This week in AI is a story told in two languages — capital and concrete — and the loudest sentences came from China and from the floors where robots now clock in.
Twelve stories sat at the top of the pile, weighted on purpose toward China, Africa, and the machines moving into the physical world. The facts come first, then my take. Roughly 80% reporting, 20% opinion.
1. China is putting an 8,500-strong robot army on its power grid
On 27 April 2026, the State Grid Corporation of China set aside 6.8 billion yuan — close to US$1 billion — to buy around 8,500 AI-powered robots this year, Caixin reported. About 5,000 of the machines are quadruped "robot dogs" sent to inspect substations and patrol transmission lines across remote terrain.
The fleet also includes humanoid and dual-arm robots for live-line work on ultra-high-voltage cables — jobs that get people killed. State Grid named Unitree, Deep Robotics, AgiBot, UBTech, and Fourier Intelligence as suppliers. Counting China Southern Power Grid, sector spending on embodied robots should pass 10 billion yuan in 2026, the South China Morning Post reported.
💡Our take
A power grid is the least forgiving classroom a robot can train in. China is not piloting embodied AI — China is buying it by the thousand and pointing it at the most dangerous work first. The race for robotics will be won on factory floors and substation gantries, not on demo stages.
2. Unitree filed to go public and wants to ship 20,000 humanoids
Unitree Robotics — the Hangzhou firm whose somersaulting machines went viral at China's Spring Festival gala — had an IPO application accepted by the Shanghai Stock Exchange on 20 March 2026, seeking to raise about 4.2 billion yuan (US$610 million), per the South China Morning Post.
Founder Wang Xingxing wants to ship as many as 20,000 humanoid robots in 2026, up from roughly 5,500 in 2025, eWeek reported. Unitree's filing claims a 32.4% share of global humanoid shipments and 2025 revenue of 1.708 billion yuan, up 335% on the year.
💡Our take
Five years ago a humanoid robot was a research demo. Unitree is now asking public markets to fund a production line for twenty thousand of them. The somersaults were marketing; the prospectus is the real story — China is industrialising the humanoid before the West has agreed what one is for.
3. Moonshot AI raised $2bn as China’s open-weight bet paid off
On 7 May 2026, Beijing-based Moonshot AI closed a $2 billion round at a $20 billion valuation, TechCrunch reported, making Moonshot China's most heavily funded large-language-model startup. Annual recurring revenue topped $200 million in April, driven by paid subscriptions and API demand.
Moonshot's Kimi K2.6 — a trillion-parameter model released under a permissive open licence — sits alongside Alibaba's Qwen, MiniMax, and Z.ai in a wave of Chinese open-weight releases closing the gap on American labs faster than almost anyone predicted, as MIT Technology Review documented.
💡Our take
The American AI story gets told in closed models and trillion-dollar valuations. The Chinese AI story is increasingly told in open weights anyone can download and run. For builders across the Global South priced out of frontier APIs, a capable open model from Beijing is not a geopolitical statement — the model is a budget line that finally works.
4. Baidu’s robotaxis crossed 190 million driverless kilometres
Baidu's Apollo Go has now driven more than 300 million autonomous kilometres, 190 million of them fully driverless, and passed 20 million paid trips by late February 2026, CnEVPost reported. The service runs across more than a dozen Chinese cities.
Apollo Go is going global too. Baidu launched commercial robotaxis in Dubai in April 2026 with a plan to scale from 50 vehicles to over 1,000, and is preparing London pilots with Uber and Lyft, CleanTechnica reported.
💡Our take
Robotaxis are the most public test of physical-world AI. A model that hallucinates a citation is embarrassing; a model that hallucinates a pedestrian kills someone. Baidu has quietly logged more driverless kilometres than most Western rivals will manage this decade — autonomy, like manufacturing, is migrating east.
5. Microsoft’s $1bn Kenya data centre stalled on power
A $1 billion AI data centre that Microsoft and Abu Dhabi's G42 planned to build in Kenya has stalled, Semafor reported on 6 May 2026, after talks broke down over who guarantees payment for the electricity. President William Ruto put the scale plainly.
Kenya's entire installed capacity is roughly 3,000 megawatts. The geothermal project in the Rift Valley was meant to begin at 100 megawatts and scale toward a gigawatt, DatacenterDynamics reported. Talks have not collapsed outright, but the deal is on ice.
We would need to switch off half the country to keep it running.
— Kenyan President William Ruto, on the proposed Microsoft and G42 data centre (reported by Semafor)
💡Our take
Here is the catch nobody puts on the keynote slide: AI is a physical industry, and the physics is electricity. The same compute hunger straining grids in Virginia hits harder where the grid is thinner. Kenya’s dilemma — host the future or keep the lights on — is the one most of Africa will meet next.
6. Cassava switched on “Africa’s first AI factory”
Strive Masiyiwa's Cassava Technologies began deploying what Cassava calls Africa's first Nvidia-powered AI factory in South Africa, installing 3,000 Nvidia GPUs by June 2026, DatacenterDynamics reported. A further 9,000 GPUs are planned across Kenya, Nigeria, Egypt, and Morocco over four years, in an estimated $720 million build-out.
The facility sells GPU-as-a-Service and AI-as-a-Service, so African developers can train and run models on the continent instead of routing through data centres abroad, Cassava said. The build extends the firm's CAIMEx model exchange, launched in 2025.
💡Our take
Cassava and the stalled Kenya deal are the same story from opposite ends. Compute sovereignty is not a slogan — sovereignty is whether an African researcher can rent a GPU without a credit card in another country’s currency. Masiyiwa is betting Africa builds its own floor rather than renting someone else’s ceiling.
7. The AfDB says AI could add $1 trillion to Africa’s economy
An African Development Bank report, "Africa's AI Productivity Gain", projects that inclusive AI deployment could add up to $1 trillion to the continent's GDP by 2035 — lifting output from $4.23 trillion to $5.23 trillion, the bank said. The study estimates 35 to 40 million new digital jobs.
Five sectors — agriculture, retail, manufacturing, finance, and health — would capture 58% of the gains, about $580 billion. The bank sets a three-phase roadmap, opening with an "ignition" phase that runs from 2025 to 2027.
💡Our take
A trillion-dollar projection is easy to print and hard to earn. The AfDB number matters only if the power, the chips, and the skills land — which is exactly why the Kenya and Cassava stories sit on either side of it. The opportunity is real; the bottleneck is concrete, not imagination.
8. 1X opened orders for NEO, a $20,000 home humanoid
The robotics firm 1X opened pre-orders for NEO, a humanoid robot built for the home, The Robot Report noted, priced at $20,000 outright or $499 a month, with first deliveries to US homes in 2026. NEO weighs 66 pounds, lifts more than 150, and runs at 22 decibels — quieter than a fridge.
From day one NEO can open doors, fetch items, and switch off the lights, with new abilities promised through software updates, 1X said. Markets beyond the US are slated for 2027.
💡Our take
The factory humanoid and the home humanoid are diverging fast. NEO is the first credible attempt to put one in a living room at a car-sized price. Whether households want a soft-bodied robot folding laundry is the open question — but the hardware is no longer the thing standing in the way.
9. Isomorphic Labs raised $2.1bn to design drugs with AI
On 12 May 2026, Isomorphic Labs — the Google DeepMind spin-out commercialising AlphaFold — raised $2.1 billion in a Series B led by Thrive Capital, Isomorphic Labs said, with Alphabet, GV, MGX, Temasek, CapitalG, and the UK Sovereign AI Fund joining. The money funds a drug-design engine called IsoDDE.
Isomorphic now aims to put an AI-designed drug into clinical trials before the end of 2026, working with Novartis, Eli Lilly, and Johnson & Johnson, Bloomberg reported. Hiring spans London, Cambridge in Massachusetts, and Lausanne.
💡Our take
Here is physical-world AI of a different kind — molecules, not motors. If a model can design a working drug, the proof is not a benchmark but a patient. Isomorphic moving toward the clinic is among the most consequential AI deadlines of 2026, and almost nobody is watching the calendar.
10. The US will now test frontier models before they ship
On 5 May 2026, the US Commerce Department's Center for AI Standards and Innovation signed pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI, CNBC reported, adding three labs to a programme that already covered OpenAI and Anthropic. All five frontier labs now submit models for government testing before public release.
The agreements are structured as formal research collaborations and have run more than 40 assessments to date, Axios reported. The move marks the Trump administration stepping deeper into AI oversight.
💡Our take
Pre-deployment testing is the governance idea I have argued for repeatedly — check the work before shipping, not after harm. The detail to watch is independence: an evaluator paid in access rather than authority is a referee who can be benched. Structure matters more than the signing ceremony. <a href="https://humphreytheodore.com/writing/governance-not-models-may-2026-ai-pattern" style="color:#D97706;text-decoration:underline;text-underline-offset:3px">I wrote about this pattern earlier in May</a>.
11. Google answered with cheaper models and a personal agent
At Google I/O on 19 May 2026, Google introduced Gemini 3.5 Flash — a lighter model priced at roughly a third of comparable frontier systems — and Gemini Spark, a general-purpose agent that reasons across a user's connected apps, CNBC reported. Spark reaches Google AI Ultra subscribers first.
The announcements are Google’s bid to keep pace with OpenAI and Anthropic while undercutting both on price. Cheaper capable models widen access well beyond wealthy markets.
💡Our take
The frontier is splitting into two races: who is smartest, and who is cheapest. Google just made a loud move on price, and price is the lever that actually reaches the Global South. A model at a third of the cost is not a small story in Lusaka or Lagos.
12. DeepSeek tied its new model to Huawei’s chips
DeepSeek released a V4 model on 24 April 2026 at aggressively low prices and, more tellingly, in close integration with Huawei's homegrown AI chips, Fortune reported. The pairing lets Chinese developers run a frontier-class model without Nvidia silicon.
V4 undercut Western APIs by a wide margin, continuing the price deflation DeepSeek set off a year earlier. The model is open-weight, like most of its Chinese peers.
💡Our take
Here is the story underneath the stories: China is assembling a full stack — model, chips, and the grid robots to power and patrol the data centres — that does not depend on the West. Export controls were meant to slow that down. The DeepSeek and Huawei pairing suggests the controls accelerated it.
What I make of it, in one line
Twelve stories, two clear currents. China is building the physical stack of AI — chips, robots, autonomy — at industrial scale, while Africa fights to own its compute rather than rent it. The software race takes the headlines; this week belonged to concrete, electricity, and silicon.
For the longer argument that intelligence only matters once it touches the physical world, see The Body Gap.
Frequently Asked Questions
These are the questions readers have been asking since this week’s news broke. Short answers follow, drawn from the primary reporting linked above and the underlying company and government announcements.
What is embodied AI, and why does China lead it?
In short, embodied AI is artificial intelligence that acts in the physical world through robots and vehicles rather than only generating text. The answer to China’s lead, simply put, is scale: data from the State Grid procurement shows a single Chinese utility buying 8,500 robots in one year, and research in Unitree’s prospectus reveals a 32.4% global share of humanoid shipments.
How does the stalled Kenya data centre affect Africa’s AI plans?
The Kenya project shows that power, not ambition, is Africa’s binding constraint. According to Semafor, Microsoft and G42 paused a $1 billion build because Kenya’s roughly 3,000-megawatt grid could not guarantee the load. The key is that Cassava’s South African AI factory and the African Development Bank’s $1 trillion projection both depend on solving the same electricity problem.
Why is open-weight AI from China significant for the Global South?
Open-weight models can be downloaded and run without paying a per-token frontier fee. According to TechCrunch, Moonshot raised $2 billion on exactly that demand, and DeepSeek’s V4 shows prices still falling. In other words, a capable model from Beijing is often the only one a developer in Lusaka or Lagos can actually afford to run.
Who is funding the AI drug-discovery push?
Isomorphic Labs, the DeepMind spin-out, is the clearest case. The data shows a $2.1 billion Series B led by Thrive Capital, with Alphabet, Temasek, and the UK Sovereign AI Fund joining. The answer to why the round matters is the deadline: Isomorphic aims to reach human clinical trials before the end of 2026.
What are the real risks in this week’s AI news?
Analysis of the week reveals three durable risks. First, concentration — China’s full-stack push and America’s five-lab testing club both narrow who controls AI. Second, energy — the Kenya stall shows compute outrunning the grid. Third, oversight that looks firmer than the evidence supports, where a signed testing deal matters less than the independence of the testers.
Sources and further reading
Every story above links to its source at first mention. The full list, in order:
CAISI’s frontier-model testing deals: CNBC and Axios.
Google I/O’s Gemini announcements: CNBC.
DeepSeek’s V4 and the Huawei pairing: Fortune.
