Mountain ViewGoogle plans $180–190B in capital spending this year, up from $31B in 2022.
CEO Sundar Pichai said token throughput hit 3.2 quadrillion monthly; Gemini 3.5 Flash claims 4x rival speed at a third the cost.
Gemini Spark, a 24/7 cloud agent, opens to Ultra subscribers next week with Model Context Protocol tool support to follow.
Sources: TheRegister.com, THE DECODER, The Wall Street Journal
THE CAPEX ARMS RACE
Google's $190B 2026 capex would exceed the entire annual GDP of Hungary. Four hyperscalers — Google, Microsoft, Meta, Amazon — are collectively spending over $500B/year on AI infrastructure, the largest concentrated industrial buildout since the postwar interstate highway system.
WHAT A TOKEN IS
A token is roughly three-quarters of an English word — the atomic unit a language model reads and writes. 3.2 quadrillion tokens per month means Google's models are processing the equivalent of every book ever published, several times over, every day.
THE INFERENCE ECONOMY
Training a frontier model is a one-time cost of hundreds of millions. Serving it — inference — is a recurring cost that scales with users. The economic gravity has flipped: by 2026, inference is roughly 80% of compute spend at the major labs, not training. Whoever serves tokens cheapest wins.
THE MCP STANDARD
Model Context Protocol, introduced by Anthropic in late 2024, is an open standard that lets an AI agent connect to external tools — databases, APIs, file systems — through a uniform interface. It is becoming the USB-C of agentic AI: every major lab now supports it, breaking the walled-garden lock-in that defined the first wave of chatbots.
THE POWER CONSTRAINT
A single Nvidia GB200 rack draws 120 kilowatts — roughly 100 homes. Google's planned datacenter expansion will require an estimated 15 GW of new electricity capacity by 2030, equivalent to fifteen nuclear reactors. The bottleneck is no longer chips or capital; it is the grid.
THE BUBBLE QUESTION
Comparisons to the 1999 telecom buildout are unavoidable: WorldCom and Global Crossing laid millions of miles of dark fiber on the bet that demand would absorb it. Most went bankrupt; the fiber was eventually used. The question for AI capex is not whether the infrastructure gets used, but whether the companies building it survive long enough to monetize it.