The money follows the policy. In May 2026, the Pentagon announced direct agreements with Nvidia, Microsoft, AWS, Alphabet, OpenAI, SpaceX, and others to deploy AI on classified networks via GenAI.mil, its official AI platform. Separately, CDAO awarded up to $200 million each to four frontier AI model vendors. And the $9 billion Joint Warfighting Cloud Capability (JWCC) program — with JWCC Next solicitation targeted for early 2027 — is the pipe through which AI inference workloads run. The compute layer beneath all of this is not abstracted away. It is very specific hardware.
Who Cashes In
NVDA is the structural winner. Every hyperscale and government cloud cluster running AI inference or training on classified networks is overwhelmingly GPU-dependent, and NVIDIA's H100, H200, and now Blackwell architecture are the default. The Pentagon's direct agreement with Nvidia in May 2026 is a signal, not an exception — the DoD is now a named end-customer, not just a downstream beneficiary.
AVGO benefits through a different but equally durable mechanism. Broadcom's custom ASIC (XPU) business serves Google, Meta, and now OpenAI — all of which are JWCC/GenAI.mil vendors deploying into DoD environments. Every inference cluster those companies run for the Pentagon requires Broadcom's custom networking silicon (Tomahawk switching, Jericho routing) to function at scale. AVGO's $73 billion AI revenue backlog is a commercial-first story, but defense AI adoption converts commercial cloud capacity into government workloads.
MU is the third-order play. AI inference at classified-network scale is memory-bound. High-bandwidth memory (HBM) and DDR5 content per server rises with each new GPU generation, and Micron is one of two companies supplying HBM at volume alongside SK Hynix. A sustained DoD AI compute buildout is a structural HBM demand signal.
Who Is Exposed
INTC faces the wrong architecture at the wrong moment. Intel's GPU (Gaudi) and data center products have failed to take meaningful share in the AI accelerator market, and the NDAA's mandates around AI compute infrastructure play directly to GPU- and custom-ASIC-first architectures where Intel is structurally behind. More risk: any further budget scrutiny on legacy Pentagon IT — where Intel server CPUs dominate — could compress Intel's defense footprint without a replacement revenue stream.
AMD is in an awkward middle. Its MI300X GPU is competitive on paper but has not secured the hyperscale or government-cloud adoption needed to become a named partner in DoD AI programs. AMD benefits marginally from any GPU demand expansion but is not a structural winner of this particular policy cycle.
The Play
Watch the JWCC Next solicitation (expected early 2027) and any DoD data center energy and capacity disclosures required under the FY2026 NDAA's Section 1531. Those documents will name the compute infrastructure standards DoD expects vendors to meet — and whatever architecture dominates that language is the chip stack that gets built out at scale. NVDA is already inside the perimeter. AVGO wins every time the cloud vendors it supplies win government task orders.