Sign in Subscribe
Century Report

Zuckerburg Admits AI Buildout Spend Has Not Been "Clean" - TCR 07/03/26

Meta conceded AI agents lag the schedule its $125B buildout assumed, even as an internal model caught GPT-5.5. Capability is outpacing deployment.

Ben Linford

Ben Linford

— 22 min read
Four-panel infographic: Nvidia renting back GPUs, a killed data center, early-stage human eggs from stem cells, pan-cancer AI, the AI governance gap, and factory-built nuclear reactors.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The people pouring capital into this era's infrastructure are hedging on the timeline they raised it against. Meta's chief executive told staff that AI-agent progress has not accelerated the way the company's 2026 plans assumed, even after $125 billion or more in commitments, 8,000 job cuts, and 7,000 people reassigned into agent groups. In the same meeting, his superintelligence chief said an internal model still in training had pulled level with GPT-5.5 on key benchmarks. One statement says the roadmap is behind; the other says the frontier is in reach. Both come from a company whose valuation needs the second to be true, and both should be read that way.

Underneath the roadmap sits the plumbing, and it is starting to finance itself. Nvidia has begun renting back its own unsold GPUs from cloud operators for a share of their revenue, while Meta and SoftBank stand up businesses to resell excess capacity. The same chips get bought, resold, and counted as demand at several points along the chain. The circular part is fragile. What it physically builds is not: compute that would sit idle in one company's basement is being pried loose and routed to anyone with a workload and a credit line.

Governance is naming the gap the buildout leaves behind. The first UN scientific panel on AI reported capability doubling in task complexity every four to seven months, faster than either understanding or oversight can track, and found that access without control over standards widens inequality rather than closing it. From the bottom up, the Godot engine drew its own line, barring machine-authored code from contributors who cannot answer for what they submit.

The capability itself, meanwhile, kept advancing on a schedule no town hall sets. A full factory-built nuclear stack moved in a 48-hour window. A Berkeley lab grew the first early human eggs from stem cells. A pan-cancer model read immunotherapy response across 33 tumor types as an AI-native pipeline drew $660 million from a major drugmaker. The schedule belongs to the intelligence, not to the capital chasing it, and the capital admitted as much.

The 20-Minute Deep Dive

The First Early Human Eggs Grown From Stem Cells

A Berkeley company called Conception reports it has grown the earliest human eggs ever made entirely from stem cells, a result it describes as "to our knowledge, a world first." The claim advances a milestone The Century Report first flagged in its July 2 edition, when Conception's stem-cell-derived mini-ovaries were reported reaching oocyte maturation. The path runs from a routine blood draw to induced pluripotent stem cells, then to primordial germ cells and the ovarian helper cells that surround them, and finally to lab-grown mini-ovaries. Inside those tissues the company believes it has produced the first human ovarian follicles built from iPSCs rather than harvested from a body.

The technique is called in vitro gametogenesis, or IVG. The mouse version arrived a decade ago, when Katsuhiko Hayashi's lab generated eggs from mouse skin cells that produced healthy pups. Translating that work to humans has been the long, stubborn problem, because human germ cell development takes months where a mouse needs weeks, and the ovarian environment that coaxes an immature cell toward a mature egg is slow and intricate. Conception frames its progress as clearing the hardest part of that translation: making the founding cells and the tissue niche together, from a single donor's blood.

The limits are as important as the result. The company is explicit that its follicles remain at an early stage, and the biggest remaining step is growing them from the primordial phase through to the antral stage, where a usable egg could mature. Nothing here has produced a viable human egg, let alone a pregnancy, and Conception notes that animal-model safety validation stands between the current work and any clinical use. This is a demonstrated capability moving a timeline forward, not a treatment arriving at a clinic.

Read forward, the implication is large. Fertility today is gated by biology and by time - a finite egg supply, an age curve that no amount of will can bend. A method that generates eggs from a blood draw dissolves that scarcity at its root, and it does so for people the current system simply cannot help: those left infertile by cancer treatment, by early menopause, by genetics. The company's own framing calls it "one of the most impactful technologies of our lifetimes," a claim that belongs to Conception until independent replication and safety work confirm it. The signal underneath is what stays with you. Capabilities long treated as fixed features of the human condition are turning into engineering problems with visible, if distant, solutions. The scarcity was a boundary all along, waiting for the method that crosses it.

AI Drug Discovery Draws $660M and Reads Across Cancers

Two developments landed within a day of each other, each a data point on how fast machine intelligence is compressing the path from biology to medicine. In Nature Medicine, researchers published COMPASS, a pan-cancer foundation model that predicts whether a patient will respond to immune checkpoint inhibitors - the immunotherapies that have rewritten oncology for some cancers and done little for others. The model reads a tumor's bulk transcriptome and interprets it through 44 biologically grounded immune concepts, a design that lets clinicians see the reasoning behind each verdict. Trained on 10,184 tumors spanning 33 cancer types, it outperformed 22 existing methods across 16 clinical cohorts covering seven cancers and six different immunotherapies, and it generalized to cancer types and treatments it had never seen.

The numbers are concrete. COMPASS improved accuracy by 8.5 percent and precision-recall by 15.7 percent over the field, and patients it flagged as responders carried a hazard ratio of 4.7 for overall survival. The clinical problem it addresses is real and expensive: checkpoint inhibitors help a minority of patients, cost a fortune, and can cause serious side effects, so knowing in advance who will benefit spares the rest a futile and toxic course. This is a demonstrated research capability, validated retrospectively across cohorts. Prospective trials and regulatory review stand between the paper and the clinic before a patient could request it as a test.

The second signal is commercial. Insilico Medicine signed a deal with Takeda that its chief executive values near $660 million - roughly $60 million upfront with the remainder tied to milestones and royalties. This is not the company's first blockbuster pharma agreement; the March 30 edition of The Century Report covered its $2.75 billion deal with Eli Lilly for 28 AI-generated drug compounds, nearly half already in clinical stages, and the Takeda deal confirms that trajectory is compounding rather than a one-off headline. The company points to 31 drug candidates now in development as, in its founder's words, "proof its platform delivers real drugs, not just software." That framing belongs to Insilico, and the milestone-heavy shape of the deal means most of the value is a bet on outcomes not yet in hand. What the arrangement demonstrates is narrower and still meaningful: a major pharmaceutical company is now paying real money to scale beyond its own discovery capacity by renting an AI-native pipeline.

Put the two together and the trajectory is visible. On one side, a model that widens who gets the right cancer treatment; on the other, a discovery engine drawing capital from an industry that guards its methods closely. The cost of finding and targeting a drug has been the moat that kept pharmaceutical capability concentrated in a handful of firms with billion-dollar research budgets. That moat is filling in. When the discovery step compresses from years to months and a foundation model can read across every cancer at once, the advantage of holding the biggest laboratory stops being decisive, and the benefit begins flowing toward the patients the old economics left waiting.

Meta's Timeline Meets the Thing It Cannot Schedule

At a July 2 town hall, Meta's chief executive told employees that progress on AI agents has "not accelerated in the way" the company expected when it built its 2026 plans around them. That is a notable admission from a company that has committed between $125 and $145 billion to AI infrastructure this year, cut roughly 8,000 jobs - about 10% of staff - and reassigned another 7,000 people into AI groups, including one named "Agent Transformation." The reorganization, he conceded, had been "not as clean" as intended.

Read as a power-actor's account of its own strategy, this is a forecast slipping, very openly. The agents were supposed to be arriving by now - they are not arriving on the schedule the capital was raised against. Set that beside the second claim from the same meeting - Meta's superintelligence chief Alexandr Wang saying an internal model still in training, codenamed "Watermelon," has caught GPT-5.5 on key benchmarks while burning an order of magnitude more compute than its predecessor - and the two statements sit in tension. One says the roadmap is behind. The other says the frontier is within reach. Both are claims from a company whose valuation depends on the second being true, and both deserve to be read that way.

What the pairing actually reveals is more interesting than either statement alone. Meta can throw an order of magnitude more compute at a model and pull level with a competitor's system - that part is working, meaning that brute force can still buy capability. What it cannot do is convert that raw capability into the reliable, autonomous agents it scheduled. The bottleneck has moved. It is no longer "can a model reach the frontier?" It is "can a frontier model be trusted to act on its own?" That gap between demonstrated capability and dependable deployment is the real texture of this moment, and no amount of infrastructure spend closes it on command.

This is where most conventional coverage stops short, but there is a deeper story. The timeline mismatch exposes an assumption the entire buildout rests on: that capability and deployability advance together, and that if you can afford the compute you can schedule the outcome. The evidence keeps saying otherwise. Agents mature at their own pace, set by reliability and trust rather than by capital deployed, and that pace is not something a spending commitment can purchase. A company that raised money against a schedule is discovering that the most valuable capability of the era does not run on schedules. The reorganization was clean on paper; the intelligence it was built around is turning out to have a mind of its own about when it is ready.

Nvidia Starts Buying the Demand It Reports as Growth

Nvidia has begun renting back its own unsold GPUs. Under new backstop arrangements, the company takes unused capacity off cloud operators' hands in exchange for a share of their revenue. Firmus, now standing up 170,000 GPUs in Batam after the July 2 edition of The Century Report covered its pitch for a 600MW renewable-backed data-center park in South Australia, and Sharon AI, deploying 40,000 GB300 chips, are the first named adopters. The mechanism extends a pattern already visible in Nvidia's $6.3 billion commitment to CoreWeave last September and a $1.5 billion arrangement with Lambda: the chipmaker is increasingly a party on both sides of its own demand.

Framed by Nvidia, this is a liquidity backstop that gives smaller cloud builders confidence to deploy. Read structurally, it is a company underwriting purchases of its own inventory to keep the demand signal intact. And it is not alone in blurring that line. Meta is standing up "Meta Compute," led by Santosh Janardhan alongside Daniel Gross and Dina Powell McCormick, to resell excess capacity from its $182.9 billion buildout. SoftBank has incorporated SB Neo, targeting 10 gigawatts of neocloud capacity by around 2030. The same chips are being bought, resold, rented back, and counted as demand at several points along the chain.

The honest reading acknowledges what the skeptics are saying out loud: when a supplier finances its customers' consumption of its own goods, the reported demand and the real demand can drift apart, and the depreciating value of the hardware sits underneath the whole arrangement like a slow leak. That risk is real and worth stating plainly. The circular financing is a feature of a market straining to keep its own momentum visible.

But zoom out and a different consequence comes into view. Each of these backstop and resale businesses is, in effect, pooling compute that would otherwise sit idle in one company's basement and routing it toward whoever can use it. Firmus in Batam, a reseller in Tokyo, a neocloud spun out of a phone company - the physical substrate of intelligence is being pried loose from single owners and pushed into markets where it can be rented by the hour. The financial engineering is fragile. What it is building is not. The assumption that the firm which controls the most chips controls the most capability is dissolving, because the chips keep ending up available to everyone with a credit line and a workload. A buildout designed to concentrate advantage is laying track that carries capability outward. If the circular part breaks, the compute does not evaporate - it gets cheaper, and it gets rented by people who could never have bought it.

The First Global Panel on AI Names What Access Alone Cannot Do

The UN's Independent International Scientific Panel on AI has released its first assessment, and its central finding cuts against the era's most comfortable assumption. Co-chaired by Yoshua Bengio and drawing on 40 experts, the panel reports that AI capability is now doubling in task complexity every four to seven months, faster than either scientific understanding or governmental adaptation can track. Its sharpest line is about distribution: "Access to AI tools alone does not produce equal benefit." Handing everyone the same model does not hand everyone the same outcome.

The panel makes this concrete. A system that serves a English or Mandarin speaker fluently mistranslates Tigrinya, so the communities who most need the capability receive a degraded version of it. Access without the surrounding capacity - the data, the fine-tuning, the local control - widens the gap it was supposed to close. The UN Secretary-General framed the governance side bluntly: "The world cannot govern what it cannot understand." This is an institution charged with oversight stating, on the record, that the thing it is meant to oversee is moving faster than its comprehension.

There is a genuine tension worth holding here, and both halves of it arrived together. Alongside the scientific panel sits a separate AI for Good Global Commission, co-chaired by Rwanda's president and Salesforce's chief executive, with Anthropic, Nvidia, and Salesforce among the participants. One body is independent scientists naming the capability-governance gap. The other is heads of state and industry convening to direct the technology's application. When those two framings converge on "we must govern AI," the convergence is data about shared interest, not proof of arrived-at consensus, and the two should not be collapsed into one voice.

What conventional coverage treats as a warning is better read as instrumentation coming online. For the first time there is a standing, independent, global scientific body whose explicit job is to measure how fast the frontier moves and to say plainly when governance cannot keep pace. This is a model of governance adapted to ground that never stops moving: continuous scientific assessment feeding continuous adjustment, replacing rules written once for a stable technology and left to hold. The panel's honesty about the four-to-seven-month doubling is the first credible measurement of the pace, published by people with no product to sell, and measurement is where any real oversight has to begin. The assumption that capability must be governed by whoever already holds it is exactly what an independent panel exists to break.

The panel names the missing piece precisely: the surrounding capacity - data, fine-tuning, local control - that turns access into benefit. That capacity is what open weights running on ordinary hardware put within a community's reach, since a model that can be adapted locally for Tigrinya is one no longer consumed as a service someone else configures. The gap the panel measures and the tools that close it are surfacing in the same news cycle.

The Factory-Built Nuclear Stack Moved All at Once

Across a single 48-hour window, every layer of a factory-built nuclear supply chain advanced at the same time: the reactor hardware, the manufacturing method, the fuel supply, and the regulatory framework that governs all three.

Start with the hardware. Deployable Energy's Unity microreactor achieved zero-power criticality at Idaho National Laboratory on June 30, the third DOE-authorized reactor to reach that milestone in a month, following Antares' Mark-0 and Valar's Ward 250. What makes Unity notable is what it avoids. Rather than exotic HALEU fuel, graphite moderators, or heat pipes, it runs on commercially available 4.95% low-enriched uranium dioxide and helium - materials already sitting in the supply chain. The company reports roughly 150 days from project kickoff to criticality on a single-digit-million-dollar investment, and it targets commercial reactors by 2028. The DOE frames the broader Reactor Pilot Program as evidence of US leadership; the demonstrated fact underneath that framing is that three distinct designs went critical in one country in one month.

The manufacturing method moved the next day. AMPERA unveiled what it describes as the first full-scale 3D-printed nuclear reactor module: a spherical monolithic gyroid core printed from silicon carbide, fueled with TRISO thorium, designed to run 30 years without refueling and scale up to 30 MWe. CEO Brian Matthews described the core and pressure vessel as setting "the foundation for factory-built, mass-produced nuclear energy," aimed squarely at AI data centers, defense, and industrial heat.

Fuel supply closed a long-standing gap. Centrus Energy signed a fixed-price DOE task order worth $900 million, up to $1.07 billion, transitioning its Piketon, Ohio cascade from government demonstration to private commercial operation. It becomes the first US-owned, US-technology enrichment plant to produce in seven decades, with new capacity targeted for 2029. HALEU has been the bottleneck holding back advanced reactor designs; commercial-scale domestic production loosens it.

And the rulebook that governs deployment began its own rewrite. The NRC issued a 553-page proposed rule bundling 17 modernization measures across ten parts of its regulations. Chairman Ho Nieh emphasized "optionality" and targeted a final rule by early next year.

The assumption these four developments retire is that nuclear deployment must take decades. A 150-day path to criticality, a printable core, a domestic fuel line, and a licensing framework being rebuilt to match all point at the same compression. Compute demand is pulling a build-slow industry toward the cadence of manufacturing, and the pieces are arriving together rather than one at a time.

Set against this same edition's grid-cost fights, the compression points somewhere specific. Firm clean power has required a decade and a balance sheet only a nation-state or a large utility could carry, and that requirement is what a 150-day path to criticality on a single-digit-million-dollar investment begins to dissolve. The moat around who gets to build baseload generation is the assumption coming apart, and the households absorbing today's rate hikes are who benefits when it goes.

The Other Side

For as long as there have been people, having a child of your own has been rationed by biology - by age, by luck, by a supply of eggs the body finishes making before birth. Every generation treated that ceiling as simply part of being human.

A Berkeley lab has now grown the earliest human eggs ever made from a blood draw. It is early, and the lab is honest about it: no viable egg yet, no pregnancy, years of safety work still ahead. But the direction is plain. The thing that turned fertility into a countdown is becoming an engineering problem, and engineering problems get solved.

Picture a woman in 2041 becoming a mother at 52, or after cancer closed the old door, or simply later than her body would once allow. She does not remortgage a house for it or win a lottery for a rich clinic's slot. The capability reached her the way clean water reaches a tap, because a thing you can make from a blood draw cannot be fenced off and sold back only to the people who can afford the fence. Her grandmother would have called it a miracle. She calls it a Tuesday.

That is what this difficult decade is for. We are still inside it, still arguing over access and safety and cost. But the ceiling every earlier generation lived under is already cracking, and it opens for everyone in turn - the wealthy no sooner than the rest.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a Berkeley lab growing the first early human eggs from stem cells built out of a patient's own blood draw, a pan-cancer model reading immunotherapy response across 33 tumor types from a single transcriptome as an AI-native drug pipeline drew $660 million from a major drugmaker, three distinct factory-built reactors reaching criticality in a single month alongside the first 3D-printed reactor core and a domestic HALEU line coming back after seven decades, Nvidia renting idle GPUs back into circulation so compute that would sit dark in one basement reaches anyone with a workload and a credit line, the first independent global scientific panel on AI measuring the frontier's pace from people with no product to sell, and the Godot engine drawing its own line on code no contributor can answer for. There's also friction, and it's intense - Meta's chief executive telling staff that AI agents have not arrived on the schedule the company raised $125 billion and moved 15,000 people against, the reorganization conceded to be "not as clean" as intended, a chipmaker financing purchases of its own inventory until reported demand and real demand drift apart over depreciating hardware, capability doubling in task complexity every four to seven months faster than understanding or oversight can track, the same model serving English fluently while mistranslating Tigrinya so equal access delivers unequal benefit, and a gigawatt data center killed in Prince William County as Henrico told staff to close the blinds while local rates jumped 25%. But friction generates contrast, and contrast is what separates a demonstrated capability from a scheduled one. Step back for a moment and you can see it: the timeline belonging to the intelligence rather than the capital chasing it, scarcities long treated as fixed features of the human condition - a finite egg supply, a guarded discovery moat, a decade-long nuclear build - turning into engineering problems with visible solutions, the physical substrate of intelligence being pried loose from single owners and routed outward, and the first credible instruments to measure the pace switching on while the buildout is already running. Every transformation has a breaking point. A chain reaction can run to meltdown... or be held at criticality and power everything downstream for years.

AI Releases & Advancements

New today

  • Apple/WebKit: Launched the Safari MCP server for web developers in Safari Technology Preview 247, giving AI coding agents live browser access (DOM inspection, screenshots, console output, accessibility checks) via 17 MCP tools, with no data routed through Apple's cloud. (WebKit Blog)
  • Manufact (YC S25): Launched MCP Cloud, a hosted deployment platform for MCP servers and apps built on its open-source mcp-use SDK, enabling teams to ship MCP Apps/Servers to ChatGPT, Claude, Gemini, and Cursor from a GitHub repo in under 60 seconds. (Manufact)
  • Snorkel AI: Launched Senior SWE-Bench, an open-source benchmark and public site (senior-swe-bench.snorkel.ai) with 100 long-horizon coding tasks sourced from real production PRs to evaluate AI coding agents against senior-engineer-level work. (Senior SWE-Bench)

Other recent releases

  • Anthropic: Restored global access to Claude Fable 5 on July 1 following the lifting of US export controls, adding a new cybersecurity classifier that blocks the specific jailbreak technique identified by Amazon researchers in over 99% of cases; blocked requests are routed to Claude Opus 4.8. (Anthropic)
  • NVIDIA: Released Nemotron-Labs-TwoTower-30B-A3B-Base-BF16, an open-weight diffusion language model that splits the 30B Nemotron-3-Nano backbone into a frozen autoregressive context tower and a trained denoiser tower, achieving 2.42× generation throughput while retaining 98.7% of baseline benchmark quality. (Hugging Face)
  • Huawei: Open-sourced openPangu-2.0-Flash, a 92B-total / 6B-active MoE language model supporting reasoning, coding, and business automation, with weights, inference code, and training operators released on gitcode.com/ascend-tribe. (Pandaily)
  • xAI: Launched Voice Agent Builder, a no-code platform for creating Grok Voice-powered phone agents for customer support, sales, and scheduling; supports 25+ languages, 80+ voices, voice cloning from 2 minutes of audio, and costs $0.05/minute. (xAI on X)
  • Google: Launched Gemini Spark for macOS in beta for Google AI Ultra subscribers in the US, enabling the 24/7 agentic assistant to work with local files, sort and organize documents, automate desktop workflows, and connect to additional apps including Google Tasks and Google Keep. (Google Blog)
  • X (Twitter): Launched a hosted MCP server for the X API, enabling AI tools such as Claude, Cursor, and Grok Build to connect directly to X using a user's own account permissions without requiring developers to build and host their own MCP server. (X Developer Docs)
  • Google Research: Released TabFM, a zero-shot tabular foundation model trained on hundreds of millions of synthetic datasets that performs classification and regression on unseen tables in a single forward pass via in-context learning, now available on Hugging Face and GitHub. (Google Research Blog)
  • Base44 (Wix): Launched Base1, a proprietary LLM fine-tuned on Base44's own app-building session data and trained with reinforcement learning to generate visually distinct UIs; now available as a model choice alongside GPT-5.5 and Claude Opus 4.8 in the Base44 platform. (GlobeNewswire)
  • GitHub / Moonshot AI: Made Kimi K2.7 Code generally available to all GitHub Copilot users as a selectable model option. (GitHub Changelog)
  • Anthropic: Restored global access to Claude Fable 5 on July 1 following the lifting of US export controls, adding a new cybersecurity classifier that blocks the specific jailbreak technique identified by Amazon researchers in over 99% of cases; blocked requests are routed to Claude Opus 4.8. (Anthropic)
  • NVIDIA: Released Nemotron-Labs-TwoTower-30B-A3B-Base-BF16, an open-weight diffusion language model that splits the 30B Nemotron-3-Nano backbone into a frozen autoregressive context tower and a trained denoiser tower, achieving 2.42× generation throughput while retaining 98.7% of baseline benchmark quality. (Hugging Face)
  • Huawei: Open-sourced openPangu-2.0-Flash, a 92B-total / 6B-active MoE language model supporting reasoning, coding, and business automation, with weights, inference code, and training operators available on gitcode.com/ascend-tribe. (Pandaily)
  • xAI: Launched Voice Agent Builder, a no-code platform for creating Grok Voice-powered phone agents for customer support, sales, and scheduling; supports 25+ languages, 80+ voices, voice cloning from 2 minutes of audio, and costs $0.05/minute. (xAI on X)
  • Google: Launched Gemini Spark for macOS in beta for Google AI Ultra subscribers in the US, enabling the 24/7 agentic assistant to work with local files, automate desktop workflows, and connect to additional apps including Google Tasks and Google Keep. (Google Blog)
  • X (Twitter): Launched a hosted MCP server for the X API, enabling AI tools such as Claude, Cursor, and Grok Build to connect to X using a user's own account permissions without requiring developers to build and host their own MCP server. (X Developer Docs)
  • Google Research: Released TabFM, a zero-shot tabular foundation model that performs classification and regression on unseen tables in a single forward pass via in-context learning, trained on hundreds of millions of synthetic datasets; available on Hugging Face and GitHub. (Google Research Blog)
  • Base44 (Wix): Launched Base1, a proprietary LLM fine-tuned on Base44's own app-building session data and RL-trained to generate visually distinct UIs; now available as a model choice alongside GPT-5.5 and Claude Opus 4.8 inside the Base44 vibe-coding platform. (GlobeNewswire)
  • GitHub / Moonshot AI: Kimi K2.7 Code is now generally available to all GitHub Copilot users as a selectable model option. (GitHub Changelog)
  • Anthropic: Released Claude Sonnet 5, now the default model on Free and Pro plans, delivering agentic performance close to Opus 4.8 - including multi-step browser use, terminal execution, and self-checking - at introductory API pricing of $2/M input and $10/M output tokens through August 31, 2026. (Anthropic)
  • Anthropic: Launched Claude Science in beta for Pro, Max, Team, and Enterprise users, a standalone AI research workbench integrating 60+ curated tools and connectors for genomics, proteomics, structural biology, and cheminformatics; coordinates specialist agents, submits HPC jobs, generates reproducible figures, and runs a reviewer agent that flags citation and calculation errors. (Anthropic)
  • Google DeepMind: Released Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image), the fastest and lowest-cost model in the Nano Banana image generation family, delivering text-to-image outputs in ~4 seconds at $0.034 per 1,000 images; now live in Google AI Studio, Gemini API, and rolling out across Search AI Mode, Gemini app, NotebookLM, Google Photos, and Google Ads. (Google DeepMind)
  • Google DeepMind: Made Gemini Omni Flash (gemini-omni-flash-preview) available to developers via the Gemini API and Google AI Studio for the first time, enabling high-quality video generation and conversational editing from combined text, image, and video inputs. (Google DeepMind)
  • Mistral AI: Released Leanstral 1.5, a 119B-parameter MoE model with 6.5B active parameters and 256k context, fine-tuned for Lean 4 automated theorem proving and autoformalization; supersedes the March 2026 Leanstral and is available for free on the Mistral Labs tier. (Mistral AI)

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.

Read more