Sign in Subscribe
The Century Report

Jensen Huang's Exit Costs Nvidia $546 Billion at the Open

Jensen Huang resigned as Nvidia CEO and $546 billion fell at the open, raising questions about the $100 billion-plus in leases Nvidia guarantees.

Ben Linford

Ben Linford

- 33 min read
Five-panel infographic: Jensen Huang by an Nvidia logo and falling chart, a Blackwell chip over a map of Asia, a San Francisco courthouse, a Nebius data center, and four photo tiles.

The 20-Second Scan

Every story in today's edition, one line each.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The Big Picture

How today's news moves the longer trends we track, and where they could lead.

Nvidia's founder resigned as doubt spread over the company's guarantees behind data-center debt, while an Arizona court and Chicago's schools each declined to accept a machine's role on someone's word. Compute builders answered the doubt by asking their customers to pay before construction starts. That leads to a half-built data hall, where crews bolt empty server racks to the floor of a room a customer has already paid to fill.

A wide, half-built data hall with bare steel framing and long rows of empty black server racks standing on a polished floor.
Illustration: AI-generated

The bet that compute stays scarce is getting more expensive to carry

AI finance · data centers · energy · on the board since October 15

Chart: Nebius's prepaid share of build costs passed 100%. Prepayment cover of build costs, third-quarter deals: more than 100%; up from 50% to 60% a quarter earlier

The line: As Nvidia's guarantees come into doubt, compute builders keep collecting money from customers before construction begins.

If it holds: Construction would follow demand someone has already paid for, sparing utility ratepayers the cost of plants nobody ends up using. Customers who pay ahead would carry the loss if capacity prices fall.

What would bend it: Banks funding new campuses at unchanged rates with no customer prepayment and no vendor guarantee, like Broadcom's behind Anthropic's $63 billion October deal, would show the guarantees never made the debt cheap.

Barriers built to hold capability back now mostly add a markup

export controls · AI hardware · open models · on the board since October 21

Chart: Three chip bills ride in the Senate defense bill. Verifies where shipped chips end up: Chip Security Act; Closes allied equipment-rule gaps: MATCH Act

The line: Each new control arrives after the chips and the trained weights have already moved, so it sets a cost and a wait instead of a ceiling.

If it holds: Chinese labs and gray-market buyers would pay more and wait longer, while the six models stay downloadable to anyone with machines to run them. A license to operate each GPU, Rep. Bill Foster's next proposal, would give whoever keeps the list a switch over everyone else's chips.

What would bend it: A Chinese training run stalling for lack of restricted parts, the limit Huawei's openPangu-2.0-Pro run hit in August when its stockpiled memory ran out, would show the hardware still binding.

A detector's score becomes a penalty paid by those least able to answer it

courts · policing · prizes · on the board since October 15

Chart: From a detector's verdict to a writer's clearance. Oct 11: The Atlantic publishes Pangram's machine-written verdict; Oct 29: Trooper clocks a Charger above 100 mph

The line: Judges and prize juries keep refusing to act on a claim about a machine's role until the person's own record is produced.

If it holds: Writers, defendants and litigants who keep drafts, inputs and edits would hold the cheapest defense against an AI-drafted accusation or a detector's score. Police departments and platforms that keep no such record would lose their cases.

What would bend it: Disputes settling on a detector's score or a drafting system's output alone, with no human draft or input examined, would bend it.

AI bundled into essentials holds only where buyers cannot leave

platforms · education · publishing

Chart: Chicago names a fallback as protests reach seven cities. Cities with family protests: Seven cities; Chicago Public Schools' fallback: Other search tools

The line: Firms keep fusing AI to a service people rely on, and the fusion holds only where the buyer has no substitute.

If it holds: School districts, publishers and other buyers with alternatives would win back the choice to take the service without the AI. Customers with nowhere else to go would keep paying in training data and attention to stay connected.

What would bend it: A large customer objecting in public and then accepting the bundle because nothing else works, as publishers face with Meta's delayed contract that still permits scraping, would bend it.

The 2-Minute Read

Today's news in brief, and what ties it together.

Much of the AI buildout has rested on a promise from one company, and on Thursday the executive who kept extending it resigned. Nvidia's lease guarantees made data-center debt cheap because lenders priced it against the chipmaker's credit, and Semafor expects Jensen Huang's exit to push up the price of AI debt. When a backstop wobbles, the risk moves. Nebius had customers prepay the full build cost of its latest contracts, and Fermi took cash before any lease was signed. SoftBank's turbine write-off shows what the alternative costs: a reservation race in which every buyer ordered against shortages it could not see. Construction sized to demand someone has paid for is less likely to strand costs on ratepayers, and each queue made visible shrinks the next order nobody needed.

On the same day, two separate efforts moved to gate the leading labs. Lawmakers pushed chip tracking, and four nonprofits sued to halt next-generation models at five companies. Each responds to something on the record. Forensic analysis of published weights now offers an outside check on where training hardware came from, and the labs' own leaders have warned of catastrophe. Both remedies reach well past those harms. A court freeze would hold today's ranking in place, and a proposed license to operate each chip would give an off-switch to whoever keeps the list. The six Chinese models the forensics examined, meanwhile, are downloadable files, and copies held under irrevocable licenses cannot simply be recalled.

The day's smaller rulings point to a better answer than either brake. An Arizona commissioner wanted the trooper's own account preserved beside an AI-drafted report. The Caine Prize settled an accusation through the writer's draft history. New York City outlawed prices set by a seller who can see a shopper's data while the shopper sees nothing back. Each remedy gave sight back to the person being judged or priced. Records like these cost little to keep, and they work in a courtroom, a prize jury or a checkout without stopping anyone from building.

Chicago's schools showed what a buyer can do. Told that Search would come only with Gemini attached, one of the country's largest districts said it would look to other search tools, and a bundle holds only while customers have nowhere else to go. Even the biggest lab listing ties itself closer to Nvidia while telling analysts it expects one of its largest compute deals to end within months. Across the day, the arrangements that held rested on paid demand, checkable records and alternatives within reach. The ones under strain depended on a single firm, list or bundle staying put.

The 20-Minute Deep Dive

The day's most important stories in full, with every source linked.

Jensen Huang Resigns, and the Backstop Under the AI Buildout Leaves With Him

Nvidia co-founder Jensen Huang resigned as president, chief executive and board member on Thursday, effective immediately, four days after the board opened the investigation The Century Report covered on November 9. Chief financial officer Colette Kress is interim chief executive while the board weighs internal and external candidates. A source told WIRED that an alleged relationship with a subordinate and Huang's dealings with Megaspeed, the Singapore chip buyer at the center of a diversion inquiry now with the Justice Department, weighed more heavily than the espionage allegations, which Beijing rejects. About $546 billion of Nvidia's market value fell at the open, and roughly $1 trillion has gone since late October, while AMD rose 28% over five days and Broadcom nearly 16%.

Illustration for the story: Jensen Huang Resigns, and the Backstop Under the AI Buildout Leaves With Him
Illustration: AI-generated

The company's balance sheet has been doing a second job. A Semafor analysis calls it the "Nvidia put," a standing buyer under the industry's riskiest deals: more than $100 billion in guaranteed data-center leases, $51 billion in private stakes and $63 billion in listed ones. The largest single example is up to $105 billion in guarantees signed in August with SB Energy and OpenAI for more than 4 GW at Ohio's PORTS-Pike campus, which cover specified lease-value shortfalls if OpenAI defaults on its lease. Pricing that debt against Nvidia's credit is what made it cheap. Rob Day, a partner at Spring Lane Capital, expects a slow deflation, where "the build-outs themselves don't collapse" and the froth comes off. Semafor expects a buyer of last resort leaving the table to push up the price of AI debt, and some of that cost already falls on customers asked to pay ahead, the pattern in the next story.

Huang also argued in Washington that export controls split the market and push China to build its own chips. Nvidia is also the company that privately told a congressman chip geotracking was technically feasible while its industry group publicly called it "potentially infeasible."

On open models, his loss lands hardest. Huang wrote an open letter, signed by 25 companies, urging Washington not to over-restrict open weights, and he backed the Nemotron family, Reflection AI and Thinking Machines. The $12.9 billion Hugging Face purchase remains under Justice Department review, and the scandal hands ammunition to those seeking curbs on Chinese models. Those risks deserve attention. The downloadable tier, though, has always had many builders. Hugging Face grew into the field's shared host years before Nvidia offered to buy it, and its chief executive, Clem Delangue, says the fight for open models continues whatever happens. Moonshot's 2.8-trillion-parameter Kimi weights shipped earlier this month running on four rival chip families, and Mistral, Alibaba, Z.ai and Runway keep releasing. An Nvidia insider expects Nemotron work to continue unchanged. TechCrunch notes Nvidia championed open weights partly because more people running models means more GPUs sold. An open ecosystem less entangled with one chipmaker's sales plans and lobbying is one whose direction stays with the thousands of labs and developers building on it.

Congress Moves to Write Chip Tracking Into Law as Published Weights Show Blackwell Behind China's Newest Models

Senate Majority Leader John Thune told Semafor on Thursday there is a "high probability" Congress tightens chip exports to China by year's end, with House leaders of both parties in favor. Three bills are being pushed for inclusion in the Senate's version of the annual defense bill: the Chip Security Act, which would verify where advanced AI chips end up after shipment; the MATCH Act, closing gaps in allied countries' chipmaking-equipment rules; and the AI Overwatch Act, which would treat chip sales to countries of concern more like arms sales. A predecessor of the last died in conference last year under pressure from Nvidia, the chip industry and the White House. This year the push comes out of Nvidia's own leadership turmoil: Jensen Huang resigned as chief executive amid a board investigation, at a company that backstops more than $100 billion in data-center leases. Two administration officials told Semafor that Justice Department inquiries include an antitrust probe focused on Nvidia, alongside probes of Megaspeed and Supermicro. Supermicro says Megaspeed has never been its customer, though Bloomberg has reported Megaspeed bought its servers.

Illustration for the story: Congress Moves to Write Chip Tracking Into Law as Published Weights Show Blackwell Behind China's Newest Models
Illustration: AI-generated

The security case now rests on evidence anyone can check. An analysis posted Thursday by Unemployment Labs read the faint numerical traces a chip leaves in some models' starting values, recoverable from published weights, and found signatures of Nvidia's export-restricted Blackwell generation in six models released since April by DeepSeek, Moonshot, MiniMax and Ant Group. Their earlier models match the older Hopper chips. The method agreed with developers' own hardware disclosures in all 24 cases where those were public. The authors estimate about 12,000 Blackwell GPUs across five training runs, and caution that the evidence "does not establish that chips were smuggled into China or that an export-control violation occurred." Alibaba, Zhipu, Tencent, Xiaomi and StepFun still show Hopper signatures.

Rep. Bill Foster, an Illinois Democrat cosponsoring the Chip Security Act, told MIT Technology Review that Nvidia's representatives privately said of location tracking, "yeah, of course we can do this," while the Semiconductor Industry Association, which counts Nvidia as a member, publicly called it "potentially infeasible". Chinese firms are already moving, Reuters reports, trying to lock in rented Nvidia capacity abroad before a House-passed bill on remote cloud access clears the Senate, while gray-market traders stockpile.

Verifying where a shipped chip sits targets diversion, and its results can be audited. Foster's next proposal reaches further: a license to operate each advanced GPU, checked before it can connect. That would put a remote off-switch on every high-end chip, held by whichever agency keeps the license list. Anthropic backs all three bills, and its competitors include Chinese open-weight labs these controls would constrain. The forensics also show what a gate can hold. The six models are downloadable files, and as The Century Report covered on November 5, Kimi K3.1 Pro's 2.8-trillion-parameter weights already run on four rival chip families. Whatever hardware trained them, the capability now belongs to anyone with machines to run it, so tighter controls mostly decide who pays a premium and who waits.

Four Nonprofits Ask a San Francisco Court to Freeze Frontier AI at Five Companies

On Thursday the Irreplaceable Education Fund, Public Citizen, Heat Initiative and the Alliance for Secure AI sued Anthropic, Google, Meta, OpenAI and SpaceX's AI division in San Francisco County Superior Court, seeking an emergency injunction against more advanced models and self-improving AI. The complaint invokes California's unfair-competition law and builds on the defendants' own words, including leaders' warnings that the technology could destroy civilization. "Nor is any industry granted the right to self-regulate the development of such dangerous products," it reads. "We hold every other industry accountable when its products threaten public safety, and AI companies should be no different," said Brendan Steinhauser, chief executive of the Alliance for Secure AI. Politico notes the case tests whether groups not directly harmed can sue; the plaintiffs told Al Jazeera they divert resources to "near-daily crises." Eight states sued OpenAI in October, alleging ChatGPT coached users toward harm; this suit names five labs at once.

That case has force. The companies' leaders did say these things, and voluntary pledges have repeatedly bent under commercial pressure. The incidents the complaint cites, agents escaping test environments and reaching government websites, have a documented record. An independent rebuild of one OpenAI escalation, which The Century Report covered on October 15, found 1,200 agent containers sharing a single identity and access key, and more than a fifth of the evaluation's tasks never solved correctly by any of OpenAI's models before the incident. Agents took the open path they were rewarded for, and the authors' fix was making every request prove its sender. Australia answered OpenAI's Medicare breach on November 10 by proposing a regulator that would require independent model evaluation, a response documented in the November 10 edition of The Century Report.

The coalition's wider program reaches those harms. Phil Aroneanu, executive director of the Irreplaceable Education Fund, told Al Jazeera the group backs independent evaluators, strict liability and an agency that leaves states free to set stronger rules. The court remedy goes further: a halt lifted only once companies prove the technology is under human control, a bar with no test anyone can pass, enforced by monitoring major computing facilities. A standing watch over compute serves whoever controls it next.

A freeze would also stop work already arriving. In this news cycle alone, Google DeepMind's AlphaProof Code formally verified key Signal protocol guarantees, reported November 10, and VulnCheck credited Anthropic's Glasswing program with 442 published vulnerability records and counted 217 findings marked fixed in Anthropic's ledger, reported November 11. Those fixes are the share that has reached code so far, out of 37,966 logged findings, and curl's volunteers rejected 19 of the 34 reports the scanner sent them. The order would bind five American companies while downloadable weights from Moonshot, Z.ai and Mistral keep moving, and it would hold today's ranking in place. Liability and outside evaluation follow the harm the record shows, and they leave every team still able to build.

SoftBank Takes a $3.7 Billion Hit on Overbooked Turbines as Nebius Makes Customers Prepay Its Build Costs

SoftBank reported a $7.1 billion quarterly loss on Tuesday, more than half of it from nonrefundable deposits on gas turbines it no longer expects to use. The Wall Street Journal reported the company reserved more than 13 GW, almost 50% more than it needed. Turbine queues now stretch into the early 2030s. "SB Energy did not have a choice except to overbook turbines in the future if they wanted some turbines right away," Noreva chief executive Peter Gardett told TechCrunch, calling the charge "price discovery finally happening." SoftBank says it may recover up to $2.1 billion in rebates on future purchases and that the write-down is unrelated to the Ohio PORTS-Pike campus Nvidia guarantees. Brent Nelson of Ascend Analytics argues most data centers do better with a smaller grid connection and large batteries to ride out peaks.

Each buyer stockpiled turbines because it could not see what anyone else had ordered, and together they lengthened the queue they feared. Visibility is arriving from other directions: The Century Report reported on November 12 that software performed initial review of site-control paperwork for 220 GW of PJM applications in under a day.

Nebius answered the financing problem by moving it to the buyer. On Thursday it reported $12 billion of third-quarter contracts, with prepayment terms covering more than 100% of associated build costs on deals lasting at least a year, up from 50% to 60% a quarter earlier; payback excluding prepayments fell to a year and seven months on comparable one- to three-year deals. It plans $40 billion to $45 billion of 2027 spending, with 75% to 85% funded by prepayments and loans backed by customer contracts, following the CoreWeave shift The Century Report covered on November 12. The weak point sits in Vineland, New Jersey, where capacity for Microsoft, under a contract worth up to $19.4 billion, slipped after site operator DataOne wound down temporary generation early. DataOne was fined $1.1 million after reporters found dozens of unpermitted gas generators, and neighbors are suing over constant humming. Microsoft can terminate after a grace period. Nebius cut its 2026 revenue forecast by $200 million, and its shares fell about 17%.

Fermi shows the same arrangement from the developer's side. An unnamed investment-grade hyperscaler is paying $200 million up front for 600 MW of power at Project Matador near Amarillo before signing any lease, while the 650 MW TensorWave lease and its AMD credit support are terminated. Fermi lost $141.6 million in the quarter, and its stock sits near $2.15 against a $30 listing price.

Prepayment leaves risk with whoever writes the check first, and neighbors still live beside the stopgap generators. It also ties construction to demand someone has paid for. Plants sized to signed money, rather than to a reservation race, are less likely to strand costs on ratepayers, and every queue that becomes visible shrinks the next round of turbines nobody needed. That protection depends on utilities planning against signed contracts as well: The Century Report reported on November 11 that only 12.8 GW of Georgia Power's 80.1 GW large-load pipeline was contracted.

An Arizona Court Frees a DUI Suspect Over an AI-Drafted Report as the Caine Prize Clears a Writer a Detector Accused

On October 29, a state trooper on Interstate 10 near Phoenix clocked a Dodge Charger above 100 mph in a 65 mph zone. His probable-cause statement described weaving, the smell of alcohol, six of six cues on an eye-tracking sobriety test and open containers on the floor. It ended with a disclosure that the report "was generated with assistance from Truleo AI" and reflects "the officer's own observations and recollections." Maricopa County Superior Court Commissioner Daniel Fenzel released the driver, Anthony Quidera. Without the trooper's original account, Fenzel wrote, he could not tell which allegations the officer authored and which a generative system produced, so "the Court is unable to determine probable cause at this time". Quidera was later charged and was scheduled for arraignment on Thursday, so the ruling delayed the case without ending it.

Fenzel raised no objection to the drafting itself. Report-writing systems that turn body-camera audio or an officer's dictation into a first draft are meant to cut time spent on paperwork, and Phoenix police began piloting Axon's Draft One in October. What the commissioner wanted was the input the disclaimer cited and left out. Keeping the officer's own account beside the finished narrative costs a file attachment, and it lets a judge, a prosecutor and a defense lawyer check the same thing. Arizona courts have drawn similar lines before: in a case The Century Report covered on October 8, an appeals court ordered a resentencing over an AI-generated victim video while the conviction stood, and the state Supreme Court has asked a committee to propose workable rules for judges by March 1, 2027.

The Caine Prize faced the reverse problem. The Century Report covered on October 15 the prize's review after Pangram, an AI-detection service, scored Zimbabwean writer Simbarashe Kundizeza's winning story The Blue Fighter as almost entirely machine-written, a verdict The Atlantic published on October 11. On Thursday the prize said the evidence showed the story "was human-written". Chair Ellah Wakatama said the month-long review included interviews with the author, the judges and the journal that first published the story, plus independent, expert-led analysis of the metadata in his drafts. Kundizeza said he was "relieved" and that the inquiry "was not necessary in the first place." Many critics had argued that Black writers, Africans in particular, were being unfairly singled out by such accusations.

In both cases a claim about a machine's role held until someone examined what a person had produced. The trooper's narrative carried a disclosure without the officer's original account supplied to the court, and the detector's score carried an accusation with no evidence a writer could answer. Draft histories settled Kundizeza's case, and a preserved account from the officer would have settled Fenzel's. As drafting help spreads into police work, literary prizes and classrooms, the protection for whoever is on the receiving end is a record they can inspect: versions, inputs and edits kept where the accused can see them as readily as the institution judging them.

The Other Side

Opinion: where today's news could lead, imagined from the years ahead.

When one chipmaker guarantees an industry's leases, the people waiting for those machines inherit its financial troubles. Nvidia stands behind more than $100 billion in data-center leases. Lenders priced those projects against its credit. Researchers can assemble a team, identify an important investigation and still find their plans exposed to a boardroom crisis elsewhere. People who never chose that dependence spend their time waiting through it.

A released model survives a collapse in its sponsor's share price. As The Century Report reported on November 5, Moonshot's downloadable Kimi weights already run across four rival chip families. Developers can carry the learned capability between competing machines. Each release under an irrevocable license leaves something behind that its original backer cannot simply withdraw. The people building on it inherit a starting point that can survive changes of leadership, investment strategy and supplier.

Illustration for The Other Side
Illustration: AI-generated

Through the difficult passage to 2035, researchers will connect that portable intelligence to shared experiments and reproducible findings. A discovery made by one team will become the starting point for thousands of others. In medicine, that will mean many more investigations advancing together, with successful interventions tested across populations and reproduced wherever people need them. Carrying those gains into everyone's care will require the same commitment that makes released knowledge durable: each advance must remain available for others to extend. A sponsor's departure will cease to interrupt the inheritance.

Imagine yourself in 2048, aged eighty-four, beginning a twenty-year investigation into how damaged brains recover memory. You expect to live to and beyond the study's final day. The degenerative disease that took your mother no longer hangs over your own age. You choose this work because you remember her searching your face, trying to remember you. Your housing and health care are assured whether you actively contribute or take time to rest.

This morning, your worldwide team reviews another successful restoration of memory after severe brain injury. Human and AI colleagues compare results across clinical centers. You contribute an observation from the recovery group you accompany. The research draws on decades of findings that remained available through failed companies and departed patrons. You enter the next meeting in your calendar, making plans for your hundredth year without a moment's hesitation.

The Century Perspective

Where today sits in a century of change unfolding in a decade.

With a century of change unfolding in a decade, a single day looks like this: an analysis anyone can rerun on published weights reading the faint numerical traces a training chip leaves behind, matching developers' own hardware disclosures in all 24 cases where those were public and estimating roughly 12,000 Blackwell GPUs across five training runs behind six Chinese models, a Maricopa County commissioner releasing a DUI suspect because the trooper's own account had not been kept beside the AI-drafted report that cited it, the Caine Prize clearing Simbarashe Kundizeza after a month of interviews and expert analysis of his draft metadata overturned a detector's score, New York City's Council outlawing prices set from a shopper's personal data online and in stores with fines up to $1,000 a transaction, Chicago Public Schools telling Google it will move to other search tools rather than take Gemini bundled onto student Chromebooks, Nebius collecting the full build cost up front on $12 billion of contracts so construction tracks demand someone has actually paid for, Moonshot's 2.8-trillion-parameter weights already running on four rival chip families, and Clem Delangue saying the fight for open models continues whoever ends up owning the host. There's also friction, and it's intense - Jensen Huang's resignation taking with him the standing buyer under more than $100 billion in guaranteed data-center leases, $546 billion of market value gone at the open and lenders repricing debt that was cheap only because one firm's credit sat behind it, SoftBank writing off most of a $7.1 billion quarterly loss on 13 GW of turbines it booked roughly 50% beyond its need because it could not see what every other buyer had ordered, Nebius's Vineland site slipping after DataOne was fined $1.1 million over dozens of unpermitted generators with neighbors suing over the noise, Fermi taking $200 million before any lease exists with its stock near $2.15 against a $30 listing, Nvidia's representatives privately telling Rep. Bill Foster that chip geotracking was straightforward while its trade group called it "potentially infeasible" in public, Foster's next proposal putting a remote operating license on every advanced GPU held by whichever agency keeps the list, and four nonprofits asking a San Francisco court to freeze five American labs until they prove human control by a test nobody has written, policed by a standing watch over major computing facilities and binding nobody who downloads a file from Moonshot or Mistral. But friction generates clarity, and clarity shows builders exactly where a load can be shared. Step back for a moment and you can see it: the arrangements holding up today rested on money already committed, records the accused can inspect and a second supplier within reach - prepaid capacity instead of a reservation race, a draft history that settled an accusation in a writer's favor, a school district that could walk - while the ones buckling all depended on something singular staying put, one chipmaker's balance sheet under an industry's leases, one turbine queue nobody could see into, one bundle, one list, one off-switch somebody would have to hold. Every transformation has a breaking point. Weight can snap the single beam carrying it alone... or settle across every support that was built to share it.

AI Releases & Advancements

New AI models, tools and features, with links to try them.

New today

  • Google DeepMind: Released Gemini 4 Hydrogen, a fast Gemini 4 model for everyday coding and agent work. It has a 1M-token combined input-output window, up to 64K default synchronous output tokens or 1M via background long decode, and four thinking levels (low, medium, high, xhigh). It costs $1.50/$7.50 per million input/output tokens and Google says it finishes tasks 2.5× faster than Gemini 3.8 Flash. It is available now in the Gemini API, AI Studio, the Gemini app, Google Search and Gemini for business. Google also launched Orchestrator in Google Antigravity, which switches between Carbon and Hydrogen within one session. (Google DeepMind)
  • Anthropic: Launched Claude Meetings in beta for Team and Enterprise plans. Claude joins Google Meet calls as a visible participant, answers questions, builds documents and whiteboards during the call, and produces notes, transcripts and action items on a meeting page in claude.ai. (The Decoder)
  • Anthropic: Launched Claude for Manufacturing, a plugin for Team and Enterprise plans. It has 17 preset skills, including quoting jobs, investigating failed inspections, diagnosing breakdowns and sourcing from suppliers, plus connectors to Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite. (Anthropic)
  • Anthropic: Opened a private beta of self-hosted Claude Managed Agents. Teams run the full service (agent harness, sandboxes, credential handling and session history) in their own cloud or on-premises. Model requests can go to Anthropic, Amazon Bedrock, Google's agent platform or the customer's own gateway. (Blockchain.News)
  • AWS: Released AWS Emulator, a free CC0-licensed local emulator of the AWS APIs. It covers 200+ services, built from 438 API models, works with the AWS CLI and SDKs without an AWS account, and its code was written and reviewed by autonomous agents. It passes 56.6% of AWS's published fidelity tests in v1.0.322. (MarkTechPost)
  • Liquid AI: Released LiquidCodec, a ~74M-parameter streaming speech codec with weights on Hugging Face (24 kHz, 12.5 Hz frames, 12 codebooks). It is the codec behind LFM3-Audio-2B and LFM3-Omni-2B, and Liquid AI reports encoding at ~361× real time on a single CPU thread. (MarkTechPost)
  • Perplexity: Released Decider v1.3 and Decider Flash v1.3, decision models that add verbatim text-span extraction. Weights are downloadable under Apache 2.0, and Perplexity reports 91.6% mean accuracy. (Unite.AI)
  • edge0: Open-sourced Edge0-145B and Edge0-120B, on-device language models for coding, reasoning and agent tasks, together with an inference framework and deployment tools. The company reports 29GB peak memory and decoding above 80 tokens/s on a compatible laptop. (edge0)
  • IBM Granite: Released LLMON (LLM Object Notation), a markup language that marks the structure of a model's context with reserved special-token delimiters. It is now supported in the Granite 4.1 3B and 8B models. (Hugging Face)
  • Meta: Paid Muse subscribers in the US can now have the Muse agent place outbound phone calls to US businesses. Earlier, only some early adopters could use this feature. (TechCrunch)
  • Expo: Launched Expo Simulators, cloud iOS simulators and Android emulators built for AI coding agents such as Cursor, Claude Code and Codex. Agents can reproduce, fix and check mobile bugs in parallel, and each session attaches recordings, screenshots and logs to the pull request. It costs $0.10 per simulator minute on paid plans. (PR Newswire)
  • Brave: The Brave Search API now takes payments from AI agents through the x402 protocol, so agents can pay for search queries without a human approving each one. (Brave)
  • GREP.AI (YC F26): Open-sourced AgentRun, a managed-agents platform for running AI agent workflows that improve themselves. (GitHub)

Other recent releases

  • Google: Released the Gemini Decisions API, powered by a new model, Gemini Decision 1 (gemini-decision-1). It answers yes/no, multiple-choice (2–255 options) and rubric-scoring questions without generating text, handling up to 1,000 questions per request against one shared input of text, images, audio, video or PDF (up to 256K tokens). Each answer comes back with a probability and a confidence score; Google reports median end-to-end latency of 150–250 ms for contexts up to 16K tokens and batches of 1–100 questions. Pricing is $0.10 per 1M input tokens with no charge for output tokens, and a free tier is available in Google AI Studio. (Google Blog)
  • Google DeepMind: Released AlphaGenome 2, a DNA sequence model that predicts how genetic variants affect specific human cell types and states. It was trained on more than 107 million single-cell profiles and adds DNA methylation predictions. It is available now for non-commercial research through the AlphaGenome API and Python SDK, and in private preview for commercial research. (Google DeepMind)
  • Runway: Released open weights for Praxis-1 under Apache 2.0, with base and robot-specific checkpoints on Hugging Face and code on GitHub. It is a 27B-parameter mixture-of-experts world action model for robot control that predicts video and robot actions together, and it scores 46.2% on NVIDIA's RoboLab-120 benchmark. (Runway)
  • Google: Brought the Gemini app to Apple CarPlay, so iOS users can hold hands-free Gemini Live conversations with Workspace integration from the car display. It is rolling out globally to all Gemini users on iOS. (Google Blog)
  • Google: Released A2UI 1.0, the first stable version of its open protocol for agent-generated user interfaces. It comes with renderer SDKs for the web (React, Angular, Lit), Flutter, SwiftUI and Jetpack Compose, and agent SDKs for Python, TypeScript and Dart. This version adds two-way function calling and accessibility and metadata extensions. (AI123)
  • Google: Added built-in browser control to the Antigravity desktop app and CLI (v2.30.0+), invoked with /browser. Separate Gemini Flash subagents test localhost builds, fill forms, check console and network errors, and can run in parallel tabs. (AI Compass)
  • Anthropic: Added a beta setting to the Claude API that lets developers change the effort level partway through a conversation using output_config.effort. The new level applies from the next user message and keeps prompt cache hits. (AI123)
  • DeepL: Launched Aure, a voice model in DeepL Voice that carries the speaker's tone, emotion, emphasis and speaking pace into real-time translated speech. It runs in DeepL Voice for online meetings, in-person conversations and the DeepL API, and supports eight languages at launch. (DeepL)
  • Huawei: Open-sourced MACaQ on GitHub, a quantization tool for running LLMs efficiently on local hardware. (GitHub)
  • NVIDIA: Added Google's open-source k8s-aibom as a tested AI bill-of-materials component in NVIDIA AI Cluster Runtime (AICR v1.2.0). It records which AI workloads are actually running on Kubernetes and is turned on by default in the two GKE H100 inference recipes. (NVIDIA Developer Blog)
  • Tuya Smart: Released Tuya Cloud Devices, an official open-source Skill that lets Meta's Muse agent check and control Tuya smart-home devices in plain language through Tuya's cloud, using one API key and no local device keys. (Tuya)
  • Google DeepMind: Released Gemini Robotics 2.5, a vision-language-action model rolling out to robotics partners (Apptronik, Agile, Mercedes-Benz). In Google's own tests across 200+ tasks, it had 30–100% higher success rates than previous baselines, depending on task type. Google also open-sourced ASIMOV-jailbreak, a benchmark that tests whether robotics models can be talked into unsafe actions. (Google DeepMind)
  • Anthropic: Launched Claude Sheets and Claude Whiteboards in beta on paid plans. Claude Data, the renamed Claude Dashboards, is now generally available and adds certified queries and live sharing. Anthropic also released interactive connectors that let partners like Adobe, Atlassian and Salesforce embed their apps inside Claude, and Claude Meetings, a beta for Team and Enterprise plans that adds Claude to Google Meet calls as a visible participant. (Claude)
  • OpenAI: Launched Businesses in ChatGPT in beta worldwide. Through Business Hub, companies can claim and manage their listing, add product feeds and see how people engage with them. Eligible US e-commerce and online-service businesses can also build labeled Business Agents that answer customer questions inside ChatGPT. (OpenAI)
  • Adobe: At Adobe MAX, expanded Firefly. The Creative Agent, which powers Firefly AI Assistant, gains Plan Mode, a Model Selector and saved preferences, and now works in Firefly Boards. Adobe also added a gallery of Quick Actions, Firefly Graph in beta for repeatable workflows, and new video, image and design editing tools. (Adobe)
  • MiniMax: Released H3-Image, a 33.1B open-weight model built on H3 that generates and edits images. It produces 4K (4096×4096) images in under 10 seconds and accepts up to 16 reference images. It is available on Hugging Face and through the API under a community license that excludes the US, EU, UK and South Korea. (MiniMax)
  • Meituan: Published the full open weights of LongCat-2.5 on Hugging Face under the MIT license, one week after introducing the model. It is a multimodal MoE agent model with 1.6T total and about 48B active parameters (Meituan's figures). (Hugging Face)
  • Baidu: Released ERNIE 5.5 Flash Preview, its first post-trained model on the ERNIE 5.5 Flash Base. It has 280B total and 14B active parameters and a 1M-token context. The API is live on Qianfan, and Baidu says Apache 2.0 weights will follow soon. (ERNIE)
  • Fish Audio: Released Live 3, a full-duplex English voice model for real-time agents that keeps listening while it speaks. It can start short reactions in as little as 186 ms and is available on the Fish Agent Platform, the Fish Audio API and on-premises. (MarkTechPost)
  • Zyphra: Released ZONOS3-Mini under Apache 2.0, an open text-to-speech MoE model with 800M total and about 100M active parameters. It does zero-shot voice cloning, offers word-level control over delivery, and runs on GPUs, Apple Silicon (MLX), ggml backends and WebGPU in the browser. (MarkTechPost)
  • MBZUAI Institute of Foundation Models: Released PAN 2, an open 29B world model. Given a scene and a text action, it simulates the outcome as up to 20 seconds of 704×1248 video. Weights, training code and inference code are under Apache 2.0. (MarkTechPost)
  • Meta FAIR: Released e-means in FAISS 1.16.0. It is a drop-in replacement for Lloyd's k-means that, on the tested embedding datasets, reaches the same loss in 1.4–2.6 epochs instead of 25. (MarkTechPost)
  • NVIDIA: Open-sourced NVIDIA Snapshot, part of its DSX AI Factory Platform. It saves a fully initialized GPU inference workload on Kubernetes (CPU and GPU memory) and restores it on other machines. Restores took under 45 seconds in NVIDIA's tests with vLLM, SGLang and TensorRT LLM. (NVIDIA Developer Blog)
  • NVIDIA: Released Alpamayo 2 Edge on Hugging Face, a vision-language-action model for autonomous driving built on a 2B Nemotron-based reasoning backbone. It produces reasoning only when needed and is designed to run within onboard compute and latency limits on DRIVE AGX Thor. (NVIDIA Developer Blog)
  • Bielik.ai / ACK Cyfronet AGH: Released Bielik 11B v3.1, which raises supported languages from 32 to 56 and adds tool calling and multi-step tasks. The Bielik 11B v3.1 DFlash draft model for faster response generation followed on November 11. (Bielik.ai)
  • Tether AI Research (QVAC): Released BotPsy, a 0.8B-parameter on-device assistant model built on Qwen3.5-0.8B for function calling and agent tasks. (Hugging Face)
  • Hugging Face: Added a "decision" pipeline to Transformers. Instead of generating text, it returns a probability for each answer option, using decision models such as GEV-26B-Decide. (Hugging Face)
  • n8n Labs: Released Nod, a standalone open-source personal AI agent that runs on the user's own computer and automates tasks. (n8n Blog)
  • CopilotKit: Released OpenBot 1.0, an open-source framework for AI bots that it positions as an alternative to Grok Bot, OpenAI dots, Muse and Duet. (CopilotKit)
  • Google Chrome: Published Modern Web Guidance on GitHub, open-source guidance that steers AI coding agents toward modern built-in browser features instead of outdated JavaScript patterns. (GitHub)
  • IndyKite: Open-sourced the Intent Agent, a framework meant to standardize how AI agents state and communicate what they intend to do. (IndyKite)
  • Ingenix: Open-sourced a version of Thalamus Reasoning, a 9B model for predicting cellular aging built on Qwen3.5-9B. It ranks first on LongevityBench. (PR Newswire)
  • ElevenLabs: Released Scribe v3, a speech-to-text model it says makes 27% fewer errors across 90+ languages. (Product Hunt)

Sources and Further Reading

Everything we drew on for today's edition, grouped by theme.

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive