Essay: The Kobe Problem

I had Kobe beef in Kyoto last December. I looked forward to it, it was on my todo for that trip.

It was extraordinary. The kind of meal you count yourself lucky to experience if only once. The difference between Kobe and regular beef isn't subtle, it's an order of magnitude. In price, in texture, in the experience of eating it.

Everyone who's heard of Kobe understands the economics instinctively: it costs what it costs because of what it is. You can't serve it at regular steak prices and run a viable restaurant. The math doesn't work. Not even close.

Which is exactly what's happening with AI right now.

Kobe At Regular Steak Prices

AI companies are raising capital in the trillions. Revenue is in the tens of billions. That gap doesn't close on the adoption curves we're seeing. There's no hockey stick visible that bridges the distance between what's being spent and what's coming in.

We don't know the exact numbers, and that's not an accident. Some of the biggest AI companies are still private and they are not required to show us what inference actually costs versus what they're charging. They're not disclosing the real relationship between token consumption, infrastructure spend, and revenue. They're not telling us, and it's my view they're not telling us because the numbers are brutal.

What we do know is this: the investment cycle underway in AI infrastructure is unlike anything we've seen. Data centres, chips, energy, the capital commitments are extraordinary. And the revenue required to justify those commitments is nowhere near where it needs to be. As in, NOWHERE near where it needs to be.

So we're being served Kobe at regular steak prices. And we're getting addicted to it, by design.

It Gets Worse

Here's what I was told recently by a director at an AI-native software startup that I am an advisor to.

Their lead engineers routinely open four parallel instances of their AI coding tool. They prompt each agent slightly differently, same problem, different approaches. The agents go off, do the work, and come back with results in plan mode. The engineer reviews all four, picks the best one, and discards the other three.

Four Kobe steaks cooked. One served. Three in the bin.

And here's the part that makes this genuinely alarming: each of those agents is being charged at regular steak prices, when the actual cost is closer to Kobe. So you're not just cooking four steaks to serve one. You're cooking four Kobe steaks, charging the customer for one regular, and throwing the rest away.

This is being promoted as a best practice.

I don't say this to criticise the engineers. If the owner of the restaurant tells the chef they can use whatever ingredients they want, in whatever quantities, with no accountability for wastage, the chef might cook four slightly different Kobes and serve the single best... why not right? It's free! That's human behaviour. The broken signal is the economics, not the people responding to them.

But it means the real cost of AI consumption is being amplified far beyond what the headline pricing suggests. And nobody, not the companies, not the consumers, not the engineers, is being truly held accountable for it.

Nest Feathering: Jensen's Expectation

NVIDIA's Jensen Huang has been publicly suggesting engineers should be spending a ridiculous and unsustainable number of tokens every year. Quotes like "If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed,"

I'll let that land for a moment.

The person who manufactures the (very scare, over priced?) chips that power AI inference is setting an industry expectation that engineers should consume tokens at a rate equivalent to at least half of their own compensation.

Of course he is.

Jensen's business depends on it. The more tokens burned, the more chips needed, the more NVIDIA sells.

This isn't vision. It's demand generation. And in an environment where nobody's seeing the real bill, it's remarkably easy to sell, but it's hooking a lot of the technology industry in.

Don't be fooled, the 1/2 comp expectation on engineers burning tokens is nothing in the scheme of things in the overall cost of sale and I suspect it's actually coming out of NVIDIA's marketing budget.

The Bill Is Coming

Here's what I believe, and the mathematics support it:

Repricing is coming.

I don't know exactly when. I don't know the magnitude, partly because the companies won't tell us, and partly because the correction will depend on competitive dynamics we can't fully model from the outside. But the current pricing is not sustainable, the investment cycle has to be serviced, and the only mechanism available to close the gap is price.

When that happens, it won't just be a financial shock. It'll be a cultural one.

Engineers who've been told to use all the AI they can, to burn tokens freely, to parallelise, to iterate without cost constraint, are going to be told to stop / dramatically dial it back.

Timelines committed to boards and customers on the assumption of current pricing will need to be renegotiated. Revenue forecasts will need to be reset, guidance updated... Markets will respond...

We've seen Uber's internal AI token budget consumed in four months of a twelve-month year for 2026. That's not an edge case. That's a preview.

The Only Strategy That Can Work

There are several ways organisations can approach this. After modelling it out, I think all but one end badly.

You can be unaware repricing is coming. You'll be blindsided.

You can know it's coming but have no plan. You'll scramble.

You can have a plan but hold back on using AI now, waiting for the economics to stabilise. You'll fall behind while competitors advance aggressively on subsidised pricing.

These won't work.

Use the subsidy hard. And plan for when it ends.

Move fast now. Get ahead. Build the capability. Let your engineers use the tools.

Burn these AI companies subsidised addictive inducing tokens while they are available, everyone else is and you don't want to look that gift horse in the mouth.

But plan for the correction. Understand which workloads are genuinely valuable versus which are just cheap to run right now. Architect for flexibility. Don't lay off all your engineers in the belief that AI has replaced them, only to find you can't afford the AI when the bill arrives and you have no one left to do the work.

That's the trap. And it's a real one.

What We Don't Know

The honest answer is we don't know how bad the repricing will be. We don't have access to all the numbers.

Most of these companies are private. They're choosing opacity. They won't discuss the relationship between inference costs and revenue. They won't tell us how the unit economics actually work. And when people like Sam Altman are asked, the conversation gets shut down.

That opacity is itself a signal. Companies with good numbers talk about their numbers.

When these companies go public, transparency becomes mandatory. And when the real cost structure becomes visible, to investors, to analysts, to the market, the repricing conversation will accelerate.

That's probably where the correction begins. And it's probably closer than most people think.

I only had Kobe once on my last trip to Japan. It was delicious and worth every yen I paid. I love a lot of the AI I'm using, but at the same time I know I'm paying for a fraction of the price it's costing OpenAI or Anthropic and I know a time of reckoning is coming.

If OpenAI and Anthropic were public now, anyone shorting?

A Note On Inference Location / Open Source Models

Inference location is and will impact cost over time, and will, if planned for and executed well start to do this soon. Local workstation inference (think Mac M architectured shared memory), local inference servers in private data centres / IaaS and the like running open source models will provide some offsets, but definitely as a step back from managed services models.

This will still be at the margins however, with the ever present frontier model chasing behaviours we see etched in agentic AI culture.

This is the first in a pair of essays. The second explores what happens when repricing hits a market that's been deliberately designed to prevent you from switching — and why we've seen this pattern before, solved it before, and are choosing to ignore it now.