Platform · 12 November
Nvidia’s $12.9 billion Hugging Face deal is agreed but not closed, and nobody will say which way it goes
Huang was the most senior advocate for open weights inside the largest AI company by market value, and he made the case in commercial terms as well as philosophical ones: open models let Nvidia sell chips to states, enterprises and anyone who wants to own their own weights, which diversifies a customer base that frontier labs are steadily trying to design out by building their own silicon. What is now in question is whether his successor inherits the argument. Nvidia agreed in September to buy Hugging Face for $12.9 billion, the second-largest acquisition in its history behind Groq’s assets at $20 billion, and the deal has yet to be finalised. It has also agreed to pay $6 billion for a non-exclusive licence to Poolside’s technology, and Huang personally pushed the Nemotron family of open-weight models built for long-running agentic work. One company insider told TechCrunch they do not expect the leadership change to affect Nemotron development. Nathan Lambert, who tracks the open-model field closely, was blunter: “Jensen no longer being CEO would be very very bad for the open model cause in the US.”
Source: TechCrunchDelangue: “we’ll keep fighting”
Policy · 12 November
The Democratic inquiry is aimed at Commerce, not at Huang, and that is the point
Neal and Wyden wrote to Commerce Secretary Howard Lutnick asking the department to preserve records on how the H200 arrangement was built and licensed, and their account of it is specific: the president agreed to permit the sales on condition the Taiwan-made chips were routed through the United States first, a tariff took a share as they passed, and the department then issued the export licence. They note that a 15 per cent revenue share on China sales had already been agreed with Nvidia and AMD in August 2025, and they ask why Megaspeed was allowed to keep buying chips after US intelligence flagged concerns about its executive Alice Huang and her relationship with Nvidia’s chief executive. The inquiry is therefore about the administration’s licensing conduct, not only about the company’s. It is a request, not a subpoena: the signatories are the ranking members of the minority on Ways and Means and Senate Finance, and the committees they sit on cannot compel a cabinet secretary to answer them.
Source: TechCrunchHuang agrees to testify
Models · 12 November
Gemini 4 Hydrogen is now the default in AI Mode, and it answers with tables, maps and working simulations
Google put the model into Search globally on Thursday, and the pitch is comprehension rather than benchmark position: it parses questions with several parts, hard-to-read typos and highly specific constraints, addresses each part, and runs simultaneous queries across the relevant subtopics so that a single answer carries a wider set of links than before. The output is more visual — comparison tables, graphs, maps and, in one demonstration, an interactive animated simulation explaining how a phone works out which lane of a highway a car is in without a signal. Thinking levels now run low, medium, high and a new “xhigh”. It ships in Search, the Gemini app, all tiers of AI Studio, the Gemini API, Gemini for business and Antigravity for Pro and Ultra subscribers. Robby Stein, who runs product for Search, said it is “now the new default model in AI Mode & being used for your hardest questions in AI Overviews”. Google claims it matches or beats Claude Sonnet 5.5 and GPT-6.1 Terra in its class, which is a claim about parity at the middle of the market, not at the top.
Source: GoogleThe Verge on the launch
Infrastructure · 12 November
Nebius wants to spend $40 billion next year and is already missing the deliveries that pay for it
The neocloud used its third-quarter call to lay out $40 billion of planned 2027 spending, far more than Wall Street expected, to be funded by customer prepayments and loans secured against customer contracts — building the data centres out of their own future revenue. The model depends on delivery dates, and one of them has slipped. Microsoft, one of its largest customers, has a contract worth up to $19.4 billion for capacity at a Vineland, New Jersey site that is now months behind schedule; only just over half of that capacity is due by the end of this year, with the rest moved to the second quarter of 2027. The delay came from the site operator winding down temporary generation faster than planned while switching to Bloom Energy fuel cells, a transition that was not entirely voluntary: the investigative outlet Floodlight found at least 45 unpermitted gas generators running at the site, and New Jersey regulators fined the operator $1.1 million. “Building at this scale is hard, and delays are part of our industry right now,” said chief executive Arkady Volozh. “We are not immune.”
Source: TechCrunch
Economy · 12 November
SoftBank booked nearly half again as many gas turbines as it needed and wrote off $3 billion
The group reported a $7.1 billion quarterly loss on Tuesday, more than half of it from power-generating equipment. Beginning about two years ago, its SB Energy subsidiary placed non-refundable deposits on more than 13 gigawatts of turbines, close to 50 per cent more than it needed, on the theory that whoever controlled the queue would control the price of power for AI. The queue now runs into the early 2030s and the bet has been marked down instead. “This charge is just that price discovery finally happening,” said Peter Gardett of Noreva, who argued SoftBank had no choice but to overbook if it wanted turbines immediately. The more useful signal is what operators are concluding: Brent Nelson of Ascend Analytics told TechCrunch that a smaller grid connection plus large batteries is the better answer for most data centres than islanded gas, because gas carries winter curtailment risk that the grid does not. SoftBank may have found the ceiling for gas as an AI power strategy. It has not found the ceiling for AI’s power demand.
Source: TechCrunch
Product · 12 November
Meta’s Muse will now telephone the cable company on your behalf, if you subscribe
After testing the capability in September, Meta has turned on outbound calls to US businesses for paid subscribers of its consumer agent, which passed 10 million users last month. Alexandr Wang, Meta’s chief AI officer and head of Superintelligence Labs, described the target as the calls people put off rather than the ones they enjoy: the cable company, the DMV, a restaurant that does not take online bookings. The interesting part is the direction of travel rather than the feature. This week United Airlines barred third-party AI agents from its booking systems, citing cybersecurity risk; a voice agent that dials a business is a different fight, over disclosure and consent rather than terms of service, and it arrives with no industry agreement about who is liable when an agent misrepresents the person it is calling for.
Source: TechCrunch
Research · 12 November
Forty-eight agents found a Maya site that a journal had published last year
Pietro Miotti, a Google researcher working on what he describes as a personal side project with no Google compute or resources, set 48 AI agents to comb NASA lidar data for unrecorded archaeological sites. The agents flagged numerous shapes, including one major village that did not appear in existing registers. He asked Mexico’s National Institute of Anthropology and History whether it knew of a registered site there, using the agents to translate the correspondence, and was told it did not. “It took less than 72 hours to go from a random thought while traveling to an official reply from the Mexican government,” he wrote. “This is by far the best time ever to be alive.” Archaeologists then pointed out that the site had been described in an academic journal last year, and that local communities generally know what is in their own landscape. The failure is not the model’s: lidar pattern-matching is a reasonable thing for an agent to do. It is that a plausible shape, a government’s silence and three days of enthusiasm were treated as evidence, when the step that was missing is the one a specialist would have taken first.
Source: The VergeMiotti’s own writeup