Alphabet CEO Sundar Pichai was at the White House on Thursday for what officials described as cybersecurity meetings. The New York Times DealBook newsletter, citing people familiar with the discussions, reports that the real agenda was far more specific and more alarming: the US government does not have enough AI computing power to maintain its own defenses.

That sentence should stop you cold. The government that spends more on defense than the next ten countries combined is telling Google it needs help getting access to AI chips.

The compute gap is real

This visit lands in a very specific context. Last week, the Pentagon signed classified AI deals with seven companies (Google, OpenAI, Nvidia, Microsoft, AWS, SpaceX, and Reflection AI) under "any lawful use" terms. Anthropic was excluded for refusing those terms. The Pentagon allocated $54 billion toward autonomous weapons systems.

But signing contracts with AI companies is one thing. Having enough compute to actually run their models at military scale is another. The NYT scoop suggests the Trump administration is privately grappling with a reality that the AI industry has been warning about for years: the demand for AI processing power is outstripping supply, and government agencies are competing for the same scarce GPU time as every tech company on the planet.

Samsung and SK Hynix both told analysts last week that the memory shortage runs through at least 2027. HBM is consuming 23% of total DRAM wafer capacity. Apple warned investors about memory constraints hitting consumer products. If even Apple can't get enough chips, what chance does a government procurement office have?

Why Google specifically

Google is one of very few companies that both builds frontier AI models and manufactures its own AI chips (TPUs). That dual capability makes Alphabet uniquely positioned to offer the government something the hyperscalers can't: a vertically integrated AI stack that doesn't depend entirely on Nvidia.

Google Cloud already committed 5 gigawatts of compute capacity over five years as part of its investment in Anthropic. Its 8th-generation TPUs, designed for what Google calls the "agentic era," launched this year. The White House meeting may signal interest in securing dedicated government compute capacity on Google's infrastructure, separate from commercial cloud.

The bigger picture

The DealBook report noted the mood at the Milken Institute conference in Los Angeles this weekend was "shockingly bullish" on AI. By a show of hands, virtually nobody polled said they believed AI was in a bubble. Andrew Ross Sorkin posed the right question: is that confidence a positive sign of staying power, or the ultimate crowded trade?

Either way, the US government appears to be a believer. And the fact that it's going to Google to discuss insufficient AI capacity tells you something about where power actually sits in 2026. The companies that build AI infrastructure don't just sell to the government. They negotiate with it as near-equals.

We have been tracking this thread for weeks: the Pentagon's seven AI vendor deals, the Anthropic blacklist, the $54 billion autonomous weapons budget, the White House blocking Anthropic's Mythos expansion over compute concerns. Now the administration is sitting down with Pichai to figure out how to get enough chips to run its own AI defenses.

The US doesn't have a compute problem because it lacks money. It has a compute problem because the demand curve for AI processing power is steeper than any supply chain in history can match. Every new model, every new agent, every new military application makes the shortage worse. Thursday's White House meeting wasn't about cybersecurity. It was about whether the world's most powerful government can keep up with the technology it helped create.