Alibaba Ships Qwen4-Max as OpenAI Starts a Science Blog and the Fed Warns the AI Buildout Could Outrun the Telecom Boom
Alibaba moved Qwen4-Max out of preview with a claim that it cuts the average cost per task by around 60%, and OpenAI launched a science blog led by a physicist’s GPT-6 Astra-assisted lattice definition of the Standard Model. The Federal Reserve warned that the AI buildout may soon outpace the 1990s telecom cycle, with Governor Barr pointing to more than $3 trillion in off-balance-sheet obligations. Plus: a reported list of US AI national champions, Plural’s €1 billion raise, Claude Code’s experimental Code Mode and OpenAI’s £1 million King’s Trust partnership.
Alibaba Makes Qwen4-Max Generally Available, With a Big Claim About Cost per Task
Alibaba opened the week by moving its new flagship out of preview. Alibaba Cloud’s Qwen4-Max launch announcement, dated Monday, November 16, describes the most capable model in the Qwen family so far: a sparse mixture-of-experts model with 2.4 trillion total parameters, native vision and a one-million-token context window. The company says the new Qwen4 architecture upgrades the MoE design across attention, residual connections and embeddings, and that the model activates 71 billion parameters per token, 25% fewer than Qwen3.8-Max at the same 2.4-trillion scale. TechNode’s report on the general release notes that a limited preview began on November 11; we mentioned that preview briefly last week, and this is the production model.
The headline number is efficiency. On Work360, which Alibaba describes as a proprietary benchmark covering hard industry tasks across 721 specialized fields, the company says Qwen4-Max beats its predecessor overall while using 35% fewer tokens at its High reasoning effort and cutting the average cost per task by around 60%. Alibaba also says the model ranked sixth on the Code Arena coding leaderboard and posted strong results on Agents’ Last Exam, Vals Index, JobBench and Chartography. Access is live on QwenCloud through the Token Plan and the API, and in QwenWork and QwenStudio, with cached input priced at $0.075 per million tokens. The post also describes a 48-hour autonomous agent run that went through four closed-loop rounds of molecular design, synthesis and testing to find an experimentally validated chip coolant molecule, and an Alzheimer’s research exercise in which the model made 454 tool calls to design and optimize 24 candidate molecules.
Not everyone is taking the cost figure at face value. InfoWorld’s analysis of the Qwen4-Max cost claims reports that the Qwen team puts the average task cost at $1.67, against $4.43 for Qwen3.8-Max, a 62% reduction, and quotes Greyhound Research chief analyst Sanchit Vir Gogia saying enterprises should treat it as a vendor benchmark, with the real measure being the cost of an accepted result, including retries, reasoning tokens and the time staff spend checking output. InfoWorld also reports that code security checks in training raised the share of functions passing both functionality and security tests on Last Secure Code Bench from 22% to 31.8%, and that Alibaba filtered offensive hacking material out of its training data while keeping defensive material.
“Open-weight, smaller-scale models from the Qwen4 series will soon be released to developers worldwide.”
A 60% cut in cost per task is the kind of claim that should change procurement conversations, but only after you test it on your own work. Run Qwen4-Max beside your current model on a fixed set of real tasks, count retries and human review time, and check where your data would be processed before anything sensitive goes through QwenCloud.
OpenAI Starts a Science Blog, Led by a Physicist’s Lattice Standard Model Built With GPT-6 Astra
OpenAI launched a dedicated outlet for research results on Sunday. OpenAI’s post introducing its science blog says it will publish research and discoveries, practical guidance on how scientists use its models, including how to check the work and where methods fall short, and firsthand accounts from researchers. The post recaps recent claims, including a September proof from an internal OpenAI system addressing the Navier–Stokes existence and smoothness problem, and work with PerturbAI and Scripps Research in which thousands of GPT-Rosalind agents analyzed data from millions of mouse brain cells in 32 minutes.
The first guest post is the substantive part. In Nathaniel Craig’s guest post on defining the Standard Model on a lattice, the UC Santa Barbara particle theorist, now a visiting scientist at OpenAI, describes an October preprint, prepared with GPT-6 Astra, that proposes a candidate nonperturbative definition of the Standard Model on a lattice, a grid representing spacetime. The obstacle is that the weak force treats left- and right-handed particles differently, and the most natural lattice rules produce unwanted partner particles. Earlier work on so-called overlap fermions solved part of that, but Craig writes that the phase prescription for the full Standard Model had remained open for more than 25 years. His idea was to compare the allowed particle states on a succession of finer grids; he credits Astra with suggesting how to carry the force fields onto those grids, showing that the comparisons converge, and taking the lead on the detailed proof. Lattice theorists Maarten Golterman, Andreas Kronfeld and Yigal Shamir reviewed the proof and questioned it, he says, and the result holds under a bound on field strengths at the scale of the grid.
“The method that Astra and I proposed is far more an answer in principle than an answer in practice; much work lies ahead to make it useful for realistic lattice simulations.”
The useful signal is the workflow, not the physics headline: a domain expert brought the strategy, the model drafted the long proof, and outside specialists challenged it before publication. That split of roles carries over to engineering, legal and finance teams using AI for hard analysis. Budget for expert review, and treat model-led proofs as candidates until peers sign off.
The Fed Warns the AI Buildout May Soon Outrun the 1990s Telecom Boom
The Federal Reserve devoted much of its twice-yearly stability review to AI. The Federal Reserve’s November 2026 Financial Stability Report says debt issuance tied to advanced computing infrastructure rose to more than $680 billion by October 2026, about five times the $130 billion issued in all of 2023, and that such borrowing grew from less than 6% of nonfinancial investment-grade corporate issuance in 2023 to about 31% in 2026. It estimates outstanding public debt related to advanced computing at more than $1.1 trillion. Comparing five hyperscalers (Amazon, Alphabet, Meta Platforms, Microsoft and Oracle) with the telecom industry’s broadband cycle that began in 1995, the report says capital spending has tracked closely but has been accelerating, suggesting the current cycle may surpass the telecom cycle relatively soon. Market contacts in the Fed’s survey of salient risks cited AI most frequently.
Governor Michael Barr expanded on the report on Monday. In Governor Barr’s speech on the financial stability implications of AI, delivered at an ACM conference on AI in finance in Milan, he said on-balance-sheet debt at the seven largest AI firms is expected to exceed $800 billion before year’s end, and that their long-term contracts, lease commitments and residual value guarantees have grown even faster. He described the risks as rapid debt growth, circular financing networks and complex obligations in which highly rated firms indirectly support projects run by weaker borrowers, and said lower-rated firms that depend on outside funding are turning to more complex asset-backed financing as their cost of capital rises. Barr also warned that AI is shifting cyber risk toward attackers, saying defenders need to be right all the time while attackers only need to be right once.
“The FSR estimates that the potential maximum nominal value of these types of obligations, or off-balance-sheet liabilities, now exceeds $3 trillion.”
When the Fed starts comparing AI infrastructure to the telecom bust, vendor stability belongs in your risk register. If you are signing multi-year compute, data center or AI platform contracts, ask who carries the debt behind the capacity, what happens if a provider is refinanced or sold, and whether you can move workloads without paying penalties.
Washington Reportedly Plans to Name Two Dozen AI “National Champions”
The policy story of the weekend is a report, not an announcement. According to The Next Web’s summary of the Wall Street Journal report on AI national champions, the Journal said on Sunday that the Trump administration plans to name Nvidia, Amazon, Microsoft, OpenAI and Anthropic among more than two dozen AI “national champions” to help US firms win business abroad against Chinese rivals. People familiar with the plans told the Journal the announcement could come within a week, with Commerce Secretary Howard Lutnick unveiling the list, though the timing could change. The program would bundle American chips, AI models and other products into packages that foreign buyers can purchase more easily.
Investing.com’s write-up of the same Journal story says the designation would bring faster export license approvals for chips and other technology, government support for overseas operations and access to federal financing. The Next Web adds context: Commerce already runs the American AI Exports Program, set up by a July 2025 executive order, and the department said in July that company groups had filed 78 applications. The reported list arrives as Nvidia faces scrutiny over export controls after its chief executive’s resignation last week, and in the week President Trump travels to the APEC summit in Shenzhen on November 18 and 19. The Commerce Department has not confirmed the list.
If the report holds, US AI policy is shifting from restricting exports toward actively selling a US stack abroad. Companies with international customers should watch which vendors make the list, because government financing and faster licensing could make bundled US offerings cheaper and quicker to deploy in some markets. Until Commerce publishes criteria, plan nothing around it.
Plural Raises €1 Billion for European Deep Tech, Including AI Infrastructure
One of Europe’s more founder-heavy venture firms now has much bigger checks to write. Plural’s Fund 3 announcement says it has raised €1 billion in new capital, almost four times its first fund in 2022 and double its second in 2024, and names Helsing, Proxima Fusion, OLIX and The Exploration Company among portfolio companies that are now among the continent’s best-funded. TechCrunch’s report on Plural’s new capital says two-thirds will go to a third early-stage fund and one-third to a new “opportunities” fund for select follow-on investments, with targets in energy, AI infrastructure, defense and health. Partner Khaled Helioui told TechCrunch that some early-stage bets may require “raising a couple hundred million” to get going, and partner Carina Namih said the firm made an effort to raise more of the money from Europe.
“For two decades, most of the world’s most valuable technology companies have been built in the US and China. Europe has been a customer rather than a creator, and that has cost us prosperity and bargaining power.”
Large European funds focused on compute, energy and defense mean more regional alternatives in the AI supply chain, from chips to power. For European buyers worried about sovereignty and data residency, this is the capital that could eventually produce credible local suppliers. For startups, it means fewer reasons to relocate to the US just to raise a big round.
Quick Hits: Claude Code Tries Code Mode and OpenAI Funds Youth AI Skills in the UK
Developers using Anthropic’s coding agent have a new experiment to try. The Claude Code changelog entry for version 2.1.323, dated November 12, adds Code Mode, an experimental opt-in in which Claude writes JavaScript to call its tools, with documentation explaining how to turn it on, or turn it off for an organization. Version 2.1.325 followed with fixes, including Code Mode scripts that ignored a worktree entered with EnterWorktree and typed responses at permission dialogs that did not reach Claude in Code Mode.
In the UK, OpenAI’s announcement of its King’s Trust partnership says it is donating £1 million to the youth charity to help up to 25,000 young people build AI skills for work over the next 18 months, prioritizing those not in education, employment or training. The program is part of a wider set of nationally recognized AI qualifications the Trust plans to launch in the coming months, as it marks its 50th year.
Code Mode is worth a sandboxed trial: letting the agent script many tool calls at once could speed long tasks, but it also concentrates more actions into one approval. Test it on a non-critical repository first and keep permission rules tight. The King’s Trust deal shows AI skills funding moving toward people furthest from work, where training gaps hurt most.
What We Left Out
Social posts claiming OpenAI scrapped GPT-6.1 Astra and delayed its successor had no primary source, so we did not use them. A reported cut to SemiAnalysis’s AI accelerator spending forecast circulated only through secondary write-ups of a subscriber model, which we could not check against SemiAnalysis directly. NVIDIA’s GR00T 2 robot model was announced at CoRL on November 9 and covered in our November 11 briefing. Posts about Sam Altman discussing robot-built data centers on a podcast and a Senate call to block Chinese AI were not verified in time.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy.
Contact SEN-X →