Aside from Alphabet’s own forays into AI, they’re also selling pickaxes to most other AI prospectors.

It’s a story deSouza likes to tell in numbers. In a conversation with this editor, he notes several times that nine out of the top 10 AI labs use Google’s infrastructure. He also says that nearly all generative AI unicorns run on Google Cloud, that 60% of all GenAI startups worldwide have chosen Google as their cloud provider, and that the company has lined up $58 billion in new revenue commitments over the next two years, which represents more than double its current annual run rate.

Asked what percentage of Google Cloud’s revenue comes from AI companies, he offers instead that “AI is resetting the cloud market, and Google Cloud is leading the way, especially with startups.”

The strategy extends beyond simple customer acquisition. Google offers AI startups $350,000 in cloud credits, access to its technical teams, and go-to-market support through its marketplace. Google Cloud also provides what deSouza describes as a “no compromise” AI stack — from chips to models to applications — with an “open ethos” that gives customers choice at every layer.

The approach reflects both opportunity and necessity. In a market where companies can go “from being a startup to being a multibillion-dollar company in a very short period of time,” as deSouza puts it, capturing future unicorns before they mature could prove more valuable than fighting over today’s giants.

“Companies love the fact that they can get access to our AI stack, they can get access to our teams to understand where our technologies are going,” deSouza says during our interview. “They also love that they’re getting access to enterprise-grade Google class infrastructure.”

Google’s infrastructure play got even more ambitious recently, with reporting revealing the company’s behind-the-scenes maneuvering to expand its custom AI chip business. According to The Information, Google has struck deals to place its tensor processing units (TPUs) in other cloud providers’ data centers for the first time, including an agreement with London-based Fluidstack that includes up to $3.2 billion in financial backing for a New York facility.

Competing directly with AI companies while simultaneously providing them infrastructure requires … finesse. Google Cloud provides TPU chips to OpenAI and hosts Anthropic’s Claude model through its Vertex AI platform, even as its own Gemini models compete head-to-head with both. (Google Cloud’s parent company, Alphabet, also owns a 14% stake in Anthropic, per New York Times court documents obtained earlier this year, though when asked directly about Google’s financial relationship with Anthropic, deSouza calls the relationship a “multi-layered partnership,” then quickly redirects me to Google Cloud’s model marketplace, noting that customers can access various foundation models.)