild CEO Mary Rasenberger gave her reading to The New York Times. She points to bad recordkeeping and a confusing settlement process, and told the paper she sees no "grab by the publishers."

Literary agencies are filing too. Strauss said complaints have reached her about a number of them making claims, which she called surprising since "agents are not rightsholders in the books that they sell." Author Courtney Milan, the pen name of former law clerk and law professor Heidi Bond, put it on Bluesky: "Apparently some agents are trying to claim percentages on the Anthropic settlement, and I do not REMOTELY think they should do this, what the fuck, stop that shit!"

A date decides part of the argument. For an author to claim 100 percent of a title, the rights reversion has to have happened before August 10, 2022, the download date named in the settlement.

Microsoft rolls out GPT-6 Astra in Copilot Cowork and Copilot Studio

Microsoft announced on September 4 that GPT-6 Astra, OpenAI's flagship model, is rolling out that same day to users in Copilot Cowork and Copilot Studio.

What the company leads with is the size of the unit of work a user can hand over. Instead of breaking work into smaller pieces and guiding it step by step, Microsoft writes, customers can delegate larger tasks and spend their time reviewing results and making decisions.

The other half of the pitch is the context the model can reason over. Work IQ grounds GPT-6 Astra in a user's files, meetings, chats and business data, and Microsoft says that grounding stays within existing permissions.

Access depends on the organization. Availability may vary by region and by organization, and admins manage access and configure availability through the Microsoft 365 admin center.

Anthropic's Claude Fable 5.1 joined the same lineup on September 1.

Other Developments

Models

  • OpenAI product lead Thibault Sottiaux posted on X on September 6 that GPT-6 Astra's "low" setting outperforms GPT-5.6 Sol's "high" setting, and suggested that users happy with Sol on high move to Astra on low or medium. He said the same day that token consumption on the long tail of heavy tasks for subscription users has been cut by up to three to four times. OpenAI says GPT-6 Astra's "low" beats GPT-5.6 Sol's "high" (ITmedia AI＋)

Research

Products

  • Tesla Japan will give the Cybercab, the two-seat robotaxi vehicle with no steering wheel or pedals, its first public showing in Japan. The tour runs at OMAKASE Aoyama Garden by GMO in Tokyo from September 11 to 14, Tesla Grand Front Osaka from the 17th to the 20th, Tesla Nagoya Noritake on the 22nd and 23rd, and Tesla Shinjuku from the 26th to the 28th. Cybercab service began in parts of Austin, Texas on September 3. Tesla shows the Cybercab robotaxi in Japan for the first time, in Tokyo, Osaka and Nagoya (ITmedia AI＋)
  • Highlanders, a Tokyo startup spun out of the University of Tokyo, announced the HLQ EDGE quadruped robot on September 7, built and developed in Japan. Its in-house AI uses a world model of physical space to adjust foot placement and posture to the terrain as it walks. The robot runs three hours on its internal battery, carries up to 60 kilograms, has 12 degrees of freedom and 330 newton-meters of maximum joint torque, and runs its AI on an NVIDIA Jetson Orin NX. Tokyo University startup Highlanders shows a domestically built quadruped robot with in-house AI (ITmedia AI＋)
  • Fukushima Prefecture opened AI tools to about 6,000 staff for fiscal 2026. Its fiscal 2025 pilot ran on 100 accounts, 50 on an NTT East service and 50 on Microsoft 365 Copilot, with roughly 80 percent of participants reporting efficiency gains and about 70 percent reporting better work quality, while usage split along how interested each employee was. The prefecture caps token use on the newest models, leaves older models unlimited, and logs inputs while putting active use ahead of tight surveillance. Fukushima Prefecture's contrarian strategy for M365 Copilot adoption (ITmedia Keyman's Net)
  • Hitachi announced four new services in its physical AI line. HMAX Data Center, which manages a data center's power, IT and cooling as one system, is slated for January 2027. For the security service HMAX Cyber, Hitachi says its White Hacker team will use Claude Mythos Preview through Anthropic's Project Glasswing security program. Hitachi unveils four physical AI services (ITmedia Enterprise)
  • GMKtec unveiled the EVO-X5 Pro mini PC at IFA 2026. It pairs a Ryzen AI Max+ PRO 495 with up to 192GB of LPDDR5X-8533 memory, which the company says runs 300-billion-parameter-class LLMs locally and offline, alongside an integrated Radeon 8065S GPU and an NPU rated at up to 55 TOPS. GMKtec's EVO-X5 Pro packs 192GB for local 300B-class LLM inference (PC Watch)

Business

  • Bankruptcies are rising in Japan's design industry. Tokyo Shoko Research counted 40 in the first half of 2026, up 166.6 percent year on year and the first time the half-year figure has passed 40 since 2012. The firm groups the causes in three: clients bringing design work in-house with generative AI and cutting outside orders; design firms centered on print media such as magazines losing orders as clients shift to digital; and a rise in homebuilder failures reaching interior design firms. Bankruptcies in Japan's design industry rise 166 percent, and generative AI is only part of it (ITmedia Business Online)
  • OpenAI announced a program on September 7 with the publishing group WAN-IFRA and AIRPPU, Ukraine's association of independent regional newspapers, to build practical AI capability at independent Ukrainian news publishers. The program covers newsroom AI projects and business transformation support. Supporting independent journalism in Ukraine (OpenAI)
  • Google DeepMind selected 16 organizations across the Asia-Pacific region for the first cohort of its AI for the Planet accelerator. The startups, nonprofits and research teams, including participants from Japan, work on biodiversity, sustainable agriculture and carbon solutions, and the three-month program gives them access to Google's frontier models and mentorship from company experts. Backing 16 green AI projects in Asia-Pacific (Google)

Source: selected by the editors from the AI news inbox collected on September 8, 2026 (25 items, 2 primary and 23 secondary).