What the hell is tokenomics? Let's talk HEWYs ...

Say goodbye to tokenomics ... it's time to talk HEWYs (human equivalent work years - pronounced 'Hughy' but it's a work in progress).

Right at the nexus of some of my favourite things - real estate, software, finance and AI - is the data centre debate. More broadly, it is a question of the economics underpinning the entire AI industry (and by extension, the markets) today.

1 gigawatt (GW) of compute currently costs around $50B. 40-50% of that is the data centre itself (and the associated infrastructure) and 50 - 60% of that is the servers, GPUs, networking etc.

That's a lot.

Except that revenues from this 1 gigawatt could be in the region of $20B. That's what Nebius says it is signing contracts at for slightly longer contracts, but Anthropic is paying XAI the equivalent of around $50 billion on spot pricing (!).

So how much is 1 GW worth to Anthropic or OpenAI? Well, based on NVIDIA's Blackwell benchmarks and SemiAnalysis's figures, 1GW can produce around 3 - 5B tokens per second (!). Or 90 - 150 quadrillion tokens per year.

Let's use 100 quadrillion (i.e. 100,000 trillion tokens).

So, Anthropic is paying roughly $0.20 per million tokens ($20bn/100Q) and charging $20/$50 per million output tokens for Opus 5 & Fable 5 respectively.

That seems pretty good.

Actually, that seems incredibly good. That is minimum 99% margins for inference. So, the question then becomes what is their utilisation and what is their CapEx and OpEx to realise these potential profits. Even with 50% utilisation and an average sale price of just $1/M tokens that is $50B. Per year.

Well, given the outages and limits experienced by me, utilisation seems above that (and while margins on me would not be 99%, that's not really the point).

So, CapEx then .. well, assuming 2 major model generations per year and a $500M final training run per generation (and 5 - 10x this in R&D costs), that means CapEx is around $10B. Add in another $2B for labour and $1B for SG&A we are at $50B - ($20B + $10B + $2B + $1B) = $17B (34% margins).

So far, so f*%king amazing for the model makers.

So how long can this last or how far could this go?

Well, this is the interesting (but a bit more fuzzy) bit. Thankfully, it is starting to come into focus (if you squint).

How many tokens of work does a human produce?

What a stupid question.

Now, obviously words or tokens is not the way to measure human output. However, METR does measure AI capability in terms of its completing tasks. Triangulating between this, AgencyBench and OfficeVal I would propose a (broad) range of 500k - 4M tokens per hour.

So, how many employees could 1GW replace? Between 14M and 111M. Let's use 30M - 60M as the 'meaty' part of this range. At a fully-loaded cost for knowledge workers of around $120k ... this translates to $3.6 - 7.2T. er year

Wait, we can replace a (minimum) $3.6T of costs with $20BN?

Sign me up.

Let's say companies can only actually recreate 30% of tasks using AI (end to end), and even then only use the compute 50% of the time and we are off by 50%. That's still $270B.

And we are starting to see evidence of these sorts of returns being possible (at small scale). arXiv have reported that for every $0.03 increase in AI spend, companies are spending $1 less on outsourced labour costs. So that 33x return is considerably better than my 13.5x return and done without full end to end capabilities from the models.

So, where is the disconnect coming from? Why is everyone so concerned about these prices?

Part of it is the question of how long until this gets 'priced in'. Part of it is that there is not a lossless transfer of labour from human to digital. Part of it is there there will be a significant increase in GDP (I am still on the hook for a 10% print). Part of it is that there will be hugely increasing quality and quantity of goods & services delivered. Part of it will be (major) deflation.

This brings me back to my starting point.

It is no longer about tokens. What is going to become important is HEWYs (Human Equivalent Work Years) produced per gigawatt of compute.

NVIDIA is going to improve throughput and those token numbers are going to go up. Intelligence is going to get commoditised. Price setting will come from land, proprietary data, trust, regulation and capital. Those are each worth an article ...

But in the meantime, the future looks bright for those that are investing in AI.

#ai #workingcleverly