Price the Build Before You Decide You Cannot Afford It

Ask a founder why they are renting the most important component in their product and you get a version of the same answer. It costs millions to build. We are not a research lab. Nobody our size does that.

Then ask what the build would actually cost for the thing they specifically need. Almost nobody has the number. They have a number for something else, usually the largest version of the thing, usually quoted by someone selling the rental.

A price you did not calculate is not a constraint. It is a rumour you organised your company around.

In late 2025, a lab in Kingston called Maestro AI Labs finished training a large language model from scratch. Not a fine-tune. Not a wrapper over somebody else's weights. An original model, trained on data whose origin they recorded, in a country with no AI industry ten years ago. It is called Maestro, it is still in red-team testing, and it is not going to beat a frontier lab on hard reasoning. The team says so themselves.

That last part is why the example is useful rather than inspirational. They did not out-spend anyone. They worked out what they actually needed, priced that, and found the number was not the one everybody repeats.

The number everybody repeats is for a different product

When DeepSeek published its V3 technical report, it put the final training run at roughly US$5.6 million in GPU hours, and stated plainly that the figure excluded prior research, data work, ablations and failed runs. Both halves of that got misread. Half the internet decided frontier models now cost five million dollars. The other half decided the number was a lie.

Neither reading is the useful one. The useful one is that the cost of building depends entirely on which model you need, and almost every founder quoting a build price is quoting the cost of a model far larger than the one their product requires.

The same distortion runs through every build-versus-rent decision you will make, and it is not specific to AI. The quoted price of building a payments stack is the price of building Stripe. The quoted price of building a data warehouse is the price of building Snowflake. You do not need Stripe. You need to take card payments in two countries.

Run the arithmetic you have been avoiding

Take the component you have already decided you cannot build. Model calls, a data pipeline, a scoring engine, whatever carries the most value in your product.

Write down what you pay for it today. Say it is US$9,000 a month. That is US$108,000 a year, and US$324,000 over three years, and that number grows with revenue, because your rent scales with your usage.

Now write down what building the version you actually need would cost. Not the frontier version. The narrow one that does the one job your customers pay for. Two engineers for four months at US$9,000 a month fully loaded is US$72,000. Add US$40,000 of compute and data work and you are at US$112,000, plus something like US$3,000 a month to run and maintain it, which is US$108,000 over the remaining period.

Build: about US$220,000 over three years, front-loaded, and the asset is yours. Rent: US$324,000 over three years, back-loaded, growing, and it belongs to somebody else.

I have made that arithmetic land in favour of building, because it is the case founders never model. Run it with your real numbers and it will often land the other way, and that is a legitimate result. Renting is the right default for most startups, most of the time. The failure is not renting. The failure is renting without ever having written the second column.

Three triggers that flip the answer

The comparison changes character when any of these arrive, and they tend to arrive at the worst moment.

The rent becomes the margin. When the component's cost is a large share of gross margin and scales with revenue, growth stops improving your economics. You are running a business whose best outcome is a bigger payment to your supplier.

Someone asks a question your vendor cannot answer. A regulated customer asks where the data is processed, or what the model was trained on. If your answer is that you do not know and cannot find out, the deal walks on a fact about your architecture, not on your price.

The capability can be withdrawn. In June 2026 a frontier model was suspended worldwide in an afternoon on a foreign government order, with no notice and no migration path. Every product that had wired it into production found out the same day. No service-level agreement binds the government sitting above your vendor.

The cheap version of building is not building everything

The answer is rarely all or nothing. It is usually a hybrid: something you control handling the routine volume, and the expensive rented capability called only for the fraction of work that genuinely needs it.

For an AI product that means a smaller model you host doing the classification, extraction and drafting that makes up most of your calls, with a frontier model reserved for the hard tail. You keep the capability where it earns its price, you cut the variable bill on the boring 80 percent, and when a vendor disappears your product degrades instead of stopping.

Where this argument is weakest: a hybrid is two systems, and two systems cost more attention than one. A four-person team that splits its focus to save US$40,000 a year has probably made a bad trade, and I would not fault a seed-stage founder for staying on pure rent until the numbers are big enough to be worth the complexity. The discipline is to know which side of that line you are on, with numbers, rather than assuming.

Frequently asked questions

Should my startup train its own model?

Almost certainly not. Renting a frontier model is the correct default for nearly every startup, because the capability is genuinely better and the cost is variable rather than fixed. The point is that you should reach that conclusion by comparing two numbers you wrote down, not by repeating a price you heard at a conference. The founders who get hurt are the ones who never ran the comparison and then discover at scale that the rent was the whole margin.

How much does it cost to train a large language model from scratch?

There is no single number, which is the problem with the one everybody quotes. DeepSeek reported the final training run of DeepSeek-V3 at roughly US$5.6 million in GPU hours, and was explicit that the figure excluded prior research, data work and failed runs. Smaller models trained on narrower corpora cost dramatically less. Frontier runs cost dramatically more. Your build quote depends on the model you actually need, not on the largest model anyone has trained.

When does renting become the wrong call?

Three triggers. When the rent line becomes a large share of gross margin and grows with revenue. When a compliance, residency or provenance requirement arrives that your vendor cannot answer. And when the vendor can withdraw the capability with no notice, which stopped being hypothetical in June 2026. Any one of those turns a convenience into a structural dependency.

What is the cheapest version of building?

A hybrid. Run a smaller model you control as the always-on baseline for the routine volume, and call the frontier model only for the fraction of tasks that genuinely need it. You keep frontier capability where it pays for itself, you cut the variable bill on the boring 80 percent, and if the frontier disappears your product degrades instead of stopping.

Sources

1. DeepSeek-AI, "DeepSeek-V3 Technical Report," on the reported final training-run cost in GPU hours and its stated exclusions.
2. Maestro AI Labs, Kingston, Jamaica, on Maestro, trained from scratch in late 2025 and in red-team testing: maestroailabs.com.
3. The June 2026 worldwide suspension of a frontier model under a national-security export-control directive, as reported at the time.