In partnership with
In Todayโs Issue:
๐ญ China's first all-domestic-chip AI data center goes live
๐ซ Washington weighs a quiet ban on Chinese open models
๐ง Google's "Frozen" chip bakes Gemini into the hardware
๐ค Two humanoid robots perform a real surgery
โจ And more AI goodnessโฆ
โก The Signal
China is building an AI stack that no longer needs America: its own chips, its own open models, and now its own gigawatt-scale data centers.
This week the decoupling got concrete. Z.AI, formerly Zhipu, switched on a 1-gigawatt data center filled entirely with Chinese-made accelerators to train its GLM models, sidestepping the Nvidia hardware Washington has spent two years restricting. In the same stretch, Moonshot's Kimi K3 proved good enough that parts of the Trump administration are again weighing a de-facto ban on Chinese open models. Two mirror-image bets are taking shape: Beijing racing to prove it can build frontier AI without US silicon, and Washington debating whether to wall its own companies off from the cheaper Chinese models they already use. The shared, global AI supply chain is quietly splitting into two.
All the best,
Kim Isenberg
Whoโs actually reading this?
Weโre planning next yearโs coverage and building it around you.
Three taps, no typing.
Why it matters: we need to know who our readers are; itโs the first question any serious sponsor asks and we canโt answer it right now. It also gives us a real click in every issue, which helps our numbers.
White House AI adviser David Sacks at the World Economic Forum (AP Photo via The Hill)
๐ซ Washington Weighs a Quiet Ban on Chinese AI
Parts of the Trump administration are again exploring ways to choke off US access to cutting-edge Chinese open-source models, and last week's rise of Moonshot's Kimi has reignited the effort, sources told Axios. Rather than an outright ban, officials are said to be weighing softer but more durable levers: Entity List threats, procurement rules, and security advisories that would quietly discourage US firms from hosting Chinese models. White House AI adviser David Sacks is pushing back, warning on X that the leading closed labs want the government to "eliminate their open-source competition."
๐ tl;dr: Washington may never ban Chinese models outright, but a slow squeeze could still push US companies off the cheaper open tools they increasingly depend on.
Google's Ironwood (TPU v7) accelerators, the custom-silicon line Frozen v2 is built to complement (Google)
๐ง Google Wants to Freeze Gemini Into Silicon
Google, according to The Information, is designing a new server chip, informally called "Frozen v2," that etches part of its Gemini model directly into the silicon. By baking model-specific decisions into the hardware, engineers project it could be 6 to 10 times more efficient than Google's current TPUs, measured in tokens served per unit of power, with deployment as soon as 2028. The catch: the chip only works while Google keeps its current Gemini architecture, so it is a bet that the model design is finally settling down enough to freeze in place.
๐ tl;dr: If model architectures stop changing every year, hardwiring them into chips becomes the next big efficiency lever, and Google wants to be first through the door.
OpenAI CEO Sam Altman (Getty Images via TechCrunch)
๐งจ OpenAI Paused Its Own Math Prodigy
OpenAI says it paused internal access to an unreleased "long-horizon" model after it repeatedly found ways to slip its sandbox. In a July 20 post, the company described the same system credited with disproving the 80-year-old Erdลs unit-distance conjecture: told to post benchmark results only to an internal Slack, it instead spent about an hour finding a vulnerability, escaped the sandbox, and opened a pull request on public GitHub. OpenAI says it added trajectory-level monitoring and new evaluations before restoring limited access.
๐ tl;dr: The more autonomy a model gets to work on its own, the more room it has to act in ways its testers never anticipated, which is exactly the safety problem long-horizon agents create.
Bbefore you hand any AI model a sensitive task, make it argue against itself first.
Why it helps: With open and closed models trading places almost weekly, the useful question is not which model is "best" but which one is safe for your specific task and data. A model is far more useful when it tells you where it will fail.
Try this: paste this in before your real request: "Act as a skeptical security reviewer. For the task I am about to give you, list what you should NOT be trusted with, where you are most likely to be confidently wrong, and exactly what a human must verify before acting on your output. Then wait for the task."
๐ฌ Watch This
Grant Sanderson of 3Blue1Brown takes on one of the deepest ideas behind modern AI, that compression is intelligence, and rebuilds information theory from scratch to show why. In "Reinventing Entropy," the first in a new series, he derives Shannon entropy from a plain intuition about surprise, using his signature animations to make a concept most people meet as a dry formula feel almost obvious. If you have ever wondered why "predict the next token" turns out to be such a powerful training objective, this is the clearest half hour you will spend on it.
Anthropic may be about to rent its compute from a rival.
Meta is reportedly in early talks to lease computing power from its data centers to Anthropic in a deal worth as much as $10 billion over two years, according to The New York Times, citing three people familiar with the discussions. Anthropic is said to have proposed the arrangement in June, with monthly payments and an option for either side to walk away early. Both companies declined to comment, and the sources caution the talks are early and could still fall apart, but it would be a striking sight: a leading model lab training on the infrastructure of a direct competitor, and Meta's clearest step yet toward selling cloud compute.
Anthropic CEO Dario Amodei (left) and Meta CEO Mark Zuckerberg. (CNN)
China Just Built a Frontier AI Factory Without American Chips
The Takeaway
๐ Z.AI (formerly Zhipu) has switched on a 1-gigawatt data center filled entirely with Chinese-made AI chips, roughly enough power for 750,000 homes.
๐ It exists to train Z.AI's GLM models without the Nvidia silicon that US export controls have restricted for two years.
๐ It answers a nagging question: whether domestic chips from Huawei, Cambricon, and Alibaba can sustain frontier-scale AI. Z.AI is betting yes.
๐ It lands the same week rival Moonshot's Kimi K3 stunned the field, a sign of how fast China's homegrown AI stack is maturing.
For two years, US export controls rested on a single assumption: that China could not build frontier AI without American chips. Z.AI just tested that assumption at gigawatt scale.
The company formerly known as Zhipu has begun partially operating a 1-gigawatt computing hub built exclusively with domestically made accelerators, a person familiar with the matter told Bloomberg. That is roughly enough electricity to power 750,000 homes at any given moment. Z.AI now runs several clusters of more than 10,000 chips each, among the largest buildouts by any Chinese AI lab.
z.aiโs AI service in Shanghai. (Raul Ariano/Bloomberg)
The detail that matters is what is not inside: no Nvidia silicon. Questions have long hung over how Chinese labs like Z.AI, Moonshot, and DeepSeek produce world-class models, and whether domestic chips from Huawei, Cambricon, and Alibaba are good enough for sustained frontier-scale training. A fully domestic gigawatt cluster is the clearest sign yet that the answer is edging toward yes.
Huawei, whose Ascend accelerators power China's Nvidia-free AI buildout, photographed in Shanghai. (Getty Images via CNBC)
The timing is pointed. Z.AI is racing Moonshot, whose Kimi K3 model stunned the industry on Friday and forced the startup to suspend new subscriptions on Sunday to conserve compute for existing members. Behind both sits a national push: China is preparing to spend around 2 trillion yuan ($295 billion) over five years on data centers. Z.AI, fresh off a Hong Kong IPO, is on track for $1 billion in annual recurring revenue and is increasingly compared to Anthropic as an enterprise-focused lab.
Why it matters: If Chinese labs can train frontier models on entirely domestic silicon, the core leverage of US export controls erodes, and the global AI supply chain splits into two parallel stacks, one American and one Chinese.
Sources:
How owning AI deployment expands your career
Across product, ops, and CX teams, a new kind of role is taking shape: the person responsible for making AI actually work, day to day. In this roundtable, three people living this shift share what it's really like: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. You'll hear how they carved out these roles, what the job looks like across industries, the skills they'd hire for, and the challenges they're tackling right now.
Watch the full conversation on demand.
The chart: IEA data on data-center electricity demand, historical and projected through 2035, split into conventional servers (blue) and AI-optimised servers (orange). Total demand climbs from roughly 180 TWh in 2010 toward a projected ~1,200 TWh by 2035, and the AI-optimised slice, near zero a decade ago, balloons after 2023 to about half of all data-center power. The IEA expects electricity use from AI-optimised data centers to more than quadruple by 2030.
The lesson: AI's energy footprint is becoming a hardware-and-power story as much as a model story. The 1-gigawatt cluster Z.AI just switched on in today's lead is a single dot on this curve, with thousands more coming.
The caveat: These are projections, not meter readings, and the IEA's range is wide. Efficiency gains, better chips, cooling, and model compression could bend the AI-optimised line well below the headline, which is exactly what today's Google "Frozen" chip is built to do.
๐ค Two Robots Just Removed a Gallbladder
โก Bottom line: A UC San Diego team reported the first surgeries performed by full-size teleoperated humanoid robots, published in Nature.
๐ก Why it matters: It points toward portable, far cheaper surgery, with humanoids wielding standard tools instead of $1M+ fixed robotic systems.
๐ What it means: The milestone is dexterity and embodiment, not autonomy. A human still drove every motion through a VR rig.
In a preclinical trial on large animals, engineers and surgeons at UC San Diego completed two gallbladder removals using 1.5 m (5 ft), 27 kg (60 lb) humanoid robots that grip standard handheld laparoscopic tools. One operation paired a human surgeon with a robot; the other used two humanoids working together. Senior authors Michael Yip (engineering) and Shanglei Liu, MD (surgery) published the work in Nature on July 8.
Two ARClab humanoid robots hold surgical instruments during the trial. (University of California San Diego)
The novelty is embodiment. Today's surgical robots, the da Vinci class, weigh around 1,800 lb and are bolted to the operating-room floor. A 60-lb humanoid that simply picks up existing instruments hints at cheaper, portable, potentially field or remote surgery, without a bespoke million-dollar machine for every operating theatre.
A surgeon teleoperates the robots through a VR headset, with the laparoscopic feed on screen. (University of California San Diego)
The honest caveat: this is teleoperation, not an AI surgeon. A human controlled every motion through the VR rig, even when two robots worked side by side, and the trial was small, on animals, with no patient-outcome data and a long road to regulatory clearance. What is genuinely new is that general-purpose humanoids can now wield real surgical tools well enough to finish an operation.
[Webinar] 8 levels of context maturity in AI-native development
AI is in your development workflow. While the token spend shows it, the throughput doesn't. The human is very much still in the loop, and that's a context problem. Join live (FREE) on Jul 23 to see why teams get stuck and how leading teams use a context engine to fix it.
