Top stories

IBM CEO Arvind Krishna said open-weight AI models can deliver 90 percent of the performance at 10 percent of the price in on-premises deployments. He said the systems can fit in two or three racks, making the combination of cost, performance, and compact infrastructure difficult to beat.

OpenAI CEO Sam Altman said that distributing limited DRAM capacity across personal computers for infrequent use is less efficient than concentrating memory in cloud data centers. In an interview with Cheeky Pint, Altman said many people expect open-source AI models to make local inference practical and reduce the need for data-center memory, but argued that cloud infrastructure can achieve higher utilization when memory supplies are constrained. Altman said the industry should seek “maximum utilisation in the cloud” and characterized DRAM sitting in individual computers and used only part of the time as a poor global allocation of a limited resource. The source says memory prices have risen alongside AI infrastructure demand and that Micron announced plans to leave consumer memory production, with the company describing AI data-center needs as essential compared with consumer use. Altman also said OpenAI is designing a custom AI processor called Jalapeno. He said the chip is being designed by AI for AI and argued that constrained global silicon capacity should be directed toward AI systems that could produce major technical advances, including potential efficiency gains or discoveries comparable in importance to the transistor.

Amazon Web Services has booked most of its cloud-computing capacity through 2027, and sizable customer commitments already extend into 2028, according to Amazon CEO Andy Jassy. Speaking on a video podcast hosted by Andreessen Horowitz’s Ben Horowitz and Erik Torenberg, Jassy said AI-infrastructure demand is “insatiable” and that AWS capacity for 2026 is effectively sold out. AWS is reserving large blocks of infrastructure years before delivery as enterprise demand for AI computing exceeds available supply. Jassy said AWS’s constraint is the speed at which Amazon can construct data centers and procure advanced chips, rather than a shortage of customers. Amazon has indicated roughly $220 billion in capital expenditure for 2026, and Jassy said the company could spend more. AWS continues to hold some computing capacity for on-demand customers and services including Bedrock, its managed AI-model access service, and EC2 virtual servers, but Jassy said most large long-term capacity commitments are already booked. Delays in data-center construction or chip supply could postpone delivery of reserved capacity.