Salesforce CEO Marc Benioff called for National Guard troops in San Francisco from his private plane, shocking his own PR team. The theatrical Trump embrace—timed before his big conference—tests whether loud loyalty beats quiet accommodation.
Coursera courses now live inside ChatGPT—world-class instructors summoned mid-conversation. No app switching, no browser tabs. EU users blocked by privacy laws while Wall Street bets big on AI-powered education. The learning shortcut arrived, but not for everyone.
Google's 1.3 quadrillion token milestone sounds massive—until you see growth rates halving and realize tokens measure server intensity, not customer demand. The slowdown reveals something uncomfortable about AI economics.
Google's 1.3 quadrillion token milestone sounds massive—until you see growth rates halving and realize tokens measure server intensity, not customer demand. The slowdown reveals something uncomfortable about AI economics.
Google says its AI now chews through 1.3 quadrillion tokens a month; growth, however, is already decelerating and the metric itself measures compute burn more than customer uptake.
We now have three clean data points from Google: ~480 trillion tokens in May, ~980 trillion in July, and 1.3 quadrillion in October. The absolute jump is large, but the rate of increase has halved—from roughly +250 trillion per month in spring to about +107 trillion per month since July. That’s the tension behind the headline.
The company bundled the token milestone with the launch of Gemini Enterprise, an “AI front door” for workplace agents and data. That framing suggests surging adoption, yet tokens are primarily a unit of computation: slices of text, audio, or pixels the model ingests and emits. They are not users, seats, or solved workflows.
The Breakdown
• Google's token growth halved from 250 trillion monthly in spring to 107 trillion since July despite hitting 1.3 quadrillion total
• Tokens measure computational load per request, not user activity—reasoning models burn 17x more tokens than earlier versions
• Environmental claims use light-prompt medians that exclude heavy reasoning, multimodal, and agent tasks driving actual token growth
• Watch how aggressively Google steers customers to lite models and caps context windows—that reveals true unit economics
Reasoning models inflate token counts because they think more internally. As providers add chain-of-thought-style steps, larger contexts, and multimodal inputs, each request consumes extra tokens even when user activity stays flat. That’s why token totals can soar while growth cools and revenue lags. Analysts at The Decoder made the same point this week, calling the figure “window dressing” for back-end scale. They’re right about the direction of travel.
Pricing changes reinforce the picture. Google’s own updates this summer simplified 2.5 Flash pricing and removed separate “thinking vs. non-thinking” rates—an implicit admission that internal reasoning tokens are a real cost driver that customers notice. Efficiency tweaks can trim invoices; they do not change what tokens fundamentally represent.
Google’s blog post touting Gemini Enterprise also includes the quadrillion-token milestone, tying it to AI permeating Cloud, Search, and Workspace. Take that as confirmation of deployment breadth, not proof of clear ROI. Tokens prove that systems are busy, not that they deliver dependable outcomes at an acceptable unit cost. The deceleration visible in the May→July→October series is the tell.
On environmental claims, the gap is wider. Google’s technical paper pegs a median Gemini text prompt at 0.24 Wh of electricity, 0.03 g of CO₂e, and 0.26 ml of water—numbers that sound tiny because they measure the lightest, shortest prompts. The same write-up sidesteps heavier cases like long-context analysis, video, and agentic browsing, which are exactly the scenarios pushing token totals higher. Critics flagged the scope issue within hours. The frame to use
Think of tokens as server-side intensity, not end-user demand. A quadrillion tokens can reflect three forces at once: more users, more work per request, and more modalities per session. Only the first is durable demand. The other two are architecture choices and product bets that can reverse as vendors de-emphasize expensive reasoning, turn on caching, or push cheaper “lite” models for routine tasks. Watch how often Google steers enterprises to Flash-class models and how aggressively it caps context windows and agent steps by default. Those settings reveal the true economics.
None of this means the milestone is meaningless. It proves Google can provision capacity at global scale and keep latency tolerable while moving customers into agent workflows. But the slowdown in monthly additions suggests either supply constraints, deliberate throttling for unit-economics, or simple maturation after a launch burst. Efficiency gains will also push the token curve down without any demand change. Read the chart accordingly.
The key question is whether Google can translate server-side intensity into net-new revenue per employee for its customers. If Gemini Enterprise consistently replaces license seats, call minutes, or analyst hours, decelerating token growth can coexist with rising sales. If not, a quadrillion tokens is just a very large power bill expressed in syllables.
Q: What exactly is a token in AI terms?
A: A token is the smallest unit an AI model processes—roughly word fragments or syllables. "Hello world" might be three tokens. Models break text, images, audio, and video into tokens to process them. More complex inputs and reasoning steps create more tokens, even for the same user request.
Q: How much does processing 1.3 quadrillion tokens actually cost Google?
A: Google doesn't disclose internal costs, but industry estimates suggest large-scale AI inference costs $0.10-$0.50 per million tokens depending on model complexity. At the low end, 1.3 quadrillion tokens could represent $130-650 million monthly in compute costs—before accounting for data center overhead, power, and cooling.
Q: Why would Google deliberately throttle growth if they're bragging about scale?
A: Unit economics don't work at scale if compute costs exceed revenue per user. Google may be rate-limiting free users or capping expensive features because reasoning models burn tokens faster than customers generate revenue. The company acknowledged capacity constraints through Q4 2025 in earnings calls, suggesting supply limits matter too.
Q: What makes reasoning models use 17 times more tokens than regular models?
A: Reasoning models perform internal chain-of-thought steps before answering, generating tokens users never see. A simple question might trigger dozens of intermediate calculations, self-corrections, and verification steps. These "thinking tokens" drive up processing costs even when the visible answer stays short.
Q: Are other AI companies reporting similar token growth slowdowns?
A: Most competitors don't publish token metrics publicly, making direct comparisons impossible. OpenAI, Anthropic, and Microsoft report usage in vague terms like "queries" or "active users" rather than computational load. Google's transparency on tokens is unusual—which makes the visible deceleration from 250 trillion to 107 trillion monthly growth particularly telling.
ASML fills an 18-month CTO vacancy as the chip monopoly races to expand Eindhoven capacity, deploy AI tools, and navigate China export bans—while every AI model depends on machines only it can make. Succession meets systemic risk.
Microsoft licenses Harvard Medical School content for Copilot health queries while training models to replace OpenAI's infrastructure. The healthcare play addresses a billion-download gap and builds switching costs where credibility trumps speed.
Sam Altman laid out OpenAI's plan: one AI assistant that follows users everywhere, backed by a trillion-dollar compute buildout. The vision is coherent. The execution surface is vast, spanning chips, power, and partner risk across simultaneous bets.
OpenAI opened ChatGPT to third-party apps that render inside conversations, built on an open standard but distributed through algorithms it controls. The trade: reach 800M users, lose funnel ownership. Platform economics arrive at chat.
Get the 5-minute Silicon Valley AI briefing, every weekday morning — free.
