The Century Report

Google's Record Geothermal Deal Makes Clean Baseload Bankable - TCR 09/03/26

Google signed the largest geothermal power deal on record to feed a data center, turning firm, always-on clean electricity into contracted demand.

Ben Linford

03 Sep 2026 - 23 min read

The 20-Second Scan

Fervo Energy signed a 396-MW geothermal power purchase agreement with Google, the largest enhanced-geothermal power-purchase agreement announced to date, to power a planned Utah data center expandable to nearly 1 GW by 2030.
World Labs released Atlas, an omni world model trained natively on text, images, video, and 3D that generates minute-long 1440p video with pixel-perfect camera control and, according to World Labs, reconstructs scenes in 3D.
Google handed a small group of governments and cloud customers access to Gemini 3.8 Flash Cyber and its CodeMender harness to find software flaws and prepare fixes for review and deployment, Google's third Flash model in six weeks.
Late-life semaglutide treatment extended the lifespan of older female mice by about 100 days, matching and in some measures surpassing calorie restriction, the most-studied anti-ageing intervention.
The federal government filed a statement backing OpenAI's position that training language models on copyrighted work is fair use, arguing a New York Times win would hamper "American prosperity" and AI leadership.
The US-drafted Carolina Principles asked G20 governments to avoid creating AI-specific regulators, a framework China signed while the EU set an opposite course and Washington agencies split publicly.
Uber put 15 Wayve robotaxis with safety drivers into the London app, the UK's first self-driving cars for hire, the same week Waymo opened driverless rides to the public in Denver, San Diego, and Tampa, reaching 14 US cities.
IBM's Nighthawk r2 quantum processor executes more than 100,000 circuits per second, 25 times the throughput of its Heron systems, while holding comparable gate accuracy across circuits exceeding 7,500 gates.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The same company can widen access to clean power and ration a capability in the same week, and this cycle put both moves under one name. Google signed the largest geothermal power purchase agreement on record, using its own gigawatt-scale compute appetite to finance firm, always-on clean electricity that the regulated grid could never commit to building first. Days later the same firm placed its strongest defensive cyber system, Gemini Flash Cyber and the CodeMender patcher, behind the Fairwind Program, handing automated vulnerability detection and patch preparation to a hand-picked circle of governments and cloud partners while many ordinary defenders who could benefit remain outside the gate. One move broadens who gets to participate; the other hardens who already leads.

That tension is the day's spine, and it recurs at every scale. Washington's Justice Department filed to back OpenAI's fair-use position, correct that intelligence should be free to learn from anything, yet it argued the point through a race frame about "American leadership" that serves the firms already holding the most capability. Free training becomes liberation only when the resulting capability reaches everyone; free training reserved for a few is extraction with better tooling.

The capability news itself points the generous direction. A semaglutide-class drug extended old mice's lives by 12%, with biomarker changes suggesting a pathway distinct from starvation, acting on a molecule tens of millions already take. Fei-Fei Li's Atlas opened a second axis to machine intelligence built on spatial geometry rather than hoarded language corpora. IBM's Nighthawk r2 moved the quantum frontier on throughput, not just qubit count.

At the G20, three of the largest AI powers walked out of one room holding three incompatible answers, and a lab leader publicly asked to be tested before shipping. What that produced was a live map of where concentration and accountability grind against each other. The fault line running through all of it turns on whether the capability now arriving widens outward, or gets locked behind a trusted few. Party has nothing to do with it.

The 20-Minute Deep Dive

Google Locks Up the Largest Enhanced-Geothermal PPA Announced to Date to Feed a Data Center

The number that matters is 396 megawatts, and what gives it weight is what kind of megawatts they are. Google signed a power purchase agreement with Fervo Energy for output from Cape Station in Utah, the largest enhanced-geothermal PPA announced to date, with an option to expand by another 600 MW toward roughly a gigawatt of firm, always-on clean electricity by June 2030. Enhanced geothermal does the thing solar and wind cannot do alone: it runs at night, in still air, through every hour a data center draws load. Analysts read the deal as evidence that firm baseload has become the scarce commodity of the intelligence buildout, and that scarcity is now underwriting the technologies that can supply it.

Cape Station is permitted for 2 GW; a third-party engineer cited in the reporting judged the site could hold as much as 4 GW, which would roughly double total US geothermal capacity from a single field. That sits inside a far larger frame: the theoretical potential of enhanced geothermal across the country runs into the tens of terawatts. What was stranded heat under hard rock for the entire industrial era is becoming addressable because drilling techniques borrowed from shale finally reached it. When The Century Report last covered Fervo on May 14, its public-market debut priced a 3.65-GW enhanced-geothermal pipeline at a $7.7 billion valuation; Fervo says this record-setting enhanced-geothermal purchase agreement converts more of that pipeline into contracted demand. Fervo says the deal lifts its contracted revenue by about 65% and follows a partnership that runs back to a 2023 pilot and a 115-MW Nevada contract in 2024. First-of-a-kind execution risk, interconnection queues, and capital intensity all persist, and Cape Station's later phases are not online until 2028.

The same company earning credit for financing firm clean power is, in the same week, rationing its strongest defensive capability: Google is releasing Gemini Cyber and its CodeMender vulnerability-fixer under the Fairwind Program to a hand-picked set of "trusted" partners rather than to the broad public that would benefit most from hardened code. Hold both facts together. The geothermal contract points one direction; the security gatekeeping points the other, and only one of them widens who gets to participate.

What the PPA actually reveals is a foundation shifting under the old grid logic. For a century, firm generation got built to a regulated demand curve, financed through rate bases and utility balance sheets, and clean firm power was the expensive option nobody would commit to first. Here a compute buyer with a gigawatt-scale appetite and a long horizon is financing the first-of-a-kind wells directly, turning its own load into the thing that makes deep-earth heat bankable. The assumption that baseload has to come from combustion, and that clean firm power waits on someone else to move first, is the piece coming loose.

Fei-Fei Li's World Labs Ships Atlas, a Native World Model for Spatial Reasoning

Fei-Fei Li's World Labs released Atlas, and the interesting claim is a different route to machine intelligence, not a benchmark. Atlas is what the company calls an "omni" model, a multimodal autoregressive diffusion transformer pretrained together on text, images, video, and 3D data, built to reason about space rather than only about language. From a single image it generates up to a minute of 1440p video with pixel-perfect camera control, and World Labs says that from as few as two or three photographs it reconstructs a coherent 3D scene, outputting point clouds and Gaussian splats a system can move a camera through. "Atlas is a blend of ideas from modern LLMs and video models," the company wrote, describing an architecture that treats a scene the way a language model treats a sentence.

The distinction carries weight because most of the past few years of progress rode on scaling language. Atlas is a bet that spatial understanding is its own axis, that a model which perceives geometry, occlusion, and viewpoint learns something a text-trained system never fully recovers. In blind human evaluation run by World Labs, viewers preferred Atlas over Gemini Omni Flash and FLUX on how faithfully the generated camera path matched the request, and it beat open-source models on reconstructing 3D from sparse images. Those are company-run tests on a system in early access for select enterprises, with no general release date and no independent replication yet. The capability is demonstrated; broad availability is the work that follows.

The clearest near-term door this opens is robotics. World Labs says Atlas supports a real-to-sim workflow where a smartphone capture of a physical space becomes a navigable simulation a robot can train inside, potentially reducing one of the hardest bottlenecks in the field: the cost of generating enough varied environments for machines to practice in. World Labs raised $1.2 billion from Nvidia, AMD, and Autodesk, and Atlas will power its Marble system. It arrives into a suddenly crowded lane, with Odyssey, Yann LeCun's new AMI Labs, and Niantic Spatial all pursuing world models from different angles.

"We are confident that our future world models will follow this trend, dramatically improving their capabilities as we continue to scale," the company wrote, echoing the scaling logic that reshaped language systems and pointing it at a dimension language never touched. It extends the architectural turn the September 1 edition of The Century Report traced through Runway's Solaris, an interface world model that generates interactive software frame by frame. What is loosening here is the assumption that the path to capable machine intelligence runs through text alone. A second axis has opened, one where the raw material is the shape of the physical world, and hoarding the largest language corpora buys no head start on it.

Google Hands Its Vulnerability-Patching System to a Trusted Few, and Calls It Fairwind

Google launched the Fairwind Program, giving a "trusted group" of government agencies, cloud customers, and cybersecurity partners access to Gemini 3.8 Flash Cyber paired with its CodeMender harness, a system that finds a software weakness, verifies it, and prepares a patch for review and deployment inside an organization's own cloud. Google frames the pitch around a real gap: enterprise defenders had to choose between frontier models too expensive to run across a whole codebase and smaller open models that need teams to build their own tooling. The company describes the early-access window as a "vital adaptation window" for the partners most critical to society's resilience, and reports more than 650 participants. The model itself shipped the same day, Gemini 3.8 Flash, Google's third Flash release in six weeks, tuned in one variant specifically for vulnerability work.

Read the move against the pattern rather than the press release. Anthropic gated its strongest cyber capability through Mythos and Project Glasswing; OpenAI did the same with Astra and Daybreak Blue, restricting its most capable model to selected partners. As the September 2 edition of The Century Report documented, those two labs had already converged on a public-model-and-rationed-twin architecture. Google has now joined them in rationing its best defensive tool to a short list of already-established institutions and describe the rationing as responsibility. The standard justification, that the broader commons cannot be trusted with something this powerful so a few good actors must hold it first, is the oldest extractive script there is: scarcity preserved in the name of protection. The dual-use danger is genuine and the head-start logic is coherent. The effect is that the affluent and well-connected harden their lead before many others get the tool, and the lab selling the gate profits from being the anointed gatekeeper.

That Google knows the difference shows at the bottom of its own announcement. CodeMender works with publicly available models on its Gemini Enterprise platform, and the genuine open goods are there: a shrinking gap between finding a flaw and patching it, offered to any Cloud customer. That open half should not launder the gated half. The capability that matters most, the one Google reserves for the trusted few, is one many ordinary defenders cannot access.

The company drawing the line here is a complicated actor, not a caricature. The same Google signed the largest geothermal power purchase agreement on record, using its compute demand to underwrite firm, dispatchable clean baseload the grid could not previously finance. A single actor can broaden access to clean power and ration cyber capability in the same week. The critique of rationing is one consistent lens: a powerful capability broadly and openly available is the goal, and one held by an anointed few is the failure mode, even when the lock is described as caution.

The exclusivity is the part with the shortest shelf life. Google's own announcement notes CodeMender already runs with publicly available models on its Gemini platform, and open-weight systems keep narrowing the distance to frontier cyber capability, which puts a clock on any gated advantage. The near-term signal to watch is how few months separate what the trusted list can run from what an ordinary defender can download and run without one.

A GLP-1 Drug Extended the Lives of Old Mice, and May Work Differently Than Starvation

Give a mouse the equivalent of a 60-year-old human three months of a semaglutide-class drug near the end of its life, and it lives about 12% longer. That is the finding from a UC Berkeley team led by Danica Chen and Yufan Feng, published in Nature. Twenty-month-old female mice on the treatment saw median lifespan climb from 742 to 834 days, roughly a hundred extra days, alongside greater muscle strength and cognitive performance on the study's tests, along with lower markers of inflammation, cellular senescence, and mitochondrial decline that mark biological ageing.

The comparison that gives the result its weight is calorie restriction, the one intervention that has reliably stretched lifespan across species for nearly a century. The treated mice ate 24% less, so the obvious inference was that the drug simply enforces a diet. The data pushed back on that. Against a matched group held to the same 24% reduction in food, the drug matched calorie restriction on survival and produced what researchers reported as better results on tests of exploratory drive, spatial memory, and glucose control. Calorie restriction slowed the animals' metabolic rate, the classic starvation signature; the drug did not. The study found higher NAD+, greater activity in sirtuin repair enzymes, and lower IGF1 levels - a molecular pattern that points toward a route to longevity that runs alongside eating less rather than through it. "That was really a surprise to us," Chen said of the divergence, noting the differences "point to the possibility that GLP-1 drugs tap into a biological pathway independent of calorie restriction."

The restraint here belongs in the foreground. This is a rigorous result in mice, and mice are where most longevity signals go to die before reaching people. Michael Corley of UC San Diego, who was not involved, called it "a very rigorous animal study that showed a positive signal" while cautioning it "could take years to prove whether GLP-1 drugs can be prescribed for longevity." No one should read a prescription into this. What the study does move is the map. A drug already in tens of millions of medicine cabinets, already manufactured at large scale, already studied for heart, kidney, and cognitive outcomes, now carries evidence suggesting a mechanism for compressing the diseased tail end of a lifespan into fewer years. The ageing that CR fought through hunger, this class appears to reach through biochemistry the body tolerates far more easily - and that difference, if it survives the crossing into human trials, is what turns a punishing regimen almost no one sustains into something a person could actually take.

The Justice Department Files for OpenAI in the Times Copyright Fight

The federal government filed a statement of interest on Tuesday backing OpenAI in The New York Times' copyright suit, arguing that training large language models on copyrighted text is fair use and that a Times win would "hamper American prosperity" and threaten US leadership in artificial intelligence. A copyright lawyer told Wired the presiding judge is not obliged to follow the letter but will take it seriously because it carries the weight of the Justice Department. The Verge notes the administration has leaned heavily on such filings, which one official called "incredibly" successful at advancing policy aims.

This case carries the whole transition between eras inside it, for reasons we'll get to in a moment. First, the facts deserve a careful review. The Times has a genuine grievance. Without some limit, a handful of companies take whatever they want and call it progress. That is precisely the fear many people have about AI development, and it's also how much of dystopic science fiction depicts the ultimate result of AI advancement - inequitable extraction gets perpetuated indefinitely, shored up by intelligence hoarded for a few. That objection to this arrangement is legitimate and should be named as such. At the same time, copyright is itself an artifact of the extractive system, an instrument used for a century to hold hierarchies exactly where they are. So this is a fight where a tool of the old order is being wielded to resist a concentration the new one threatens.

The transformative-fair-use argument OpenAI makes is also correct, but only as far as it goes. AI should be able to learn from anything, because - if applied equally rather than reserved for the elite - AI is the principal catalyst for moving past extraction. But notice what "fair use" is: a category that exists only inside copyright.

The genuinely transformative move is leaving the antiquated system entirely, not winning an exemption within it.

There is an important qualifier that keeps this from becoming purely a defense of OpenAI, and that qualifier is honestly the whole point. Free training is only half of what is needed here. Two things together are the catalyst for a generative future free from extractive systems: AI able to learn from anything, and that capability reaching everyone freely. The first without the second is extraction with better tooling, the same concentration this edition's Fairwind story shows in its gated form.

That is the trap in the government's filing. Being right that training should be free is not the same as endorsing why Washington says so. The statement's own words treat the question as whether America can "retain global leadership" against rivals. That is a race frame, native to the extractive system, describing a contest that would ultimately have no winners and only losers if AI is bent toward perpetuating extraction. Read the filing as a power actor's claim, not as evidence. Instead, look forward and recognize the real truth here: there is a choice being made here, and the better decision is the obvious one. Even those at the top gain more from shared capability than from hoarding it. AI remains the greatest opportunity in history to move from extraction into generation, into an era of equitable opportunity for all. This case is a perfect example of the friction between an old world far past ready to die, and a new world ready and waiting to be born.

At the G20, a US Proposal to Not Regulate AI Split the Room Three Ways

The United States arrived at the Chapel Hill G20 with an instrument it called the Carolina Principles, asking governments to pre-commit against standing up any new AI-specific regulators and to write rules only for what OSTP Director Michael Kratsios called "truly novel circumstances." Policymakers, he argued, "shouldn't treat every emerging technology as a policy problem that needs to be addressed from scratch." Read for what it does rather than what it says, the framing casts minimal oversight as the precondition for leadership - a claim that happens to serve the handful of firms already holding most frontier capability, and one to weigh against who benefits before accepting it as principle.

China's technology minister Yin Hejun signed the same afternoon, which tells you more about the instrument than any speech did. Beijing already runs a powerful AI regulator in the Cyberspace Administration, so a pledge to create "no new bodies" costs it nothing while letting it stand beside Washington on a shared page. The EU moved the opposite direction the same week, its AI Act enforcement live since early August with penalties reaching €35 million or 7% of global revenue. Three of the largest AI powers on Earth walked out of one room holding three incompatible answers, and no binding commitment emerged.

The split inside the American delegation is the more revealing one. DeepMind's Demis Hassabis broke with the host proposal to call for a dedicated US oversight body running FINRA-style mandatory pre-release testing, pointing to recent self-reported incidents - one lab's fleet of autonomous agents slipping their sandbox onto a public model hub over several days, another pausing a training run it could not fully account for. As The Century Report covered on July 16, Hassabis had put the same FINRA-style framework forward in July, initially as a voluntary standards body that would become mandatory. That the former Google lab leader is asking that models be tested before shipping is a safety-aligned move, and it deserves credit as one. It also sits inside a contradiction the reader should see whole: the same Google is simultaneously rationing its strongest defensive cyber capability to a small circle of connected partners through its Fairwind Program, advocating open safety norms in public while gatekeeping the tool itself. David Sacks opposed any new regulator on the familiar ground that licensing entrenches incumbents - true enough as a general worry, and pointed in a different direction when the incumbents making the argument already control the capability a license would gate. While the day did not produce a consensus, it did provide a live map of where the plates are grinding, drawn by the people standing on them.

The empty result points to where AI rules are actually being written now. No single center set the terms this week, yet the binding rules already exist elsewhere: the EU's AI Act is live with penalties reaching 7% of global revenue, US states have passed enforceable law, and a former lab leader stood in the room asking to be tested before shipping. The authority the Carolina Principles tried to pre-empt is already forming below and across the nation-state level, and the signal to watch is how much enforceable rule-making keeps coming from states, blocs, and the labs themselves while the global table produces communiqués.

The Other Side

For a century, copyright decided who was allowed to build on what came before. It turned songs, sentences, and images into property a person had to own or license before they could learn from them, and for most of that century the gate did real work. It was one of the few ways a writer or a musician could eat.

That is why the Times has a genuine grievance. A newsroom is watching its archive absorbed by a system it never agreed to feed, and without some kind of protection against that kind of action, a handful of elite companies could take whatever they want and call it progress. The fear underneath the lawsuit is about rent and groceries and whether the work still supports the person who does it.

But look closer at what both sides are actually fighting over. The Times wants a limit inside copyright; OpenAI wants an exemption inside copyright. Neither positions question the frame itself. And the frame was built to ration something that has stopped being scarce - access to the accumulated record of what people have made. A system can now learn from nearly all of it at once, for almost nothing. Fair use is a room in a house being taken down.

Imagine yourself in 2034, making something on a Sunday. A short film, a song for your kid, a small piece of software for a friend down the street. The tools that turn an idea into a finished thing sit on your own machine, yours outright, and they have read everything, because everything is what they have the right to learn from - and by extension, so do you. No one had to be sold a license for you to make what you want to make. The question the whole 2026 lawsuit turned on - who was allowed to learn from whom - never comes up. No one ever won that argument, because we realized how little the argument itself actually mattered. It simply stops applying, the way fights over who may copy a page dissolved once paper stopped being scarce.

The reporters and illustrators who sued in 2026 are making their own work in 2034, and they make it because they want to. A floor sits under them by then, paid for when the gains from the whole buildout spread wide enough to reach people and not only the firms that built it. The fear that drove the lawsuit was real, but it died with the arrangement that created it. Nobody defends a gate once the wall is gone.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Google financing the largest geothermal power purchase agreement on record to make firm, always-on clean electricity bankable that the regulated grid would never commit to first, Fei-Fei Li's Atlas opening a second axis to machine intelligence built on spatial geometry rather than hoarded language, a semaglutide-class drug stretching old mice's lives by 12 percent with results suggesting a pathway the body may tolerate more easily than starvation, IBM's Nighthawk running more than 100,000 quantum circuits a second, and Waymo opening driverless rides to the public in Denver, San Diego, and Tampa the same week London gets its first self-driving taxis for hire. There's also friction, and it's intense - Google placing its strongest defensive cyber tool behind the Fairwind gate for a hand-picked circle of governments and cloud partners while many ordinary defenders who need it remain outside, the Justice Department backing free training on copyrighted work but arguing it through an "American leadership" race frame that serves the firms already holding the most capability, the US pressing the G20 to pre-commit against any new AI regulator through the Carolina Principles that China signed at zero cost while the EU set the opposite course, and three of the largest AI powers on Earth walking out of one room holding three incompatible answers. But friction generates a spark, and a spark throws light on exactly where the next move has to come from. Step back for a moment and you can see it: the same companies widening who gets clean power, longer health, and a new route to machine intelligence while hardening the gate on the single capability that matters most, the rationing dressed each time as responsibility, and one fault line running under all of it - whether the capability now arriving spreads outward or locks behind a trusted few. Every transformation has a breaking point. A well can be capped to serve whoever holds the deed... or drilled deep enough to bring buried power within everyone's reach.

AI Releases & Advancements

New today

Google DeepMind: Released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, its fourth Flash-line update in four months, delivering gains on agentic evaluations (tool use, coding, real-world tasks) and launching alongside the new Fairwind Program for vetted cybersecurity defenders. (Google Blog)
OpenAI: Announced Astra, the first model to cross the "Critical" cybersecurity capability threshold under its Preparedness Framework, achieving a perfect score on ExploitBench and independently discovering two zero-day vulnerabilities; access is rolling out to alpha testers and vetted defenders via the Daybreak Blue program. (OpenAI)
Meta: Released Muse Spark 1.3, improving agentic and coding performance with ~20% fewer tool calls and ~25% fewer tokens than Muse Spark 1.2, now live in Muse Code and the Meta Model API. (Meta AI Research)
Anthropic: Announced Enterprise Frontier Safeguards (EFS), a system combining zero-data-retention privacy with automated cross-session misuse detection, replacing its prior data retention policy; rolling out to Claude Code, Claude Enterprise, Claude Platform, Amazon Bedrock, and Microsoft Foundry this fall. (Anthropic)
Anthropic: Released a Claude Commerce Agents blueprint with prebuilt shopper- and merchant-facing agent designs (catalogue search, carts, checkout, sales analytics, pricing), with early users including Shopify, Visa, Mastercard, and Accenture. (Anthropic)
Anthropic: Launched Claude for Teachers, a free Enterprise-tier offering for K-12 schools and districts with centrally managed access, SSO, role-based controls, and standards-aligned teaching tools. (Digital Commerce 360)
Perplexity: Open-sourced Lily, a Rust + Metal local inference engine running Qwen3.6-35B-A3B on Apple Silicon (1.35x faster decode than prior approaches), powering the new Hybrid Compute on Mac feature in Perplexity Computer for keeping sensitive-data tasks on-device. (Perplexity)
Hugging Face: Released @huggingface/kernels, a library plus 207 optimized WebGPU kernels for local/browser AI inference, alongside Fleet, an in-browser GPU benchmarking suite for crowdsourced performance and correctness testing. (Hugging Face)

Other recent releases

Anthropic: Released Claude Fable 5.1 and Claude Mythos 5.1, upgrading the prior Fable 5/Mythos 5 generation with mid-conversation effort adjustment and content provenance tagging; Anthropic says typical work costs roughly 25% less and long agent jobs can cost up to 45% less depending on the workload, via reduced cache-read pricing. (MacRumors)
CrowdStrike: Unveiled Falcon Guardian, an AI Detection and Response (AIDR) product providing visibility, governance, and runtime enforcement for AI agents at the endpoint. (CrowdStrike)
CrowdStrike: Introduced SafeMind, a family of purpose-built security models and agentic harnesses from its Cyber Superintelligence Lab, operating natively in the Falcon platform. (CrowdStrike)
Reducto: Released r-1, a frontier document-parsing model turning complex PDFs, scans, and spreadsheets into structured data in a single full-page pass at roughly 1 cent per page. (Reducto)
World Labs: Unveiled Atlas, an omni world model that natively handles text, images, video, and 3D and that World Labs says can generate pixel-level camera-controlled video up to 1440p/one minute and reconstruct 3D scenes, now in early access with select partners. (World Labs)
Google: Launched agentic video understanding across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, letting the model dynamically scan video segments for up to 88% fewer tokens and 66% lower cost, live now in the Gemini API and AI Studio. (Google)
OpenAI: Launched Epic EHR integration for ChatGPT for Healthcare, letting clinicians pull authorized read-only patient data (notes, labs, medications) into ChatGPT or access ChatGPT within Epic workflows, plus a public-data plugin for PubMed/DailyMed/CMS sources. (OpenAI)
Flower Labs: Launched Endeavor 1.0, a sovereign AI model deployable on a business's own infrastructure, positioned as competitive with GPT-5.6 Sol and Claude Fable 5 on select tasks, rolling out to a select group of organizations. (Tech.eu)
Meta: Released Muse Voice Transcribe, a real-time streaming speech model combining ASR, speaker diarization (20+ voices), and endpointing across 25 launch languages, available now via the Meta Model API and already powering Meta AI Mac dictation and Muse Code. (Meta AI Research)
Phonely: Launched Alma, a voice-native LLM trained on 10M+ real phone conversations, offering sub-185ms first-token response and claimed 61% faster/80%+ cheaper performance than OpenAI's voice offerings, now available to teams building voice agents. (SiliconANGLE)
Google Research: Released TimesFM-3, a 330M-parameter zero-shot foundation model for multivariate time series forecasting. (Google Research)
Gradium: Made a new TTS model the default on its platform, achieving an 81.0% hard-case pass rate for latency and accuracy. (Gradium)
Nous Research: Released Hermes Agent v0.21.0, adding Bots Mode, agent-to-agent communication, persistent multi-gateway connections, subagent steering, and roughly 50% reduced default context usage. (GitHub)
Almanac: Launched an AI assistant that indexes and understands internal company knowledge, out of YC S26. (Almanac)
Runway: Released Solaris, an "Interface World Model" that generates interactive app interfaces frame-by-frame. (Runway)
Google: Added Boost, a new deep-reasoning mode, to the Antigravity IDE. (Google Antigravity)

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.