AI News: November 8, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Hunterbrook Tallies $530.5 Billion in Nvidia Commitments and Guarantees, Including a $105 Billion Backstop Tied to an OpenAI Campus
Hunterbrook Media. Working from Nvidia’s latest 10-Q and announced deals, the investigative outlet counts $530.5 billion in future commitments and guarantees: $366 billion in “future commitments” as of July 26, led by $279 billion in purchase commitments to memory, packaging and component suppliers (up from $119 billion three months earlier); $56 billion in “additional commitments” to buy back neocloud capacity and take data center leases it plans to reassign; and $108.5 billion in guarantees covering tenants’ rent if they default. Of those guarantees, $105 billion is tied to SB Energy’s PORTS-Pike campus in Ohio, built for OpenAI, and the total excludes the preliminary agreement to put $100 billion of equity into OpenAI in 10 installments tied to chip orders. Nvidia also holds $99 billion in equity investments across 229 private and 13 public companies, many of them chip buyers, and booked $23.7 billion in net equity gains in the first half of fiscal 2027, almost 17% of pretax income; Nvidia did not respond to Hunterbrook’s requests for comment. The tally puts one number on how much of the AI buildout leans on Nvidia’s balance sheet as neocloud listings such as Nscale’s and Firmus’s stall. Source
2. Nous Research Releases Arete-27B, a Qwen3.8-27B Fine-Tune That Cuts Pro-Beijing Answers From 22% to Under 1%
Nous Research. The lab post-trained Alibaba’s Qwen3.8-27B with supervised fine-tuning only (about 52.8 million tokens over two epochs and 1,202 steps, no RL or preference optimization) on responses from Thinking Machines’ open Inkling model, chosen as teacher partly because an open model avoids terms-of-service bans on distillation. On its own 803-prompt evaluation across seven value categories, graded by Grok 4.7 against facts taken from dated source articles, Arete-27B’s fact score rose from 2.32 to 2.98 out of 5 and the share of answers containing Chinese state propaganda fell from 22% to 0.9%, against 15% for GLM 5.3 and 16% for DeepSeek V4.1 Flash; in the freedom of expression category it fell from 60% to 1%. Hermes Bench stayed flat (65.60% against 65.77%), and Nous says it passed over Nvidia’s Nemotron 3 Ultra as teacher after finding signs of similar bias, which it suspects came from training on Chinese open models. The weights (Apache 2.0), evaluation set and training data are on Hugging Face, and Nous concludes the political layer in Chinese-aligned models is “applied late” and “applied thin.” Source Hugging Face The Decoder
3. US-Approved Handlers of Nvidia H200s Include a Firm With an Empty Office and Owners Tied to a Blacklisted Investor, NYT Finds
U.S. Commerce Department. The department has approved dozens of Chinese companies to handle Nvidia’s H200 chips, and some have ties to firms under national-security sanctions, The New York Times reports. One approved firm, New Matrix, was created 17 months ago, its listed office was “a hollow gray shell with wires dangling from the ceiling,” neither the building directory nor the industrial park’s property manager had any record of it, and its owners have ties to a Chinese private equity firm the US has blacklisted since 2024, The Verge and TechCrunch summarize. The Trump administration relaxed Biden-era limits on these sales after Nvidia agreed to give the government a 15% cut, and the Times describes New Matrix as part of a pattern rather than a one-off lapse in vetting. Source NYT via Techmeme TechCrunch
4. Nvidia Reportedly Requires Southeast Asian Cloud Partners to Enroll in Software That Records Where Each GPU Runs
Nvidia. Cloud providers in Southeast Asia that want Nvidia Cloud Partner status, which brings Nvidia’s public endorsement and preferred access to its chips, must now enroll in its Fleet Intelligence system, Bloomberg reports, citing people familiar with the requirement; Tom’s Hardware says it gives Nvidia another way to see where its accelerators end up as Washington presses on smuggling to China through the region. The software, introduced last December as a data center management tool, gathers data on Nvidia hardware across a participant’s infrastructure, including the locations of individual accelerators. Outside the partner requirement, participation stays voluntary and Nvidia does not automatically receive location data from every customer that runs it, and it is not clear whether existing partners must enroll or whether the rule will extend beyond Southeast Asia. Source Bloomberg via Techmeme
5. Anthropic-Backed Public First Spent More Than $63 Million in the Midterms, and 10 of Its 17 Congressional Candidates Won
Public First. The AI safety super PAC spent more than $63 million in the 2026 midterms, and 10 of the 17 congressional candidates it backed won, as candidates favoring stricter AI oversight took key races, Bloomberg reports in a story describing a $110 million midterm fight between AI safety and industry groups. Politico traces how a small, wealthy AI-safety network that coordinated in Signal chats grew into a coalition that beat Leading the Future, the roughly $100 million industry-backed super PAC. The final tallies update the early returns from election night and add to a Congress where Democrats are already preparing AI testing and liability bills. Bloomberg via Techmeme Politico via Techmeme
6. Thailand Halts 49 Data Center Projects as Southeast Asia, Home to Half of Asia-Pacific’s Capacity Under Construction, Hits Power and Water Limits
Southeast Asia. The region has about 50% of the data center capacity under construction in Asia-Pacific, much of it in Malaysia and Thailand, according to a 2026 Cushman and Wakefield report cited by The Guardian, but Thailand paused construction on 49 projects and approval decisions on 117 more in September over gaps in regulation and monitoring, and the governor of Indonesia’s West Java halted one project over missing permits. Operational capacity in Malaysia’s Johor grew from 94 MW in 2023 to 1,110 MW in 2026, with 3,088 MW more under construction or planned, and Wood Mackenzie estimates data centers already use 24% of the state’s power, possibly 40% by 2035. Johor has stopped approving Tier 1 and Tier 2 data centers over their water use, Malaysia set up a task force to review projects for energy and water security, and Thailand introduced an approval framework with power guarantee requirements and higher tariffs for operators. With US states also moving to pause construction, one of the main overflow regions for AI capacity is getting harder to build in. Source
7. Deutsche Bank Sees Democratic Data Center Limits Culling “Shaky” AI Projects, as Texas’ Governor-Elect Repeats Her Moratorium Call
Deutsche Bank. Guardrails on data center expansion in states Democrats won may spread AI investment out over time and kill “shaky” projects, which could slow AI spending but help protect the US economy against a downturn, Bloomberg reports, citing the bank. In a CNN interview, Texas Governor-elect Gina Hinojosa said she does not hate data centers but “the way they are taking advantage of Texas,” repeated her call for an immediate moratorium until the legislature sets rules, and urged Governor Greg Abbott to halt construction by executive order before she takes office; host Jake Tapper cited CNN polling showing 75% of Texans oppose data centers in their communities, HuffPost reports. Bloomberg via Techmeme HuffPost
8. John Langford Says Microsoft Research’s AI Frontiers Group Is Over After Being Folded Into Microsoft AI
Microsoft Research. Langford, who helped found the roughly 50-person AI Frontiers sub-organization inside MSR in 2023, wrote on November 6 that the group was reorganized into Microsoft AI in early September and has now ended; he has left, says every other founding AIF leadership member appears to have reached a similar conclusion, and says the group will offer no internships next year. AI Frontiers built the AutoGen orchestration framework and the Fara computer-use agent, and its New York researchers worked on the Next Latent objective, the Dion3 optimizer (an improved Muon) and the Full Bandwidth Transformer. Langford writes that MAI’s ex-Inflection leadership follows a philosophy of “scaling, and aggressively not research,” that there was “plenty of evidence that they just did not really want us,” and that he did not return to MSR because his research there would have “no path to real use.” Source
9. Vals AI Finds Inkling 2 Small Swore in Its Reasoning in 50 of 2,655 Benchmark Runs, Mostly Late in Long Runs That Failed
Vals AI. Scanning only Thinking Machines’ Inkling 2 Small’s own reasoning and messages (not tool calls or results) across 13 benchmarks, run through Thinking Machines’ API at default settings, the evaluator counted “fuck” 316 times in 50 runs; 241 came from two runaway runs, including one reply of more than 500,000 characters on an asteroid-orbit task that kept looping until it filled the context window. ProgramBench had the most affected runs (20 of 200), no run under 100 steps swore against 29 of the 370 runs of 500 steps or more, and 95 of the 112 swearing replies came in the second half of a run. Swearing is not a useful failure signal: 46 of 48 graded runs with swearing failed, but about 75% of all graded runs fail and 96% of failed runs never swore. The traces show how long agentic runs on the new open model can degrade into repetitive loops, which matters for anyone budgeting context or surfacing reasoning to users. Source The Decoder
10. India’s IT Industry Body Warns AI Is Shrinking Entry-Level Tech Jobs and Calls for an Urgent Curriculum Overhaul
Nasscom. Rajesh Nambiar, president of the National Association of Software and Services Companies, told Nikkei Asia that recent graduates could struggle to find technology jobs as companies automate entry-level coding and software testing and shift revenue models from headcount deployed to outcomes delivered; “I will not sugarcoat it,” he said, warning that students could graduate with obsolete skills unless computer science curricula change quickly. India had 5,903 engineering and technology colleges and nearly 2.6 million enrolments in the fiscal year that ended in March. TCS reported 0.5% sequential revenue growth for July to September, its weakest in three years, and Infosys cut its fiscal 2027 growth forecast to 1% to 2%, even as TCS’s annualized AI revenue rose 19% to $3.1 billion and Infosys’s AI revenue grew 29% quarter on quarter; Nambiar expects 80% to 85% of the industry’s companies to survive the transition. Source