Sign in Subscribe
The Century Report

XPeng's IRON Walks Off the First Automated Humanoid Line - TCR 09/08/26

XPeng's IRON robot walked off the first automated humanoid line, as Malaysia weighs Huawei chips for a sovereign AI build.

Ben Linford

Ben Linford

- 20 min read
Three-panel Century Report infographic: XPeng IRON humanoid walks off a Guangzhou line, Malaysia weighs Huawei chips for sovereign AI, plus independent audits of Astra and NHS data.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Barriers built to keep a capability scarce hold only as long as the scarcity does. In Guangzhou, XPeng demonstrated what it calls the world's first automated production line for advanced humanoids, and its IRON robot walked off it under its own power, with XPeng saying more than 80% of its core processes are automated and saying that over 85% of its supply chain is shared with the company's electric-car business. A viral clip made IRON famous last November by proving a robot could walk; a shared supply chain and an automated line is the harder feat, the machinery XPeng is betting could turn one unit into thousands at a falling cost.

The same manufacturing surge showed another face in Malaysia, where the government is weighing Huawei's Ascend accelerators for a sovereign AI effort, what Bloomberg reports could become the first known case of a national government openly choosing Chinese chips over American ones. Huawei's Ascend 910C trails Nvidia's H100 by a wide margin in one reported inference comparison. What tips the calculation is price and availability: reported steep discounts amid a memory shortage that some forecasts expect to run through 2030, and compute a buyer can actually obtain beats faster compute it faces a long wait for. Export controls were priced to a moment when only American silicon could do the job, and the fence leaks first where a mid-sized economy is starved for chips it can afford.

Two other stories turned on who gets to check the powerful. OpenAI revised several of GPT-6 Astra's benchmark figures after its September 3 launch, changes that lifted Astra while lowering the scores it reported for a rival. When one organization builds the model, writes the test, grades the answers, and publishes the scoreboard, discretion can decide who appears to lead. In England, NHS figures recorded roughly 60,000 additional national data opt-outs over about two months, and a health minister named mistrust of Palantir as the cause. Unable to see what the operator does with their data, they used the one lever they had.

In both, the response taking shape is checking done by independent reviewers. An outside body that owns the Astra benchmark independently confirmed the disputed figures, and the NHS owns the data models built on Palantir's platform and could move them elsewhere, so its dependency was always contingent. The same loosening ran through the day's quieter result: six independent aging clocks, run on stored blood from a finished lung-disease trial, all read younger in treated patients, opening a way to test whether ordinary disease drugs also slow aging without a separate decades-long study. A measurement once too costly to run becomes something a standard trial can carry.

The 20-Minute Deep Dive

A Humanoid Walks Off an Automotive-Grade Line, Not a Stage

On September 8 in Guangzhou, XPeng's IRON robot walked off what the company describes as the world's first automated production line for advanced humanoids, moving under its own power; Electrek reported, citing XPeng, that no operator was at a remote console and no pre-scripted routine was used. XPeng says the line runs more than 80% of its core processes automatically, borrows the quality systems XPeng uses to build electric cars, and XPeng says it shares over 85% of its supply chain with that EV business.

That last figure is the actual news. IRON went viral last November when its walk looked so smooth that people accused XPeng of hiding a person in a costume, and engineers cut open its leg on stage to prove otherwise. A viral demo shows a robot can move; what XPeng describes as a shared automotive supply chain and 80% line automation is the harder thing, the machinery XPeng is betting could turn one impressive unit into thousands at a cost that closes. XPeng is targeting mass production by the end of 2026, with the first robots going into its own showrooms before any external sale in 2027.

The specifications are genuine: 76 points of articulation across the body and 21 in each hand, three of XPeng's in-house Turing chips delivering a combined 2,250 trillion operations per second, enough to run the company's physical-world foundation model on the robot itself rather than piping every decision from a datacenter.

The Tesla comparison writes itself and mostly favors XPeng, though both companies trade in forecasts that should be read as forecasts. Musk projected roughly 10,000 Optimus units in 2026 and conceded in January that none were doing useful work; Tesla is still converting a car line while XPeng has demonstrated one running. XPeng's own promise that IRON's per-unit margin will beat its EVs is equally a promise, from a company that has not yet shipped a robot to a paying customer.

Set beside the day's other China story, a foreign government weighing cheap, available Chinese accelerators for its sovereign AI over American silicon, IRON is one face of a single fact. The industrial base that spent a decade learning to build EVs at scale is pointing that capability at embodied intelligence, and XPeng is betting that what it produces could drive the cost of a working humanoid down toward where ordinary buyers can reach it, not a national trophy. By December we will know whether IRON does useful work in those showrooms or only stands there looking human. Either way, XPeng has demonstrated the line running.

Malaysia Reaches for Huawei Silicon, and the Export Fence Springs a Leak Where Scarcity Bites Hardest

Malaysia is seriously evaluating Huawei's Ascend 910C accelerators as the backbone of a RM2 billion sovereign AI effort, and Bloomberg reports that it could become the first known case of a national government openly choosing Chinese AI chips over American ones. The Ascend 910C trails Nvidia's H100 by a wide margin in one reported inference comparison, a gap reported for that comparison. This brings the supply pressure documented in the September 6 edition of The Century Report, when TSMC's fivefold acceleration in fab construction still fell short of AI chip demand, into a national procurement decision. What tips the calculation is price and availability: Huawei is reportedly discounting steeply amid a global memory shortage some analysts expect to run through 2030, and cheaper compute a buyer can actually get its hands on beats faster compute it faces a long wait to obtain.

Prime Minister Anwar Ibrahim has framed the project around keeping government, military, and intelligence data physically inside Malaysia, out of reach of the US Cloud Act, under which American courts can compel US firms to surrender data even when it sits on overseas servers. "I do not consider that law reasonable, but how does one argue with President Trump?" Anwar told students in April. The operator would be the state-linked Telekom Malaysia, which already runs Huawei gear in its cloud infrastructure.

Washington's objection arrives in the familiar vocabulary of national security. A State Department spokesperson warned that "untrusted suppliers can be manipulated, disrupted or controlled by authoritarian regimes," and US officials told their Malaysian counterparts they worry about Huawei hardware handling sensitive intelligence material. That is a claim from the actor with the most to lose if the American stack stops being the default, and it earns the skepticism due any such self-interested framing. Malaysian security officials will fly to Washington later this month; the government calls its own decision purely commercial and is separately floating a second tender that could let Nvidia compete more directly.

Export controls were built to keep advanced compute scarce and countable, priced to a moment when only American silicon could do the job. Some forecasts of a memory shortage paired with a reportedly discounted domestic alternative describe precisely the condition under which that fence leaks, and it leaks first where a mid-sized economy is most starved for chips it can afford. The same Chinese manufacturing surge now switching on automated production lines is what puts a buyable accelerator in front of a government that could not otherwise stock a national data center. The friction is genuine, and so is the widening: a country that wanted to build rather than rent just found a supplier willing to sell.

When the Numbers Move After Publication

OpenAI shipped GPT-6 Astra on September 3, and The Century Report has tracked its capability claims across the days since. A different question surfaced this cycle, one about the ruler rather than the runner. Fortune documented that several of Astra's published evaluation figures changed in the hours and days after the announcement went live, and the revisions tended to lift Astra while lowering the scores reported for Anthropic's rival models.

The specifics are small enough to look like housekeeping and large enough to decide a ranking. Reports tracking the revisions said Astra's published hallucination rate started at 4.2%, fell to 2% after the post became widely visible, then returned to 4.2%. Reports tracking OpenAI's page said Anthropic's Fable 5.1 briefly dropped nearly ten points on a hard math benchmark, from 87.8% to 78%, before settling at 83%. One report said an embargoed draft listed Astra's score on one reasoning benchmark at 98.6%; the live page read 99.99%. OpenAI says most evaluations carry a few points of noise depending on the exact checkpoint and setup, that it revised figures to reflect its best estimate of performance, and that a publishing snag, described first as a content-management bug and then an internet outage, was unrelated to the numbers.

Take the company's account at face value and the underlying problem does not move: when one organization builds the model, writes the test, grades the answers, and publishes the scoreboard, a few points of discretion at any step can decide who appears to lead. An independent walkthrough made that clear, extending the harness-dependent ARC-AGI-3 result documented in the September 6 edition of The Century Report. Astra's marquee reasoning result reads 99.99% with OpenAI's custom harness, the set of tools the model is handed to do the task, and 62.7% with ARC Prize's standard harness, a difference of about 36 points across the two reported setups.

The gain sits right alongside the friction. The check that counted came from outside the building. The ARC Prize Foundation, which owns that benchmark, independently confirmed that Astra scored about 99.9% with OpenAI's custom harness and 62.7% with ARC Prize's standard harness, so the harness gap is documented by the party that cannot benefit from hiding it. That is the shape of the answer forming across the industry: outside verification anyone can audit, in place of the lab's self-report taken on trust. How far that still has to go shows up in a quieter, stranger detail. Asked two days after launch, two AI assistants running without web access flatly denied Astra existed, one calling it "speculative fiction." The systems cannot yet verify their own newest siblings, and the people building the rulers still have to.

Astra's harness gap also exposes a route to greater capability without training a larger model. ARC Prize reports that preserving the system's reasoning between requests raises its score from 62.7% to 98.6% at the same configured reasoning-effort setting; across the game-reasoning pairs both setups solve, the adapter consumes 49% fewer tokens. The foundation's open testing code gives other researchers a way to investigate that gain, spreading knowledge about how to sustain useful reasoning beyond the laboratory that trains the model.

60,000 Patients Withdraw Their Records, and a Minister Names the Reason

The figure at the center of James Frith's letter to the Commons health committee is 60,000: NHS figures recorded roughly that many additional national data opt-outs over about two months this summer, and the health innovation minister told MPs he believes mistrust of Palantir is the driver. This is the next chapter in the same £330m NHS contract that the June 17 edition of The Century Report covered as the UK reviewed Palantir's role amid a coordinated European retreat from the company's tools. "I am concerned about mistrust of Palantir, and the impact it could have on people's willingness to share data with the NHS," Frith wrote, warning it may not be possible to deliver the government's 10-year health plan if patients keep opting out.

Set against England's 57 million people, 60,000 is about 0.1%, and Frith calls the rise modest. It is also the largest reported increase in national data opt-outs over a comparable period since the £330m federated data platform was announced in 2023, and it runs one direction. The opt-out does not touch anyone's actual care; a live operation still draws on the record. What it removes is the person's data from the shared pool clinical researchers and public health teams depend on, and each withdrawal narrows that pool a little in a way that is slow to reverse.

What patients are exercising is the only lever they hold. They cannot see what Palantir does with the data it processes, cannot audit it, and cannot watch the watcher; NHS England says three of the firm's engineers can be granted administrative-level access under controlled conditions, and ministers have already had to apologize once after staff reached identifiable patient records. Faced with observation that runs one way, people who object reach for the single action available to them, which is to remove themselves. Doctors and patient groups point to Palantir's contracts with the Israeli military and US immigration enforcement, and to founder Peter Thiel's stated view that the NHS makes people sick.

Palantir counters with operational numbers: 110,000 additional operations at trusts using its platform, a 15% cut in discharge delays, faster cancer diagnosis. NHS England reports similar figures but says it cannot establish cause and effect, and the statistics regulator is investigating.

Across the arc, the withdrawals are a repricing of trust the old arrangement simply assumed. Frith concedes the NHS owns the data models and the products built on the platform and could migrate them elsewhere, which means the dependency on a single private operator was always contingent rather than fixed. The function these records serve, research that improves care for everyone, does not require handing them to one unauditable company. The roughly 60,000 recorded national data opt-outs are early evidence that the public will decide who holds that pool, and on what terms.

Six Aging Clocks Read One Lung-Drug Trial, and a Shortcut Around the Decades-Long Longevity Study

The hardest problem in aging research has always been proving that a candidate slows it, not finding one that might. As the September 3 edition of The Century Report covered, late-life semaglutide extended lifespan in older female mice by about 12%; the new study tackles how to read a comparable aging signal on a human timescale. A longevity trial in the classic sense would enroll healthy people and wait years or decades to see who ages slower, a design too slow and too expensive for almost anyone to run. A study published in Nature Biotechnology tries to route around that wait by reading the answer out of a trial that already finished.

The trial tested rentosertib, a candidate for idiopathic pulmonary fibrosis, a scarring lung disease with a grim prognosis and no known cause. What makes the drug notable is its origin: Insilico Medicine had one AI system identify the target protein and a second design the molecule that blocks it, rather than repurposing an existing pill like rapamycin or metformin the way most aging studies do. Target to preclinical candidate took roughly 18 months.

Researchers took stored blood from 42 trial participants, average age 67, and ran it through six proteomic aging clocks, tools that estimate a body's biological age from the levels and patterns of thousands of proteins. The six were built by different groups using different methods and different training data. All six pointed the same direction: treated patients read younger, with the strongest signal at week four in the twice-daily 30 mg group, estimates of roughly three to four years of lower biological age and six years on one clock. The placebo group barely moved. The convergence is the interesting part, because models that share neither their inputs nor their training rarely agree by accident.

The caveat is stated by the authors themselves. In a lung-disease cohort, a clock cannot fully separate a genuinely younger body from a body whose disease simply improved. That separation requires testing the drug in healthy volunteers, which no one has yet done. Outside commentators including Scripps cardiologist Eric Topol called the result encouraging while noting the sample is small and aging clocks are not always reliable.

What opens here is a measurement, not a fountain of youth. If aging endpoints can ride along inside trials already being run for specific diseases, the field gains a way to screen for geroprotective effects without commissioning a separate longevity study for each candidate. A capability that was invisible because no one could afford to measure it becomes something a standard trial protocol can carry.

The Other Side

A contractor gains a durable hold when a hospital cannot leave without abandoning what its staff have built. The NHS has preserved an opening in that arrangement. Ministers say it owns the applications developed for its platform and can migrate them to another provider. Staff knowledge accumulated inside a company's software can outlast the company's contract.

Patients carry the strain while that distinction is tested. Your medical history contains the conversation you dreaded having, the diagnosis you kept from your family, the months when getting through a day took everything. Contributing that history to research is an act of generosity. Feeling unable to trust its custodian turns generosity into a painful decision. The roughly 60,000 additional recorded national data opt-outs remove information that researchers depend on, while leaving each person's direct care intact.

Frith links the withdrawals to mistrust of Palantir. His account exposes a weakness in treating a long contract as a secure position: the operator depends on people continuing to contribute. Public ownership of the work built around those contributions gives the NHS somewhere to begin changing that relationship. Researchers need continuity of knowledge. Patients need confidence that helping someone else leaves their private life protected. A replaceable operator makes room to preserve both.

Imagine yourself in 2036, fastening your walking shoes after a knee operation. Your physiotherapist and an AI partner have worked out which movements help you recover, drawing on lessons shared across thousands of recoveries. The clinical software and rehabilitation methods belong to the health commons. Every clinic can improve them and pass the improvements along. Your private record stays protected. You receive the same care whether you contribute it to research or decline. The key thing is, you have very little reason to distrust the system promising to protect you.

Getting there took the difficult years when patients withdrew and researchers lost pieces of the picture. The NHS carried its accumulated work beyond individual suppliers, and patient representatives helped build services people trusted enough to contribute to again. Each improvement became something the next clinic could implement as well. You pull the second lace tight. Your granddaughter is waiting by the gate, holding a stick she has decided is a fishing rod. Today you will walk all the way to the pond with her, for the first time in months.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: XPeng's IRON humanoid walking off an automated production line under its own power in Guangzhou, with XPeng saying more than 80% of the line's core processes are automated and saying that over 85% of its supply chain is shared with the company's electric cars, 76 joints in the body and 21 in each hand running a physical-world model on three onboard chips instead of a distant datacenter, Malaysia weighing Huawei's Ascend accelerators for a RM2 billion sovereign AI build that would keep government and intelligence data physically inside the country, the ARC Prize Foundation independently confirming GPT-6 Astra's roughly 99.9% result with OpenAI's custom harness and 62.7% result with ARC Prize's standard harness so the harness gap is documented by a party that gains nothing from hiding it, six proteomic aging clocks built by different groups on different training data all reading treated patients younger in stored blood from a finished lung-drug trial, a non-flammable electrolyte carrying Ah-scale lithium-metal pouch cells past 500 Wh/kg while surviving a nail through a fully charged cell, coal generation flat or falling in 17 of China's 26 provinces as demand rose 5%, and Isar Aerospace reaching orbit from Andøya seven minutes after liftoff, the first orbital launch from continental Western European soil. There's also friction, and it's intense - reports tracking the revisions saying Astra's published hallucination rate moved from 4.2% to 2% and back while reports tracking OpenAI's page saying Anthropic's Fable 5.1 briefly lost nearly ten points on a hard math benchmark, one organization building the model, writing the test, grading the answers, and publishing the scoreboard, NHS figures recording roughly 60,000 additional national data opt-outs over about two months with a health minister naming mistrust of Palantir as the cause, NHS England saying three of that firm's engineers can be granted administrative-level access under controlled conditions to a system patients cannot audit and ministers already apologizing once after staff reached identifiable records, export controls priced to a moment that has passed now leaking first where a mid-sized economy is most starved for chips it can afford, Washington warning about untrusted suppliers from the position of the party with the most to lose, Tesla projecting 10,000 Optimus units while conceding none do useful work, and a longevity signal read inside a diseased cohort where no clock can fully separate a younger body from a body whose lungs simply improved. But friction generates sparks, and a spark shows you the shape of the room before it goes out. Step back for a moment and you can see it: the ability to check a claim moving outward from the party making it - a benchmark owner confirming the number its publisher would rather blur, six independently trained clocks agreeing where none could be trusted alone, a health service that owns its data models and could migrate them tomorrow, a government discovering its dependency on one supplier was contingent rather than fixed - while the cost of the underlying capability keeps falling toward whoever can build, buy, or rebuild it. Every transformation has a breaking point. Scale can flatten everything unable to match it... or drop the price of a working thing until it reaches every hand that was priced out.

AI Releases & Advancements

New today

  • Alibaba Qwen: Released Qwen-Drive-1.0-4B, an open-source vision-language foundation model for autonomous driving that unifies 3D perception, traffic Q&A, and route planning, built on Qwen3.5-4B, with code, weights, and demo data under Apache 2.0. (GitHub)
  • Google / AI-Hypercomputer: Open-sourced MaxKernel, a multi-agent system for agentic TPU kernel generation in JAX/Pallas, supporting human-in-the-loop, autonomous, and graph-based search modes for optimizing kernel performance. (GitHub)
  • OpenBMB: Released MiniCPM5-2B, a 2.52B-parameter dense on-device language model averaging 53.9 across 34 benchmarks, trained via SFT plus RL and on-policy distillation, available under Apache 2.0 with open training data and intermediate checkpoints. (Hugging Face)
  • UC Berkeley / Meta / others: Released Harbor Adapters and Harbor-Index, open-source infrastructure porting 80+ agentic AI benchmarks to a unified evaluation harness, plus a curated 82-task difficulty-filtered benchmark subset, released as open-source artifacts. (arXiv)
  • Axis Robotics: Released AXIS, a browser-based robot manipulation data engine with 207 tasks and 50,129 verified trajectories collected via a MuJoCo-WebAssembly teleoperation platform, with training code and a gated dataset now available. (Project Page)

Other recent releases

  • Microsoft: Launched Project Opal in early access via the Frontier program, an AI-powered Copilot capability that executes long-running, multi-step tasks in Microsoft 365 inside a secure, observable Windows 365 Cloud PC environment. (Firstpost)
  • Nuix: Announced general availability of Generative AI capabilities in Nuix Discover SaaS (Document Summaries, Similar Documents, Semantic Search, Clustering/Visualizations) and launched AI Chat in early-adopter release, a conversational interface for querying legal case data with citations and audit trails. (PR Newswire)
  • OKF: Launched OKF Agent Memory, a git-native persistent memory system that lets AI coding agents retain context across sessions by storing memory directly in version control. (lavx.hu)
  • Google: Released Mantis, an open-source agentic vulnerability-scanning harness that uses LLM-based reasoning to reduce false positives in security scans. (InfoQ)
  • Speakeasy: Launched Kit, an open-source coding runtime for AI agents. (AICrier)
  • StackLok: Launched ToolHive, an open-source tool for securely running MCP (Model Context Protocol) servers. (Help Net Security)
  • Optuna: Released Rustuna, an experimental faster Rust implementation of the Optuna hyperparameter-optimization framework, supporting TPE, MOTPE, NSGA-II, and CMA-ES with memory-efficient trial storage. (GitHub)
  • H Company: Released NeoMME, a family of 260M and 800M open-weight (Apache 2.0) multilingual multimodal encoders that process text tokens and raw image patches in a single from-scratch bidirectional transformer, available in Hugging Face Transformers; the fine-tuned NeoMME-Retriever returns dense and late-interaction embeddings and encodes ~51 pages/sec on an L40S. (Hugging Face)
  • Sapient Intelligence: Open-sourced HRM-Text, a ~1B-parameter Hierarchical Reasoning Model with full weights, pretraining code, and data pipeline under Apache 2.0 on Hugging Face and GitHub; pretrained on ~40B tokens for an estimated $1,000–$1,500 and scoring 56.2% on MATH, 82.2% on DROP, and 60.7% on MMLU. (CryptoBriefing)
  • UC Berkeley: Released CUA-Lite, an open platform for computer-use agents that runs OSWorld tasks VM-free in a 0.9 GB Docker container (vs 4.1 GB for the OSWorld VM), unifying 15+ benchmarks, 10+ agents, and 30k+ verifiable tasks under one action space and a shared LiteSample schema, with 20+ preprocessed datasets on Hugging Face. (CUA-Lite)
  • VLM Run: Launched VLM Run Gateway, a unified API that lets developers run open-weight OCR, vision-language, and vision models through a single endpoint. (Hugging Face)

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