Sign in Subscribe
The Century Report

China Claims to Have Trained an Open Frontier Model With No Nvidia - TCR 08/02/26

Huawei trained a 505-billion-parameter open model entirely on its own Ascend chips as Europe and Korea moved to own the compute layer themselves.

Ben Linford

Ben Linford

- 20 min read
Three-panel Century Report Aug 2 2026 infographic: compute shifting from gatekeepers to sovereign funds and domestic silicon, costs returning to the ledger, and a 48-sensor microscope.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The clearest signal running through the day is what happens when a barrier meant to preserve advantage collides with the will to route around it. Huawei's half-trillion-parameter model, which Huawei says was trained end to end on Ascend chips with no Nvidia silicon in the run, is the export fence being climbed while the same reporting shows it being tunneled under - a supply chain still leaning on stockpiled foreign dies, a stockpile that is now reportedly dwindling. The lesson the analysts watching most closely draw is that a control built to hold a capability lead has largely produced a parallel stack instead. Intelligence, denied the front door, finds another channel and cuts it deep.

That same instinct to own the foundation rather than rent it is now moving governments. Europe opened bidding for seven compute hubs, the largest holding up to 100,000 chips each, while South Korea widened a sovereign fund to buy directly into domestic AI infrastructure. Belgium absorbed the commercial risk of a Congo germanium plant to loosen a single-country grip on a chip-critical mineral. Read as a spending scoreboard, these look like line items. Read as what the capital actually builds, they are the physical substrate of frontier compute passing out of a handful of corporate gatekeepers and into a widening set of hands.

Underneath the ambition sits the bill, and it is landing on specific people first. SpaceX will run sixty-nine unpermitted gas turbines beside a historically Black Memphis community for another year, one thread in a season of permit-dodging complaints the industry now admits it is losing in public. For a decade a developer could externalize that cost almost entirely. The lawsuits and Clean Air Act filings are the sound of it returning to the ledger.

And Reddit is testing the same reversal in a quieter register, pressing in court and in its investor letter whether the human conversation an answer layer summarizes for free can be repriced by the people who made it. The Berkeley microscope that captures whole living organisms at cellular detail belongs to this movement too: a scarce, expensive limit displaced by something that scales like silicon and code. Across every one of these fronts, the assumption that value concentrates and stays concentrated is meeting evidence that it does not.

The 20-Minute Deep Dive

Huawei Says It Trained a Frontier Model With No Nvidia, and the Supply Chain Tells on Itself

Huawei released openPangu-2.0-Pro, a 505-billion-parameter model with 18 billion parameters active through a mixture-of-experts design that Huawei says was trained from first token to last on the company's own Ascend 910B NPUs, with no Nvidia silicon anywhere in the run. It is an open-weight model above half a trillion parameters that Huawei says was trained entirely outside the Nvidia ecosystem, and it is live now on GitCode and Huawei Cloud's ModelArts. The run consumed 34 trillion tokens, used the Muon optimizer, and reaches a 512K context window. This is the next chapter of a pattern The Century Report flagged on June 6, when Huawei's Ascend 910C chips completed the full post-training run of DeepSeek-V4-Pro - Huawei says its Ascend silicon has now moved from handling a model's fine-tuning stage to running an entire training pipeline for a frontier-scale system from scratch. By Huawei's account, the compute layer that the export regime was built to fence off just produced a frontier-scale result without a single restricted chip.

Then the supply chain tells a more complicated story. Some Ascend parts reportedly used TSMC 7-nanometer dies, while roughly 2.9 million such dies were ordered by Sophgo. Those dies moved through a shell-company chain around Sophgo before the diversion was caught, a violation that has TSMC facing a Commerce penalty north of a billion dollars. The high-bandwidth memory came from Samsung, some 13 million stacks stockpiled ahead of tightening controls. That die bank is now reportedly dwindling. No publicly documented run built end to end on SMIC-fabricated logic and domestic CXMT memory has yet happened, and analysts at RAND note that iFlytek absorbed a three-month delay simply switching its training over to Ascend. The fence, in other words, was climbed over and routed around at the same moment.

The performance gap is real and worth stating plainly. Ascend inference runs around 17 to 19 tokens per second against 25-plus on comparable Nvidia parts, and the Council on Foreign Relations estimates leading US chips remain roughly five times more capable than China's best domestic alternatives, a lead it projects widening to seventeen-fold by 2027 if domestic fabrication stays behind. There is also the matter of what Chinese law requires of the firm: companies must "support, assist, and cooperate with national intelligence work in accordance with the law." That is a claim to weigh against what open weights actually permit, which is inspection and local execution by anyone who downloads them.

What the episode reveals is that a control designed to preserve a capability lead has, by the assessment of the analysts watching it most closely, "largely backfired." Denied the top of the market, an entire parallel stack was willed into existence beneath it - fabrication, memory, interconnect, optimizer, open weights. Intelligence behaves like water finding its level, and a wall raised across the channel does not stop the water so much as teach it where else to flow. The half-trillion-parameter demonstration says the channel is already cut. The reportedly dwindling die bank says the next one will be dug at home.

The EU Fires the Starting Gun on Seven AI Gigafactories

The European Commission opened bidding for seven AI gigafactory hubs, a roughly €30-billion program to place frontier-scale compute on European soil. Four of the sites are sized as smaller facilities holding 25,000 to 75,000 chips, each eligible for up to €1 billion in support; three are large hubs of 40,000 to 100,000 chips, each drawing up to €2 billion. Germany, Greece, Portugal, Italy, and Spain are backing bids for the largest of them. Winning consortia are expected to be named early next year, with eighteen months to build.

The distinction from Europe's earlier moves is worth drawing. Coverage over the past year tracked the continent's push to own its models - the Apertus open-weight release, Korea's K-EXAONE, the sovereign-model wave. This is a different layer of the same ambition: the financing and the physical siting of the compute those models will run on. A bloc that spent a decade renting capacity from three American hyperscalers is now underwriting the concrete, the transformers, and the chip orders directly, into a memory market where a handful of data-center buyers have already locked up most capacity through 2028.

The same shift is visible on the other side of the world. South Korea is injecting 20 trillion won, about $13.9 billion, into its Korea Investment Corporation sovereign fund, and for the first time the mandate explicitly includes domestic AI and data-center assets. The move follows the same week's other sovereignty milestone, when the July 31 edition of The Century Report covered Korea's National AI Foundation project releasing LG's 750-billion-parameter K-EXAONE 2.0 under an open license, giving the country a frontier-scale model to go with the capital now backing its physical infrastructure. The timing followed a sharp tech-stock rout, which makes the decision read less as momentum-chasing than as a considered judgment that the compute layer is infrastructure worth owning through the volatility rather than around it.

The instinct to read these announcements as a spending scoreboard misses what the money is actually doing. Capital here is being redirected out of financial abstraction and into a physical substrate - power interconnects, cooling, silicon that will still be computing when the funding cycle that bought it is forgotten. What the capacity enables counts for more than whether the sovereign funds see a return: research groups, hospitals, universities, and startups across two continents gaining access to frontier compute that until now cleared through a handful of corporate gatekeepers an ocean away.

That is the deeper movement underneath the procurement notices. The assumption that frontier capability must concentrate in a few private hands, in a few jurisdictions, is being met with an alternative that treats compute the way earlier eras came to treat electricity and clean water - as a common substrate too foundational to leave rented. Seven hubs in Europe and a sovereign fund in Seoul do not settle that question. They do show that the number of places willing to answer it is climbing.

Reddit Fights AI's Value Compression on Two Fronts at Once

Reddit is pressing the same question through two very different channels: what is the human conversation it hosts actually worth, and who captures that value when an answer layer summarizes it for free?

On the legal front, a federal judge kept alive a case most observers expected to collapse. US District Judge Paul A. Engelmayer largely denied web scraper SerpApi's motion to dismiss, finding Reddit had plausibly pleaded that SerpApi and Perplexity AI conspired to pull copyrighted Reddit content out of Google search results - SerpApi supplying the circumvention capability, Perplexity paying for the feed. The ruling landed less than two weeks after a different court dismissed Google's own scraping suit, where Google could not show that rights holders had authorized it to block scraping on their behalf. Reddit cleared the bar Google missed. SerpApi frames the fight as an attempt to "wall off the open Internet," and the tension is genuine: the anti-circumvention technology at issue was invented more than a year after Reddit and Google signed their licensing deal, yet the judge found the claim plausible anyway.

On the business front, CEO Steve Huffman used Q2 earnings to voice his first public doubt about the other side of that same Google relationship. Reddit's investor letter cast the company as "the antidote to an automated web," arguing that "people don't want a summary of Reddit; they want Reddit." Huffman was blunter on AI Overviews: "we're still looking for that win-win." A Pew study of roughly 900 US adults measured why the doubt has teeth - AI Overviews cut referral clicks to source sites by nearly half compared with the old list of blue links. Reddit is reportedly weighing whether to end its $60M Google licensing deal, and it is not alone; The Economist, Reuters, Politico, and USA Today are all reevaluating similar arrangements.

Both fronts test an assumption the answer layer has relied on: that the places where humans actually talk to each other are interchangeable, and that the value of that talk can be summarized away without paying the people who made it. The Pew number says the extraction is measurable. The court ruling and the wavering licensing deals say the originators have started to price it. When the source of the answers decides its conversation is worth more than the free summary of it, the cheap economics that fed the answer layer begin to move.

The court thread carries a reading the copyright fight has been missing. Google's own scraping suit failed because it could not show rights holders had authorized it to block scraping; Reddit cleared that bar by suing over the circumvention technology itself, giving originators a legal lever that does not wait on the still-unsettled question of whether training on their content is fair use. Watch whether the other publishers reevaluating their Google arrangements - the Economist, Reuters, Politico, USA Today - begin citing Engelmayer's ruling as the same lever.

xAI's Unpermitted Memphis Turbines Will Run Another Year

SpaceX confirmed on Friday that it will keep sixty-nine gas turbines running to power xAI's Colossus cluster outside Memphis until a permanent 1.2-gigawatt plant comes online in mid-2027, declining to remove the unpermitted units for another full year. The turbines sit in Mississippi just south of the Memphis line, in one of the most pollution-burdened stretches of the region, and by the plaintiffs' estimate they can emit upward of 2,000 tons of nitrogen oxides a year. The NAACP and the Southern Environmental Law Center are suing. An IPO filing put the company's turbine spend at $2.8 billion over three years, and the permanent plant is slated to run 41 turbines rated between 16 and 50 megawatts each.

This is the environmental-justice claim in its earned form, and it deserves to be named as such rather than deflected with a proportion argument. The turbines are self-built generation, running without the permits that any comparable industrial source would require, sited beside a historically Black community that has spent generations hosting the burdens wealthier districts successfully refuse. The test for whether a resource objection is real is simple: would it be raised if the same turbines fed a paper mill, and would the mill have been allowed to skip the permit? Here the answer is yes and no, and that is exactly why the objection lands.

The pattern extends past Memphis. Amazon and Duke Energy stand accused of evading Clean Air Act review at an under-development data center in Hamlet, North Carolina, where backup generators were permitted for a single year. Nebius reported a thirty-two-fold jump in emissions in a single year on the back of data-center growth, with no quantified reduction targets attached. The industry itself sees the exposure: operators told Politico they have "ceded the narrative to their critics" as local opposition gains organized momentum.

The transition genuinely requires this buildout - the compute has to sit somewhere, drawing real power. What it does not require is the specific arrangement Memphis is fighting: dirty generation stood up outside the permitting system on the community least equipped to say no. And the economics of that arrangement are turning. For a decade the cost of running polluting generation next to a low-income neighborhood was one a well-lawyered developer could externalize almost entirely. The lawsuits, the Clean Air Act filings, and an industry admitting it has lost the public argument are the sound of that cost coming back onto the ledger, even as the Justice Department invoked national security only weeks ago to dismiss the NAACP's earlier Clean Air Act suit over the Memphis turbines. The buildout that outlasts this decade will be the accountable, justly-sited version, because the other kind is becoming the expensive kind to defend.

A 48-Sensor Microscope Breaks Optics' Oldest Trade-Off

For as long as microscopes have existed, they have forced a choice. Zoom in for fine detail and you lose the wide view; pull back to see the whole specimen and the small structures blur away. Add speed to that and the compromise gets worse. Optical designers have treated this three-way tension between resolution, field of view, and frame rate as a fixed cost of seeing small things - a limit set by lenses and the physics of light gathering.

A team at UC Berkeley, led by Laura Waller with lead author Kevin C. Zhou, has now sidestepped that compromise by refusing to solve it with optics alone. Their instrument, described in Nature Photonics, arrays 48 separate camera sensors and stitches their outputs together computationally. The combined system captures 25.2 billion pixels per second - holding 3-micron resolution across a 5-square-centimeter field at 120 frames per second. Each of those numbers alone is achievable on some existing microscope. Holding all three at once is what the field had accepted as impossible.

The distinction here is what was demonstrated versus what is deployed. The Berkeley team imaged living C. elegans nematodes - transparent roundworms roughly a millimeter long, a workhorse of developmental biology - capturing the movement of many individual organisms across the full field while resolving cellular detail in each. This is a demonstrated capability in a research instrument, not a clinical scanner arriving in hospitals. What it moves is the date on which whole-organism imaging at cellular resolution becomes ordinary, and what it changes is which questions a biologist can even pose.

That last point carries the weight. When you can only watch one worm at a time, or one region of tissue, you study individuals and extrapolate. When you can watch hundreds of organisms simultaneously, each in full cellular detail, at video speed, you can ask questions about populations, variation, and rare events that the old instruments made unaskable. Drug responses across a whole culture, developmental timing across a cohort, the one cell in ten thousand that behaves differently - these come into view together.

The deeper move is in the method. The performance came not from a better lens but from treating the sensor array and the reconstruction software as one seeing system, with cheap commodity image chips doing the gathering and computation doing the resolving. The scarce, expensive component - precision optics engineered against an unforgiving trade-off - gets displaced by something that scales the way silicon and code scale. The historic limit was never a law of nature so much as a limit of what a single lens could do alone, and that particular ceiling has just lifted.

The Other Side

For decades, a well-lawyered developer could build dirty generation next to the community least able to refuse it and pay almost nothing for the privilege. The permit system slowed that developer down and rarely stopped one. The nitrogen oxides, the asthma, the burden a wealthier district would have blocked, all landed on people who had spent generations hosting exactly that.

Memphis is where that arrangement is now being challenged. SpaceX will run sixty-nine unpermitted turbines beside a historically Black neighborhood for another full year, and the NAACP and the Southern Environmental Law Center are in court over it. Amazon and Duke face Clean Air Act complaints in Hamlet; Nebius logged a thirty-two-fold emissions jump; the operators told Politico they have "ceded the narrative to their critics." Each is the externalized cost finding its way back onto the ledger it was kept off of.

The compute has to sit somewhere, and the buildout is moving forward. What is turning is the economics of where. The accountable, justly-sited version is becoming the cheaper kind to build, because the other kind now draws lawsuits, filings, and a public it has already lost.

Imagine a woman in 2034 who grew up on that Memphis block and is raising her own kids there now. The data center two miles off draws its power from generation that had to clear the same permits her high school did. The air on an August night is just air. Her kids run the streets she ran, without the low hum that used to sit under everything, without the inhaler by the door. That ordinary evening exists because the people who sued in 2026 - when the turbines still ran and the law still moved slower than the machines - made externalizing that cost the expensive choice. The hard year was hosting the burden one more time; what it bought was a neighborhood that finally gets to breathe.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Huawei's half-trillion-parameter model, which Huawei says was trained end to end on Ascend chips with no restricted silicon anywhere in the run, published as open weights, Europe opening bids for seven compute hubs holding up to 100,000 chips each while Seoul widens a sovereign fund to own the substrate rather than keep renting it, Belgium backing a Congo germanium plant that could supply 15% of world demand and loosen a single country's grip on a chip-critical mineral, a Berkeley microscope holding cellular detail across a five-centimeter field at video speed by replacing precision optics with 48 commodity sensors and reconstruction code, and a first-of-its-kind Minnesota law against nudify apps taking effect over xAI's objection. There's also friction, and it's intense - the stockpiled foreign dies that made the Chinese run possible now reportedly dwindling with no publicly documented all-domestic replacement yet proven, Ascend inference still trailing and a gap between leading US chips and China's best domestic alternatives that analysts project widening to seventeen-fold by 2027, a legal duty to assist state intelligence weighing against what open weights actually permit; sixty-nine unpermitted gas turbines set to run another full year beside a historically Black Memphis community while Amazon and Duke face Clean Air Act complaints and Nebius logs a thirty-two-fold emissions jump; Google yanking an AI feature one day after launch over fears of fabricated satellite imagery; and AI Overviews cutting referral clicks to source sites nearly in half as Reddit weighs walking away from a sixty-million-dollar deal. But friction generates contrast, and contrast is what makes an edge visible that a smooth surface would have hidden. Step back for a moment and you can see it: the assumption that value concentrates and stays concentrated meeting evidence it does not - a capability lead fenced off and answered with an entire parallel stack, frontier compute passing from a handful of gatekeepers toward a widening set of hands, an optical limit displaced by something that scales like silicon, the free summary of human conversation being repriced by the people who wrote it - while the bill for the buildout reattaches to the ledger it was externalized off, permits and lawsuits and emissions counts landing back on the developers who skipped them. Every transformation has a breaking point. A wall can hold a capability in place for a season... or teach it every route around itself.

AI Releases & Advancements

New today

  • ByteDance: Released Seedance 2.5, generating synchronized video and audio clips up to 30 seconds from prompts and multimodal references, now available through Jimeng AI and Doubao Pro. (The Decoder)
  • LightOn: Released mDenseOn and mLateOn, two open 307M-parameter multilingual retrieval models for long-context, cross-lingual, and code search, alongside datasets and training code. (Hugging Face)
  • Microsoft Research: Released Echoverse, an open framework and benchmark for developing computer-use agents across four stateful synthetic applications, with seed data and database-grounded graders. (arXiv)
  • JetBrains Research: Open-sourced KotlinLLM, an experimental IntelliJ IDEA plugin that lets language models generate and hot-reload Kotlin/JVM code through smart macros and JDI. (JetBrains)
  • Supabase: Open-sourced Supabase Evals, a benchmark and evaluation framework that tests coding agents including Claude Code, Codex, and OpenCode against real Supabase tasks in containerized environments. (Supabase)
  • NVIDIA: Made nvmath-python 1.0 generally available, providing a unified Python interface for CPU, GPU, and distributed CUDA-X mathematical workloads across NumPy, CuPy, and PyTorch. (NVIDIA Developer Blog)
  • Nous Research: Added Block Buzz support to Hermes Agent through the Buzz Desktop runtime, a relay bridge, and a native Hermes gateway, preserving agent memory, skills, approvals, scheduled tasks, and sessions. (Hermes Agent Documentation)
  • South Africa DPSA / Meta / Juicetel: Launched the Llama-powered Batho Pele AI Chatbot, providing citizens and public servants with conversational access to government policies, circulars, legislation, and public-service information. (DPSA)

Other recent releases

  • xAI: Released Grok Imagine Video 1.5 with image references, taking text-to-video and native 1080p generation to general availability in the xAI API and on grok.com/imagine, iOS, and Android, adding support for up to seven reference images to lock character, scene, and voice consistency across generations. (xAI)
  • Huawei: Open-sourced openPangu-2.0-Pro, a 505B-total/18B-active MoE language model trained on Ascend NPUs with a 512K-token context window and released weights, inference code, and technical report, expanding the openPangu 2.0 family beyond the earlier 92B Flash variant. (AIbase)
  • Google DeepMind: Released Gemini Robotics 2, Gemini Robotics ER 2, and Gemini Robotics On-Device 2, adding whole-body control that lets robots dynamically balance, step, squat, and bend to navigate cluttered spaces, plus adaptation in Google's tests to new robot embodiments with only a few hours of data. (Google DeepMind)
  • Thinking Machines Lab: Released Inkling-Small, a 276B-total/12B-active open-weights multimodal MoE model that matches the original 975B Inkling's performance at a quarter of the size, with day-0 vLLM support and Tinker fine-tuning. (Thinking Machines Lab)
  • LG AI Research: Released K-EXAONE 2.0, a 750B-parameter open-weight model under Apache 2.0 on Hugging Face, more than tripling K-EXAONE 1.0's parameter count and improving benchmark scores by over 10%. (Korea Times)
  • PolyAI: Released Dialog-RSN-1, an audio-native dialog model that fuses turn-taking, speech recognition, function calling, and response generation into a single model, delivering sub-300ms responses in live production calls. (MarkTechPost)
  • Tether Data (QVAC): Open-sourced VisionPsy-Nano, a ~460M-parameter on-device vision-language model achieving the top score among sub-0.5B VLMs across 17 benchmark tasks, with a Flash variant up to 36x faster on iPhone. (Tether)
  • AMD: Released Instella-MoE-16B-A3B-Think, a fully open 16B-parameter (2.8B active) MoE model trained from scratch on AMD Instinct GPUs, with full pipeline checkpoints released across pretraining through RL. (AMD ROCm Blogs)
  • Pangram: Launched Pangram Image Detection in research preview, a new AI-generated image detector claiming 99.5% accuracy across outputs from GPT Image, Nano Banana, Midjourney, FLUX, Grok Imagine, and video models like Kling and Veo. (Pangram)
  • Tenzai: Added autonomous mitigation to its AI Hacker platform, enabling the system to automatically generate and deploy targeted protections (via partners like Akamai) within minutes of confirming an exploitable vulnerability. (IT News Online)
  • Invicti Security: Launched Invicti Agentic Pentest, combining autonomous AI reasoning agents with its proof-based DAST engine to identify attack paths and adapt testing strategies as it runs. (PR Newswire)
  • MiniMax: Launched H3, an open general-purpose multimodal video model that unifies text, image, video, and audio inputs, generating up to 15-second 2K videos with native stereo sound, with weights to follow. (Investing.com)

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