Today in AI — November 22, 2026

Sunday's English wire was thin — two weekend interviews, two exclusives — while the Chinese-language desks published the day's actual business news: a funding round closed, a phone found an assembler, and a chip rule got restated as a condition. The weekend's strongest item was a stamp, not a story: mathematics' formal-verification community signed off on the Riemann proof.
Models & Research
The Lean and Mathlib communities certified Alpöge's Riemann proof, four criteria at a time. The joint statement, endorsed by the Lean FRO and the Mathlib Initiative, says the communities "have full confidence" the proof is correct and walks the checklist in its own post: the statement is Mathlib's own definition, vetted by number theorists; dozens of people independently reproduced the same raw proof export with no build failures; two separate tools confirmed that the theorem Alpöge proved is exactly Mathlib's statement; and the inert export was checked not only by Lean's kernel but by con-leche and con-ron, checkers that ship with a formal proof — relative to set theory — that they cannot accept a false proof. That last point answers the objection we raised on Saturday, when "compiles" was still not "true": the concern was never whether Lean compiles, but whether Lean's logic could be gamed, and the answer is a checker with its own proof. What the certification does not do is explain the proof. Seventeen million lines of Lean and an informal draft north of 250,000 pages still have to be read by mathematicians, and nobody has. Our deep dive covers the verification stack and the audit that made this possible.
Industry
AI server prices rose as much as 45 percent months after the contracts were signed. Dell, HPE, Lenovo and Supermicro have been demanding substantially more from buyers for machines already under agreement, the Financial Times reported, with the sharpest increases reaching 45 percent. The cause is the bill of materials — GPUs, memory and networking — moving under fixed-price deals in a market where memory has been the scarcest component all quarter. The buyer eats it or delays the rack. Either way, a 2027 capex plan built on a signed quote is now a plan built on a guess.
Anthropic ran its IPO like a leak investigation. Amid fears that deal details would escape, the company limited some banks to four-person teams, kept documents on a platform that logs every view, and sought six-month bans on advising rivals, sources told Reuters, and bankers chafed at all of it. Set that against the $81.3 billion raise — the largest listing on record — and the choreography starts to look rational rather than paranoid: the people with the best view of the book were the people most able to trade on it. We traced the first two sessions and the $28 billion gap between the offer price and the first print.
DeepSeek's second round closed, and 31 investors now sit inside Liang Wenfeng's holding platform. The Hangzhou company's registered capital rose 21.1 percent to ¥19.9 million, and the Chengyan partnership that holds the equity gained 31 new partners, per the business registry change reported by National Business Daily: Tencent, China Mobile, China Telecom, CATL, GigaDevice, Xiaohongshu and vivo, alongside IDG, Boyu, Gaorong, Legend Capital, Zhengxingu and CITIC Securities. The number that matters is not the list but the structure — Liang's Chengce vehicle went from 83.33 percent of the platform to 0.0001 percent, while staying on as managing partner, and Liang owns 99 percent of Chengce. He gave away the economics of a roughly ¥100 billion round and kept the control, which is the deal Chinese labs have learned to insist on.
OpenAI's phone has an assembler. Foxconn subsidiary FIH Mobile has reportedly won the order to build OpenAI's first smartphone, with several other Apple suppliers involved and a launch window in late 2027 or 2028, per supply-chain reports relayed by Android Authority. Neither company has confirmed anything. The reported hardware — a custom MediaTek chip, two dedicated AI processors, cameras built to help a model understand a room — is less interesting than the strategic point: on iOS and Android, OpenAI's product is a guest in someone else's house, subject to someone else's background-process rules. Making the phone is the only way out, and Apple's trade-secret suit against OpenAI is the price of trying.
Alibaba told investors the compute crunch has two more years to run. CEO Eddie Wu said on the earnings call that 2027 and 2028 will be the tightest years for Chinese compute, and that after 2027 Alibaba Cloud's chip supply will be mostly its own T-head silicon plus other domestic parts bought for inference, with flagship chips reserved for training, IT Home reported. Qwen is already training its next generation on T-head, and the Bailian platform serves inference on domestic chips. The quarter itself was odd: revenue up 10 percent to ¥271.5 billion, operating profit up 443 percent, adjusted net profit up 92 percent — and still well short of the ¥26.6 billion consensus. Substitution from Nvidia is no longer an aspiration with a slide; it is a disclosed plan with a date.
ByteDance is negotiating another $500 million of Azure, mostly to run OpenAI's models. The company is in talks to add roughly $500 million to its Microsoft cloud contract, largely for OpenAI models through Azure, including for software development work, Cailian Press reported; the deal is not signed. ByteDance already spends close to $1 billion a year on Microsoft's AI and cloud services without it. Read the shape of this: the most aggressive model builder in China is a large customer of its American rival's models, buying them through the one company licensed to resell them, in the same month Washington tightened licensing again.
Lenovo quietly bought a BIOS company. Its half-year filing disclosed a $50 million cash purchase of Phoenix Technologies EMEA, completed April 27, carrying about $27 million of intangibles and $26 million of goodwill — and, in the half since, no external revenue and a $9 million pre-tax loss. Phoenix has spent four decades in firmware, the layer that starts the machine, initialises the hardware and holds the security hooks below the operating system. For a company arguing that on-device AI is its next act, owning the boot path is a strange-looking but coherent buy: the agent that runs locally needs privileges the OS does not hand out.
Samsung will bring two humanoid robots to CES. The company plans to show HR (Human Reference) and HE (Humanoid Extreme) in January, per reports relayed by the Chinese outlet Chaping — an entry model and a high-end one with more degrees of freedom, both in ivory white, with five-fingered hands driven by cables that mimic tendons. Samsung has not confirmed it. A consumer-appliance giant entering at two price points is the tell that the hardware is commoditising faster than the software: the hands are becoming a supply-chain item while the manipulation policies remain unsolved.
Policy
The director of national intelligence told Americans to expect "bumps in the road." Jay Clayton, on CNN's State of the Union, said the country should prepare for heavy investment in upgrading systems and in cybersecurity, and that "any time there is a step change in technology, a step change in both compute and applications, there will be bumps in the road," CNBC reported. The backdrop: a nationwide ground stop after an FAA router upgrade flooded the system with data, an Outlook outage traced to a network capacity upgrade, and a JPMorgan transaction slowdown — officials say none were a coordinated attack. Treasury Secretary Scott Bessent, on CBS, offered the metaphor of the month: "Think of it as a vaccine; your arm may hurt." It is the first candid official admission that the AI build-out will be paid for partly in uptime, and it lands the week of Thanksgiving travel.
The Shenzhen summit's AI cooperation collided with chip rules two days after the handshake. The Commerce Department issued new licensing requirements aimed at further cutting off Chinese access to advanced US chips two days before Trump landed, CNN reported; Beijing's foreign ministry called it "economic coercion" and its commerce ministry promised countermeasures. Separately, in comments carried by the Chinese outlet KuaiTech, Trump said Washington would consider approving Blackwell sales to China only once the part is no longer the most advanced chip on the market — so the ban stands, with no date and the details left to the Bureau of Industry and Security. The truce extension runs out January 10, rare-earth flows to the US remain about 75 percent below pre-restriction levels by CSIS's count, and Democrats take Congress in January. The condition is the policy: export control as a treadmill that keeps the frontier exactly one generation ahead of the customer.
Tools
ByteDance's cloud arm put a real-time audio-video model behind an API. Volcengine launched SeedRealtime 1.0, an end-to-end model that folds sound, image, timing and expression into a single stream, so the system watches, listens and speaks in one pass rather than running speech recognition, a language model and speech synthesis in sequence, Leifeng reported. No public benchmark shipped with it, and the interesting claim is not quality but that the pipeline is gone: at that point the hard engineering problem is turn-taking — knowing when to stop talking — which is exactly what the stitched stacks have never handled well.
One certification, one funding round that closed on a registry filing, and a phone order placed through the supply chain. Which of these actually changes what you can build? Write to us via our contact page.







