SoftBank's Son says AI will need $5 trillion per year by 2040, dismisses bubble talk ↗
Masayoshi Son claimed AI development will eventually require $5 trillion in annual investment, arguing that warnings of an investment bubble are simply absurd.
He expects AI revenue to represent 20% of global GDP and power demand to reach 3 terawatts. Son did not explain how he calculated those figures, though... a rather important loose screw in the machine.
AI startup Reflection signs over $1 billion computing deal with Nebius ↗
Reflection secured more than $1 billion in computing capacity from Nebius, including access to Nvidia's latest chips. The agreement adds to a separate capacity deal with SpaceX.
Founded by two former Google DeepMind researchers, Reflection is building open models positioned against closed offerings from OpenAI and Anthropic. Compute is becoming the true moat here - or the very expensive castle wall.
US to launch AI and cybersecurity coordination group, White House says ↗
The US is creating a coordination group linking AI developers with hospitals, banks, energy operators and other essential-service providers.
Members will share vulnerabilities discovered by advanced AI systems and coordinate fixes rather than duplicate work. Open-model developers are included too, somewhat unexpectedly, although the participating companies were not identified.
DeepMind CEO calls for an independent standards body to regulate frontier AI ↗
Demis Hassabis proposed an independent, technically staffed organisation that would test frontier models and establish release standards.
Labs could initially submit models for review up to 30 days before launch. The body would be government-backed but industry-funded - independent, apparently, though that funding arrangement is a slightly wobbly chair.
New York becomes the first state to enact a data center moratorium ↗
New York paused new environmental permits for data centres exceeding 50 megawatts, giving officials up to 12 months to develop stronger rules.
The review will examine electricity prices, water consumption and air quality. A separate legislative proposal goes further, using a lower 20-megawatt threshold and adding renewable-energy and labour requirements.
Meta accused of using biased AI targeting for mass layoffs ↗
Twenty-six former Meta employees allege that internal AI systems helped rank workers for dismissal without properly accounting for parental or medical leave.
They claim this disproportionately penalised people exercising protected rights. Meta says the allegations lack merit and insists workforce decisions were made by humans, not AI - a distinction the lawsuit sharply disputes.
The real AI race may no longer be at the frontier ↗
Chinese open-weight models reportedly captured 41% of Hugging Face downloads during the spring, while six Chinese models occupied OpenRouter's top six positions.
Open models also handled nearly a third of AI requests measured by Vercel. The data covers only parts of the market, admittedly, but cheaper, customisable models are clearly consuming more of the production workload.
Regulatory action on chips, AI is coming, Commerce official says ↗
A US Commerce Department official said further regulation covering artificial intelligence and semiconductors is coming.
The administration reportedly does not intend to replace the existing AI diffusion rule, which limits how many advanced chips certain countries can receive. So, tougher action perhaps - but built on the old scaffolding.
FAQ
Why does Masayoshi Son believe AI investment could reach $5 trillion per year?
Masayoshi Son expects artificial intelligence to become a major force in the global economy, demanding vast investment in chips, data centres, energy infrastructure and model development. He also predicts that AI power demand could reach 3 terawatts. However, he did not provide a detailed calculation to support the $5 trillion annual investment estimate, leaving considerable uncertainty around the projection.
Why are AI startups signing billion-dollar computing deals?
Training and operating advanced AI models requires access to large volumes of specialised computing hardware. Reflection’s agreement with Nebius gives it capacity powered by Nvidia’s latest chips, adding to another deal with SpaceX. For startups competing with established laboratories, securing compute early can support model development, reliability and growth, though it also creates substantial financial commitments.
How could the new US AI cybersecurity coordination group work?
The proposed group would connect AI developers with organisations that operate hospitals, banks, energy systems and other essential services. Participants could share vulnerabilities identified by advanced AI tools and coordinate fixes, rather than duplicating the same work across separate organisations. The initiative is expected to include open-model developers, although the participating companies and detailed operating structure have not yet been identified.
What would an independent frontier AI standards body do?
The organisation proposed by DeepMind CEO Demis Hassabis would test powerful frontier models and establish standards governing their release. AI laboratories could initially submit models for assessment up to 30 days before launch. The body would be backed by government but funded by industry, raising practical questions about how its independence, technical authority and accountability would be protected.
Why has New York introduced a data centre moratorium?
New York has paused new environmental permits for data centres exceeding 50 megawatts while officials develop stronger rules. The review will examine their effects on electricity prices, water use and air quality. A separate proposal would apply requirements to facilities above 20 megawatts and could introduce renewable-energy and labour conditions, potentially creating a stricter framework for AI infrastructure projects.
Are open AI models becoming more important than frontier models?
Open models appear to be handling a growing share of practical AI workloads because they can be cheaper and easier to customise. Chinese open-weight models reportedly gained substantial visibility across Hugging Face and OpenRouter, while Vercel data also showed meaningful open-model usage. These figures cover only parts of the market, but they suggest that competition is increasingly shifting towards deployment, cost and flexibility.