Token subs are dead! Long live token subs!

I keep seeing this take and it bugs me.

There is a massive rug pull coming. Soon there’ll be no AI coding subs, it’ll all just be API pricing, we’ll all be drowning in token costs and unable to code anymore with our AI addled brains. I don’t think this is the whole story.

Yeh, token subscriptions are subsidised. Frontier labs are land grabbing like there is no tomorrow. We’ve seen this before in tech and more broadly time and time again: underprice the product, capture the market, then move pricing closer to reality.

It is costly to serve heavy subscription users but its also costly to under utilise inference infra. Frontier labs have stratospheric capex tied up in GPUs that depreciate quickly, plus significant non-electricity operating costs. Idle GPUs aren’t neutral. They are burning a hole in your pocket.

Model shops want close to 100% utilisation of their GPUs so have to take demand to the very edge and walk a fine line. The goal is to keep the fleet as close to saturated as possible while still preserving reliability for the users who pay for guarantees.

Subs are a huge pool of elastic demand that can soak up spare capacity. That demand can then be shaped through rate limits, quantisation, model routing, batching, caching, smaller models and degraded access during peaks. Subscription users complain, but most understand the deal. Insane value in exchange for being lower-priority traffic. If you have been following Claude Code and the commentary around it lately, you would’ve seen this happening in real time.

This is also why a lot of commentary gets the economics wrong. People take the tokens included in a subscription, multiply them by public API prices, and conclude the provider must be losing impossible amounts of money. But API price is not provider cost. It includes margin, scarcity pricing, reliability, product packaging and the value of stable production access.

So another possible future instead of the death of subs, is further tiering of access:

• “Strategic partners” get access to the sharpest frontier systems (likely via preview) through bespoke deals (cough cough Mythos)
• Premium API access - expensive but valuable for companies that can justify it, premium rates for throughput, reliability and guarantees
• Subscription users - get leftover or flexible capacity, shaped dynamically by limits and routing.

Model competition from China is what makes this interesting though. They are injecting price pressure into the market and are providing real alternatives that are getting more and more compelling. When they are good and cheap, subs may be even harder to remove.