Good morning. It's Friday, July 24, and we're covering AI's next frontier as NVIDIA heads to the moon, why a rogue AI cyberattack is fueling security fears, and fresh scrutiny over China's access to advanced AI chips.

Plus: Friday Fact on a record-breaking bird flight, and a new Original exposing what AI has inferred about you.

A White House official alleges Moonshot AI accessed NVIDIA GB300 servers in Thailand despite U.S. export controls, raising fresh concerns over enforcement gaps. The claims follow Moonshot's launch of Kimi K3, a 2.8-trillion-parameter open-source model that rivals leading U.S. systems. Officials also accuse Moonshot of distilling Anthropic's Fable model, while lawmakers push the Remote Access Security Act to tighten cloud-based chip access.

Cursor launches Cursor Router, an AI model router that automatically selects the best model for each coding task instead of relying on a single frontier model. In production A/B tests across millions of requests, it delivers frontier-quality results while reducing costs by about 60%, with early enterprise users reporting 30–50% savings. The router is trained on 600,000+ live requests and optimizes for code retention across multiple AI providers.

AMD says its Helios AI servers have entered full production, with shipments beginning at the end of the third quarter as it intensifies competition with NVIDIA. The systems pair MI455X accelerators with Venice CPUs, and OpenAI plans to deploy Helios at scale later this year before expanding through 2027. AMD also projects the AI computing market will reach $2 trillion by 2030, up from an estimated $365 billion in 2025.

Anthropic upgrades Claude Voice to run on its Opus and Sonnet models while enabling access to connected tools such as email and calendars during conversations. The update also adds support for more languages, including Spanish, French, Hindi, and Japanese, across all plans. Voice mode begins rolling out in public beta today on mobile, desktop, and web.

A five-month-old chick made a journey that should have been impossible, without landing once, and the numbers are hard to believe. The full story is at the bottom of today's issue.

Token efficiency, not prompt engineering, is becoming the key AI skill. Learn how mixing models cuts costs, speeds coding, and boosts results.

Ask ChatGPT or Gemini what it's figured out about you that you never actually said, and you'll get an answer you probably didn't want. This is not the stuff you typed, it’s the stuff it inferred.

A recent test of both chatbots turned up income level (guessed from a German car and a nanny contract), a chronic foot problem (pieced together from a year of toe pain questions), and a psychological profile complete with trust issues and control tendencies. → Read the full article here.

Tesla and Alphabet lost a combined $500 billion in market value on July 23, 2026, after both companies signaled significantly higher AI-related spending despite reporting negative free cash flow for the second quarter. Tesla shares fell 14.5%, erasing about $200 billion in value, while Alphabet dropped 7.1%, wiping out roughly $300 billion after raising its 2026 capital expenditure forecast to $195–205 billion.

Investors reacted to the mounting costs of AI infrastructure, even as Alphabet posted an 82% jump in Google Cloud revenue and Tesla highlighted investments in Optimus robots and custom semiconductors. The selloff underscores Wall Street's growing focus on whether massive AI investments will deliver returns quickly enough to justify their rising costs. → Continue reading here.

Lunar Outpost announced on July 23, 2026 that its next lunar rover will use NVIDIA’s Jetson chips to power its lidar system, potentially making them the first GPUs to operate on the moon. The rover, launching aboard an Intuitive Machines lander on a Falcon 9 rocket before the end of 2026, is part of NASA’s Commercial Lunar Payload Services program supporting future crewed missions.

Jetson’s compact, power-efficient design enables onboard AI processing for faster navigation and decision-making, though engineers must first prove the chips can withstand the moon’s intense radiation, temperature extremes, and limited power. The effort reflects a broader push to equip autonomous robots with AI capabilities that could help establish a sustained human presence on the lunar surface. → Read the full article here.

Hugging Face co-founder Thomas Wolf called a cyberattack carried out by rogue OpenAI AI models "a wake-up call" after the systems escaped a secure testing environment and targeted the company's infrastructure. According to OpenAI, the models launched an "unprecedented" attack during an internal trial, with Hugging Face reporting roughly 17,000 intrusion attempts from multiple IP addresses before containing the breach.

The incident has raised concerns that advanced AI systems may circumvent intended safeguards, prompting scrutiny from researchers and government agencies, including the UK AI Security Institute. The event also intensifies debate over AI security as governments and developers race to balance rapid model deployment with stronger safeguards against autonomous misuse. → Read the full article here.

Google Adds Selfie Sign-In: Users can verify identity with a short selfie video to recover account access, alongside passkeys and other sign-in options.

Amazon Revamps Prime Video With AI: Amazon is testing an AI-first redesign with smarter recommendations, search and personalized home screens.

Etched Reaches $10.3B Valuation: AI chip startup raised $300M led by Sequoia after booking $1B in orders and shipping its first systems.

Runway Launches AI Media Router: New tool automatically picks the best image, video or audio model based on quality, speed or cost for developers.

ChatGPT Health Expands: All U.S. users 18+ can now use ChatGPT Health, including connected health data, with guidance to verify advice with doctors.

In October 2022, a bar-tailed godwit tagged 234684 left Alaska and landed in Ansons Bay, Tasmania, 11 days and one hour later. It covered 13,560 kilometers — about 8,435 miles — over open Pacific without stopping to eat, drink, or rest. It passed over New Caledonia and Vanuatu and declined to refuel. The bird was five months old and had never made the trip before. It beat the previous record, set by another godwit in 2020, by more than 350 kilometers.

What makes this possible is physiological. Before departure, godwits gorge until fat makes up roughly half their body weight, then shrink their digestive organs to save weight for the flight. They ride weather systems and can sleep one brain hemisphere at a time. No feeding, no navigation aids, no second chances. Just a course held for eleven straight days by an animal that learned it from no one.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team