May. 19 at 6:57 PM
$GOOGL CEO:
$GOOG now processing 3.2 quadrillion tokens per month [up ~7x y/y]
This is a massive scale-up: up from ~480T (0.48Q) a year ago & 9.7T two years ago
A token is the basic unit LLMs process — roughly ¾ of a word or ~4 characters in English
Google does ~255B searches/month. Even if each used 500+ tokens, it would total ~127T tokens/month or just 4% of the 3.2Q total
Back of napkin:
Google charges customers ~
$0.075–
$15+/million input tokens (higher for output/reasoning) depending on the Gemini model tier
At an blended average of ~
$1–
$3/million tokens (very rough, mixing cheap/fast vs. heavy reasoning), 3.2Q tokens could theoretically equate to
$3.2B-
$9.6B/month in "equivalent" rev value if all billed externally
Note: Note all tokens are equal & most consumption is internal (Search, Ads, YouTube)
Search grew +19% y/y in 1Q26. If baseline (no-AI) growth was +10% to +12% (typical pre-AI acceleration), the extra +7% to +9% is largely AI-driven → Extra ~
$4.2B–
$5.4B in 1Q26 Search revs attributable to AI features or
$1.4B to
$1.8B/month. Cloud (direct AI monetization):
$6.68B/month & growing +63% y/y.
Google reports a median Gemini per prompt uses ~0.24 watt-hours (Wh) of energy (full stack: servers, cooling, etc., w/ efficient PUE of ~1.09). At 3.2Q tokens, Google's AI inference alone draws roughly 768GWh to 1530GWh of electricity/month
At standard industrial data center electricity rates (~
$0.07/kWh), the raw power bill for this token volume is
$54M/month
Direct operating cost (power + hardware depreciation) for 3.2Q tokens is likely in the low-to-mid billions/month range (rough order of magnitude), but it's embedded in Alphabet's overall ~
$15B+/month capex/opex run rate on AI infra
Note: Unlike "expense" spending (which vanishes), this Capex is building long-term assets (data centers, TPUs)
Cloud has strong backlog (
$460B+)
$AVGO $CLS $TSM.X
