Alibaba · Qwen · AI infrastructure
Alibaba would like ten trillion parameters, half a million chips and several power stations
Alibaba did not introduce a single AI product. It described a stack. At the model layer, the company says its next training target is 5 trillion to 10 trillion parameters. Architecture, data, inference-time reasoning and software matter as much as raw scale, but ten trillion is not a subtle target.
At the hardware layer, Alibaba calls the Zhenwu V900 China's most powerful AI chip and says it delivers three times the performance of its predecessor. The company says the design can scale to clusters of 500,000 accelerators; production and commercial availability are planned for the first quarter of 2027. Those performance and scale figures remain Alibaba's claims until independent benchmarks and deployments arrive.
Below both sits the power budget: more than 20 gigawatts of global cloud capacity by 2032. Generative AI began publicly as a box on a website. Three years later, one of the world's largest technology companies is planning models in trillions of parameters, machines in hundreds of thousands of processors and infrastructure in tens of gigawatts. AI has made the power station visible again.
Source: Reuters ↗Meta · Muse · Consumer agents
Meta released an agent and Wall Street rediscovered enthusiasm overnight
Meta's Muse assistant launched in the United States about two weeks ago and has climbed to the top of free-app rankings. Sensor Tower estimated 730,000 downloads over five days. Meta shares closed more than 11% higher Monday, helping revive technology stocks after a week dominated by safety warnings.
Muse matters because it is presented as more than a chatbot. Reuters describes it as capable of autonomous consumer tasks such as sending email, booking travel and helping sell a car. Those actions move the software from producing information toward changing state in the world: it sends the message, makes the reservation and posts the listing.
Early app rankings do not establish lasting use or profitable demand. They do show why slowing the AI race is hard. After a week of arguments about rogue agents and restraint, Meta shipped an agent many consumers wanted to try and investors rewarded it. Useful autonomy is worth money, even while its risks remain unresolved.
Source: Reuters ↗Payments · Shopping agents · Identity
The payment network would like to know whether you or your robot bought the television
AI-assisted shopping is becoming measurable. John Lewis says searches arriving through AI agents rose from 0.3% to 2.5% in a year. Most agents still help with discovery rather than autonomously completing purchases, but the payment industry is preparing for the next step.
Visa, Mastercard and Ant International have begun developing a common “Know Your Agent” framework so card networks, wallets, marketplaces and agent platforms can identify trusted purchasing agents. The idea addresses a basic problem: existing fraud systems assume a person is on the other end of a transaction. Agentic commerce adds software acting under some form of delegated authority.
The hard questions arrive after the futuristic demo. Did the user authorize that exact item and price? Was the agent manipulated? Which data did it reveal? Who disputes an accidental purchase? A useful system needs identity, authorization, audit trails and chargebacks. Giving the robot a wallet is easy. Explaining the charge to the bank is the product.
Source: Reuters on Know Your Agent ↗Source: Reuters on John Lewis ↗