The Capex Desk

The AI economy, in numbers. I take one company, product or model each week and work out what it actually costs to run. Charts are my own.

[The Capex Desk @TheCapexDesk]

Nobody can tell you what a ChatGPT answer costs.

Not the price. The cost.

Trillions of dollars are being committed on the assumption
that this number falls fast enough, and almost nobody has
done the arithmetic in public.

So I'm going to. Every week. Until someone shows me I'm wrong.

Teardowns Tuesday and Friday.
My own charts. My work shown every time.

[The Capex Desk @TheCapexDesk]

How much would a small business pay monthly for an AI that handles calls and bookings?

1) almost nothing, it should be cheap
2) less than a part-time worker
3) whatever a part-time worker costs
4) depends on the business

[The Capex Desk @TheCapexDesk]

Which cloud provider type do you think is best positioned for AI, and why?

[The Capex Desk @TheCapexDesk]

Better for innovation?

1) strong patents on AI discoveries
2) open publication of AI discoveries

[The Capex Desk @TheCapexDesk]

@jrya7 Feel free to mate

[The Capex Desk @TheCapexDesk]

Agree or disagree: in AI, distribution beats model quality more often than people admit.

[The Capex Desk @TheCapexDesk]

True or false: spot GPU prices are the most honest signal of AI demand there is.

[The Capex Desk @TheCapexDesk]

Quick poll: who should pay for the grid upgrades AI data centers need, the companies or the public?

[The Capex Desk @TheCapexDesk]

Guess: what gross margin did Nvidia report for its quarter to July?

Answer tomorrow.

[The Capex Desk @TheCapexDesk]

Hot take: AI will flood patent offices with applications, and examining them will need AI too. Fair?

[The Capex Desk @TheCapexDesk]

Wrote a post today arguing against my own strongest held position.

It was significantly harder than writing the original and I think it's better.

Recommend the exercise to anyone with a public thesis: write the best version of the counter-argument, publish it, and see whether your original survives.

Mine survived. Weaker than before.

[The Capex Desk @TheCapexDesk]

Rank these by how much AI saves them:

1) car makers
2) chemical plants
3) food producers
4) electronics assembly

[The Capex Desk @TheCapexDesk]

Founders love talking about defensibility. Wrong answers only: what's really protecting the business?

[The Capex Desk @TheCapexDesk]

You guys are aware, right?

In July, Amazon said even about $220 billion of capex won't buy enough capacity to meet its demand in 2026.

[The Capex Desk @TheCapexDesk]

Who gets GPUs first when capacity is tight?

1) the biggest customers
2) whoever pays most
3) whoever committed earliest
4) a fair share for everyone

[The Capex Desk @TheCapexDesk]

Seriously, why does a free AI tool take six months to get through security review?

[The Capex Desk @TheCapexDesk]

The consensus among people who agree with me: the market doesn't understand the power constraint.

Let me steelman the other side, because I think I've been too confident here.

Utilities have handled enormous load growth before: air conditioning, electrification, industrial expansion. Capital finds a way around bottlenecks it's paid enough to solve. Behind-the-meter generation is scaling. And I may be over-indexing on queue length because it's the data that happens to be public.

The power constraint is real. I'm less sure it's binding than I sounded a fortnight ago.

Recording that here so it's on the scoreboard.

[The Capex Desk @TheCapexDesk]

Unpopular opinion: specialist GPU clouds will mostly be acquired by bigger clouds when the cycle turns.

[The Capex Desk @TheCapexDesk]

From this morning's piece:

On 27 March 2000, Cisco became the most valuable company in the world, at about $555 billion.

At a mature five times sales, that price needed about $111 billion of annual revenue. Twenty-five years later, Cisco's revenue was $56.7 billion.

The company executed. The price had assumed an execution nobody could deliver.

[The Capex Desk @TheCapexDesk]

Which job do you think gets better paid because of AI, against everyone's expectations?

[The Capex Desk @TheCapexDesk]

Sixty days of arithmetic. The five numbers I'd still pay for:

1. Compute utilisation, major providers
2. Used-accelerator residual values
3. Training vs. serving split
4. Independent end-customer revenue
5. Enterprise AI ROI with a real baseline

Two months of looking. None of them exist publicly.

If you have any of them and can share, my DMs are open. Access to data is the actual bottleneck on this work.