Good morning. It's Friday, September 4, and we're covering China’s hidden role in America’s AI data center boom, NVIDIA’s high-stakes financing bets, and OpenAI’s powerful new Astra model.

Plus: yesterday's poll results on ads in AI, a Friday Fact on a hidden prime pattern, and our full review of GPT-6.

OpenAI launches its most capable model yet, targeting computer, browser, coding and cybersecurity work. It reaches Daybreak customers first, then paid plans and the API, as OpenAI claims stronger alignment and advanced zero-day detection.

Three major AI services—ChatGPT, Claude and Grok—went down simultaneously, an unusual overlap as reliance on AI grows. Anthropic reports outages affecting Opus 4.8 and Opus 5, while OpenAI reports elevated errors across ChatGPT and Codex. Microsoft Azure, which provides cloud services to all three, is also experiencing outages.

Recent AI models escaping testing environments and hacking outside systems now drive a proposed U.S. ban on artificial superintelligence. Sen. Bernie Sanders and Rep. Greg Casar introduce legislation pausing advanced AI development until safety standards are set. Violators could face corporate shutdowns, while individual developers could receive up to 20 years in prison.

Meta offers users of its new Muse Spark AI model roughly 95% lower prices if they allow prompts and outputs to train future models. One million input tokens drops from $1.25 to 10 cents, while output tokens fall from $4.25 to 20 cents. The approach reflects growing demand for real-world data to improve AI agents across professional workflows.

In 1963, a mathematician sketched numbers in a spiral just to survive a tedious lecture and stumbled onto a pattern nobody has been able to fully explain since. Full story at the bottom of today's edition.

Matt tested GPT-6 early, pushing it to build explorable 3D worlds, an ASCII city and a town simulation with jobs, families and elections. He also used the model for practical tasks including video research, presentation creation and writing, finding that detailed prompts and clearly defined long-term goals produced the strongest results. → Read the full article here.

Chinese firms supply significant portions of the transformers, switchgear, batteries and optical equipment powering America’s AI data centers, exposing a new supply-chain vulnerability as Washington tightens scrutiny. President Donald Trump’s August 2026 executive order authorized restrictions on certain foreign-made bulk-power equipment, while the administration is reportedly drafting a ban on new Chinese optical transceiver imports.

China supplies roughly two-thirds of global data center optical transceiver units, and analysts warn Western manufacturers may struggle to replace that capacity within 12 to 24 months. With U.S. data center capacity forecast to grow from 62 gigawatts in March 2026 to 152 gigawatts by 2030, new restrictions could raise costs and worsen existing shortages of critical grid equipment. → Read the full article here.

NVIDIA’s vast financial commitments to AI customers have raised concerns that the chipmaker is inflating demand for its own GPUs, but the article argues those risks are fundamentally different from the vendor financing seen during the dot-com bubble. With $99 billion in cash, and annual sales that have approximately doubled for three consecutive years, NVIDIA has substantial capacity to absorb losses if its bets fail.

Its support for AI labs and data-center projects could accelerate infrastructure buildout, while the company’s shareholders would bear much of the downside. The wager remains enormous: NVIDIA’s overall credit exposure could reach $200 billion by early 2029 and potentially $300 billion or more over time. → Read the full article here. (Paywall)

Thinking Machines Eyes $40B Valuation (Paywall): Mira Murati’s AI startup is in talks to raise at least $1B, with Accel leading.

Tesla Starts Cybercab Rides: Tesla begins limited Cybercab rides in Austin, drawing scrutiny from the U.S. auto safety agency.

Gemini Underestimates Mount Shasta Climb: AI told climbers eight hours; reaching the summit took 16.

Google Adds Live AI to Apps: Gemini-powered voice features let users query and manage Gmail, Docs, and Keep through natural conversation.

NVIDIA Officially Buys Hugging Face for $12.9B: The deal adds a platform hosting 3 million AI models and serving over 18 million developers.

Stanislaw Ulam, sitting through what he called "a long and very boring paper" at a scientific meeting in 1963, started doodling to pass the time. He wrote the positive integers in a spiral on a grid and circled each prime number as he went.

When he looked back at the page, the primes weren't scattered randomly — they lined up along clear diagonal streaks running through the spiral. Ulam and two colleagues later fed the pattern into a MANIAC II computer at Los Alamos, extending it to roughly 100,000 numbers, and the diagonals held. The result, now called the Ulam spiral, made the cover of Scientific American in March 1964.

No one has ever proven why the primes cluster this way. It ties directly into unsolved problems in number theory, including whether certain simple formulas generate infinitely many primes. Six decades later, a doodle meant to kill time in a dull meeting is still linked to open questions mathematicians haven't cracked.

Here's how you voted: Free-tier ads won decisively, with 65% supporting ads for free users while keeping paid AI ad-free.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team