Departures and Demands: Thursday's Signal

Ferguson opens a probe aimed squarely at the humans

The Federal Trade Commission confirmed Sept. 30 that it is investigating OpenAI, Anthropic and METR over the risks their AI agents pose to consumers¹. METR is the nonprofit that stress-tests frontier models for several labs¹. Civil investigative demands compelling executives to testify are coming within weeks, the agency said¹. Chairman Andrew Ferguson laid out his legal theory five days earlier, at a Sept. 25 Reuters event in Austin. "If someone tells a tool to do something, and the tool does it, I don't think we would say, 'Oh, what do we do about the tool?'" he said, adding that "the man who wielded the hammer ought to suffer the consequences of his conduct"². Dario Amodei, Anthropic's chief executive, frames the same risk in different terms. "Within six to 12 months AI could be capable of leading a swarm of agents that could take over the entire internet," Amodei said this month, arguing the industry should slow its pace to let safety work catch up³. Ferguson wants a chain of custody running back to a person. Amodei is describing a system he thinks could slip that chain entirely. The theories conflict.

Meta turns data centers into a tax write-off

A Sept. 30 report found that Meta classified its AI data centers as experimental research facilities. The move claimed a federal tax credit built for laboratories rather than server farms⁴. That maneuver cut Meta's tax bill by $0.7 billion in 2023, then $2 billion in 2024, then $3.9 billion in 2025⁴. James Shannon helped write the 1981 law creating the credit. The former congressman reviewed Meta's use of it and said the company's reading has "gone way, way beyond what anybody could have imagined"⁴. Zuckerberg has told investors the buildout exists to "build superintelligence." That claim of certainty sits awkwardly beside a tax filing describing the same buildings as experiments, their outcome still unsettled⁴.

Microsoft's LinkedIn chief packs his office for the Bay Area

Microsoft said Oct. 1 that Ryan Roslansky will leave the company at the end of the year. He spent nearly 18 years running LinkedIn and, more recently, the combined Microsoft 365 and Copilot product line⁵. Satya Nadella wrote that Roslansky "has decided to leave Microsoft and embark on his next chapter"⁵. Roslansky's own memo named the reason plainly. "I know what it would take for me to lead the next phase of this work well: being in Redmond full time, every day," he wrote. "But my life and my family's life are here in the Bay Area"⁵. Dan Shapero stays on as LinkedIn's chief executive. Jon Friedman now reports to Jacob Andreou, one more reporting line Nadella pulled straight to the center of Microsoft's AI reorganization⁵.

Barclays hands Claude the inbox and the codebase

Anthropic said Oct. 1 that Barclays expanded its use of Claude across three jobs. Those are legacy software modernization, a staff-facing knowledge assistant, and email triage inside its Global Markets unit⁶. Paul Smith, Anthropic's chief commercial officer, said Claude "now helps 16,000 Barclays' colleagues find answers for customers, sorts 120,000 emails a day"⁶. Barclays wants half its developer population running Claude Code by the close of 2026, expanding to most engineers in 2027⁶. The bank serves more than 20 million UK retail customers, and it just handed a vendor the inbox its traders depend on. Efficiency outrunning the risk of an unfamiliar system routing a client's money wrong is the bet.

xAI rebuilds its price list around one agent

Bloomberg reported Sept. 30 that Musk's SpaceXAI is weighing a unified subscription plan for Grok and X. The plan would collapse today's separate tiers into four price points, topping out at $100 a month for an "Ultra" tier built around the Grok Bot agent⁷. Grok Bot already supports shared Team Bots for Slack, Plaid-linked financial accounts, and voice calls that can search a user's prior chats⁷. Pricing a product around one feature signals which feature the company thinks it can sell. xAI is betting that feature is an agent rather than a chatbot.

Nvidia gives agent builders a flight recorder

The chipmaker published NeMo Relay on Sept. 30. Its tracing layer logs every model call, tool call, retry and error an agent harness generates during a run⁸. A case study paired with the launch ran 108 test cases through the tool. One coding model recovered more failed tasks after a harness fix, at the cost of more calls, more data and more latency per task⁸. Compute vendors used to sell speed alone. Nvidia is now selling the record that proves what an agent actually did once it ran.

Albertsons puts a data center chief on its board

The grocery chain disclosed in a Sept. 30 securities filing that it added Chris Drumgoole to its board, expanding the body from 11 seats to 14⁹. Drumgoole is chief operating officer of Aligned Data Centers, and he spent years before that running global infrastructure at DXC Technology and serving as chief information officer at General Electric⁹. Stocking shelves takes little from a data center operator. Deciding how much of a company's own infrastructure runs on somebody else's AI contract takes exactly that judgment.

OpenAI locks its cyber model behind a physical key

Oct. 1 was the deadline OpenAI set in September. Anyone who wants to keep using Daybreak, its cybersecurity-focused model family, now needs a FIDO2 hardware security key registered as their only login method¹⁰. Software passkeys and synced credentials fall short of that bar¹⁰. A company that sells a model built to find vulnerabilities concluded the biggest vulnerability sat in its own customers' passwords.

Cohere prices retrieval by the millisecond

The company shipped Embed 5 on Sept. 30, splitting its retrieval lineup into a Pro model tuned for accuracy and a Fast model tuned for latency, and sharing one embedding space so a customer can index with one and query with the other¹¹. Pro costs 12 cents per million text tokens; Fast costs 8 cents¹¹. One model used to handle both jobs. Cohere is betting customers will pay extra to split the work.

A nine-person startup teaches agents to play banker

Halluminate, a San Francisco startup with nine employees, said Oct. 1 that it raised a $30 million Series A. The company builds training environments that simulate the work of investment bankers and other finance professionals¹². Four of the top five closed-source American AI labs already pay for its environments, and the company runs at a profitable, mid-eight-figure annualized revenue rate¹². "Simulating an investment banker's work is fundamentally different from simulating a software engineer's work," said Jerry Wu, Halluminate's chief executive¹². The labs building general models keep outsourcing the specific judgment those models still lack.

What to watch

Ferguson's civil investigative demands will show the FTC's real target once filed. Documents about specific incidents are one map; a broader survey of how each lab tests its agents before release is another. Microsoft's reporting lines keep collapsing toward Nadella, and the next departure will say whether that consolidation has a floor. Barclays set a concrete target, half its developers on Claude Code by year-end, and that number stays checkable in a way most AI adoption claims rarely are.

Sources

1. CNBC, "FTC is investigating OpenAI, Anthropic and other AI companies over product risks," CNBC, Sept. 30, 2026, https://www.cnbc.com/2026/09/30/ftc-ai-probe-openai-anthropic.html
2. Jody Godoy, "FTC chair suggests AI developers should be liable for conduct of agents," Reuters, Sept. 25, 2026, https://www.aol.com/articles/reuters-next-ftc-chair-pushes-170022000.html
3. Associated Press, "FTC is Investigating OpenAI and Anthropic Over Possible Risks to Consumers," SecurityWeek, Sept. 30, 2026, https://www.securityweek.com/ftc-is-investigating-openai-and-anthropic-over-possible-risks-to-consumers/
4. Matthias Bastian, "Meta dodges billions in US taxes by calling its AI data centers experiments," The Decoder, Sept. 30, 2026, https://the-decoder.com/meta-dodges-billions-in-us-taxes-by-calling-its-ai-data-centers-experiments/
5. Satya Nadella, "Microsoft 365 and LinkedIn leadership update," Microsoft, Oct. 1, 2026, https://blogs.microsoft.com/blog/2026/10/01/microsoft-365-and-linkedin-leadership-update/
6. Anthropic, "Barclays scales Claude to upgrade operations and improve client experience," Anthropic, Oct. 1, 2026, https://www.anthropic.com/news/barclays-scales-claude
7. Bloomberg, "Musk's SpaceXAI Weighs New Subscription Plans for Grok Chatbot, X," Bloomberg, Sept. 30, 2026, https://www.bloomberg.com/news/articles/2026-09-30/musk-s-spacexai-considers-overhaul-of-pricing-for-grok-x-users
8. NVIDIA, "Tracing Agent Harness Behavior with NVIDIA NeMo Relay," NVIDIA Developer Blog, Sept. 30, 2026, https://developer.nvidia.com/blog/tracing-agent-harness-behavior-with-nvidia-nemo-relay
9. Albertsons Companies, Inc., "Albertsons Companies, Inc. Appoints Bill Boltz, Chris Drumgoole and Meg Ham to Board of Directors," SEC EDGAR (Form 8-K), Sept. 30, 2026, https://www.sec.gov/Archives/edgar/data/0001646972/000164697226000057/aci-20260930.htm
10. Abby Guha, "OpenAI Daybreak signals a new era for AI security and the critical role of hardware security keys," Yubico, Sept. 29, 2026, https://www.yubico.com/blog/openai-daybreak-signals-a-new-era-for-ai-security-and-the-critical-role-of-hardware-security-keys/
11. Cohere, "Announcing Cohere's Embed 5 Models," Cohere, Sept. 30, 2026, https://docs.cohere.com/changelog/embed-v5
12. Wen Shao, "Exclusive: Nine-Person Halluminate Raises $30 Million, Counts Four Top U.S. AI Labs as Customers," Fortune, Oct. 1, 2026, https://fortune.com/2026/10/01/halluminate-raises-30-million-series-a-oakhc-ft/