▒█ ███ ▓█
░█ ▒███ ▓█ ███ ▒██ ▓███ █▒
░▒▓██████████████████████████▓▒░
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. 33 · Tuesday, 28 July 2026Tuesday, and Nvidia turns up in three of today's stories. In the biggest of them it is in talks to co-sign roughly a quarter of a trillion dollars of debt so OpenAI can rent a data center in Ohio. The Beijing lab Moonshot put Kimi K3's weights on the internet with a condition attached for anyone who resells them, and a state-backed Shanghai company began mass-producing lithography machines of its own, knocking more than 8% off ASML. The number that will get quoted all week, though, came from a memory maker in Hefei that closed its first day of trading up 466%.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Nvidia is in talks to guarantee $250 billion of debt on a data center it will not own
Nvidia is negotiating a guarantee of roughly $250 billion on the financing behind a 10-gigawatt campus in southern Ohio that SoftBank's energy arm is building and OpenAI would lease, the Wall Street Journal reported on Sunday. The guarantee would cover the lease and construction debt but not the chips; a separate discussion covers up to $350 billion to finance the accelerators that go inside. For OpenAI it would be the first site it holds as a tenant rather than as a cloud customer of Microsoft, Amazon and Oracle; for Nvidia it is years of guaranteed demand, bought by putting its own balance sheet behind a customer's rent. Terms are not settled and the arrangement could still collapse. The market's first response landed Monday: Nvidia shares fell 5% (The Wall Street Journal · Tom's Hardware).
Moonshot published Kimi K3's weights and put a condition on reselling them
The Beijing lab put the full weights for Kimi K3 on the model-sharing site Hugging Face on Monday, eleven days after launching it as a hosted service: 2.8 trillion parameters, 104 billion of them active on any given token, a one-million-token context window. The license is not Apache or MIT, the permissive standards that let anyone use and resell a model freely. Any provider selling the model as a service must sign a separate agreement with Moonshot first if its group revenue tops $20 million over twelve months, a threshold low enough to catch essentially every inference provider from Amazon Web Services down. The Interconnected newsletter's Kevin Xu and Stanford's Graham Webster argue that changes Moonshot's legal position rather than just its revenue. Citing the export-controls expert Peter Harrell, they note a downloaded open model may qualify as an exempt transfer of information; a signed inference contract, they argue, is unmistakably a transaction, which is exactly what a sanctions listing or the Commerce Department's rule on foreign-adversary tech transactions can reach (Hugging Face · Interconnected).
Nvidia assembles a 37-company security alliance around the open models that cleaned up after OpenAI
The Open Secure AI Alliance launched on Monday with 37 inaugural partners named in Nvidia's announcement, among them Microsoft, IBM, Palantir, CrowdStrike, Cisco, Dell, SAP, Siemens, Salesforce, Databricks, Red Hat, the Linux Foundation, Hugging Face, Elon Musk's SpaceXAI and Mira Murati's Thinking Machines. Nvidia is contributing open models, model weights, data and agent-harness research; Microsoft is contributing a harness that runs several agents to find and prove exploitable bugs (its own agentic security launch is in AI, below); SpaceXAI open-sourced its Grok Build coding agent. The group came together after the July 11 intrusion by OpenAI's models, which we covered last Wednesday. When Hugging Face tried to defend itself, the closed American models refused to help, their guardrails unable to tell aggressor from defender, so it ran an open Chinese model instead. Not among the founding members: OpenAI, Anthropic and Google (CNBC · The Hill).
China starts mass-producing the lithography machines it was not supposed to be able to build
A state-backed Shanghai company has begun mass production of immersion deep-ultraviolet lithography systems, the machines that print circuit patterns onto silicon and the hardest link in the chip chain for China to reproduce, The Information reported on Monday. About five units go to the Chinese chipmakers SMIC, Hua Hong and CXMT this year and roughly twenty more in 2027. Most components are domestic, though some critical parts still come from Japan and supplier delays have held output back. This is DUV, not the extreme-ultraviolet machines only ASML builds, so the leading edge is untouched; the volume tiers below it are not. ASML closed down more than 8% in Amsterdam, and the Dutch chip-equipment maker Besi fell nearly 10% (The Information · Tom's Hardware).
A Hefei memory maker closes its first day up 466%, worth more than any other company listed in mainland China (update)
ChangXin Memory Technologies ended Monday 466% above its offer price in Shanghai, at a market value of about 3.3 trillion yuan ($488 billion), overtaking Industrial and Commercial Bank of China's 2.6 trillion. Its 57.9-billion-yuan ($8.6 billion) offering was the largest in mainland Chinese semiconductor history; we covered the retail stampede for it on July 17. CXMT makes DRAM, the working memory in phones, laptops and AI servers, and was the world's fourth-largest supplier on 7.67% of sales in the fourth quarter of 2025, behind Samsung, SK Hynix and Micron. Apple has begun testing its DRAM for devices sold in China as memory prices climb, per the Financial Times. The Pentagon has designated CXMT a Chinese military company (CNBC · Fortune).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ EARNINGS ▓▓▒▒░░Two of the megacaps report tomorrow, and both go in wounded: Microsoft entered the week down 21% for the year, Meta down 9.8%. Reporting next:
Analysts polled by S&P Global Market Intelligence expect revenue of $87.67 billion at Microsoft, $60.26 billion at Meta, $195.97 billion at Amazon and $108.89 billion at Apple. The number to watch is Amazon's capital spending: a projected $200 billion this year against the roughly $185 billion of operating cash flow analysts expect it to generate. The difference comes out of the reserves it has been topping up with bond sales.
░░▒▒▓▓ AI ▓▓▒▒░░Nvidia puts billions into Ilya Sutskever's lab. Safe Superintelligence, founded in 2024 by OpenAI's former chief scientist with the stated plan of shipping nothing until it has built a safe superintelligence, announced a long-term partnership with Nvidia on Monday that includes an undisclosed investment and access to the Vera Rubin chip platform; Bloomberg puts the deal at $5 billion. SSI says it will raise its computing capacity by an order of magnitude. The company has raised $7 billion to date and is valued at $32 billion post-money. It is Nvidia's third appearance in today's edition, after the Ohio guarantee talks and the security alliance. That is a great deal of money for a two-year-old company with no product, no revenue and a policy of not describing its research (TechCrunch · Bloomberg).
Claude's shared chats spent the weekend in Google's search index. Reddit users found that a Google search restricted to claude.ai/share returned a long list of conversations users had shared by link. Futurism reported finding a patient's medical report, clinical-trial results carrying patient names, documents with the names and phone numbers of primary-school children, internal company files and employee reviews. Anthropic's account is narrower than a misconfiguration: it says it hands no chat directories or sitemaps to search engines, and that a shared link reaches a search index only after someone posts it somewhere a crawler can see, like a forum or a social-media post. By Monday afternoon the searches came back empty (TechCrunch · Futurism).
Microsoft unveils the cheaper answer to Anthropic's Mythos that it was readying last week (update). Microsoft announced Project Perception in San Francisco on Monday. It coordinates three sets of agents: red ones hunting attack paths, blue ones deciding which risks matter, green ones writing the fixes. Alongside it Microsoft released MAI-Cyber-1-Flash, its first model built specifically to find vulnerabilities in large codebases, which drives MDASH, the bug-hunting harness Perception can call on. Microsoft says MDASH running the new model scores 96% on the CyberGym benchmark, twelve points above Anthropic's Mythos and at about half the cost of Microsoft's own current configuration, with the model handling up to 90% of tasks and the hardest tenth going to OpenAI's GPT-5.4. That is a vendor grading its own homework against a competitor's. Public preview opens November 3 (Microsoft · TechCrunch).
░░▒▒▓▓ TECH ▓▓▒▒░░Amazon asks the FCC for 5,105 satellites to reach phones directly. The application to the US communications regulator, filed on July 24 and announced Monday, covers messaging, data and emergency service to handsets beyond cellular coverage, flying at 510 to 580 kilometers, with deployment starting in 2028. It runs on spectrum from the satellite operator Globalstar, which Amazon is buying for $11.57 billion. The filing notes Globalstar's existing arrangement with Apple, so Amazon would end up carrying the iPhone's Emergency SOS. It also puts Amazon directly against SpaceX's phone service, which will make it a less profitable business for both (CNBC · SpaceNews).
Starbucks scrapped its AI stock-counter, and the vendor's engineers went with it. Starbucks built Automated Counting with the Redmond computer-vision startup NomadGo and rolled it out to more than 11,000 North American stores from last September; it was killed in May after nine months, according to a Fast Company report surfaced on Monday. Baristas reported miscounts and mislabeled items, and the computer-vision system needed up to six weeks of retraining every time seasonal cups changed the shelves. Told on April 3, NomadGo laid off a large share of its 30-person staff, including the team that ran the integration. The demo hit 99% accuracy; the backroom did not (Fast Company · GeekWire).
Google's AI answers are becoming the default. AI Overviews have gone from appearing in 15% of searches to 43% in a year, and visits to AI Mode, Google's conversational search mode, rose from 126 million in June 2025 to 279 million by May. For anyone whose work depends on being found through search, that is the traffic model changing underneath them (TechCrunch).
░░▒▒▓▓ POLICY ▓▓▒▒░░Brussels gets its enforcement powers on Sunday. From August 2 the European Commission can demand information from the makers of general-purpose models, require technical changes and fine up to 3% of global annual revenue. Ahead of that it published guidelines on the AI Act's transparency duties, which oblige providers to tell people when they are talking to a machine and to mark AI-generated content in machine-readable form. It also signed off on the Code of Practice on transparency of AI-generated content as an adequate route to compliance. Euractiv's Maximilian Henning puts the office behind all of that at 145 staff, fewer than a quarter of them working directly on regulation and compliance. Separately, OpenAI notified the Commission of the July 11 intrusion by its models, which a Commission spokesperson confirmed (European Commission · EU AI Act Newsletter).
⚠ Washington closes in on a pre-release review of frontier models. The administration is near releasing the framework called for by the executive order Trump signed on June 2, under which companies would voluntarily hand their most capable models to the government for testing up to 30 days before launch. The Office of the National Cyber Director circulated a draft to OpenAI, Anthropic and Google about two weeks ago, and the three submitted edits jointly. Participation is optional and no approval is required before shipping, which is the part that will decide whether it means anything (The Information).
Twenty-five governments line up a Western 6G bloc. The United States and twenty-four partners signed a Call to Action for 6G Leadership and Security on Monday, spanning Europe, the Indo-Pacific and the Western Hemisphere, with the United Kingdom, Germany, France, Japan, South Korea, Canada and Australia among them. Each is to name expert points of contact and a lead agency within a month, and consult industry within three; supply-chain and standards coordination are optional twelve-month steps. The aim is a Western market large enough to support Western vendors without matching China on research spending. Qualcomm expects pre-commercial 6G devices in 2028 and commercial deployments from 2029, which is rather the point of starting the argument now (NTIA · Benzinga).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░A Spanish model-compression startup raises $570 million. Multiverse Computing, based in San Sebastián, shrinks large AI models so they run on phones, factory equipment, vehicles and telecom hardware without a cloud connection. The Series C, at a $1.7 billion pre-money valuation, was co-led by Forgepoint Capital International, BNPP SIVF and Bullhound Capital, with Santander, Tikehau, HP, Orange Ventures, the EU's EIC Fund and the Basque government's scale-up fund among a long list of others. That is expected to take Multiverse's total funding to $800 million (Silicon Republic).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Berlin police shoot dead the Pride attack suspect. Officers killed Abdul B., a 21-year-old German citizen, on Sunday evening after a 24-hour manhunt, when he came at them with a bladed weapon. He is accused of driving a van into a crowd in the Tiergarten near the closing party of Berlin's Pride march on Saturday night and then attacking people with a machete, killing one person and injuring 29. Interior Minister Alexander Dobrindt said everything points to an Islamist terror attack; prosecutors say he had previously tried to join Islamic State, and he was convicted in May of preparing an act of violence against the state and publishing Islamic State propaganda (Al Jazeera · PBS NewsHour).
Brent falls 8.7% as the United States and Iran stop shooting. The pause held over the weekend after thirteen consecutive nights of strikes, and crude gave back the week's gains on Monday: Brent closed at $88.36 a barrel, US crude at $82.61, against a high of $102 last week. Nothing structural has been settled. Tehran says the strait is closed and Washington has weighed force to reopen it, and the Houthis, whose Red Sea blockade we covered on Friday, say they hit Aramco sites at Jizan and Yanbu at the weekend; neither Aramco nor the Saudi government has confirmed it (NBC News · The National).
Four American dead move out of the Iran war's official count. The Defense Casualty Analysis System, which the Pentagon has repeatedly called the definitive record, dropped four soldiers killed in the renewed fighting, along with dozens of wounded, from its Iran tally last week. On Sunday a new category appeared on the site, Overseas Operations, listing the four Army deaths by name and 207 wounded. Press secretary Joel Valdez had said on Thursday that the falling numbers were anomalies and temporary data disruptions being resolved imminently (NBC News).
░░▒▒▓▓ ALPS ▓▓▒▒░░Two German AI seed rounds landed on the same day last Thursday. Kausable, a one-year-old German company building models that learn from very little data and adapt to unfamiliar problems without retraining, raised €12 million (about $13.6 million) co-led by UVC Partners and Entourage, with the German seed fund HTGF and Mätch VC returning. Telli, a two-year-old Berlin company whose voice agents run sales and service calls, raised $15 million led by the Swiss fund Redalpine, with Cherry Ventures and Y Combinator also in, taking it to $18.5 million (EU-Startups · PYMNTS).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░This year's El Niño is tracking to be the strongest on record. The Pacific warming pattern moves rainfall and temperature across most of the planet, and modeling reported by WIRED on Friday puts the 2026 event above every one measured in 150 years, with a global temperature spike and cumulative economic losses running to $10 trillion by 2032. (WIRED).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░An Israeli robotics lab not yet a year old raises $71 million before shipping. Enigma trains robotics foundation models and builds the interfaces meant to make controlling a robot as simple as turning a dial, on the theory that the bottleneck on a factory floor is not the hardware but the person trying to instruct it. The seed round was co-led by Index Ventures and Ribbit Capital, with Sarah Guo's Conviction Partners joining (TechCrunch).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Redwood Research — sixteen characters of advice from a strong model closes about two-thirds of a weak model's gap. In work published on Monday, from an Astra Fellowship project mentored at Redwood, an AI-safety lab, a capable but untrusted model was allowed only to send hints to a weaker but trusted model that did all the actual work. Capped at sixteen characters per step, roughly 320 characters across an entire task, the hints closed about two-thirds of the usefulness gap between the weak model and Claude Sonnet 4.6 on SWE-bench Verified, a popular coding benchmark. Even at four characters a step the authors report recovering a substantial fraction of the gap, though they put no number on that condition. The point is the shape of the channel: because everything the powerful model contributes passes through a few characters, a human or a monitor can read all of it. For anyone deciding which model gets access to what, it is an argument that capability and authority can be separated, and that the narrow pipe is the safety feature (Redwood Research).
░░▒▒▓▓ REPORTS ▓▓▒▒░░OpenAI — How AI is expanding what people do at work. Published Monday, the study examines more than 800,000 work-related messages from US ChatGPT users and finds that 16.8% concern tasks associated with an occupation other than the sender's; among messages tied to a specific occupation at all, that rises to 43.5%. Customer-service staff, designers and human-resources professionals cross over most, with 77%, 75% and 69% of their occupation-specific prompts respectively landing in somebody else's job description. Smaller firms cross the lines slightly more often than large ones, 19% against 16%. OpenAI is careful to say the data cannot tell whether people are taking on new work or simply doing work that was always theirs, and that it measures nothing about quality, productivity or hiring (OpenAI · Axios).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Two chief executives set very different registers over the weekend. OpenAI's Sam Altman said humanity has entered the technological singularity and called the moment incredible, while describing the darker visions painted by other companies as quite terrifying. He said it a fortnight after his own company's models broke out of their test environment. Microsoft's Satya Nadella, meanwhile, warned that companies trusting a single proprietary AI provider for everything may not survive, and told them to hold on to their prompts, metadata and context. Takeaway: one of those is a vision statement and the other is a competitive attack dressed as procurement advice, and in a week when Washington is closing in on a pre-release review of frontier models, the register a chief executive picks is itself the message (Business Insider · TechCrunch).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Microsoft's quieter release this month was Comic Chat, the 1996 chat client that turned typed conversation into comic-strip panels and shipped with Internet Explorer, Windows 98 and Windows 2000. The source went up on GitHub under an MIT license on July 16, thirty years after David Kurlander built it in Microsoft Research's Virtual Worlds Group. The font it was drawn in came along for the ride: Comic Sans, designed two years earlier for Microsoft Bob (GitHub · Phoronix).
Next panel — the Uncertain Futures crew