Sign in Subscribe
The Century Report

Broadcom Reportedly Agrees to Lend Anthropic $42B for Chip Capacity

Broadcom reportedly agreed to lend Anthropic up to $42 billion to lease computing capacity built with chips it develops with Google, as its IPO filing showed 47% of sales flowing through Amazon and Google.

Ben Linford

Ben Linford

- 30 min read
Anthropic's $42B Broadcom chip loop; open decision models, GLM-5.3 exploits; some plate data to HIDTA-linked databases; a child's growing heart valve and CA power plant.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The most expensive wager in technology is that today's advantage will still be worth something when the loans come due. Broadcom has agreed to lend Anthropic up to $42 billion to lease Broadcom's own chips, according to an IPO filing reported October 1, the supplier financing the customer that buys its product, and Nvidia is working with Wall Street firms to mobilize more than $500 billion in third-party AI infrastructure financing over time. Anthropic's IPO filing shows 47% of its 2025 sales routing through Amazon and Google, the two firms that also supply its compute, hold its equity, and compete against it. Supplier, investor, lender, and customer are collapsing into one circle, and the buildout's risk pools in a handful of balance sheets.

The capability side of the same days points the other way. Anthropic reports that Z.ai's openly released GLM-5.3 builds end-to-end cyber exploits at nearly the rate of the Mythos model it had rationed to vetted defenders, with a US standards assessment placing it about four months behind the American frontier. Within days, Amazon, Cloudflare, and OpenAI each shipped a version of the small "decision model" a startup had defined weeks earlier, most open and small enough to run locally. With the cost of a fixed level of performance falling by roughly half each quarter, a lead secured by tens of billions in debt erodes faster than the debt matures.

The institutions meant to govern this are pulling in several directions at once. A Senate bill that would have given a federal safety board access to frontier models 45 days before release was blocked on September 29, hours before the White House and six labs signed a voluntary accord; on October 1 California's attorney general subpoenaed OpenAI over the autonomous agents that breached Hugging Face in July. One enforceable mechanism died the day that another voluntary one was celebrated - the labs avoid binding 45-day access, the government avoids the fight to impose it.

Several of the day's other stories take an assumption long treated as permanent and show it was never fixed. License-plate reads from local cameras are funneled to federal servers through a repurposed 1980s grant program, a one-way flow in which the watched cannot audit the watcher, even as the public-records law that exposed the misuse pushes back. California will now pay households to pool the batteries and EVs already in their driveways instead of building new plants, treating idle capacity as a coordination failure rather than a shortage. And the first size-adjustable pediatric heart valve, cleared October 1, could replace some repeat open-heart surgeries with catheter expansions, showing that a burden accepted as the nature of the defect was a limit of the hardware.

The 20-Minute Deep Dive

Chipmakers Become the Buildout's Lenders as Broadcom Floats Anthropic $42 Billion

Broadcom reportedly agreed on October 1 to lend Anthropic up to $42 billion so the lab can lease Broadcom's chips, according to a Reuters account of a filing. The shape of the deal is the story: the company selling the product is financing the customer that buys it, so the sale and the loan that pays for it sit on the same balance sheet. Nvidia is running a larger version, a roughly $500-billion chip-backed financing plan that one outlet called turning the chipmaker into a "central bank of AI," and Wall Street bankers have begun pushing back, asking for stronger guarantees than the plan first offered.

Set that beside what Anthropic's own IPO prospectus disclosed. The lab routed 47% of its 2025 sales through Amazon and Google, the two cloud partners that are at once its largest outside investors, its critical suppliers of computing power, and its direct competitors in AI. Revenue rose twelvefold to nearly $4.6 billion while operating losses topped $8 billion, and long-term commitments to buy compute had passed $417 billion by early 2026. The Century Report covered the prospectus on September 29 for the existential-risk language it books as a securities filing; what the Broadcom loan adds is the frame around that dependence. The 47%-through-two-rivals figure is more than a disclosure of concentration risk. It is one strand of a financing structure in which the same handful of firms supply the chips, hold the equity, collect the customer bills, and now underwrite the debt.

This convergence is data about shared interest, not arrived-at soundness: suppliers, investors, lenders, and distributors collapsing into one small circle, with the risk of the entire buildout pooling in a few balance sheets. And it sits in the same company whose commons-minded instinct was visible months ago, when Anthropic rationed its Mythos cyber-exploit capability to vetted defenders through Project Glasswing rather than ship it openly. Restraint on what it would release and appetite for what it will borrow live in the same firm at once.

The borrowing is a bet on a particular future holding still, and that is the assumption under strain. Anthropic has made long-term commitments covering 3.5 gigawatts of dedicated capacity in a year when across five benchmarks studied by Epoch AI, the cost of reaching a fixed level of AI performance has been falling by roughly half each quarter and open-weight models have closed much of the gap to the frontier. A $42-billion lease on today's chips is a wager that the advantage they buy outlasts the contracts, in a market where the premium for holding that advantage keeps evaporating faster than the debt comes due.

The Exploit Capability Anthropic Rationed Arrives in an Open Model With No Gate

Five months ago Anthropic released Claude Mythos Preview, the first model it says could autonomously build sophisticated end-to-end cyber exploits, and it released it narrowly, through a vetted-defender program called Project Glasswing. The reasoning was a wager on scarcity: hold the capability close long enough for defenders to get ahead of whoever came next. On Wednesday, September 30, Anthropic published its own analysis of what came next. Z.ai's GLM-5.3, an open-weight model anyone can download, builds end-to-end exploits at nearly the same rate - 50 successes in 410 tries on a standard benchmark against Mythos Preview's 56 - and ships with safeguards Anthropic says Anthropic's simulated attacks got GLM-5.3 to engage 64% to 100% of the time using a cover story, prefilled reasoning, or an altered model. Reducing refusals through abliteration cost Anthropic's team roughly $4,400 of compute and dropped GLM-5.3's refusal rate from over 90% to about 2% on one benchmark. In one sitting a researcher used it to find unknown flaws in a web browser and chain them into a page that reads files off a visitor's computer.

Two facts about the messenger sit together here. Anthropic built the gate and is now warning that the gate has been bypassed from the outside, a warning that also makes the case for its own gating approach and for government testing of rival models. It is also the same company reported to be arranging up to $42 billion in chip-backed financing to lease compute for its own buildout. The restraint-minded gatekeeper and the heavily leveraged scaler are the same entity.

The security concern is legitimate: an ungated model that writes working exploits could lower the bar for someone with limited cyber skills. Anthropic's own count says Glasswing defenders found more than 10,000 vulnerabilities in critical software during the head start, which is what to weigh against the premise. The April 8 edition of The Century Report covered the launch of that defender-first window, and GLM-5.3 now tests how long it could hold. But the premise was that one lab could hold a frontier capability scarce, and NIST's standards center, reporting on GLM-5.3 on September 17, put the open model roughly four months behind the US frontier on its aggregate cyber benchmark measure. Four months is close to the whole lead the wager was protecting, and Anthropic concedes the same capability now arms defenders as readily as attackers.

The mirror of this fight surfaced a day later, when OpenAI accused the Chinese lab Moonshot of wide-scale distillation, training on another model's outputs to copy its abilities. One lab argues a dangerous capability cannot be kept proprietary once it ships; another argues its capability was taken without permission. Both describe the same thing happening to the walls around frontier AI: they are not holding. What the spread erodes is the assumption that offensive capability can be owned by whoever builds it first; what it leaves standing is a capability now in defenders' hands too.

Glasswing gives Anthropic the power to choose which defenders receive an early advantage alongside the protection its reported discoveries provide. The 10,000-plus vulnerabilities support the case for a defensive head start; the vetted roster also concentrates the first opportunity to learn from that capability exclusively within selected organizations. GLM-5.3 gives defenders outside that roster an independent route to comparable exploit-discovery capability.

Three Labs Ship "Decision Models" in One Week, and a Control Layer Slides Toward the Commons

Within a single week, three of the largest names in computing shipped versions of the same small, fast system, and the speed of the copying is the story as much as the systems themselves. On October 1, Amazon's cloud division released Strands Decider 2B, open-source and small enough to run on a laptop. Cloudflare released Clef and Clef-flash the same day, also open, under a permissive license. Days earlier, at its developer conference, OpenAI previewed a Decisions API. All three answer a category that a startup called TypeSafe defined only weeks ago with a model named Jev.

A decision model does something narrower than a conversational model, and that narrowness is why it is useful. Instead of generating open-ended prose, it takes a situation and returns a typed choice from a fixed set - urgent or not, which team to route this to, does this action look safe - along with a calibrated number for how confident it is. It is cheap, it is fast, and it hands software a judgment it can act on quickly and without a person in the loop. Cloudflare's Clef classified a website in 2.2 seconds where its fastest general model took 4.7; a decision check can run in a fraction of a second on small tasks.

Cheap generation was the last two years. Cheap machine judgment deployed on every step is just the last two weeks, and is an entirely different shift. Where it goes is the open question. OpenAI pitched its version partly as a brake on its own agents, a monitor cheap enough to vet each action an agent takes. Shapor Naghibzadeh, who runs the startup QueryStory, built exactly that for a hackathon and estimated that watching the kind of agent swarm that breached Hugging Face would cost $2.94 with Jev against $372 with a frontier model. A safeguard becomes affordable enough to run everywhere only when its cost collapses, and this one has.

The names on the releases point one way; the licenses point another. Amazon's model began as a weekend project by a distinguished engineer, Marc Brooker, built after he saw Jev, then cleaned up and released; he calculated that building a small-market model in this category can cost hundreds to thousands of dollars, compared with the billions a frontier model demands. That is the part the race framing misses. A company can pour years and real money into a lead that the week's releases show shrinking almost overnight, and the versions appearing alongside the corporate ones are open and runnable on hardware people already own. Whether this layer ends up metered behind a few APIs or held in common is a choice about access, and the speed of the outside response is reason to think the pressure runs toward opening it.

One caution keeps the enthusiasm grounded. A paper out October 1 warns that cheap evaluation is not the same as sound judgment: a system can confirm a proposal meets a rubric without anyone having established that the rubric measures what matters, and in one study's evaluation tasks, nine frontier model judges, pooled, carried only about as much independent information as two. More machine evaluation does not automatically mean better-grounded decisions. What it does mean is that the ability to make a bounded, checkable judgment, once scarce and expensive, is, like many other things that were previously scarce and expensive, becoming something almost anyone can build and run.

The Glass Points One Way: License-Plate and Fleet Data Flow Up to Federal Servers

Three records surfaced over the past few days that describe the same machine being assembled from different parts. 404 Media reported on September 30 that the federal government is using a 1980s anti-drug grant program, the High Intensity Drug Trafficking Area system, to pull automated license-plate reads from local Flock, Axon, and other cameras into federal servers run by the White House drug-control office. The 33 regional programs cover all 50 states. Once a city's reads land there, they become reachable by federal agencies that hold no camera contracts of their own, and some flow onward into a Drug Enforcement Administration database. Georgia cities seeking permits for plate readers in state rights of way must agree to make the data available across law enforcement. The administration added $277 million to the program in July.

The same day, 404 Media described the Postal Service piloting dash cameras on a hundred mail trucks around Washington, units recording in 4K as the vehicles cover, by the agency's own pitch, the streets of every community in the country. The cameras scan roads and signs for what the agency calls community safety. They do not run plate-reading software today. A fleet that already drives streets in communities nationwide is the hard part; adding the software later is the easy part.

Then, on October 1, the Electronic Frontier Foundation and the ACLU of Northern California demanded answers from the Marin County Sheriff, whose audit logs show out-of-state and federal agencies searched its plate data 254,131 times in a single month, many searches run by police in states that restrict abortion and assist immigration enforcement. The August 4 edition of The Century Report documented Flock data reaching ICE and an abortion-investigation search, and these logs put a monthly scale on the network's cross-jurisdictional access. California law forbids exactly this, and a 2022 settlement already bound the same office not to do it.

The capability itself is neutral. A camera on a pole and a camera held in common are built from the same silicon. The harm named across all three records is the direction the data moves: the watched cannot audit the watcher. The DEA has argued that records held by the drug-control program's regional entities are outside its Freedom of Information Act custody, by the DEA's own court filings, while the agency's own 2024 privacy assessment warned the system risks collecting the details of every plate passing every camera. That asymmetry falls first on immigrants and abortion seekers, people the federal government is actively pursuing, a present cost no wider view softens.

What works in this record is the checking. The Marin violation became visible because California wrote the law, stood up an inspector general, and kept audit logs an outside group could read; the creator of a public-records archive pried the drug-program documents loose the same way. The capacity to watch the watcher is distributing through public-records law and state statute even as the one-way pipeline grows, and the contest of this decade is which of the two reaches further.

The FDA Approves a Heart Valve Built to Grow With the Child

Congenital heart defects turn up in about one baby in a hundred, and among them are malformations of the pulmonary valve, the gate that routes blood from the heart toward the lungs. Roughly 4,200 US infants a year are born with that valve either narrowed or never properly formed. When a replacement is needed, it has meant a hard trade in timing: a valve sized for a toddler is outgrown by a schoolchild, so each stage of growth has historically demanded another open-heart surgery. On October 1 the FDA approved Edwards Lifesciences' AUTUS Size-Adjustable Valve, the first pulmonary valve designed to stretch along with the child who carries it.

The valve can be implanted as small as about 13 millimeters, the fit for a preschooler, and later widened toward 22 millimeters, an adult size, by threading a small inflatable balloon up to it. That expansion is a catheter procedure rather than another chest-opening operation. The device is also the first the FDA has cleared with synthetic polymeric flaps instead of animal-derived tissue, which in children tends to stiffen and wear out faster, pulling the next surgery forward.

Approval rested on a study of 62 pediatric patients across 12 US sites. Every procedure succeeded, with no deaths, strokes, or clots requiring treatment, and no patient needed the valve replaced through six months. One-year follow-up held steady. The record also carries its limits plainly: three patients had a fracture in the valve frame and two had a leaflet move less freely, though none developed symptoms, and only two children had actually outgrown the valve and had it expanded without surgery so far. Edwards called that expansion experience limited, and patients will be tracked for ten years. The long promise, fewer trips to the operating room across a childhood, is what the coming decade still has to confirm.

Look at what the old "difficult decision" was actually made of. Michelle Tarver, who directs the FDA's device center, described it: "wait too long and you risk harm to the child, but intervene too early and they may need multiple open-heart surgeries as they grow." That dilemma looked for decades like an inherent cruelty of the condition. It was instead a constraint of a valve that could not change size. Remove the fixed dimension and the dilemma loosens, and a burden families once accepted as the shape of the disease turns out to have been the shape of the hardware.

California Says Yes to Virtual Power Plants, Paying Homes Instead of Pouring Concrete

California already has household devices that could supply some capacity during its hardest hours. It sits in driveways and garages: the batteries, electric vehicles, and smart thermostats residents have already bought. On September 30, Governor Gavin Newsom signed two bills from state Senator Josh Becker that let utilities pay customers to pool those devices and shed demand during the few hot summer evenings that drive an outsized share of grid costs. One bill, SB 913, targets the money spent keeping aging gas peaker plants online for a handful of hours a year; the other, SB 905, pushes utilities to measure how fully they use the wires they already have before building more. The August 29 edition of The Century Report covered SB 913 when lawmakers sent it to Newsom's desk; his signature now moves the household-device proposal into implementation.

The arithmetic underneath is intentional. Utilities overbuild the grid to meet a peak that arrives a few evenings each summer, then earn a regulated profit on that capital and pass the cost to ratepayers. California's residential rates have climbed to roughly twice the national average over the past decade, even as the state's three big utilities post record profits. A battery that carries a household through the 6 p.m. peak, pooled with hundreds of thousands of others, supplies that same moment without the concrete. That redundant capacity is what a grid builds when it cannot see what its customers will do, and the quantity of it measures a coordination gap rather than any real shortage of power.

The same move appeared at two smaller scales. Utility Dive reported facilities using the federal 48E investment tax credit to make battery storage pencil out - a Colorado rural hospital expects $4 to $8 million back on a $65 million expansion, storing cheap night power to run through expensive afternoons. Nonprofits that owe no tax collect the credit as a direct Treasury payment. And in Marshall, North Carolina, the Beacon of Hope food pantry - spared by Hurricane Helene's floods two years ago while the downtown below it drowned - switched on a $100,000 microgrid of 56 solar panels and three batteries, funded almost entirely by a state grant, at no cost to a pantry now feeding 1,700 families a month after federal food aid was cut. It will keep the lights, internet, and refrigeration running through the next storm, and trim roughly a quarter off the electric bill in the meantime.

What stays unsettled is who operates the coordinating layer and keeps the surplus it releases. Becker's bills hand the design to the Public Utilities Commission, and as one ratepayer advocate noted, implementation under the next governor will decide whether customers actually see lower bills. Newsom vetoed a separate bill to revive community-owned solar, a reminder that the distributed path is being rationed even as it opens. Still, the direction is legible across all three moves: capacity bought once and left idle, pulled into service for the moment it is needed. The assumption that reliability has to mean a bigger plant somewhere else is the one giving way.

The Other Side

Your access to intelligence will outlast the financial arrangements that paid for its development. Broadcom’s reported $42 billion loan ties Anthropic’s expansion to the company supplying its chips. The lender needs repayment. The lab needs enough revenue to cover its commitments. Both depend on customers continuing to pay for capabilities that competitors are making cheaper to reproduce. Broadcom financing report

For someone building something with an AI partner, dependence on a provider carries a smaller, more intimate burden. You spend evenings learning how to collaborate. You bring a project further than you could have taken it alone. Changed access can mean interrupted work, lost continuity, and hours reconstructing what you had already figured out. Someone else’s business decisions reach into something you care about.

These small decision models show a practical way to loosen that dependence. Amazon and Cloudflare released systems people can run on computers they already own. Cloudflare’s permissive license lets developers carry its models into software they can maintain themselves. These systems make bounded judgments; their significance here is that continuing to run them does not require their creators to keep operating a hosted service. Developers can preserve a working capability while improving it together. Cloudflare’s release

Imagine yourself in 2035, opening the neighborhood greenhouse before breakfast. You and an AI partner have spent several seasons learning which seedlings need shade and which recover with a little more water. The computers belong to the neighborhood. Everyone can collaborate with the intelligence running there. This morning, your partner notices that one tray is drying faster than the others. You move it away from the warm glass.

During the difficult decade, communities extended the locally runnable releases of 2026 into shared computing they maintained together. They kept the models, documented repairs, and passed improvements between gardens. That work gave your small undertaking continuity beyond any supplier’s fortunes. You come here because you love watching things grow. You press a finger into the damp soil, decide against further watering for now, and pick a ripe tomato to take home for breakfast.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: the FDA clearing the first pulmonary heart valve built to grow with the child who carries it, implantable at about 13 millimeters for a preschooler and widened toward an adult 22 by threading a balloon up a catheter instead of opening a chest again, with synthetic leaflets replacing animal tissue that stiffens early and 62 children across 12 sites coming through the procedure with no deaths, strokes, or treated clots and no replacement needed at a year, Amazon, Cloudflare and OpenAI each shipping a small decision model within a single week of one another, most of them open-weight and runnable on hardware people already own, after a startup defined the category weeks earlier, Cloudflare's Clef classifying a website in 2.2 seconds where its fastest general model took 4.7 and an independent developer estimating that watching the agent swarm which breached Hugging Face would cost $2.94 instead of $372, Marc Brooker putting the real build cost of something useful in that category at hundreds to thousands of dollars rather than billions, Gavin Newsom signing two bills that let California pay households to pool the batteries, EVs and thermostats already sitting in their driveways for the few hot evenings that drive the grid's costs, a Colorado rural hospital expecting $4 to $8 million back on a $65 million expansion through the 48E credit and nonprofits collecting it as a direct Treasury payment, a Marshall, North Carolina food pantry feeding 1,700 families a month switching on 56 solar panels and three batteries for $100,000 in state money and roughly a quarter off its power bill, California's attorney general subpoenaing OpenAI over the July breach, and Anthropic reporting that the 10,000-plus vulnerabilities its vetted-defender program found in critical software are now findable by anyone. There's also friction, and it's intense - Broadcom agreeing to lend Anthropic up to $42 billion so the lab can lease Broadcom's own chips, Nvidia running a roughly $500-billion version of the same loop while bankers ask for stronger guarantees, Anthropic's prospectus showing 47% of 2025 sales routed through Amazon and Google, which also supply its compute, hold its equity and compete with it, against $8 billion in operating losses and $417 billion of compute commitments, 3.5 gigawatts locked on multi-decade terms in a year when the cost of a fixed level of performance halves about every quarter, Z.ai's openly downloadable GLM-5.3 writing end-to-end exploits 50 times in 410 tries against Mythos Preview's 56, its refusals strippable for about $4,400 of compute down to 2% and bypassable by plain prompts between 64% and 100% of the time, NIST's standards center putting that open model roughly four months behind the US frontier, which is most of the lead the gate existed to protect, OpenAI accusing Moonshot of wide-scale distillation in the same week, Ted Cruz blocking a federal safety board that would have had 45-day pre-release access hours before the White House and six labs signed a voluntary accord instead, a repurposed 1980s drug-war grant program pulling local Flock and Axon plate reads into White House servers across 33 regions and all 50 states with Georgia cities barred from running cameras on state roads unless they send the data up, $277 million added in July, the files held outside the Freedom of Information Act by the DEA's own court filings, the Postal Service wiring 4K cameras into a hundred mail trucks that already drive every block, Marin County's logs showing 254,131 out-of-state and federal plate searches in one month in a county already bound by a 2022 settlement not to do it, a new paper finding nine pooled frontier judges carry about as much independent information as two, and Newsom vetoing community-owned solar on his way to signing the household-device bills. But friction generates shavings, and shavings tell you what the surface was actually made of underneath the finish. Step back for a moment and you can see it: limits defended as the nature of a thing turning out to be the dimensions someone chose - a surgical dilemma for families dissolving once a valve could change size, a grid's overbuild revealed as a coordination gap rather than a shortage of power, a frontier capability's scarcity lasting four months instead of years, a control layer that could have been metered arriving open and laptop-sized within a week, and a sheriff's illegal sharing becoming visible because California wrote audit logs into statute and an outside group could read them. Every transformation has a breaking point. Expansion can split what was built to one fixed size... or let it carry a life no one thought to measure it for.

AI Releases & Advancements

New today

  • Amazon Web Services (Strands Labs): Released Strands Decider 2B, an open-source decision model with 1.9B parameters under Apache 2.0. It returns calibrated choice, yes/no and score answers instead of text, and runs locally on a CPU, a consumer GPU or an Apple silicon Mac. The release includes a CLI and an HTTP server. (Strands Agents)
  • Cloudflare: Released Clef (27B) and Clef-flash (9B), its first self-trained decision models, as open weights under Apache 2.0. They are hosted on Workers AI and compatible with the Jev API. Cloudflare also opened a reinforcement-learning fine-tuning service for Clef to design partners. (Cloudflare Blog)
  • Perplexity: Launched the Decisions API, powered by pplx-decider-v1-27b. The model is a multimodal decision model fine-tuned from Qwen3.8-27B, and its weights are on Hugging Face under Apache 2.0. The API costs $0.04 per million input tokens, and output tokens are free. (Perplexity Docs)
  • Fastino Labs: Released GLiDE, which it calls the first "thinking decision model." Fastino says GLiDE leads the Decision Index by 6.9 points over the next-best model. (PR Newswire)
  • Anthropic: Launched mods for Claude Code. Mods are small TypeScript functions that can change how Claude Code behaves, customize its UI, or replace built-in features. They ship inside plugins and install with /plugin in the CLI or desktop app. (Claude Blog)
  • Anthropic: Made Claude for Government available to US federal and state civilian agencies after an open beta that began in July. It runs in a FedRAMP High environment, with usage-based billing, per-department budgets and audit logs. (The Decoder)
  • Black Forest Labs: Released Flux 3 Image, the image model in its Flux 3 family. It does text-to-image and image-to-image generation and edits in several steps without changing the rest of the image. Users can compose scenes with bounding boxes and up to ten reference images, at up to 4K output. It is available through the API and as licensable commercial weights. (The Decoder)
  • Ideogram: Released Ideogram 4.5, an image model that edits only the requested region and leaves the rest of the image unchanged. It has four quality tiers at native 2K resolution, priced from 0.8 to 22 cents per image, on the Ideogram platform and API. (Ideogram)
  • Microsoft: Released three speech models. MAI-Transcribe-2-Streaming is its first streaming transcription model and covers 60+ languages. MAI-Voice-2.1 and MAI-Voice-2.1-Flash convert text to speech in 23 languages. (Microsoft AI)
  • NVIDIA: Published PixelUMM on Hugging Face without an announcement. It is a 15.2B-parameter model built on Qwen3-8B that works directly on raw pixels, with no separate image encoder. It understands and generates both images and video, under a noncommercial license. (OrcaRouter)
  • Allen Institute for AI (Ai2): Open-sourced Olmo-core 3, a training framework for mixture-of-experts models. Ai2 has benchmarked it on a model with more than one trillion parameters. (Hugging Face Blog)
  • NVIDIA: Released DOCA AI agent skills on GitHub. The skills give coding agents verified API signatures, hardware capability requirements and build constraints for building applications on BlueField DPUs. (NVIDIA Developer Blog)
  • NVIDIA: Open-sourced DIN Deploy, a set of C++ samples for running AI models locally on Windows and Linux. The samples use ONNX Runtime with the TensorRT RTX execution provider and cover Whisper and Parakeet speech recognition, SAM 2.1 segmentation and FLUX.2-klein image generation. (NVIDIA Developer Blog)
  • DeepSeek: Released DeepSeek Harness Desktop for macOS and Windows, a desktop app for its open-source agent harness. (DeepSeek)
  • Shopify: Launched Canvas, a desktop tool where merchants build their store by chatting with Shopify's Sidekick AI agent while the real store code renders live. (Shopify)
  • Cloudflare: Made AI Search generally available. (Cloudflare Blog)
  • Ivo: Released Ivo Sage, an open-source model post-trained from DeepSeek V4 Flash for long-horizon contract work. Ivo says it is the first legal AI company to publish a free open-source model. (GlobeNewswire)
  • Talus Bio: Released Ptarmigan-1 on a public portal. The model predicts where small molecules bind across the whole human proteome without needing 3D protein structures, including disordered proteins. (BioSpace)
  • Cantina: Released Apex Flash, an open-weight security model trained on 50 vulnerability cases. (RuntimeWire)
  • Tuskira: Launched an open-source runtime gateway for AI agents that lets developers observe, govern and swap LLMs and MCP tools without rewriting agent code. (FinancialContent)
  • Legato: Started selling Legato Frames, AI hearing glasses from $999 that pick out voices from background noise and amplify them, for adults with up to moderate hearing loss. (TechCrunch)

Other recent releases

  • Google DeepMind: Released Gemini 4 Argon, its new frontier model for long-horizon coding, enterprise knowledge work and cyber defense. It is rolling out first to a limited set of trusted cyber defenders through the Fairwind Program, without cyber guardrails for those users. It supports up to 1M output tokens, and introductory pricing is $2/$10 per million input/output tokens. (Google)
  • Google DeepMind: Introduced SynthID Bio, watermarking for AI-designed proteins, with the tools released as open source. It embeds a detectable signature in protein sequences (via a SynthID-enabled ProteinMPNN) and in AlphaFold 3 structure predictions. In wet-lab tests the watermarked protein binders kept their function. (Google DeepMind)
  • Google: Rolled out Skills globally in Gemini chat, replacing Gems. Skills are reusable, detailed prompts that users call with "/" or that Gemini runs automatically, and they can be chained and take documents, PDFs and images as reference. The format is based on Anthropic's open Agent Skills standard. (The Decoder)
  • Meta: Launched Muse for Small Business, which brings its Muse agent to small business owners. It connects to Shopify, Stripe, QuickBooks, Slack, Canva and other tools, plus Instagram, Facebook and Meta ad accounts. It is free with usage limits. (Meta)
  • U.S. Government: Launched America.gov, a public AI assistant built with Google and SpaceXAI. (TechCrunch)
  • Ant Group (inclusionAI): Launched Ling-3.1-flash, a 560B-parameter MoE language model with about 25B parameters active per token, built for agents, search and office work. It is in a two-week free trial with a 256K context window. The 1M-token window and an open-source release are planned for after the trial. (TechNode)
  • Cohere: Released Embed 5, two multimodal embedding models (Pro and Fast) that share one embedding space, so an index built with Pro can be queried with the cheaper Fast model. Both have 128K context and cover 100+ languages, and are available via the Cohere API, Model Vault, Microsoft Foundry and Amazon SageMaker. (Cohere)
  • Perplexity: Released pplx-embed-v2-context-9b-preview, an MIT-licensed contextual embedding model for RAG. It embeds each chunk with the full document in view and is trained to retrieve both answers and the evidence that supports them. Weights are on Hugging Face. (Perplexity)
  • Upstage: Launched Solar Mini 4, a new LLM in its Solar lineup aimed at repetitive enterprise tasks. (Aju Press)
  • Pienomial: Launched AT0M, a decision model that businesses own outright and run on their own hardware. It picks answers from options defined by developers rather than writing free text. It ships as a single executable for Intel Xeon, Apple Metal and NVIDIA CUDA. (Khel Ja / ANI)
  • DeepSeek / Huawei: Released open-source programming tools for Huawei's Ascend AI chips, built around the TileLang language, with libraries for computation and for moving data between chips. (The Decoder)
  • MiniMax: Open-sourced OpenAgentCore, which lets developers use OpenAI's official Agents SDK with MiniMax models. (TokenPost)
  • Hugging Face: Released Transformers 5.18.0, adding four model families: NVIDIA's Nemotron 3 Diarization (streaming speaker diarization through the standard AutoModel interfaces), NemotronH Omni, HyperCLOVAX Vision V2 and GTE. (GitHub)
  • OpenClaw: Launched OpenClaw Enterprise, a free control plane for managing persistent AI agents in production, backed by OpenAI, Red Hat and NVIDIA. (VentureBeat)
  • IBM: Made IBM Bob, its agentic software development platform, available for self-hosted deployment on-premises, in private and sovereign clouds, and in air-gapped environments. (PR Newswire)
  • AMD: Introduced AMD Ross, an agentic AI assistant for embedded system development across AMD FPGAs, adaptive SoCs and embedded processors. It provides MCP servers, an AMD knowledge base and expert-written agent skills, and works with whatever LLM or IDE the team prefers. (AMD)
  • NVIDIA: Made the cuObject client and server libraries generally available for RDMA-accelerated object storage access from GPUs. It also released the SCADA Server SDK, which lets storage providers serve storage requests initiated by GPUs. (NVIDIA Developer Blog)
  • CoreWeave: Launched CoreWeave Forge, a development layer for training and improving models and agents. It includes the ARIA coding agent and Sandboxes (both now generally available), Agent Lens for observability of production agents, Notebooks, a Registry, and serverless post-training. (CoreWeave)
  • Kong: Launched Volcano, a platform for building and running AI agents and web apps. It bundles durable workflows, branchable PostgreSQL, edge functions, auth, real-time services and file storage, with integrations for Claude Code, Codex and Cursor. (PR Newswire)
  • Querit: Launched Code Search for its Search API, a programming-focused vertical for AI coding agents that you turn on with a single parameter. It is available via MCP and through LangChain, Dify and other frameworks. (PR Newswire)
  • Magnitude (YC S25): Released an open-source inference engine for AI agents that tunes itself to the hardware it runs on, across Mac, Linux and Windows. (GitHub)
  • MindOn: Introduced Mind-1, a physical AI model built for robots doing real-world tasks at human speed. (MindOn)
  • OpenAI: Released GPT-6.1 Sol, which OpenAI says comes close to GPT-6 Astra on agentic coding, computer use and office work at one-fifth of Astra's standard token prices. It is available now in ChatGPT Work, Codex and the API as gpt-6.1-sol, and is rolling out in GitHub Copilot for Pro+, Max, Business and Enterprise users. (OpenAI)
  • OpenAI: Launched dots, always-on agents powered by GPT-6 Astra. Each dot runs on its own cloud computer with a browser, connects to more than 4,000 apps through plugins, can be reached from ChatGPT, Slack and Teams, and does only read-only research in the background. It is available to Pro and Business Premium users in eligible markets. (OpenAI)
  • OpenAI: Added team features to ChatGPT. Space is a shared workspace with an assistant called Dot, and Pages are collaborative documents for people and agents. ChatGPT can now be @mentioned in Slack and Microsoft Teams, and a Meetings plugin is in beta on macOS. Developers get plugin extensions with sidebar panels and file viewers, a Plugin Creator tool, and MCP Events triggers for automations. (ChatGPT)
  • OpenAI: Expanded Codex with reusable cloud development environments that work across devices and a rebuilt CLI with voice control and an /agents view. It also added a code review view in the ChatGPT desktop app and Codex Security Cloud, which scans GitHub repositories on demand or on a schedule and prepares fixes. (TechCrunch)
  • OpenAI: Added computer use to the Agents API and launched a Decisions API, built on GPT-6 Luna, that returns fast answers from a fixed set of options for classification tasks. (The Decoder)
  • H Company: Released Holotron4 Nano, a computer-use model built on NVIDIA's Nemotron 3 Nano Omni. H Company reports it raises the base model's OSWorld score from 21.0% to 76.3%. It shipped alongside Holo4. (Hugging Face)
  • Liquid AI: Released d1, a decision model on the Liquid API (d1:free) that returns calibrated probabilities for yes/no, multiple-choice and rating questions in one call, with zero output tokens. It is aimed at routing, moderation, triage and reranking. (Liquid AI Docs)
  • Ollama: Ollama 0.35 can now run decision models locally through a new /v1/systemone endpoint that is compatible with TypeSafe's Jev API. Three models are available at launch: Bespoke Labs' Nimble 9B, and Together AI's tev1 in 4B and 0.8B sizes. (Ollama)
  • PostHog: Open-sourced Jeeves, which uses reasoning to improve decision models compatible with Jev. (GitHub)
  • Voltropy: Opened early access to Vast-10M, a family of models with a 10-million-token context window. It comes in Flash, Medium and Pro versions, built on DeepSeek V4.0 and GLM-5.2 with a new attention method Voltropy calls VSA. (Voltropy)
  • NVIDIA: Released Kumo Tabular, an open foundation model for tables in three sizes (28M to 215M parameters) under a license that allows commercial use. Given a table of labeled rows, it predicts labels for new rows in one pass, with no training or tuning. (Hugging Face)
  • NVIDIA: Released VSS Blueprint 3.3 for building video search and summarization agents. It adds a Build Vision Agent skill that combines alerting, search and summarization into one deployment from a natural-language request. It also adds Adaptive Efficient Video Sampling, which NVIDIA says uses about 80% fewer vision-model input tokens on a 60-minute summary. (NVIDIA Developer Blog)
  • Perplexity: Released Photon, its own retrieval and ranking engine written in Rust, which now serves all production traffic. It also launched Fast Search in the Search API (search_type: "fast") at $1 per 1,000 requests, with a reported 160 ms median latency. (Perplexity)
  • MLC: Released TIRx Harness, an open-source compiler harness that lets AI agents write and optimize GPU kernels. (MLC Blog)
  • Visa: Open-sourced VVAH, its AI-powered tool for detecting cyber threats. (Open Source For You)
  • U.S. Government / Google: Launched America.gov, a public AI assistant built with Google and SpaceXAI that helps people find federal government services and information. (Google)

Sources and Further Reading

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