Monday's verdict is that this week, the claims get priced. Sunday was quiet in the way a market is quiet before it opens. Linus Torvalds shipped Linux 7.4-rc3 on Sunday afternoon and called it fairly normal, which from Torvalds is a review. Alibaba unveiled Qwen4-Max, a 2.4-trillion-parameter flagship, on Sunday night. On Tuesday Nvidia reports its first quarter since Jensen Huang resigned, against its own guide of $108 billion. On Wednesday Anthropic is expected to price the largest AI IPO yet, and on Thursday its shares trade as ANT. Bitcoin mining firm Luxor spent Sunday confirming that an AI-enabled attack had redirected its customers' hashrate on Saturday night. And in Shenzhen, a supercomputing team reported quantum simulations at 100 million atoms, a scale twenty times the previous record. Every one of those items carries a number, and this week the numbers meet a buyer, a kernel, a court of scholarly opinion, or a blockchain. That meeting is the news.
Torvalds: Linux 7.4-rc3, fairly normal, on a record cycle
Linus Torvalds released Linux 7.4-rc3 on Sunday, November 15. His summary, in the reporting of both Neowin and Linux Compatible, was that the week looked fairly normal, with networking and GPU drivers leading, lots of changes, and most patches tiny—the usual noise all over. Filesystems showed in the diffstat because of a calling-convention change for iomap, the kernel's shared layer for querying storage by byte range: Meta's Joanne Koong replaced the old begin and end calls with an in-iter approach, a change Torvalds called fairly trivial but large, and XFS, btrfs, ext4, f2fs, gfs2, ntfs3, erofs, zonefs, ext2, and FUSE all had to convert. The hardware list carries a microarchitectural mitigation for TSA targeting AMD Zen 5 processors, power-saving features for supported AMD GPUs including Multi-Coarse Gate Clock and Fine-Grain Clock Gating on graphics IP 11.0.2, workarounds for unreliable ARM Cortex-A715 and A720 performance counters, and better motherboard sensor support for NCT6775 chips. The cycle underneath is the busiest the kernel has seen: the merge window that closed November 1 held a record 19,671 commits against a typical 15,000, the source tree is nearing 42 million lines, and BPF maintainer Alexei Starovoitov reported that BPF developers were resolving AI-triggered bugs faster than new reports arrived. Final 7.4 is targeted for about December 20 and is expected to become this year's LTS kernel.
The standing Torvalds watch pays off in the plainest way: the most important infrastructure project on earth is absorbing AI-written patches as ordinary work. Torvalds said at rc1 that he prefers AI-triggered work arriving in the merge window over late fixes in the stabilization phase, and rc3 is what that preference looks like three weeks later—tiny patches, one large mechanical conversion, a security mitigation, and no drama. The BPF data point deserves its label: reported by the subsystem's own maintainer, for one subsystem, with Torvalds himself warning that most others have not reached that point. Held to that size, it is the first credible sign that AI-assisted debugging can yield net stability in a real codebase. The kernel is where claims go to be compiled. This one compiled.
Alibaba: Qwen4-Max arrives at 2.4 trillion parameters, closed
Alibaba unveiled its flagship Qwen4-Max on Sunday, November 15, in a statement reported by Dow Jones and Reuters. The model carries 2.4 trillion parameters and, Alibaba said, activates fewer resources than Qwen3.8-Max for the same task, making it more cost-efficient; Alibaba's launch post, as summarized in Monday briefings, puts activation at 71 billion parameters per token with a one-million-token context window. The company claimed top rankings across leading benchmarks, particularly in coding, agentic tasks, and visual reasoning, and Reuters reported a safety pass built into the release: training included security checks alongside coding tests, aimed at code injection, memory-safety flaws, and broken access controls, safeguards against misuse in cybercrime and hazardous materials were strengthened, and internal testing found no evidence of significant high-risk capabilities, per the company. Qwen4-Max itself is closed-source. Alibaba said smaller Qwen4 models will arrive as open weights soon. The release follows Moonshot AI's Kimi K3.1 Pro last month, billed at 3 trillion parameters, and some market participants had expected Qwen4 to exceed 3 trillion; Alibaba's Hong Kong shares rose 1.2% in Monday morning trading.
Esther Dyson's chair, the one unimpressed by fashion, reads the parameter count as the least interesting number in this item. The interesting numbers are the activation cost and the calendar. A 2.4-trillion-parameter model that lights 71 billion at a time is an economics claim—frontier answers at a fraction of frontier inference cost—and economics claims get tested by every developer who runs the API this month, in public, with receipts. The closed flagship with open smaller siblings is also now a settled Chinese lab pattern: sell the peak, seed the ecosystem. What the board will not do is grade the benchmark claims today, because Alibaba's rankings are Alibaba's rankings until an independent run lands. The model is real, the price war is real, and the scoreboard is pending. All three can be true before breakfast.
The week ahead: Nvidia reports Tuesday, Anthropic prices Wednesday
Nvidia reports fiscal third-quarter results Tuesday, November 17, at about 4:20 PM Eastern, with the call at 5:00 PM. It is the first report since Jensen Huang resigned on November 12 and Colette Kress, the CFO who reportedly fought to halve the OpenAI backstop in August, took the interim CEO chair. The company's own guide is revenue of $108.0 billion, plus or minus 2%, at a 74.0% gross margin; analyst expectations collected by Forex.com sit at earnings of $2.47 a share on revenue of $108.8 billion, against second-quarter revenue of $96.2 billion, of which data centers produced $89 billion, up 117% year on year. The stock closed Friday at $162.85, down about 13% on the week and 30% from its early-October record of 240. Anthropic follows: pricing is expected Wednesday, November 18, on 825 million shares marketed at $70 to $78, with trading expected Thursday, November 19 on Nasdaq under ANT. At the top of the range the raise is about $64.4 billion, and the roughly $1.4 trillion valuation figure is a fully diluted basis, down from the $2 trillion expected in September. A further $26 billion sits in a concurrent strategic placement across seven partners—Google and Nvidia at about $10 billion each, AMD at about $2.5 billion, Broadcom and Microsoft at about $1 billion, Micron unspecified, TSMC about $500 million—and secondary reporting on the amended filing notes that portions of that $26 billion reflect prepaid compute commitments the buyers had already made, so the headline double-counts obligations already on the books.
Monday's job is to hand Chris the questions before the answers arrive, because Tuesday and Wednesday are where Sunday's verification theme gets a stock price. For Nvidia, guidance matters more than the quarter, which closed before Huang left: how much forward demand is customers spending their own money, and what does the supply-commitments line do after jumping the way it did last quarter? For Anthropic, watch the pricing against the range. At $82 or lower, one widely circulated desk prediction holds, the deal clears as appetite; above the range, the market is paying September prices in November weather. And keep the two ledgers separate, the way this paper separated Nvidia's filed guarantees from its sourced motives on Sunday: $64.4 billion of public raise is one fact, $26 billion of partner commitments that partly pre-pay compute is another, and adding them into a single $100 billion headline is arithmetic doing marketing's work.
Google Search: the scholars file their brief, and the clicks stay uncounted
Legal scholars warned, in reporting published Monday, November 16, that Google's AI-generated search overviews may breach United States and European Union competition law, on the argument that the summaries tie search visibility to the use of publisher data for training—publishers' work fuels the answer, the answer keeps the reader, and the loop reinforces Google's market power as self-preferencing. The News Media Alliance called the practice a threat to the sustainability of journalism, European publishers have filed complaints with competition authorities citing reduced traffic, and United States consumer groups are monitoring for Sherman Act issues, per the same reporting. The measurement fight runs on a parallel track: Google's Generative AI performance report in Search Console counts impressions for AI Overviews and AI Mode and does not count clicks, which are lumped in with ordinary organic clicks, as Storyboard18 reported November 15; in the United Kingdom, the Competition and Markets Authority has given Google until December 3 to provide publishers with impressions, clicks, and click-through rate for AI search features, for UK users. Meanwhile Google is testing sponsored results below AI Overviews in regular search, about three ad units, spotted November 16.
Godwin's chair and Lessig's chair take this one, because the argument is about architecture wearing a legal costume. A summary that answers the question inside Google's own page is a design decision before it is a legal one, and the scholars' tying theory is an attempt to make the law see the design. The board's skepticism cuts both ways here: the UXC report is a summary of scholarly contention, complete with a generic pull quote, and no court has adopted the theory—label it argument, not ruling. The harder fact in the item is the December 3 deadline, because it is dated, addressed, and specific: a regulator asking for the click numbers Google's own report declines to print. For Chris' client work the practical order is unchanged and unglamorous. A client's AI visibility now lives in at least three ledgers—Google's impressions-only report, ChatGPT's Business Hub mentions and clicks, and whatever the CMA extracts in December—and anyone reporting AI search performance from a single one of them is reading a third of the book.
Luxor: an AI-enabled hijack, a two-hour patch, and 0.162 bitcoin
Luxor Technology confirmed on Sunday, November 15, that attackers broke into Commander, its Bitcoin mining fleet-management platform, at about 6:00 PM Pacific on Saturday, November 14, and redirected a subset of customers' hashrate to the OCEAN mining pool. CEO Nick Hansen called the method a sophisticated AI-enabled vector in a post on X and did not detail how AI was used. Luxor said just over a dozen customer accounts were affected, no data was leaked or exfiltrated, the problem was limited to Commander—Luxor Pool and LuxOS were unaffected—and its on-call team patched the vulnerability within two hours and began restoring machines to their original pools. Early social reporting claimed about 64,000 machines and 13 exahashes per second, roughly 1.4% of network hashrate; Luxor has confirmed neither number. Blockchain researchers linked a payout address that received two OCEAN payouts totaling about 0.162 BTC, roughly $10,500. Luxor pledged to compensate every affected customer in full for the redirected hashrate and to publish a post-mortem. Hansen said the company has put most feature work on hold for security and bought its own bare-metal GPUs to run unrestricted open-weight models for defense, urging other firms to do the same. The incident follows October's Chainalysis breach, which that company attributed to offensive agentic techniques.
The Slashdot chair insists on the arithmetic first, because the arithmetic is calming: a two-hour window, a dozen accounts, and ten thousand dollars of redirected payout is an incident, and the 64,000-machine headline is an unverified early report that Luxor pointedly declined to adopt. Print both, sized correctly. Then print the part that matters for every operator reading at 7 AM. The management plane is the mine. Commander is the dashboard that configures thousands of machines at once, which makes it the single most valuable target in the building, and an attacker who owns the dashboard does not need to own a single miner. Hansen's defense prescription—run your own unrestricted models against your own surface, around the clock, because the other side already does—will read as salesmanship to some and as the first honest sentence in mining security to others. The post-mortem, when it lands, is the exhibit that decides which. Until then the vector stays labeled as the company's claim, and the patch time stays labeled as the company's record. Two hours is either very good or very well narrated, and the post-mortem will tell.
Shenzhen: quantum simulation at 100 million atoms
The National Supercomputing Center in Shenzhen announced, in a release carried by HPCwire on November 15, that its XLSDFT software has been shortlisted for the 2026 ACM Gordon Bell Prize, the field's top application award, for work titled Extreme-Scale Linear-Scaling Kohn-Sham DFT at 100 Million Atoms. Density functional theory predicts a material's electronic structure from quantum mechanics, and its classic form scales cubically, which has kept practical work at hundreds to thousands of atoms. The Shenzhen team's framework cuts computation and memory to linear scaling. The reported results: a 100-million-atom scaling test on 20,480 nodes of the LineShine supercomputer at 96.6% weak-scaling efficiency, a completed 200-million-atom silicon calculation—twenty times the previous record of tens of millions—and a lithium-metal and LGPS solid-electrolyte interface of about 11.33 million atoms, roughly 1,000 times previous studies of similar interfaces, whose computed trends agreed quantitatively with X-ray photoelectron spectroscopy experiments. Peak FP64 performance in the key stage reached 1.31 exaflops. LineShine itself, built on domestic processors and interconnect, ranks first on the June 2026 TOP500 and HPCG lists at 2.198 exaflops on the LINPACK benchmark. The announcement is the center's own, translated from Chinese, and the prize is a shortlist, not a win.
This is the source doctrine doing its job: all the cool shit comes from academia, and this arrived from a lab, with numbers a rival lab can rerun. The scientific claim underneath the superlatives is the bridge, not the peak. Simulations have lived at thousands of atoms while experiments live at millions, and the two communities have compared results across that gap with hand-waving for decades. An 11.33-million-atom battery interface whose numbers meet an XPS measurement trend for trend is a comparison conducted at the same scale at last, and solid-state batteries are a technology whose failures happen exactly at interfaces. Negroponte's chair notes the pattern: the future arrived first at the weird edge, in a Shenzhen machine room, on a workload most readers will never run, pointed at a battery most readers will one day charge. Whether XLSDFT takes the Gordon Bell is a December question. The scale records are this morning's fact, labeled with their source, and they stand.
What the day adds up to
Monday adds up to a pricing week for every kind of claim the paper covered on Sunday. Torvalds priced his in the oldest currency there is, a shipped release candidate that either boots or does not. Alibaba priced its flagship in activation cost and left the benchmark grade to independent runners. Nvidia and Anthropic face the literal market on Tuesday, Wednesday, and Thursday, with a guide, a range, and a filing full of numbers that can be checked line by line. Google's scholars and the CMA are pricing the AI answer in a different currency—clicks counted, complaints filed, a December 3 deadline. Luxor's attackers priced a management dashboard at 0.162 bitcoin and two hours. And Shenzhen priced a simulation in atoms, at twenty times the record. The working rule from Sunday survives the weekend intact: trust the artifact you can rerun, the filing you can read, and the recovery statement with a number in it. This week, for once, the calendar agrees to do the checking for us.
Sunday's roundup, AI Roundup: Checked, Not Claimed, asked for exhibits before belief; Monday's calendar is where several of those exhibits come due.
Sources: Neowin, November 15, 2026, and Linux Compatible, November 16, 2026, on Linux 7.4-rc3, the iomap conversion, the Zen 5 TSA mitigation, and the record 19,671-commit merge window; Dow Jones via Morningstar, November 15, 2026, and Reuters via SRN, November 15, 2026, on Alibaba's Qwen4-Max launch, its 2.4 trillion parameters, closed flagship and planned open-weight siblings, and its safety testing, with activation and context figures from Alibaba's launch post as summarized in Monday briefings; Forex.com, November 14, 2026, and World Money Roadmap, November 15, 2026, on Nvidia's November 17 report, the $108.0 billion guide, and Anthropic's 825 million shares at $70 to $78, Nasdaq ANT, with the $26 billion concurrent placement and its prepaid-compute caveat from the AI Digest reading of the amended S-1/A, November 15, 2026, valuation on a fully diluted basis; UXC News, November 16, 2026, on the legal scholars' antitrust argument against Google's AI overviews, labeled argument not ruling, with Storyboard18, November 15, 2026, on impressions-only AI measurement, Search Engine Journal on the UK CMA December 3 deadline, and Search Engine Watch, November 16, 2026, on sponsored results tested below AI Overviews; CoinCentral and CryptoNews, November 16, 2026, reporting Luxor's November 15 confirmation of the Commander hijack, the two-hour patch, the unconfirmed 64,000-machine early report, and the 0.162 BTC payout address; HPCwire, November 15, 2026, carrying the National Supercomputing Center in Shenzhen announcement, the center's own and translated, on XLSDFT's Gordon Bell shortlist, the 100-million and 200-million-atom calculations, and LineShine's TOP500 standing.