U.S. government officials held discussions with Alphabet CEO Sundar Pichai on Thursday to confront mounting worries about artificial intelligence compute capacity and the federal government's ability to secure sufficient processing power for defense-related systems.
During the meeting, administration representatives conveyed concern that Anthropic lacks the necessary computing resources to sustain broad deployment of its Mythos model. Officials indicated Anthropic may need to restrict access to Mythos, potentially even for users designated as high priority.
Those same officials expressed apprehension that, if computing constraints persist, the government could lose access to tools it relies on to remediate and patch its software systems. The potential loss of access to these tools was presented as a material operational worry in the context of maintaining and updating government software.
A technical point raised in the talks involved some of Google’s AI accelerators, known as tensor processing units or TPUs. According to the information shared at the meeting, certain TPUs cannot be deployed in specific classified settings, creating a prohibition on their use for some sensitive government workloads.
To address the shortfall, federal officials are exploring mechanisms to speed up clearance processes that would allow Google TPUs to be used in restricted environments. The objective under discussion is to expand the set of approved computing resources available to government programs that require higher assurance or classification handling.
The dialogue centered on securing enough high-performance compute capacity for defense systems and critical maintenance tasks, while balancing operational requirements that limit where some commercial AI hardware can be deployed. Participants discussed the immediate constraint on compute resources and potential administrative remedies to make additional hardware usable for classified workloads.
No new operational deployments or approvals were reported as part of the meeting. The discussion focused on identifying the capacity shortfalls, the risks to continued access to AI-backed tools for software patching, and administrative options to increase eligible compute for classified settings.