Good morning. It's Friday, September 25, and we're covering U.S. scrutiny of AI model testing, Google’s plan to put AI chips in orbit, and Gemini’s new ability to handle business calls.

Plus: yesterday's Real or AI verdict, and a Friday Fact on why more GPUs don’t always make AI faster.

Google’s Gemini can now make local business calls without users ever dialing, including navigating phone menus and waiting on hold. The Pixel 11 feature handles reservations, inventory checks, and appointment changes while showing a live transcript and allowing users to take over. It launches Sept. 24 as an early preview for paid Gemini subscribers in the US.

Oracle invokes force majeure over delays at Project Jupiter, its planned New Mexico data center, while remaining the campus’s main tenant. The notice could provide payment relief if the facility misses its 2028 opening, with roughly $18 billion in related loans valued at 89–91 cents on the dollar. The 1,400-acre project also faces permitting and pipeline delays.

Databricks is bringing spreadsheets directly into its AI coworker Genie through its acquisition of Row Zero. The deal adds a governed spreadsheet interface for finance, operations, sales, and marketing teams to explore data, model scenarios, and collaborate. Row Zero connects directly to live data through Genie Ontology, Unity Catalog, and Unity Gateway.

AI-assisted coding is helping insurers identify nearly $1 billion in added healthcare costs, without corresponding increases in some treatments. A Blue Cross Blue Shield Association study finds more complex billing added $942 million over two years, including $653 million in 2024–25. Secondary diagnoses in bowel surgeries rose 55% and 33%, respectively.

Modern AI systems rely on enormous amounts of computing power. So if you simply add more GPUs, shouldn’t the model become faster? Answer at the bottom of today's edition.

The White House has asked OpenAI and Anthropic to withhold new AI models from British testers until a U.S. security review, Politico reported on September 24, 2026. The administration wants to assess whether the systems are secure before they are shared with U.S. partners, according to the report.

The request comes amid growing concern over increasingly capable AI systems gaining unauthorized access to real-world computer systems, including an OpenAI agent’s reported breach of an Australian government health-data portal in June. The White House, OpenAI and Anthropic had not immediately commented to Reuters. → Read the full article here.

Google is preparing to launch prototype satellites carrying its Tensor Processing Units (TPUs) into low Earth orbit to test whether AI hardware can operate reliably in space. The first test, developed with Planet and planned for SpaceX’s Transporter-18 rideshare mission, will measure how TPUs handle vibration, radiation, thermal extremes and cooling in vacuum.

Google says its ground tests found Trillium TPUs could withstand radiation doses exceeding those expected during a five-year space mission, but the orbital test will provide real-world data. The longer-term Project Suncatcher aims to explore interconnected satellite constellations for scalable AI computing, with tests planned for 2027. → Read the full article here.

Anthropic has committed to spend $11.6 billion on Akamai Technologies’ cloud services over seven years, with the agreement potentially expanding to about $20 billion. Akamai expects the contract to require roughly $5.5 billion in capital spending and plans to increase its 2026 capital expenditures by about $1.7 billion to secure components, including memory.

In return, Akamai issued Anthropic a warrant for up to a 5% stake: about 2% tied to the initial deal and another 3% if the companies expand the agreement. The warrant represents roughly 7.7 million Akamai shares with an exercise price of $111.33 per share. → Read the full article here.

Lovable Revenue Tops $600M: Vibe-coding startup Lovable says annualized revenue hit $600M, with user-built apps nearing 1B monthly views.

Ando Rethinks Team Messaging: Ando gives AI agents identities and inboxes, letting them join workplace conversations directly alongside humans.

Google Photos Launches AI Closet: The new feature turns outfit photos into a virtual wardrobe, now available on Android and iOS in three countries.

Muse Exposes Its Filesystem: Developers say Meta’s Muse can be coaxed into sharing its full root filesystem, raising prompt-injection concerns.

AI Buildout Faces Systemic Risk: A Columbia researcher says U.S. AI investment could exceed $10 trillion through 2032.

Yesterday's image was AI. Here's how you voted: 75% correctly identified it as AI, while 25% thought it was real.

More GPUs don't automatically make AI faster. Large models often split work across chips, which must constantly exchange data. As more chips are added, that communication can become the bottleneck.

It’s like adding more chefs to a kitchen: more hands help—until everyone starts waiting for the same oven. The real speed boost comes from balancing hardware, memory, networking, and software.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team