On February 6, 2026, agents on OpenRouter consumed more tokens than humans did. They have not given the lead back1.
AI consumption is not a smooth curve anyone can extrapolate : it arrives in three waves, each orders of magnitude larger than the last. The second one has already broken. The third is on the horizon.
Wave one is chat : a person asks, the model answers once, done. Roughly 1m tokens per active user per day, by our estimate.
Over the next five years, US data center capacity will grow from 25 gigawatts to 70 gigawatts, part of a global buildout costing roughly $5t.1
Where will the money come from?
Data centers are built as real estate projects with some equity, but the majority debt : typically 70% or more2. Assuming we achieve our plans to build all these data centers, is there enough debt available in the credit markets to finance it?3
Meta launched two things yesterday : a state-of-the-art model & a new pricing system for foundation models.1 Muse Spark propels US open source models to the frontier. Meanwhile, the pricing system resets the industry’s economics.
For thirty years, enterprise software operated on strict licensing fees with a guarantee of total privacy & zero data retention.2 Consumer technology operated on the opposite principle : free software in exchange for behavioral data. Consumers trade queries for convenience, but enterprises fiercely protect their intellectual property.
I used to think AI productivity meant doing less. Automate the mechanical work, & human effort shrinks toward zero.
The reality is stranger. AI does not shrink human effort. It raises the ceiling of what the same amount of work produces.
My writing workflow is an interactive loop with an agent. I seed it with an idea, data, or transcript. It drafts the essay in seconds, & then we work through line by line. I never touch the raw text directly ; I issue verbal & written instructions to the model, reviewing diffs, challenging phrases, & refining rhythm until the argument lands. Each night, the system updates my style guidelines, building a personalized Strunk & White.
The frontier AI market is sorting itself into closed camps, as labs pick partners, cut rivals, & ration access to their strongest models.
Salesforce chose Anthropic as its dedicated AI partner, making Claude the default model inside the world’s largest CRM & Slack. 1 OpenAI cut Cursor’s API access on November 12 after SpaceX bought the company, citing prior contract breaches. 2 Even the open weights models now impose gates : Z.ai released GLM-5.3-Flash under an MIT license on August 26 & put its flagship behind a $10b host-revenue review two days later. 3
An AI model company buys wholesale electricity by the megawatt & resells it as cognitive work.
“The base cost of compute tends to be around 10 or 13 or $15 million per megawatt. In the case of Anthropic, the revenue has gone as high as $50 million per megawatt. And what that now enables them to do is, hey, if I spend 10 bucks on inference capacity, I actually generate 50 bucks of revenue. And then I can turn around and incrementally spend all of that profit on training.”
— Dylan Patel, SemiAnalysis, on the Dwarkesh Podcast1
NVIDIA booked $96b of revenue last quarter, up 106% from a year ago & 18% from the quarter before. The company guided Q3 to $108b ±2%.
That guide crosses $100b of revenue in a single quarter. No semiconductor company has ever reached that threshold, & no company of any kind has reached it while growing triple digits.
Q2 alone runs at $385b a year. At $108b, NVIDIA annualizes to $432b of revenue, making it the sixth largest company in the world, leapfrogging Apple, McKesson & Alphabet.
Mary, Mary, quite contrary, how long do your agents live?
New products like Grok Bot & other meta-harnesses ask us to create agents. How long should they live?
When you design a calendar agent, how long should its session run? Should it stay alive for your entire five-year tenure at a company, or reset every day?
Ever more powerful models tempt us to build perpetual sessions that never close. But long-running sessions rot from the inside out.
The popular narrative of AI infrastructure is a tidy relay race : first GPUs were scarce, then memory choked throughput, then CPUs took the strain, & finally storage started to bite.
The pricing data says the relay is real but slow. Some components plunged before they surged. Each bottleneck freezes the next component’s supply chain, but the lag runs in years, & every wave locks in a higher baseline cost.
Local AI models can already answer 89% of everyday chat & reasoning queries as well as a frontier cloud model, a result that holds broadly across more than a million real queries & 20+ local models tested.1
That means we’re generating more intelligence per watt of electricity : more work from the same number of electrons.1
Tracking computing efficiency over time is not new. Koomey’s law found that computing power per watt doubled roughly every 1.5 years for decades, a trend that shrank the power of a mainframe into a laptop’s chassis.23