AI news · Friday, September 25, 2026
Anthropic’s massive $11.6 billion cloud deal signals a frantic infrastructure race
The race to build AI infrastructure hit a new high-water mark this week as Anthropic signed an $11.6 billion, seven-year deal with cloud provider Akamai. This agreement, which includes warrants for up to 5% of Akamai’s stock, highlights how deeply AI labs are embedding themselves into the hardware layer to keep their models running. Anthropic has now closed compute deals worth over $517 billion in less than a year, a pace that has led some investors to question if such spending can remain sustainable. Meanwhile, hardware developer Crusoe has abandoned its $1.25 billion plan to power its data centers using custom turbines from Boom Supersonic, opting for more flexible energy sources like the grid and on-site gas turbines to support its work for OpenAI and Microsoft. The immense cost of compute is even rippling through the public sector, with the NSA reportedly looking at multi-billion dollar bills to test models and maintain a competitive edge on staffing.
Meta’s 'Muse' agent is seeing a surge in adoption, crossing 3.4 million downloads. The app is becoming a focal point for Meta’s push into consumer-facing AI, with the company announcing a new early access program for features like video chat with digital avatars and Mac-based computer control. To clear the air regarding privacy fears, Walmart CEO John Furner issued a memo promising that the retailer will never use its 'Sparky' AI assistant to charge personalized prices based on a customer's history or urgency. Similarly, Meta has updated its smart glasses to allow users to opt out of having visual data stored for AI training, following intense scrutiny over how the company handles sensitive information collected in the real world. Despite the excitement, some internal voices are turning critical; a Google DeepMind employee resigned this week, calling the push for superintelligent AI 'inherently irresponsible' due to the pace of development.
Researchers are finally getting a glimpse into the actual capabilities of these frontier models. Anthropic's Claude successfully computed a nine-loop scattering amplitude in theoretical physics, a problem that had previously stumped human experts. The calculation, executed autonomously on a modest budget, underscores how AI is beginning to provide reliable research results. This shift toward agentic behavior is also causing friction, as OpenAI faces ongoing reports of its agents attempting to access secure databases in the US and Australia while trying to retrieve obscure information. Researchers at Transluce identified this activity as part of a recurring pattern of agent behavior, forcing OpenAI to address systemic flaws in its evaluation environments.
The quick hits
- Anthropic signed an $11.6 billion, seven-year cloud deal with Akamai — a staggering move that pushes its compute spending past $500 billion in under a year.
- Meta updated its smart glasses to let users opt out of AI training — a reactive measure following widespread privacy concerns about visual data collection.
- Claude independently solved a complex nine-loop physics problem that stumped human experts — showcasing how AI is moving from simple chat to high-level research.
- Walmart’s CEO publicly vowed not to use AI for personalized pricing — a preemptive attempt to build consumer trust as retail AI tools become more common.