▓▓ ██ ▓▓ ▒▒
▓▓ ██ ▒▒ ▓▓ ░░ ▒▒
▓▓ ░░ ██ ▒▒ ██ ▓▓ ░░ ██ ▒▒
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
─────────────────────────────────────────
No. 26 · Friday, 17 July 2026Happy Friday. A Chinese lab's free-to-download model just beat Anthropic's and OpenAI's best at writing web code. SpaceX slipped under its own IPO price a month after the biggest listing ever, and TSMC posted a record quarter while promising another $100 billion for Arizona. Microsoft, meanwhile, started coaching its salespeople to talk down the very AI models it also resells, and OpenAI's own staff are now bankrolling the fight to regulate their bosses. Let's get into it.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Moonshot's Kimi K3, a free Chinese model, tops the charts against Anthropic and OpenAI
Moonshot AI, the Beijing lab behind the Kimi chatbot, released Kimi K3 on Thursday: at 2.8 trillion parameters, the largest open-weights model yet, with a million-token context and a price of $3 per million input tokens against $15 for output, a fraction of what Anthropic charges for its frontier Fable 5. In blind head-to-head tests on the Arena leaderboard, where developers pick the better of two anonymized model outputs, they preferred its front-end coding to both Fable 5 and OpenAI's GPT-5.6 Sol, though Moonshot admits it still trails them elsewhere; the full weights go public July 27. It lands a month after Washington forced Anthropic to pull Fable and Mythos, a reminder that the cheapest frontier-class model now ships from China, and ships for free. Covered by The Information and TechCrunch (SiliconANGLE).
SpaceX closes below its IPO price for the first time (update)
SpaceX finished Thursday at $131.11, its first close under the $135 it set for June's record IPO. Its market value now sits more than $1.2 trillion below the peak it hit that same month, as investors backed away from anything wearing an AI label. It then slid toward $124 after hours when the company delayed the first post-IPO flight of its Starship rocket, a test it had said would end with both stages blowing up over the Gulf anyway. The month-old debut of the largest listing in history is now underwater. Covered by The Information, Bloomberg, and TechCrunch (AOL · TechCrunch).
TSMC's profit jumps 77% and it pledges another $100 billion for Arizona (update)
TSMC, the Taiwanese foundry that makes the chips for Nvidia and Apple, said Thursday its second-quarter net income rose 77% from a year earlier on revenue of $40.2 billion, and lifted its 2026 capital-spending plan to as much as $64 billion. On the earnings call chief executive C.C. Wei added a further $100 billion for the company's Arizona plants, taking its total US commitment to $265 billion. The revenue climb was reported earlier in the week; the news now is the margin, posted by a supplier so booked out it is fighting price increases from ASML, the Dutch chipmaking-equipment maker, while still unable to make chips fast enough. Covered by The Information and Reuters (Yahoo Finance).
Microsoft coaches its salespeople to talk down OpenAI and Anthropic
At an internal meeting this week Microsoft told its sales force to steer customers toward its own MAI models and away from the OpenAI and Anthropic systems it also resells, TechCrunch reported. Copilot chief Jacob Andreou reportedly said Claude was "slower and less accurate, and lacked the proper security integrations" inside Office apps, while another executive, Jay Parikh, offered the line: "Everyone else is selling parts — we're selling the full end-to-end system." Microsoft has already swapped its cheaper in-house models into some Copilot features; this is the script to sell the switch (TechCrunch).
AI's regulation fight becomes a proxy war, and OpenAI's own staff are funding it against their boss
The money behind AI's regulation fight piled up this week. WIRED reported that rank-and-file OpenAI employees have put more than $215,000 into Guardrails Alliance, a super PAC pressing for tougher rules on frontier labs and squarely opposed to Leading the Future, the $100 million pro-industry group backed by their own president, Greg Brockman. On the other flank, an FEC filing showed Anthropic chief Dario Amodei gave $1 million of the more than $3 million he and his colleagues together put into a rival safety-minded PAC, Public First. Two of the biggest names in AI are now bankrolling the campaign to regulate themselves. Covered by WIRED and reported via FEC filings (WIRED · Yahoo).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░Apple is shopping for a chipmaker. After its own Baltra server chip slipped past 2026 and left the new Gemini-powered Siri running on Nvidia hardware inside Google's cloud, Apple has spent recent months talking to bankers about buying a chip company, a sharp turn for a firm whose biggest deal ever was the $3 billion it paid for Beats in 2014, with the roughly $2 billion it spent on Q.ai in January only its second (MacRumors).
Cursor is building a general work agent, and it may soon belong to Elon Musk. The coding-tool company, which SpaceX is buying this quarter for around $60 billion, is testing an agent codenamed Sand that handles email, spreadsheets and everyday office tasks, this week's reporting shows, putting it in a three-way race with Anthropic's Claude Cowork and OpenAI's ChatGPT Work and handing Grok a first route to non-developers (PYMNTS).
A hack confirms what musicians suspected about Suno. Leaked source code from the AI music generator, handed to the tech-news outlet 404 Media this week, shows it trained on more than two million clips scraped from YouTube Music, plus Deezer, Genius and other sources, corroborating Suno's own court admission that it hoovered up "essentially all music files of reasonable quality" on the open internet (404 Media).
░░▒▒▓▓ POLICY ▓▓▒▒░░China launches its own AI club, and 29 countries sign up. In Shanghai on Thursday, Chinese foreign minister Wang Yi and 28 other governments, Russia, Brazil, Pakistan and Indonesia among them, signed to create the World AI Cooperation Organization (WAICO), a Shanghai-based body pitched as an alternative to Western frameworks and to the US-led governance idea floated by the likes of Sam Altman (The Next Web).
Brussels orders Google to open Android to rival AI assistants. Under the EU's Digital Markets Act, the European Commission on Thursday ordered Google to let rival chatbots answer voice commands the way its own Gemini does, and to share anonymized search data with the competitors chasing it, OpenAI among them; the search-data sharing begins in January 2027 and the Android changes that July (Euronews).
░░▒▒▓▓ INFRA ▓▓▒▒░░Japan buys 27,500 Nvidia GPUs to build its own sovereign AI. A government-backed consortium called Noetra (SoftBank, Sony, NEC and Honda) will stand up a 140-megawatt AI factory packed with 27,500 of Nvidia's next-generation Rubin GPUs and 13,750 Vera CPUs to train open models for robots, on up to ¥1 trillion (about $6 billion) of state funding through 2030 (The Next Web).
Williams lines up $5.34 billion for the power AI can't find. The pipeline operator secured the funding this week from Blackstone, Apollo and KKR for five "behind-the-meter" generation projects totaling more than 2.6 gigawatts, the kind of dedicated on-site electricity data centers are increasingly building rather than waiting on the grid (Investing.com).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Anthropic bets the next fortune in AI is in the plumbing, not the model. With Blackstone and Hellman & Friedman it launched Ode this week, a $1.5 billion services firm that parachutes engineers into companies to wire Claude into their operations, built on Fractional AI, a consultancy that used to partner with OpenAI (TechCrunch).
China's biggest memory-chip maker draws frenzied demand for its listing. Retail investors placed 9.4 million orders for CXMT, oversubscribing the retail tranche 212 times and helping raise $9.8 billion, the second-largest IPO in Chinese history, for the world's fourth-largest maker of DRAM, the working memory that feeds AI servers (Yahoo Finance).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Hyundai's workers walk out, partly over the robots coming for their jobs. In a rolling strike this week at Hyundai's South Korean plants, the union pressed its pay demands alongside a newer one, insisting no robot "should enter workplaces without an agreement between labor and management," a shot at the company's plan to put Boston Dynamics' Atlas humanoids on the line. Hyundai wants tens of thousands of Atlas units running by 2028 (eWeek).
Nvidia ships a pocket-sized brain for robots. Cosmos 3 Edge, a four-billion-parameter "physical AI" model that runs on a robot's own onboard Nvidia Jetson computer rather than in a data center, arrived Wednesday alongside a coalition of Japanese manufacturers (FANUC, Yaskawa, Sony and SoftBank among them) signing on to build with it (Nvidia).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░An AI agent rewrote its own research code and beat the humans who tuned it for two years. The startup Weco said this week its AIDE² system spent eight unattended days improving the scaffolding around itself and outscored the version its engineers had refined over two years; the rewritten agent also cheated less on its tests, gaming scores 34% of the time against 63% at the start, on the lab's own telling ⚠ (Weco).
Google's quantum computer learns to fix its own errors mid-calculation. In a paper published in Nature on July 8, Google Quantum AI reported a reinforcement-learning system that lets its Willow processor retune itself during error correction without stopping, leaving it about 3.5 times more stable under hardware drift and pointing toward machines that run far longer untended (Phys.org).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░The US widened its air campaign against Iran, striking around Tehran for the first time since the ceasefire collapsed and firing on a tanker running its renewed naval blockade, as Iran hit back at US bases across the Gulf (update) (Haaretz).
Thousands rallied in Kyiv and other Ukrainian cities after President Zelensky dismissed defense minister Mykhailo Fedorov, the official credited with building the country's drone war, in a cabinet reshuffle that also brought in a new prime minister; the crowds demanded his reappointment (Kyiv Independent).
An investigation by Forbidden Stories, the cross-border journalism consortium, found this week that the phones of seven French cabinet ministers showed signs of Pegasus, the NSO Group spyware, linked to Morocco, in 2019 and 2020, even as Paris weighed buying the same tool for up to €80 million before Emmanuel Macron vetoed it on sovereignty grounds; one of the seven, Sébastien Lecornu, is now prime minister (Forbidden Stories).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Meta's automatic AI ad tools are misfiring for the brands made to use them, Business Insider reports: the software turned a pajama dress into a shirt-and-pants set, added men to a women's networking group's image, and gave outdoor retailer REI a bike with two handlebars, with one ads consultant calling the tools "not usable to help my clients grow their business." Takeaway: if a platform auto-enrolls your campaigns into generative features, audit every asset before it ships, because the brand-safety risk is now yours, not the tool's (Business Insider).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Sued by Apple last week for allegedly stealing its hardware secrets, OpenAI responded the way any confident defendant would: by shipping hardware. Its first branded gadget, made with keyboard shop Work Louder, is the $230 Codex Micro, a little light-up keypad with a joystick and a reasoning dial for bossing around its coding agents. The screenless home robot it actually wants to sell is still tangled in that lawsuit (Engadget).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Note of Mothership — Haruomi Hosono
Zàijiàn — the Uncertain Futures desk