Your mission—and you have no choice but to accept it—is to acquire superchips critical to the running of America’s key intelligence agencies. The need is so urgent that the White House is immediately releasing $800 million in emergency funding grabbed from other agencies to support the mission. But don’t worry. There is lots more money where that came from. The Trump Administration has approved a $9 billion request to purchase advanced AI computing chips for the Central Intelligence Agency (CIA) and the National Security Agency (NSA) in particular. The concern is that a chip shortage will impair the ability of intelligence agencies to deploy AI tools for classified intelligence work.
The $9 billion allocation will fund the development of specialized federal data centers to support NVIDIA’s Grace Blackwell superchip infrastructure. Grace Blackwell is a marriage of sorts in that it comprised of a mix of Grace CPUs and Blackwell GPUs. The resulting superchip needs its own custom build requiring massive amounts of electrical power and specialized liquid cooling. Grace Blackwell can’t run on standard government computing grids. Much of the $9 billion will fund the required Grace Blackwell infrastructure, according to New York Times reporting on the secret budget initiative.
The shortage is so dire that old grudges are being put aside. Susie Wiles, the White House chief of staff, authorized the National Security Agency to continue to use advanced Anthropic AI models even though the Pentagon has labeled Anthropic a threat to national security. Anthropic’s new Mythos model reportedly runs more efficiently on Grace Blackwell chips but can still run on a previous generation of chips. Mythos reportedly has found 10,000 critical flaws across 1,000 open sources projects in its first month which is giving it superhero status in computing circles. A new NSA contract with Anthropic reportedly will include a “carve out” to ensure its AI will not be used on Americans’ data, a framing that may be used with other government agencies.
The Grace Blackwell chip requirement also extends to other AIs. The newest version of ChatGPT reportedly requires NVIDIA’s Grace Blackwell superchip, for example. Expenditures on advanced AI superchips are likely to increase in coming years as “chip wars” increase in tempo. China’s Huawei, for example, just announced what the company called a “Logic Folding” architecture breakthrough in the production of its Ascend chips that allows it to bypass the lithography machines exclusively made by Dutch supplier ASML.
In some respects, the $9 billion allocation shows the U.S. government was caught flat-footed by underestimating the speed with which AI has become an integral part of intelligence and military operations. The $9 billion funding can be seen as making up for not allocating resources for AI in previous years. Alongside the Trump Administration’s $9 billion AI funding are investments by private enterprise. Amazon Web Services is allocating $50 billion to upgrade its government cloud computing services. The intelligence agencies make extensive use of AWS cloud networks to run their classified AI operations. AWS classified cloud data centers are physically separate from its commercial cloud operations.
