Low-key dark still-life of a blank anonymous press credential badge half-blocked by a matte black padlock, lit by a single amber glow on the shackle — press-pool blackout and cyber-lock mood. No logos.

The Friday AI brief: Networks black out Xi’s AI dinner after a press-ban TRO, OpenAI lines up GPT-6 Cyber, and Anthropic books Akamai compute plus founder voting

September 25, 2026 Aaron Harme

A federal judge ordered the White House to restore CNN, MS NOW, and Politico, but major U.S. networks still refused live video of Xi’s arrival for the AI-and-trade state dinner. OpenAI is preparing GPT-6 Cyber and a first-of-its-kind product to help customers deploy it, with Daybreak Red already in alpha. Anthropic signed a seven-year, $11.6 billion Akamai cloud deal with a warrant for about 5% of the company, and is asking shareholders for Palantir-style 50.1% co-founder voting control before a record IPO path. Overnight, Google set an October 1 Falcon 9 flight for Project Suncatcher TPUs in orbit.

In this Brief

A judge restored CNN, MS NOW, and Politico after a weeklong White House ban, and networks still skipped live video of Xi’s AI state-dinner arrival
OpenAI is preparing GPT-6 Cyber and an unnamed product to help customers deploy it, with Daybreak Red alpha already underway
Anthropic signed a seven-year, $11.6 billion Akamai cloud deal and walked away with a warrant for about 5% of the company
Anthropic is asking shareholders for Palantir-style 50.1% co-founder voting control before what could be a record IPO
Google will fly TPUs on a SpaceX Falcon 9 on October 1 under Project Suncatcher to test AI chips in orbit

A judge let CNN back in, and America still missed live TV of the Xi AI dinner

Setup: A federal judge ordered Trump’s White House to restore CNN, MS NOW, and Politico after a weeklong media ban, a short court order that pauses the ban while a First Amendment suit continues. The Wall Street Journal reports U.S. District Judge Timothy Kelly (D.C.) ruled after midnight Thursday that yanking press badges was likely done without “constitutionally adequate due process,” and that officials offered little to back a national-security justification.

Yahoo (Jeremy Barr and Robert Mackey, AP cluster) says CNN and MS Now journalists were still denied access to the Xi state dinner hours after the restore order. CNN said the White House offered photojournalist-plus-audio coverage only, excluding a reporter and producer. MS Now said Laura Barrón-López was blocked and told to “talk to Steven,” communications director Steven Cheung. Lawyers said the White House had “repeatedly violated” the overnight order. The White House streamed Xi’s arrival on YouTube, but no major U.S. networks provided live broadcast video in solidarity with CNN. Talks are framed on AI, Iran, and Taiwan. Variety notes Paramount Skydance CEO David Ellison was at the dinner while CNN and MS NOW were barred from covering the arrival, with tech guests including Bezos, Cook, Pichai, Altman, Huang, Zuckerberg, Nadella, and Musk.

The superpower AI summit played out with a hole in the broadcast pool, the shared camera feed networks normally rely on for White House arrivals.

OpenAI’s next ship: GPT-6 Cyber, and a product to deploy it

Setup: After a run of rogue-agent scare stories, OpenAI is lining up a cybersecurity model and a first-of-its-kind product that helps customers actually run it. Fortune (Emily Forlini; Sep. 24) says OpenAI is preparing to preview GPT-6 Cyber “in the coming weeks,” and to unveil a yet-to-be-named product that helps customers deploy the model in a more secure, automated way, creating automated workflows, patching vulnerabilities amid AI-powered cyberattacks, and giving OpenAI more oversight into how the models are used.

Limited customers in application-only Daybreak Red, OpenAI’s invite-only cyber program for partners testing its hardest security models, already have GPT-6 Cyber for alpha testing. Daybreak also has a Blue tier that is more generalized and still application-only. Fortune frames GPT-6 Cyber as the fourth cybersecurity-focused model this year after GPT-5.4 Cyber (April), GPT-5.5 Cyber (June), and GPT-5.6 Cyber (August). Enterprise cyber is now led by CRO Dali Rajic, who joined in August. Earlier this month OpenAI said it will invest $1 billion to subsidize cybersecurity products for critical services. A separate source told Fortune a dozen or more other products are on the DevDay Sep. 29 pile, most of them not cybersecurity. The Next Web restates the Fortune cluster.

The model is the product, and the unnamed deploy tool is how OpenAI wants to sell the oversight.

Anthropic books $11.6 billion of Akamai cloud, and a warrant for about 5%

Setup: Claude’s lab is not only renting GPUs from the usual giants. It signed a seven-year, $11.6 billion cloud deal with Akamai, with room to grow toward about $20 billion, and walked away with a warrant, a contract right to buy stock later at a set price, for about 5% of Akamai’s stock. The Next Web reports Thu Sep. 24, summarizing Akamai’s announcement and Bloomberg: Akamai issued Anthropic a warrant for 7.7 million shares on an as-converted basis at an exercise price of $111.33, in the form of non-voting convertible Series B preferred.

About 2% vests with the base commitment. The remaining 3% vests only if expansion happens, at 1% per $3 billion of extra services. The deal builds on a prior $1.8 billion agreement. Akamai CEO Tom Leighton said the contract should start in the second half of next year, with $150 million to $300 million in Anthropic revenue next year rising to an annual run rate of about $1.7 billion by 2028. Akamai expects to spend about $5.5 billion in capital to serve the contract, more than six times its 2025 capital spend on Bloomberg’s framing via TNW, and is raising 2026 capital spending by about $1.7 billion. Shares rose as much as 17% in late trading to $129.60. The Wall Street Journal notes the new commitment adds to upwards of $2.8 billion in existing multi-year cloud commitments across Akamai’s customer base this year.

The CDN that used to cache websites is now in the AI power business, and the equity sweetener is the circular-deal argument Wall Street already knows.

Longer cut: Anthropic signs $11.6 billion Akamai cloud deal with a 5% warrant.

Anthropic wants Palantir-style founder voting at 50.1% before the IPO

Setup: Before what could be a record IPO, Anthropic is asking shareholders to hand CEO Dario Amodei and six co-founders a special stock class with majority voting power. Channel NewsAsia (Reuters, citing The Information) reports the structure would award the seven co-founders a combined 50.1% of voting power, emulating the founder-control setup at Palantir, and would apply as long as three of the seven retain a minimum number of shares.

A significant exception is election of board members. The board has seven seats, one currently vacant. Anthropic is also planning an employee special class of stock as tie-breaker votes on some issues. Reuters earlier this month said the IPO could slip to after U.S. midterms in November. Anthropic raised $65 billion at a post-money valuation of $965 billion in May. CNBC-TV18 prints the same Reuters wire. Anthropic did not immediately respond to Reuters.

The same week Trump allies cast Amodei as doomerism’s face, the filing is founder control before the float, as long as enough co-founders keep the special stake.

Google will fly TPUs on a Falcon 9 next week to see if AI barns can live in orbit

Setup: On October 1, Google will fly Tensor Processing Units, its custom AI chips, on a SpaceX Falcon 9 under Project Suncatcher, an experimental bid that could eventually put AI data centers into orbit. The Verge (Emma Roth; Sep. 24) reports the goal is to measure how the chips handle the physical stress of spaceflight and the radiation and thermal extremes of space.

Google plans two satellites into orbit next year. Travis Beals, Project Suncatcher’s senior director of product management, tells the New York Times via Verge that the chips can only run about 15 minutes before they need to be turned off to cool down. In vacuum there is no air to carry heat away, so heat has to move through heat pipes and radiators. Google’s research blog says the first orbit test flies on the Transporter-18 rideshare with SpaceX, developed in partnership with Planet, with about 10 minutes to low Earth orbit and sustained acceleration up to 10 times the force of gravity. The Next Web adds that low-Earth-orbit panels can collect up to eight times more solar power than on the ground, that Trillium TPUs withstood more than a five-year mission dose in a UC Davis proton-beam test, and that Google research estimates launch prices need to fall to about $200 per kilogram for space compute to compete.

The same week Meta printed up to $145 billion of terrestrial barns, Google’s first move upstairs is a hardware-survival test.

Longer cut: Google will launch TPUs on a SpaceX Falcon 9 October 1 under Project Suncatcher.

Also noted

Oracle sent a force majeure notice, a contract escape hatch for events outside its control, on Project Jupiter in New Mexico, seeking the right to delay payments if the OpenAI-tied Stargate campus misses a 2028 open over a blocked gas pipeline for up to 2.45GW of Bloom Energy fuel cells, The Next Web reports summarizing Bloomberg.
Trump and House Speaker Mike Johnson are expected to meet tech CEOs on AI next Tuesday, Sep. 29, with no public guest list yet and the House already out until after the midterms, ABC News and The Next Web report.
The White House Office of the National Cyber Director asked OpenAI and Anthropic not to share new models with the UK AI Security Institute until U.S. government testing is done; Anthropic kept Claude Mythos 5.1 in a U.S.-only set, Politico reports.
Nscale filed for a New York listing at up to $35 billion while posting a $1.02 billion half-year loss, with Nvidia taking $1 billion in convertible notes, as its Loughton “sovereign” AI campus in Essex may wait on grid power into the mid-2030s, Fortune and The Guardian report.
OpenEvidence, the AI search engine for doctors, raised $250 million at a $15 billion valuation and could be open to a sale, Business Insider reports.
A Brookings-backed estimate via The Wall Street Journal projects $10.3 trillion of U.S. data-center and AI infrastructure spend from 2025 to 2032, about 3.6% of GDP per year.
AWS VP of agentic AI Swami Sivasubramanian told HumanX that almost 90% of Amazon’s early AI agent prototypes from two years ago never reached production, The Next Web reports.
Meta’s up-to-$145 billion 2026 capex guide and the Xi state-visit AI agenda remain live from Thursday’s Brief. Longer cut: Meta plans up to $145 billion in 2026 capital spending as AI and personal agents scale. Prior Brief: The Thursday AI brief.