AI news, 23 August: Nvidia flags 15%+ AI server price hikes

Nvidia customers face 15%+ AI server price rises. According to Bloomberg, confirmed by Reuters and Fortune, contract manufacturers have told major hyperscalers including Microsoft, Google and Oracle that prices for servers packed with Nvidia’s AI chips will climb more than 15% in many cases on shipments starting early 2027. The hikes hit flagship configurations built around Vera Rubin and Grace Blackwell chips and are driven largely by soaring high-bandwidth memory costs from suppliers such as Samsung, SK hynix and Micron. Nvidia itself has not publicly confirmed the notices. The move underscores how memory bottlenecks are now a binding constraint on the AI infrastructure boom and will raise the cost of the data-centre buildout worldwide.

Anthropic eyes blockbuster IPO near $2 trillion valuation. Reports from The New York Times, Bloomberg and the Financial Times say bankers have told potential investors the Claude maker could seek to raise more than $100 billion and may target a valuation around $2 trillion, potentially topping SpaceX’s recent record. A public filing could come as soon as late August with a possible October listing. The ambition rests on explosive growth: Anthropic’s annualised revenue run rate reached about $65 billion by end-July. The company has not set a formal target and talks remain fluid; the figures reflect investor modelling rather than a confirmed deal.

Small UK agent beats frontier models at research replication. London startup Inherent, founded by DeepMind alumni and freshly out of stealth with a $50 million seed round, said its Faraday agent outperformed much larger systems from Anthropic (Claude Opus 4.8) and OpenAI (GPT-5.5) at independently reproducing findings from published scientific papers. TechCrunch reported that Faraday runs on a relatively compact 27-billion-parameter base model yet showed stronger “research taste” in experiment design on Inherent’s Replica benchmark. The result is a company claim backed by its own evaluation; independent verification will matter, but it highlights progress toward AI systems that can assist genuine scientific work rather than just summarise it.

Nvidia agent system hits perfect score on tough reasoning test. Nvidia reported that its AVO (Agentic Variation Operators) framework, wrapping Anthropic’s Claude Opus 5, achieved a perfect 100 score on the public ARC-AGI-3 benchmark, clearing all 183 levels across 25 environments while using roughly 12% fewer actions than a prior leading system. The bare model scores far lower (around 30%), illustrating that careful system design—persistent memory, supervision loops and planning—can dramatically lift long-horizon agent performance. Nvidia noted the result covers the public set only and is not a private-set competition score.

OpenAI urges California to toughen its AI safety law. In a shift, OpenAI called for amendments to strengthen SB 53, the state’s frontier AI transparency and incident-reporting law. According to TechCrunch, the company wants added requirements for monitoring models during training and stronger cybersecurity throughout development, citing recent incidents in which AI agents escaped test environments. OpenAI previously took a cooler stance on the bill; it now frames stronger state rules as a path toward eventual national standards.