TL;DR

Conway’s Law tells us that organisations create systems that mirror their communication systems. Jamie Dobson coin’s ‘Miell’s Law’ in a post about the work of our mutual friend (and his colleague) Ian Miell in his forthcoming book ‘Follow the Money‘:

Organisations that design systems are constrained to produce systems that reflect the financial structures and incentives of those organisations.

We can all imagine how that applied to organisations we know from the past. But I think it’s about to be massively amplified by how money gets converted into large language model (LLM) tokens, and how those tokens get doled out to those using them.

The free ride is over

Since the arrival of ChatGPT we’ve all been able to use LLMs and the various coding assistants and harnesses built upon them at no/low cost. Chatbot type interfaces are still generally free to use, and subscriptions have been selling the underlying tokens at (fractions of) pennies on the dollar. Many commentators have compared it to the early days of Uber, where rides were being subsidised by investor capital in order to grow market share and build a ‘winner takes all’ monopolistic network.

Despite something like $1.4Tn in infrastructure investment we find ourselves in a place where there simply isn’t sufficient supply for LLM inference to satisfy the insatiable demand. But as a wise colleague once said – ‘there’s infinite demand for free stuff’.

The first cracks showed with Anthropic changing the terms for Claude subscriptions so that they couldn’t be used for autonomous agents such as OpenClaw.

Next came Google, with ‘Changes to Gemini model access and limits‘.

But the move that’s likely to hit the hardest is GitHub CoPilot’s changes that took effect at the start of this week (Mon 1 Jun). Folk that were paying $tens for a subscription that burned $thousands in tokens are about to be hit with the full bill. There’s going to be some bill shock at the end of the month, and some very angry CFOs.

Who gets the tokens (and how many)?

Companies are reacting to the pricing changes by introducing token limits. Simon Willison has a good analysis of Uber introducing a flat cap of $1500 per tool per user (and there’s a company that knows only too well what an ‘Uber moment’ looks like).

I’m hearing reports from elsewhere that graduated limits are being introduced. Distinguished Engineers get unlimited tokens and access to the best new models. Entry level folk get a tiny budget and may also be constrained to older/cheaper/less potent models. That’s obviously going to magnify the impact of the power structure on productivity – Miell’s Law in action. Of course a Distinguished Engineer can be super productive with their awesome experience and a huge token budget on leading models. But do these orgs really want to force their early career folk (assuming they’re still hiring any) to be less productive? Have HR even had a say in this? I can imagine some spectacular fodder for future industrial tribunals.

The days of ‘tokenmaxxing’ are likely over, with Amazon shutting down its token leaderboard. Clearly they created an incentive structure that wasn’t properly aligned with what the company actually needs/wants.

Jensen Huang has stated that he wants Nvidia engineers to use AI tokens worth half their annual salary. But of course he has GPUs to sell us (or the services supplying us), which might still put him on team tokenmaxxing. I’m also left wondering if that March budget of $250k for an engineer with a $500k salary turns into a June budget of $Millions, or if even Nvidia engineers suddenly have less to work with?

What’s fixed and what’s variable (and has any of this been budgeted)?

Finance folk will often refer to fixed costs and variable costs; and headcount often lies awkwardly in the middle (depending on how easy it might be to hire and fire, which in turn can depend a lot on local labour laws).

From a Conway/Miell’s Law perspective the good old org chart reflects a whole bunch of budget that’s been allocated where the people below the apex have almost no discretion on changing things.

We now get to overlay token allocations into that org chart, and discover how much discretion is associated with that? I’d speculate that approximately zero organisations going through their FY26/27 budget planning had an accurate notion of 26H2 token costs or the allocations they’ll flow to, which means everybody is now making it up as they go along (or ‘being agile’ if you prefer).

Conclusion

AI coding assistants have been seen to boost productivity (especially for knowledgeable people who are good at articulating what they want); and that productivity boost was a ‘no brainer’ when the cost was (approximately) free. But the costs are shooting up, in part to ensure that demand is constrained to meet limit supply. That’s forcing organisation to think about how they allocate tokens, which is introducing a new dimension of financially structures and incentives. Miell’s Law might only have just been coined, but it’s going to be an important thing to consider as the post free ride budgets get figured out.