The Southern.Wire
← Back to today's edition
Sci & Tech

Anthropic, OpenAI and Google trade blows on cash, compute and capability as the AI race tightens

Tech Desk

The contest between the AI industry's three biggest labs has shifted noticeably this week from arguments about which model benchmarks best to arguments about which company can actually afford to keep building the infrastructure the next generation of models will need. Anthropic's finances came into unusually sharp public focus after a regulatory filing revealed half-year revenue of $16.3 billion, the clearest look yet at a business that has spent years discussing its growth in relative rather than absolute terms.

That filing landed alongside news that Google has agreed to credit-backstop Anthropic's lease obligations at a data centre campus in Hawesville, Kentucky, supporting contracted lease commitments reported at roughly $19 billion across three buildings. The arrangement is a reminder that even well-funded AI labs are increasingly dependent on their cloud-provider partners not just for compute capacity itself, but for the financial engineering that makes building out that capacity possible in the first place, a dynamic that ties the fortunes of supposedly independent AI companies more tightly to a small number of hyperscale backers than their public rivalries might suggest.

OpenAI's week took a more cautious turn. The company is reported to have shrunk the target for its newest funding round from $30 billion to roughly $20 billion, at a valuation of about $1.25 trillion, a scaling-back that, whatever the specific reasons behind it, sits awkwardly alongside the scale of capital the entire industry has committed to spending on data centres and chips over the next several years. Alphabet's own shares fell more than twelve percent on separate capital-expenditure concerns in the same window, suggesting investors are growing more sensitive to the gap between AI infrastructure spending commitments and the revenue currently available to justify them.

Beneath the financial headlines, the technical and safety side of the race has kept moving too. Anthropic is developing an open-source system called FlashProve, intended to let anyone cryptographically verify which model actually produced a given output, research the company has stressed remains experimental with no current plans for deployment at Anthropic or elsewhere. Google, meanwhile, has rolled out a specialised model aimed at cyber defenders, though access remains restricted to vetted security teams rather than the public, an approach that mirrors the industry's broader caution about releasing its most capability-dense tools without guardrails.

Perhaps the most consequential development for the wider internet received comparatively little attention: OpenAI, Google, Microsoft and Anthropic have reportedly agreed to honour a shared technical signal that lets website publishers opt their content out of AI training, with rollout timelines ranging from the coming months to early 2027. If it holds, the agreement would mark one of the first instances of the industry's biggest players coordinating on a publisher-facing standard rather than competing on data access, a shift that, depending on how strictly it is enforced, could reshape the economics of how future models are trained altogether.