The Free Ride

If you've been using Claude heavily lately, you may have started seeing this message: You've reached your usage limit for peak hours. It's annoying. It's also a signal.

Anthropic is falling victim to its own success. Claude is growing faster than anyone expected — daily active users tripled since the start of this year, paid subscribers doubled, and in early March the app briefly hit number one on the US App Store, displacing ChatGPT for the first time. That's a remarkable run. The problem is that more users means more compute demand, and we don't have enough compute for everyone to be running Claude Code or Cowork all day. The infrastructure just isn't there yet.

A quick note before we go further: I write about Claude because that's what I primarily use. But everything I'm about to describe applies equally to OpenAI and every other major lab. The compute constraint, the subsidy math, the path to the IPO — ChatGPT is navigating the exact same dynamics, just at larger scale. This isn't an Anthropic story. It's an industry story.

This isn't a new insight to people who follow the labs closely. One of the main bottlenecks on AI diffusion has always been compute — the raw processing power that makes Claude work. Think of it like bandwidth. When you sign up for internet service, the plan you choose determines how much data you can move and how fast. Compute is the same thing, just for AI. Right now, we don't have nearly enough of it.

The data center crisis is real. Building new facilities takes years. Those facilities need enormous amounts of power and water, and communities near proposed sites are pushing back hard. People living next to data centers are reporting higher utility bills and water pressure issues because they're competing for the same resources. The power grid wasn't built for this. Neither was our water infrastructure. These are the same constraints showing up at the AI layer.

Anthropic's Claude Max plan at $200 a month has been reported to consume up to $5,000 in compute at heavy usage. That's a 25-to-1 ratio between what you're paying and what you're consuming. Even at more moderate usage levels, the economics are dramatically in your favor.

Anthropic knows this. They're doing it on purpose. The logic is the same logic that built Uber and Amazon Prime: subsidize adoption, create habits, monetize later. Right now, these labs are still private companies burning through venture capital. They want as many people as possible building their workflows around Claude before the economics have to make sense. The moment they go public and need to show profitability, that calculation changes.

We are in the Goldilocks zone of AI compute. Not too early that the tools are useless, not so late that the subsidies are gone. It is an extraordinary moment to be a user of these tools, and we shouldn’t take it for granted.

There are really two possible futures from here. Compute gets cheaper — advances in nuclear power, new chip architectures, data centers in orbit, whatever it ends up being — and the cost of running these models approaches zero. That's the bull case, and smart people are working on it. Or compute stays scarce and expensive, and one day these labs decide they need to actually charge what the service costs to deliver.

Both scenarios are plausible. What's certain is that pricing will move. We're at the very beginning of AI becoming a utility, the same way electricity became a utility, the same way the internet became a utility. It doesn't happen overnight, and the transition period is usually when it pays to be paying attention.

For businesses, this is already a real budget conversation. Companies running serious AI workloads are uploading corporate cards and setting token budgets for their engineers. A senior software developer might need a $500,000 token budget to ship product in a year. That's not hypothetical — that's how AI-native organizations are already thinking about compute spend. It's a line item, the way headcount is a line item.

For individuals, we're not there yet. But the habit to build now, while the stakes are low and the subsidy is generous, is using AI deliberately. Not less, more thoughtfully. You don't take pre-workout for your morning commute. You save it for when it matters.

Open Claude (or ChatGPT) and ask: "Based on our recent conversations, where am I spending the most tokens? Where am I getting the most value, and where could I be more intentional?"

You're not trying to use AI less. You're trying to develop the instinct for where it earns its keep — so that when the economics shift, you've already built the right habits.

In 2025, AI surpassed humans in annual written output. According to ARK invest data shared last week, AI now produces more text per year than every human on earth combined. To be precise: it's annual output that crossed, not the full 500-year archive of human writing. That milestone is still projected for later this decade. But the point stands. Digital noise has never been higher, which means the original thinking and genuine point of view you produce has never been more scarce or more valuable.

OpenAI shut down Sora last week. The app peaked at around a million users and was burning roughly $1 million a day in compute to keep it running. Like most generative tools, the output followed a power law: the top 10% of videos were genuinely stunning, and the other 90% never saw the light of day. AI video at scale isn't ready yet, and the economics made that impossible to ignore. From jaw-dropping demo to quiet shutdown in under two years.