· ░▒▓█████████▓▒░ ·
▒▓███▓▒ ▒▓███▓▒
░▒▓█████████▓▒░
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. 44 · Wednesday, 12 August 2026Wednesday, and it is a security-and-money sort of day. OpenAI now has a model that will answer hacking questions, behind a vetting program and, from September, a physical security key. CoreWeave's revenue doubled last quarter, and so did the cash it burned getting there. Anthropic is telling prospective investors it could go public as soon as September. Mainland Europe gets its first total solar eclipse since 1999 this evening, and Spain has the best seats.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░OpenAI gives vetted defenders a model tuned to answer the hacking questions its flagship refuses
OpenAI released GPT-5.6-Cyber on Monday and put it behind Daybreak, its vetted-access program for security teams, which now runs in two tiers: Blue, which strips the system-level cyber guardrails off GPT-5.6 Sol, and Red, which unlocks the new model for exploit validation and vulnerability research. On the company's own evaluation, GPT-5.6-Cyber completes 95% of advanced cybersecurity requests against 1.5% for the ordinary Sol model, and OpenAI says it used the model to find two previously unknown flaws in V8, the JavaScript engine inside Chrome, one of them now carrying a CVE number, the identifier given to a publicly cataloged vulnerability. Accenture, IBM, CrowdStrike, Cisco and Palo Alto Networks can build it into their own products, and from September 1 individual accounts will need a physical security key, which is a lot of ceremony for a model whose selling point is that it says yes (Axios · VentureBeat).
CoreWeave's revenue doubled last quarter, and the cash going out doubled with it
The listed AI cloud provider reported second-quarter revenue of $2.575 billion on Tuesday evening, up 112% year on year and slightly ahead of what analysts expected. The net loss widened to $626 million, capital spending ran to $6.4 billion in the three months alone, and the company lifted its full-year capital budget to between $35 billion and $39 billion. Adjusted EBITDA doubled to $1.51 billion, a measure that leaves out the depreciation and interest this particular business is built on; investors read the demand rather than the burn, and the shares rose more than 13% after hours (CoreWeave · CNBC).
Anthropic tells prospective investors it could go public as soon as September
Bankers have been putting prospective investors in front of Anthropic executives ahead of a listing the company could attempt in September or early October, according to reporting published on Tuesday. The questions coming back are about cheaper Chinese open-weight models, the company's relations with the Trump administration, and the local opposition now blocking data centers across the United States, the same week Bloomberg named it as the tenant on a 20-year, 191-megawatt Texas power lease (more in INFRA, below). Anthropic's answer to the first is that most customers want the smartest system available; it was valued at $965 billion in a private round in May, which would make this one of the largest initial public offerings anyone has attempted (Investing.com · The Decoder).
Claude's text will carry a machine-readable mark, everywhere
Claude models released on or after August 2 carry a machine-readable mark embedded in the text itself, and the company is retrofitting older ones; the mark travels with copied and pasted text and, Anthropic says, may survive some editing. It runs across the API, claude.ai, Claude Code, Cowork and Tag, worldwide, not only in Europe, and follows the EU AI Act transparency code that took effect the same day. Anthropic is careful about what a hit proves: that text may have passed through Claude, not that Claude wrote it (Anthropic · TechCrunch).
Google's Gemini app passes a billion monthly users
Sundar Pichai, Google's chief executive, said on Tuesday that the standalone Gemini app has crossed a billion monthly users, which Google calls the fastest growth of any product in its history, and the fourteenth Google product to reach that mark. ChatGPT reached it two months ago, in June. Google also says 63% of Gemini users now talk to it, and that more than 100 million of them use it on Apple hardware (Google · TechCrunch).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░A two-month-old startup founded by an xAI co-founder raised $1.1 billion. River AI, started by Igor Babuschkin after he left Elon Musk's lab, took the round from lead investor General Catalyst and AMP PBC, an AI investment firm started this year by a former Andreessen Horowitz partner, with Nvidia, AMD Ventures, Y Combinator and Singapore's state investor Temasek alongside. It sells an interface for fine-tuning open-weight models with reinforcement learning and says an enterprise can finish a complex training run in 15 to 20 minutes without keeping an infrastructure team. Babuschkin's pitch is that ownership of a model should sit with whoever uses it rather than whoever trained it (TechCrunch).
Brad Lightcap is leaving OpenAI. He joined in 2018, spent four years as chief financial officer and four more as chief operating officer, and moved to special projects in April. He told staff on Tuesday that he is going to start something new and would say more later, which costs OpenAI one of its longest-serving executives (TechCrunch).
░░▒▒▓▓ TECH ▓▓▒▒░░Google is testing a homepage with no Search button. In a version served to signed-out users, first reported on Monday, the familiar Google Search button below the box is replaced by three shortcuts, for creating images, asking about files and brainstorming, with an AI Mode button moved inside the search field. Pressing enter still returns ordinary results, and the shortcuts vanish once you sign in; the apparent target is people who do not have an account yet (Search Engine Watch · Android Authority).
Apple is building a way to prove a photo came from an iPhone. Code in the fifth beta of iOS 27, spotted on Monday, describes Apple Reference Image: a Reference mode in the camera app, after which the owner can send the raw file and its provenance data to Apple's Private Cloud Compute, which checks it against the sensor that took it and assigns the picture an identifier. Apple says it never sees the photograph, only sensor data and metadata. If it ships, an image could carry proof of the lens (9to5Mac).
░░▒▒▓▓ POLICY ▓▓▒▒░░Brussels, Meta and TikTok set up a fact-checking channel for the Ceuta border. The European Commission said on Tuesday that both platforms had agreed to an ad-hoc escalation mechanism with fact-checkers under their existing crisis protocols, after false claims that the border was open helped drive tens of thousands of people to the Spanish enclave last month, in crossings that killed around 100. All of it is voluntary, and the Commission declined to rule out compelling more (Euronews · AFP via Arab News).
France's newspaper alliance has asked the competition regulator to act on Google's AI summaries. APIG, which represents more than 300 newspaper titles, filed with the Autorité de la concurrence on Tuesday over AI Overviews, which reached French search results in late July; it leans on the media regulator Arcom's estimate that the summaries cost publishers between 33% and 38% of their traffic. The complaint turns on promises rather than copyright: Google agreed terms with French publishers in 2022 and was fined €250 million in 2024 for not keeping them (Euronews · France 24).
░░▒▒▓▓ INFRA ▓▓▒▒░░A bitcoin miner has leased 191 megawatts to a frontier AI lab for 20 years. Riot Platforms said on Monday that the contract at its Rockdale campus in Texas is worth about $9.1 billion and runs to June 2048, with 96 megawatts live by December 2027 and the rest by the following June; there are two five-year extensions on top. Bloomberg identified the tenant as Anthropic, citing people familiar with the matter, while Riot named nobody. From December 2027, the power that has been mining bitcoin at Rockdale starts going to something with a larger appetite (DatacenterDynamics).
Nvidia may guarantee up to a quarter of the $500 billion data-center financing we covered yesterday (update). Jensen Huang, Nvidia's chief executive, wrote that the company could backstop as much as 25% of any single project financed through the platforms it agreed on Monday with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. The point of housing that financing in outside vehicles was to keep the risk off Nvidia's balance sheet. A quarter of it, at most, comes back (Fortune).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Super Micro guides to as much as $72 billion of revenue next year, from $39 billion in the one just finished (update). The print yesterday's edition flagged as due landed: fourth-quarter revenue rose 93% to $11.1 billion, non-GAAP gross margin came in at 17.6% against 10.1% the previous quarter, and adjusted earnings of $1.70 a share landed far above the 65 to 79 cents the server maker had guided to. The forecast of $65 billion to $72 billion for fiscal 2027 is the number that matters, and it assumes AI server orders keep arriving at the pace of the past three months (Yahoo Finance).
Intel raised $20 billion, a third more than it set out to (update). The offering priced at $95 a share on Monday and closes today, netting the company about $19.7 billion; the sale we covered yesterday had been announced on Monday morning at $15 billion. Intel spent years unable to raise on anything close to these terms, and says the money is for general corporate purposes, which may include capital spending. The contract-manufacturing customers that spending is meant to win have mostly yet to sign (Intel).
Manus will be an independent company again after Beijing forced Meta to unwind a $2 billion deal. The Chinese-founded, Singapore-based agent startup said on Tuesday it will return to operating on its own; China's National Development and Reform Commission ordered the acquisition reversed in April, four months after Meta closed it on 29 December. Users in the affected jurisdictions have to back up anything they created since that date by 22 August, because the separation deletes it (Quartz).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░A Damascus court sentenced Bashar al-Assad to death in absentia. Tuesday's verdict convicted the former president of premeditated murder, torture and the killing of children under 15, as crimes against humanity and war crimes. Eight other former commanders and officials were sentenced to death alongside him, only one of them present in the room. Assad has been in Moscow since December 2024, and Syria's new government has asked Russia to hand him over (France 24 · The National).
Russia's Supreme Court barred the only anti-war party on the ballot from next month's elections. The court removed Yabloko on Monday on a suit brought by Rodina, a small pro-Kremlin nationalist party, which accused it of foreign funding and extremism. The Central Election Commission had unanimously approved Yabloko's registration on 29 July. The Duma vote runs from 18 to 20 September (Euronews).
Taiwan slowed its own mobile internet during military exercises. For the first time in the civil-defense part of the annual Han Kuang drills, authorities throttled mobile data across central Taiwan on Monday to simulate what a Chinese attack would do to communications; cash machines, traffic lights, emergency calls, landlines and fixed-line internet were left running. Northern Taiwan, Taipei included, gets the same treatment on Thursday (Free Malaysia Today).
░░▒▒▓▓ ALPS ▓▓▒▒░░Deutsche Telekom is selling sovereign cloud as its growth story. Chief executive Timotheus Höttges told analysts on the group's second-quarter call in early August that demand for sovereign cloud, computing kept under local jurisdiction, and for secure digitization and AI infrastructure is rising, and that T-Systems is the group's showcase for it; the unit won contracts in the quarter to build and run Volkswagen's global private cloud and to develop Germany's central citizen application with SAP. The German network business added 161,000 fiber connections in the same three months, 18% more than a year earlier (RCR Wireless).
A Swiss tech-critical blog argues that publishers have started writing for the crawlers. DNIP's Tuesday briefing points to the software developer Vincent Schmalbach's finding that a single article can serve its real text to human readers and AI-optimized advertising copy for a bank to the crawlers, and notes that TIME says bot traffic now exceeds its human traffic on most days. DNIP's reading is that the web is quietly splitting into two versions, one of which no person will ever see (DNIP).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Dyna Robotics pre-trained a robot model on a million hours of human video and no robot footage. Dyna-2, announced on Monday, learns to predict the next video frame and the next action together from roughly 170 years of waking first-person human recordings. At new customer sites, with no site-specific retraining, the company reports an 87% quality pass rate against 46% for its previous model. The claim worth watching is the curve rather than the score: performance kept improving from a thousand hours of human video up to a million, which is a supply of training data that requires owning no robots (DYNA).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Stanford HAI — Regulating Data Brokers in the Age of AI: A California Case Study. Thirty-two brokers on California's 2026 registry sell consumer data straight to generative-AI developers, which is the finding in this month's brief that travels beyond California. The brief, from Stanford's institute for human-centered AI, also tests how data brokers actually comply with the state's Consumer Privacy Act and its Delete Act, and finds that most ignore the disclosure requirements and obstruct deletion requests, leaving consumers with no direct remedy. The authors want registration laws nationwide, with automated rights requests and a private right of action (Stanford HAI).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Spotify will badge AI artists and drop them from recommendations. From mid-September, profiles representing AI-generated identities carry an AI Persona label in the banner, in search and on track rows, and their music is excluded by default from editorial and algorithmic recommendations unless a listener follows them. Artists can declare themselves from Tuesday; Spotify says it will also label photorealistic AI identities on its own, with an appeals route for anyone marked wrongly. Takeaway: the disclosure decision is moving out of the creator's hands and into a platform's classifier, and the penalty for being labeled is not removal but silence (TechCrunch).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Switzerland gets a deep partial eclipse this evening, and the run on eye protection has been thorough. Outside a camera shop in Bern on Tuesday the queue ran about 200 meters, around the block; at a quarter to two roughly 300 people were waiting some 90 minutes in 32-degree heat for certified glasses, kept going by chocolate handed out by the shop's marketing manager. Migros and Coop, the two chains that sell Switzerland almost everything else, did not stock them (SWI swissinfo.ch).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░Mind your eyes — the Uncertain Futures desk