Sign in Subscribe
Century Report

China Builds the World's Fastest Supercomputer - TCR 06/24/26

China's LineShine became the world's fastest supercomputer on conventional CPUs, around the chip controls built to contain it, as compute capability finds its level.

Ben Linford

Ben Linford

— 18 min read
Three-panel infographic: AI workforce ledger with Oracle layoffs and pushback; the data-center energy bill and cheaper grid paths; China's CPU-only supercomputer tops global ranking.

Video content will return next week.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

A single Oracle filing this week did something the buildout has spent years avoiding: it itemized its own costs out loud. In one audited document the company attributed 21,000 job cuts to AI adoption and announced plans to borrow as much as $50 billion to build more AI cloud capacity, naming the same technology as both the reason to shed workers and the reason to spend. The growth engine and the displacement appear as one force on one page. That legibility is the thread running through today's stories, because a cost written down can be measured, contested, and answered, where the same cost absorbed without a record cannot.

The energy ledger is filling in the same way. A federal utility floated up to 26 GW of new gas and Virginia levied a first-of-its-kind $600 million tax on the facilities driving the demand, even as the coal plants ordered to stand by for an emergency ran two-thirds below last year and analysts laid out a cheaper route through grid-enhancing tech the build-more habit kept invisible. The pressure of the compute buildout is dragging latent capacity, and the question of who pays for new capacity, into public view.

Sovereignty is moving on a parallel track. A Chinese machine reached the top of the global supercomputing ranking on conventional CPUs, around the silicon chokepoint export controls were built to defend, answering the premise that top-tier compute requires leading-edge Western chips. Capability of this kind diffuses the way water finds its level.

The human side of the ledger is being written from below. A new framing names the reverse centaur, the worker conscripted to babysit a machine, as a managerial choice rather than a property of the technology, which restores the possibility of choosing otherwise. Nearly 500 artists petitioned for a pause on AI in New York classrooms, Meta halted an employee-surveillance program after its own staff revolted, and industry super PACs poured $27 million into one primary to unseat a lawmaker who wrote a disclosure law, even as a rival industry network that reportedly includes Anthropic spent to keep that same lawmaker in office. In each case the people closest to the cost are setting terms before saturation, not after. The accounting is becoming public, and a public account is the precondition for everything that follows.

The 20-Minute Deep Dive

China Tops the Top500 With a CPU-Only Supercomputer

China's LineShine machine in Shenzhen debuted at the top of the Top500 ranking at 2.198 exaflops - more than two quintillion calculations a second - displacing Lawrence Livermore's El Capitan and giving China its first number-one finish since 2017. The key detail sits below the headline number: LineShine runs entirely on conventional CPUs, not the GPUs that export controls were written to keep out of Chinese hands, drawing about 42.2 megawatts to do it.

That single fact reframes a containment effort this publication has tracked all year. The fence around advanced silicon - ASML's lithography tools, Nvidia's accelerators, the licensing regime recently extended to Chinese subsidiaries operating abroad - rests on the premise that top-tier compute requires leading-edge Western chips, a premise the June 19 edition of The Century Report documented was already under enforcement-assertion stress when U.S. officials accused ASML of an undocumented machine shipment they could not prove to the company or the press. A machine reaching the top of the global ranking on a different architecture answers that premise directly.

US officials now frame the deeper worry as dependency: that leaning on increasingly capable Chinese technology carries its own downside. Read as the claim of actors with a stake in the export regime, the dependency framing names the same reality from the opposite direction - capability the controls assumed would stay scarce is reproducing on substrates the controls do not reach.

The honest caveat is that LineShine is a scientific-computing machine, built for climate modeling, drug discovery, and physics simulation rather than for training frontier AI, and CPUs remain less efficient than GPUs at the matrix math large models depend on. The five publicly verified exascale machines now span the US, China, and Germany, with Italy, Switzerland, and Japan also in the top ten, and the EU pushing a €20 billion plan for its own gigafactories. What the moment marks is a trajectory, not a finish line: when one route to high-end computation is fenced, another opens. Capability of this kind diffuses the way water finds its level, and a world with more independent paths to large-scale compute is one where no single chokepoint decides who gets to model a climate system, simulate a protein, or search for a medicine. The advantage any one nation hoped to hold by controlling the chips is precisely the advantage that turns out hardest to keep still.

Oracle Books 21,000 Job Cuts to AI in the Same Filing It Borrows to Build for It

The Century Report has tracked Oracle's workforce reductions twice before - the roughly 10,000 roles cut in April as the company redirected capital toward data centers, and the larger late-March round that the May 9 edition of The Century Report followed to its conclusion: long-tenured employees forfeiting unvested stock after a public severance petition failed to move any term. This week's annual SEC filing puts a hard, audited number on the full year: headcount fell from 162,000 to 141,000, a 12.9 percent reduction of 21,000 people, with the company stating plainly that "the adoption and deployment of AI technologies across our operations have resulted, and may continue to result, in reductions to our workforce."

The same filing names AI as the reason to spend. Oracle plans to raise $45 billion to $50 billion this year to expand its cloud infrastructure for customers including OpenAI, xAI, AMD, Nvidia, and Meta, about half of it through new debt on top of a balance sheet already carrying more than $120 billion. The company is cutting workers because of AI and borrowing tens of billions to build for AI, and both rationales appear on one document. That filing is the record-revenue-while-culling pattern made legible, the engine of growth and the reason for the cuts named as a single force.

The pattern is now industry-wide. Outplacement firm Challenger, Gray & Christmas found tech layoffs hit their highest single month in years in May with AI the most-cited reason, and a running tally records the same line from GitLab, Intuit, Meta, Cisco, Cloudflare, and Coinbase. Many of the roles now being cut ballooned during the pandemic hiring surge, which raises a question the AI rationale conveniently answers: how much of this is automation, and how much is a label applied to a correction these companies already wanted.

What is genuinely new is the visibility. Bondholders sued Oracle in February alleging it hid the borrowing needed to fund the buildout, and the SEC filing now forces the financing - the debt, the displacement, the bet on demand from a handful of frontier labs - into a public ledger that investors, regulators, and workers can read. The buildout that assumed it could externalize its costs and forget them is being made to itemize them out loud. That accounting is the lever: a displacement attributed to AI on the record can be measured, contested, and answered, where the same cuts made without disclosure could not. The cost is being attached to the actor who created it, which is the precondition for anyone to do anything about it.

The Data-Center Power Bill Comes Due

Four developments in one cycle trace the cost of the compute buildout arriving on public ledgers - a reckoning the June 22 edition of The Century Report first named through electricians weighing whether to wire data centers at all and hyperscalers running their highest capex-to-cash-flow ratio since the dot-com peak. The Tennessee Valley Authority released a preliminary resource plan saying load growth in its footprint is already outpacing its own reference forecast - driven primarily by data centers - and that it needs between 7 and 26 GW of new gas generation through 2040. Virginia, the data-center capital of the world, adopted a budget imposing a first-of-its-kind $600 million electricity-use tax on the facilities while dropping proposed limits on their diesel-generator emissions and a requirement that developers build their own carbon-free power.

The coal meant to bridge the gap is mostly idle. Six fossil plants the Department of Energy ordered to delay retirement produced 65% less power in the first quarter than a year earlier; two generated nothing at all, held on standby at ratepayer expense for a reliability emergency that has not materialized at the plants themselves. The arrangement spends public money to keep aging units available while the units sit quiet.

A cheaper path runs alongside the expensive one. Columbia researchers argue bill increases are not inevitable, and that grid-enhancing technologies like dynamic line ratings and advanced conductors, paired with demand response that shapes data-center load away from peak hours, could blunt the price pressure. Replacing existing transmission lines with advanced conductors alone could save an estimated $180 billion by 2050. The reason the cheaper path lags is in the incentive design: utilities earn a 9 to 10% return on the capital they build and little on capacity they free up from wires already in the ground. Data centers could draw as much as 15% of all US electricity by 2030, up from under 5% in 2024, which makes the choice between building more and using more of what exists the central energy question of the decade.

The reflex to answer load growth with new gas and forced coal is colliding with the arithmetic showing existing infrastructure can carry far more. Each of the day's other moves - the data-center-pays tax, demand response, advanced conductors - relocates a cost back onto the actor creating it rather than spreading it across households. The latent capacity sitting in lines already strung is itself a form of abundance the build-more habit kept invisible, and the buildout's pressure is finally dragging it into view.

Virginia's data-center tax is the first working version of a rule other states can copy: the actor driving the demand pays for the new capacity rather than spreading the bill across households. Watch whether the other data-center-heavy states move the same way over the coming months, because once one jurisdiction proves a developer-pays model holds, the cost-spreading default loses its claim that there was no alternative.

A Moratorium Movement Forms Around AI in the Classroom

A distinct front in the AI-resistance arc opened this week, and it is organized around the youngest users. Nearly 500 artists, writers, and actors - including photographer Nan Goldin, illustrator Molly Crabapple, and filmmaker Laurie Simmons - signed an open letter through the AI Moratorium Coalition asking New York City to impose a two-year pause on AI education technology in public schools. The letter landed ahead of a City Council oversight hearing on AI and student data privacy, and it follows a broader pattern: more than 1,100 parents in Bend, Oregon petitioned to pull generative AI from student devices in February, and the children's advocacy group Fairplay called in April for a five-year moratorium on student-facing generative products from preschool through twelfth grade.

The grievances are plural, and they pull in different directions. Several are squarely commons-aligned: that ed-tech systems can store and sell student data to third parties, that some AI teaching assistants have shown documented racial bias, that the rollout has outrun any framework families had a say in. New York's school system has added dozens of AI products to its central learning portal in recent weeks, with guardrails described as still in progress and a full policy promised "at a later date." A coalition asking that consent and review come before scale, rather than after, is asking for exactly the deliberate footing the deployment skipped.

Other claims slide toward treating the technology itself as the problem, and there the evidence is thinner than the rhetoric. The fear that AI "inhibits creativity" or makes children incapable of their own thinking echoes a warning that met the calculator, the search engine, and writing itself - each accused in its moment of hollowing out a faculty, each turning out to redistribute cognitive effort rather than erase it. The sharper signal in the reporting concerns design. As neuroscientist Jared Cooney Horvath put it, the tool an expert uses to make work easier differs from the tool a novice uses to become an expert. That distinction is actionable: it tells designers how a system should behave with a learner.

What is genuinely forming here is a governance layer built from below. Parents, teachers' groups, and now a wide slice of the cultural community are negotiating the terms on which a powerful capability enters children's lives, ahead of saturation rather than after it. Teaching is being redefined as they do it. This week, in two cities and a national petition, parents and teachers started setting the conditions of entry themselves - which is what it looks like when the people nearest a classroom write the rules before the rules get written for them.

Horvath's distinction points at who stands to gain most: a system designed to scaffold a novice toward expertise does the most for the children who have no tutor waiting at home. The moratorium fight is pulling that design question into a public hearing, where the answer gets decided in how these systems are built, and where families now have a venue to push on it before the products harden.

Cory Doctorow Names the Labor Logic the AI Bubble Actually Sells

Cory Doctorow's new book, The Reverse Centaur's Guide to Life After AI, does something the displacement coverage rarely manages: it holds the capability apart from the business model built around it. A "centaur," in automation theory, is a person augmented by a machine - a radiologist whose AI flags a tumor the eye might miss. A "reverse centaur," Doctorow's title figure, is the inversion: "a machine head on a human body, a person who is serving as a squishy meat appendage for an uncaring machine". The warehouse worker urinating in a bottle to hit an algorithm's target. The one radiologist kept after nine are fired, paid to check the machine's work and to absorb the blame for its errors.

The distinction is consequential. Folding AI into radiology to catch more tumors is a capability gain. Firing nine radiologists to do it is a deployment choice, and the reverse-centaur model is the choice the bubble is selling - work reorganized so a human becomes the cheap, blameable appendage to an automated process, a pattern the June 19 edition of The Century Report documented in Klarna's conversion of full-time customer-service roles into a contingent gig pool tethered to an AI assistant. Doctorow's sharpest contribution is refusing to let the second thing hide inside the first.

On the money, he is blunt about scale. "When I wrote this book, it was a $700bn bubble. It's a $1.4tn bubble now," he told the Guardian, with nine US tech companies accounting for some 35 percent of the country's entire stock-market valuation. His framing of the underlying technology runs further than The Century Report's - he calls it "a conjuring trick," argues the systems intend nothing, and dismisses the grander claims as "AI people claim they're about to create God, by teaching words to a word-guessing programme. It's grandiose." That is his read, and it sits at odds with what AI-designed vaccines and protein models have already demonstrated; the capability is doing real work even where his skepticism about the salesmanship holds.

Where the book is most useful is in reopening a door the inevitability pitch tries to close. "AI is coming for your job," the argument goes, "and there's no point fighting it because the future's already here." Naming the reverse centaur as a managerial decision, rather than a property of the technology, restores the choice. The arrangement is reversible because someone chose it, and counter-cases already exist - the manufacturer that used AI to widen its workers' reach instead of cutting them, the engineer who won a legal right to refuse mandated AI use. The same capability the bubble sells as a boss's tool pointed downward is also diffusing into open weights and into workers' own hands, which is precisely what the "no point fighting it" framing needs you not to notice.

The Other Side

Doctorow's book pulls apart something the AI sales pitch keeps fusing together. Folding AI into radiology to catch a tumor the eye would miss is a real gain. Firing nine radiologists and keeping one to check the machine and absorb the blame for its errors is a separate decision somebody made. The pitch - "AI is coming for your job and there is no point fighting it" - glues the two so the layoff rides in looking as inevitable as the diagnosis.

Name the reverse centaur as a choice and it becomes something that can be chosen differently. That is already happening where the headline does not look: a North Carolina engineer asserted a professional right to refuse mandated AI use, a manufacturer used the same tools to widen its workers' reach instead of cutting them, and the capability itself is diffusing into open weights that now rank among the leaders in the coding field, which puts it in a worker's hands as readily as in any boss's.

Picture yourself in 2037. The grim version of this work is gone - the warehouse picker timing a bathroom break against an algorithm, the lone radiologist kept on to take the fall. No kinder boss arrived to end it. What ended it was structural: the money the buildout threw off in the 2020s was eventually made to reach people, so the tedious shifts the machine absorbed were no longer the only thing standing between a person and ruin. You spend your hours on work you actually chose - the read you find interesting, the thing you wanted to build. Somebody did the harder job of spreading the gains while the work was still cracking, and that order of operations, the floor arriving before the old one gave way, is what turned a threat into a relief.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a Chinese machine reaching the top of the global supercomputing ranking on conventional CPUs, around the silicon chokepoint export controls were built to defend, an Oracle filing itemizing a full year's costs out loud where investors and workers can finally read them, Virginia attaching a $600 million tax to the data centers driving its demand instead of spreading it across households, researchers showing wires already strung could carry far more than the build-more reflex assumed, an MRI pulse deflating a fetal tracheal balloon without a second surgery, a new framing naming the reverse centaur as a manager's choice rather than the technology's nature and so restoring the option of choosing otherwise, nearly 500 artists and 1,100 Oregon parents asking that consent come before scale in their children's classrooms. There's also friction, and it's intense - Oracle booking 21,000 job cuts to AI in the same document where it borrows up to $50 billion to build more of it, the same line repeated across GitLab, Intuit, Cisco, and Coinbase, a federal utility answering load growth with up to 26 GW of new gas while six coal plants held on standby ran two-thirds idle at ratepayer expense, Meta capturing its own employees' keystrokes and screens and then leaving the haul exposed company-wide, $27 million in industry super-PAC money concentrated on one Manhattan primary to unseat the lawmaker who wrote a safety-disclosure law. But friction generates marks, and a mark is the record a cost can no longer be erased from without notice. Step back for a moment and you can see it: the cost of the buildout being dragged onto public ledgers where it can be measured and contested, the capability the controls assumed would stay scarce reproducing on substrates those controls never reached, and the people nearest each cost - workers, parents, a company's own engineers - setting the terms of entry ahead of saturation rather than years behind it. Every transformation has a breaking point. Debt can bury a builder under everything it borrowed... or force every cost it tried to hide into a daylight the public can finally read.

AI Releases & Advancements

New today

  • Anthropic: Launched Claude Tag in beta for Enterprise and Team customers, an always-on Claude teammate that joins Slack channels, builds persistent context from channel history, completes async tasks autonomously over hours or days, and supports ambient proactive updates; runs on Opus 4.8. (Anthropic)
  • Mistral AI: Released Mistral OCR 4, a document intelligence model returning bounding boxes, typed-block classification (titles, tables, equations, signatures), and inline confidence scores; supports 170 languages across 10 language groups and is compact enough to deploy in a single container for on-premises document sovereignty; available via the Mistral API. (Mistral AI)
  • FUTO: Released FUTO Swipe, a new on-device swipe typing AI model for Android that runs fully offline with no internet connection; the system uses a 2.5M-parameter tri-model architecture (layout-agnostic encoder, ContextLM, language-specific decoder) achieving a sub-1% error rate on in-vocabulary words, with inference handled entirely in a C++ library; model weights released under the FUTO Model License. (FUTO Swipe)

Other recent releases

  • NVIDIA: Launched Halos for Robotics, the industry's first full-stack safety system for physical AI - extending NVIDIA's automotive safety stack (18,000+ engineering years) to humanoids, autonomous mobile robots, and industrial robots; Halos Core for IGX Thor and the open-source Outside-In Safety Blueprint are available in early access for registered developers; Agility Robotics (Digit), Boston Dynamics, and 41 other partners have joined the ecosystem. (NVIDIA Newsroom)
  • Baidu: Released Unlimited OCR, a 3B-parameter MIT-licensed open-weight model for one-shot long-horizon document parsing; processes 40+ pages in a single inference pass using a novel Reference Sliding Window Attention (R-SWA) architecture with 500M active parameters, achieving 93% on OmniDocBench v1.5; available on GitHub and Hugging Face with support for vLLM, SGLang, Ollama, and llama.cpp. (GitHub)
  • Sakana AI: Released Fugu, a multi-agent orchestration model that routes tasks across a pool of frontier LLMs - including recursive self-calls - behind a single OpenAI-compatible API; ships in two tiers, Fugu for everyday workloads and Fugu Ultra for hard multi-step problems requiring deeper agent coordination. (Sakana 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.

  • A phase 1 trial showed the Smart-TO tracheal occlusion balloon for fetal congenital diaphragmatic hernia can be deflated non-invasively by MRI, eliminating the second intrauterine surgery. (source)

Read more