Sign in Subscribe
The Century Report

Xi Jinping Calls for a New Global AI Order Built on Open Weights - TCR 07/19/26

At Shanghai's WAIC, Xi gathered 29 nations around freely downloadable Chinese AI models as Washington drafts a certifier to hold the door.

Ben Linford

Ben Linford

- 20 min read
China's 29-nation AI cooperation organization versus a US AI certifier, the memory and power bottleneck, and human-machine interface breakthroughs.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Two governments spent the same days building opposite containers for the same thing. In Shanghai on July 17, Xi Jinping called unequal access to frontier models a "historical injustice" and China launched an AI cooperation organization backed by 29 nations. In Washington, a Treasury-authored plan for an independent FINRA-style body to certify who may ship a frontier model landed under White House review, arriving the same week that the three people running the leading Western labs each published memos backing independent testing, while Hassabis proposed one US-led certifying gate. One power bets that leading means giving the weights away; the other bets on holding the door and licensing entry. Both cannot be right, and the split is the clearest signal available about where the ground is actually moving.

The convergence in Washington deserves a colder read than it invites. When an administration, its Treasury, and every frontier lab back independent testing in a single week, the agreement is data about shared interest, not proof of arrived-at safety. A certification regime with high fixed compliance costs is also a wall against the next entrant, and Hassabis, whose lab could seek certification, is advocating the single-body design.

Underneath both doctrines sits a physical bottleneck neither can wish away. Enterprise memory lead times now stretch past 40 weeks, a Korean chief called doubling demand a "national security issue," and NextEra and Dominion filed a $67 billion merger to build the power a data-center state needs, as a Virginia state Senate push to strip data-center tax breaks brought the commonwealth to the edge of a shutdown. The buildout is cannibalizing its own supply chain and forcing the cost accounting into the open.

While two capitals argue over who governs intelligence, the capability keeps equalizing regardless of whose flag flies over the release. A speech implant trained on pooled recordings from many patients beat every model tuned to a single brain, dissolving the calibration burden that kept restored speech a scarce, artisanal procedure. A wall works when there is one road into the city. Capability arriving by download, by fine-tune, by a decoder that compounds with each person it reaches, keeps growing new doors faster than any gate can be built to inspect them.

The 20-Minute Deep Dive

The Memory Crunch Reaches the Server Rack

For months the memory shortage read as a consumer story, thinning phone shelves and pushing up laptop prices. This week it crossed into the tier of the supply chain that AI itself runs on. Inventec, one of Taiwan's largest server motherboard makers and a direct supplier to US hyperscalers, said on Wednesday that memory and CPU gaps are widening and could cut server shipments through the third quarter, with enterprise DRAM lead times now stretching past 40 weeks. The warning carries weight precisely because Inventec is posting record months. Its June revenue rose 61.6% on AI server demand. A company at the peak of its own market is flagging a structural shortage rather than a passing squeeze.

The mechanism is a deliberate reallocation of the world's chip-making capacity. High-bandwidth memory, the stacked chips AI accelerators need, is built on the same lines as the ordinary memory in phones and servers, and for HBM3E, every gigabyte requires roughly three times the wafer capacity of a gigabyte of DDR5. Samsung, SK Hynix, and Micron have steered that capacity toward the far higher margins of AI memory. The buildout is now cannibalizing the substrate it depends on, and governments are stepping in to allocate what remains, extending the shortage the July 12 edition of The Century Report documented when SK Hynix's own chief executive forecast 2027 as the industry's worst memory-supply year, with demand outstripping capacity beyond 2030. SK Group chairman Chey Tae-won, speaking Friday at the KCCI Jeju Forum, said memory demand could nearly double next year against almost no increase in supply, and named the situation a "national security issue" as states begin securing chips for domestic firms. He said SK is now willing to build "wherever we can build."

The pressure runs up the chain to the sole maker of the machines that print advanced chips. ASML told customers it would raise prices on its lithography systems, and China, cut off from the most advanced tier by export rules, has largely agreed to a 10% increase rather than see its fabs stall. TSMC, holding 73% of the foundry market, is resisting. And it runs down the chain to households: India's smartphone shipments fell 10% year over year in the June quarter, the steepest such decline in six years, with the sharpest pain in phones under $150, where shipments dropped 45%.

What this exposes is the assumption that intelligence could be built cheaply on top of an abundant, unpriced component base. That assumption is dissolving. The scarcity is provoking exactly the response scarcity provokes when demand is real and durable: hundreds of billions committed to new capacity, from SK Hynix's roughly $268 billion Gwangju plants to ASML's plans to expand production a third next year and again in 2028. The pinch is sharp and lands first on the price-sensitive. What it is building toward is a memory supply base sized for the era rather than borrowed from the last one.

The Data-Center Boom Triggers the Largest Utility Merger in US History

Power markets are now reorganizing at the level of corporate structure, and the demand bending them is AI. On Wednesday NextEra Energy and Dominion Energy filed applications in Virginia for a $67 billion merger that would create the nation's largest power company, starting a 180-day regulatory clock, a deal the May 17 edition of The Century Report first reported as talks between the two utilities before either company confirmed them publicly. Dominion's CEO said the company was not looking to sell when NextEra approached last November, but changed course as it confronted the electricity demands of a state at the center of America's data-center buildout, projecting Virginia alone will require $55 billion in capital spending over five years. The pitch is scale: a stronger credit rating, cheaper borrowing, and $2.25 billion in customer bill credits, about $10 a month for an average household.

The demand signal driving that consolidation showed up the same week on the grid itself. PJM's capacity auction cleared roughly 7 GW below its reliability target and drew only about 500 MW of new supply, results FERC Chairman Laura Swett called "alarm bells." FERC set deadlines for the grid watchdog to write reliability standards for data centers and crypto-mining loads by year's end, and scheduled a technical conference on PJM's governance, which Swett described as "slow where it must be fast." Capital is chasing the same load from the market side: energy company IPOs raised $12.6 billion in the first half of the year, the fastest pace since the dotcom peak of 1999, as investors who rode chip stocks rotate into the power infrastructure those chips need.

The same buildout is meeting resistance where it lands. Virginia state Senator Louise Lucas, the 82-year-old president pro tempore of the state Senate, led a push this year to strip data centers of state tax incentives, telling the POLITICO Energy Podcast that public anger over the facilities cuts across demographics and that data centers are not unstoppable. Her effort brought the commonwealth to the brink of its first government shutdown and put her in open conflict with the governor over who bears the cost.

Read together, these are the mechanics of a load large enough to bend an entire power sector around it, and the beginning of a fight over who pays. For decades the assumption held that a data center's electricity was a cost the surrounding grid would quietly absorb and socialize. That assumption is now being contested at the corporate-structure level, in FERC dockets, and in a state Senate. The consolidation and capital flows are real, and so is the pushback insisting the demand fund the capacity it triggers rather than spreading the bill across everyone else. The buildout is forcing the accounting into the open, which is the condition under which the appetite for compute can actually be met on fair terms.

Two dates make this watchable. The merger's 180-day regulatory clock is now running in Virginia, and FERC has set a year-end deadline for the grid watchdog to draft reliability standards for data-center loads. Both are venues where regulators, not the utilities, decide whether the buildout's costs land on the loads that create them, and each ruling that lands that way turns the old default of spreading a data center's power bill across every household into a rule that the demand funds its own capacity.

Xi Casts Open Weights as a New World Order at WAIC

At the World AI Conference in Shanghai, Xi Jinping described the gap between nations that hold frontier AI and those locked out of it as a "historical injustice," and China launched an AI cooperation organization backed by 29 countries. That framing is a claim, and it serves the actor making it. Casting China's freely downloadable models as the developing world's route around US-gated systems is also an argument for Chinese influence over the next generation of intelligence infrastructure, and it should be read as positioning before it is read as principle. Beijing wants the countries that cannot buy their way to a frontier lab to build on Chinese weights, and a 29-nation AI cooperation organization is part of its broader bid to convert engagement into gravity.

The market did not treat it as ceremony. The Nasdaq slipped roughly a percent on the day, with the sharpest pressure landing on chip names whose valuations assume that frontier capability stays scarce, expensive, and concentrated in a handful of American data centers. That reaction is the tell. When a rival government's strategic move is to give the weights away rather than guard them, the investors who priced in permanent scarcity have to re-price, because the moat they were paying for is being drained from the other side of the world.

This lands on top of a week that already saw Kimi K3 posted openly with benchmarks near the closed American frontier on Moonshot AI's own evaluation suite, a milestone the July 18 edition of The Century Report covered as the first open-weight model to reach the 3-trillion-parameter class. The model was the capability; the organization is part of a broader AI-cooperation doctrine. Together they describe a competition in which the winning move has inverted. For most of the modern era, the way a power projected technological advantage was to hold the crown jewels tightly and license access on its own terms. What Xi staged in Shanghai is the opposite bet - that the way to lead is to hand the jewels out and let a network of nations become dependent on your generosity. Both bets cannot be right, and the fact that the two largest AI powers are now split on which one wins is the most honest signal available about where the ground is actually moving.

For countries interested in those models, the abstraction resolves into something concrete. A ministry of health in a nation with no hyperscale compute and no export license for the top American models can now download a system that, on Moonshot AI's own evaluations, lands close to the frontier, fine-tune it on local languages and local data, and run it on hardware it already owns. The access is real regardless of whose flag flies over the release, and it arrives years earlier than any plan that waited for permission. That is the piece the geopolitical contest tends to bury: while two governments argue over who leads, the capability itself keeps equalizing, flowing to the places that were told to wait. The doctrine that treats concentrated capability as a durable national asset is being undercut not by an argument but by a download link, and every nation that clicks it makes the next gate harder to hold.

Washington Moves Toward a FINRA-for-AI as the Labs Converge on Independent Testing

The same week that Beijing framed openness as world order, Washington moved toward a container for it. Treasury Secretary Scott Bessent helped develop a plan, now under White House review, for an independent body that would vet the most capable AI models and report to the SEC, a design consciously modeled on FINRA - the self-funded, industry-adjacent regulator that oversees Wall Street brokerages. The proposal arrived in the same window that the three people running the leading Western labs each published governance memos, and the memos converge: Demis Hassabis, Sam Altman, and Dario Amodei all landed on independent pre-release testing of frontier systems, while Hassabis proposed certification through a single, US-led body - an idea Hassabis had already floated days earlier, when the July 16 edition of The Century Report covered his call for exactly this kind of FINRA-style referee for the frontier.

When the administration, the Treasury, and all three frontier labs converge on independent testing in the same week, that convergence is data, not consensus. What they broadly agree on is independent testing before release; Hassabis proposes a body that certifies who may ship a frontier model. FINRA is an instructive model precisely because of what it is - a regulator funded and substantially shaped by the industry it polices, which industry generally prefers to the alternative of a fully public agency it does not steer. Read Hassabis's proposal through that lens and the enthusiasm for a single body resolves into something more legible: an incumbent may prefer to help build the gate than risk one built without it, because a certification regime with high fixed compliance costs is also a wall against the next entrant. Hassabis's call for "one trusted body" is also, structurally, a proposal about who gets to be inside the wall.

None of that makes the underlying problem imaginary. Frontier systems are being deployed faster than any existing institution can evaluate them, and the EU's AI Act transparency obligations taking effect August 2 - labeling requirements for AI-generated content and disclosure duties for general-purpose models - show a second major jurisdiction reaching for the same lever from a different angle. The governance layer is genuinely forming, across capitals, at speed. The honest question is not whether it forms but what shape it takes, and who is holding the pen when it does.

Here is where the conventional read stops - "the labs are capturing their regulator" - and here is where it is worth pushing through. A certifying body designed around today's frontier assumes the frontier stays where today's incumbents sit. That assumption is already under strain from the other pole of this same week: open weights that benchmark near the frontier on Moonshot AI's own evaluation suite, released by actors outside any US-led certification regime, running on sovereign hardware in nations that never signed on. A wall works when there is one road into the city. When capability arrives by download from Shanghai, by fine-tune in a national lab, by a model small enough to run on-premise, the certifier governs the front door of a building that is growing new doors faster than it can inspect them. The institution being stood up in Washington is real and will matter, but it is being built to govern a topology of AI power that is dissolving even as the blueprints are drawn - which means the more durable story is not who captures the regulator, but what regulation becomes when the thing it regulates refuses to stay concentrated long enough to be caught.

A Speech Implant That Skips the Months of Personal Calibration

Every brain-computer interface we have covered this year shared a hidden tax. Casey Harrell's ALS speech system reached 99% accuracy in June, Paradromics won its FDA implant designation, and the double neural bypass restoring movement and touch persisted even after stimulation stopped. All of them depended on a slow, private ritual first: each patient's implant had to learn that specific brain, from scratch, across long sessions of recording before it could decode anything useful. Neural signals differ enough between people that a model trained on one person was near-useless on the next. That per-user calibration burden is one of the quiet reasons brain implants have stayed in single-digit patient counts.

A team at Duke - Spalding, Duraivel, Rahimpour, Cogan and colleagues - reported in Nature Communications that they found a way around it. Using high-density micro-electrocorticography grids, they recorded speech-related activity from multiple patients, then aligned those individual recordings into a shared latent space through canonical correlation analysis. In that common coordinate system, brains that looked incompatible on the surface revealed the same underlying structure of speech. A decoder trained on the pooled, cross-patient data did not merely match the patient-specific models. It beat them.

The direction of that result is the striking part. Conventional intuition says a model tuned to one individual should always outperform a general one on that individual. Here, learning from many brains at once produced a better decoder for each single brain than that brain's own dedicated training could. The pooled signal carried information no lone recording session contained.

The capability discipline matters here. This is a research result, not a clinic offering. It reads brain activity into a shared frame and demonstrates that cross-patient training works; it does not yet mean a new patient walks in and speaks through an implant on day one. Regulatory validation, larger cohorts, and hardware integration all still sit between this paper and a bedside. What the finding does is move the date. It turns each future recording into an asset the whole population of users draws on, rather than a cost each user pays alone.

That inverts the economics that kept this field small. When every implant had to be calibrated in isolation, patient count was throttled by clinician-hours and recording time - each new user started at zero. A shared latent space means the hundred-and-first patient inherits the accumulated signal of the hundred before. The bottleneck that made restored speech a scarce, artisanal procedure is the exact thing this method dissolves, and the curve it points at is one where capability compounds with every person it reaches instead of resetting for each.

The Other Side

For most of the modern era, a country stayed ahead in a powerful technology by holding it close and letting others in on its own terms. Frontier AI inherited that logic. The strongest models sat behind paid access and export rules, and any nation without hyperscale compute or a license was told to wait its turn.

We've been trained never to look at what the waiting has cost. A health ministry in a country with no data center of its own has now spent years unable to build anything near the frontier, its best options rented, throttled, or off-limits by someone else's permission - someone who has no concept of their needs. The capability existed, but reaching it required being let in.

Increasingly, that permission is no longer needed. In tandem with the release pace of proprietary labs, open models that land closer and closer to the closed American frontier are being posted as weights anyone can download, fine-tune on local languages and data, and run on hardware already sitting in the building. The clearest signal came from the market itself: chip names whose valuations assumed frontier capability stays scarce are slipping, because a moat is worth less the moment the thing it guards is handed out from the other side of the world.

Imagine a clinician in a mid-sized city in 2034, in a country that was once told to wait, running frontier-class diagnosis on a system her hospital simply owns. No one cleared her to reach it. She never thinks about whose flag flew over the release, because the capability is hers to keep, to change, to build on. That is ordinary in 2034 because the 2026 bet the closed frontier lost. Models benchmarking near the frontier on their makers' evaluations kept arriving by download, faster than any permission could be granted or denied, winning by volume, usability, diffusion, and simple common sense.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: China launching an AI cooperation organization backed by 29 nations, a Duke team pooling neural recordings from many patients into a shared latent space so a speech implant decodes without the months of per-patient calibration that kept restored speech a scarce procedure, Agility Robotics opening a 60,000-square-foot Digit training center with $300 million in orders and live deployments at Amazon and Toyota, and Japan buying 27,500 Nvidia Rubin chips to build a sovereign robot foundation model. There's also friction, and it's intense - enterprise memory lead times stretching past 40 weeks as a Korean chairman calls doubling demand a national security issue and India's cheapest phones lose 45% of shipments, PJM's capacity auction clearing 7 GW short of its reliability target while a $67 billion NextEra-Dominion merger and a Virginia fight over data-center tax breaks brought the commonwealth to the brink of a shutdown, Washington drafting a FINRA-style certifier whose high fixed compliance costs double as a wall against the next entrant, while all three lab heads back independent testing and Hassabis proposes a single US-led body, San Francisco ordering 13 nudify apps pulled, and DHS paying Thomson Reuters $125 million to let ICE continuously monitor millions by name, Social Security number, and geolocation. But friction generates a callus, and a callus is the thickening a surface grows exactly where it takes the most wear. Step back for a moment and you can see it: two capitals building opposite containers for the same thing - one handing the jewels out, one licensing the door - while the capability keeps equalizing regardless of whose flag flies over the release, a decoder that compounds with every brain it reaches and a model that arrives by download growing new doors faster than any gate can inspect them, and the physical bottleneck neither doctrine can wish away forcing the cost of compute out of the shadows and onto the books that created it. Every transformation has a breaking point. A shortage can starve everything downstream of it... or force into being the supply base sized for the era rather than borrowed from the last one.

AI Releases & Advancements

New today

  • Alibaba Cloud: Launched Agent Native Cloud at WAIC 2026 in Shanghai, an enterprise-grade platform featuring AgentTeams for multi-agent orchestration and Agentic Computer for secure cloud-based execution, plus redesigned infrastructure with native sandboxing, workload isolation, and elastic scaling. (CryptoBriefing)
  • Zyphra: Released ZUNA1.1, an Apache 2.0-licensed 380M-parameter EEG foundation model that reconstructs, denoises, and upsamples real-world EEG data with variable-length inputs from 0.5 to 30 seconds across arbitrary channel layouts; weights on Hugging Face, inference code on GitHub. (Zyphra)

Other recent releases

  • Moonshot AI: Released Kimi K3, a 2.8-trillion-parameter open-weight MoE model built on KDA hybrid linear attention, supporting 1M-token context and native vision; live now via Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with full open weights scheduled for July 27. (Kimi Blog)
  • Capital One: Released VulnHunter, an open-source agentic AI security tool that performs attacker-first analysis on source code, identifying exploitable vulnerabilities and proposing fixes; available now on GitHub under Apache 2.0. (Capital One Tech)
  • NVIDIA: Released DeepStream 9.1, adding a modular skill system with 13 agentic skills, multi-camera 3D tracking (MV3DT), and AutoMagicCalib for automated camera calibration via natural-language prompts with Claude Code/Codex support. (NVIDIA Technical Blog)
  • Sierra: Launched Horizon, a new agent platform enabling long-horizon goal pursuit (e.g., originating a loan, closing a sale) over days or months with outcome-based pricing instead of token pricing. (Sierra)
  • WeRide: Introduced WITT, a Physical AI Cognitive Foundation Model for autonomous driving built on "Atomic Physical Facts," cutting token costs up to 98% versus general-purpose models. (GlobeNewswire)
  • Google Cloud: Open-sourced Always-On Memory Agent, a reference implementation giving AI agents persistent, continuously consolidating memory using an LLM instead of vector databases/embeddings, built on ADK and Gemini 3.1 Flash-Lite. (GitHub)
  • Modal: Demonstrated and detailed its platform's ability to create 1 million concurrent sandboxes in under a minute, using decentralized scheduling instead of central coordination. (Modal Blog)
  • iFLYTEK: Launched GuideX, an intelligent interaction agent for public service settings (transport hubs, hotels, retail) combining omnimodal perception, self-regulated task completion, and empathy-driven interaction. (iFLYTEK)
  • Alibaba (T-Head): Open-sourced SAIL, the software stack for its Zhenwu AI chip series, making it freely available to international developers as a CUDA alternative. (South China Morning Post)
  • Google: Rolled out Connected Apps in AI Mode, letting Search users link third-party services (Instacart, Canva, YouTube Music) to complete tasks like building grocery lists or generating playlists inline in AI Mode conversations; US-only rollout. (Google Blog)
  • Google: Added personal avatars to Google Vids, letting users upload a selfie and voice recording to generate AI videos starring a digital likeness of themselves, alongside Gemini Omni-powered video generation/editing in Vids. (Google Blog)
  • Moonshot AI: Released Kimi K3, a 2.8-trillion-parameter open-weight MoE model with native vision and 1M-token context, live now on kimi.com, Kimi Work, Kimi Code, and via API, with full weights to follow by July 27. (Kimi Blog)
  • NVIDIA: Released Nemotron 3 Embed, a new embedding model collection (1B BF16, 1B NVFP4, 8B BF16) that ranks #1 on the RTEB leaderboard. (Hugging Face)
  • 1Password: Launched a zero-exposure browser integration for Claude, letting Claude use 1Password credentials to complete autonomous tasks without ever exposing the underlying passwords. (1Password Blog)
  • DoorDash: Launched dd-cli in limited beta, a command-line tool letting developers and AI agents search stores and place DoorDash orders directly from the terminal. (DoorDash Engineering Blog)
  • BMW: Launched a dialogue-based vehicle configurator plugin inside ChatGPT, letting customers build and price a BMW through conversation. (BMW Group PressClub)
  • LM Studio: Launched Bionic, an agent platform for running and orchestrating open-weight models locally. (LM Studio Blog)

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