For our bootstrapped AI project, access to compute became a serious budget constraint.
We were doing our own model work without venture capital. That does not mean our training scope or costs were comparable to frontier-model companies.
I want to be honest about why we can even attempt that, because the honest version is uncomfortable.
We got lucky, and luck isn't a strategy
Before this, we built a fintech company. It made money.
We still earn monthly revenue from a SaaS business we wound down.
That cushion is the only reason we could absorb what building AI costs.
Without it, we would not have started. Full stop.
The cost is the gate
Ideas, engineering, data and evaluation still mattered to our project.
The compute bill was another constraint we had to solve.
Fine-tuning costs depend on the model, method, dataset, hardware and iteration plan. The provider quote below was for our requested setup, not a minimum price for building with AI.
We were quoted $14,000 a week by one provider.
Editorial update, October 11, 2026: This is the author-reported provider quote; a corresponding invoice or quote document is not supplied here. The original headline was “The Barrier to Building AI Isn’t Talent. It’s Capital.” The later training postmortem reports approximately $50 in GPU compute for a particular run. That is a compute line, not an all-in engineering, data or evaluation cost, and does not make this earlier quote a general entry price.
That cost lands before you ship a single feature. Before you even know if the thing works.
So the question stops being "can you build it?"
It becomes "can you afford to find out?"
This isn't an anti-VC rant
Venture capital isn't the villain here.
For a lot of companies it's the right tool, and the best investors earn their keep many times over.
The problem is narrower, and more structural than "VC bad."
When a proposed experiment exceeds our budget, access to capital constrains which version we can attempt. Cheaper model adaptation, smaller experiments and hosted APIs can offer different paths; this account does not compare their costs.
The model doesn't get a vote. Neither does the market.
The check does.
What that does to who gets to build
Expensive training projects can narrow the field to teams with funding or existing resources. That is not the entry requirement for every AI product.
Access to capital is not a pure measure of a builder’s ability.
For those expensive projects, the budget can filter who gets to try.
The most capable builder in a mid-size Canadian city, or in Lagos, or in São Paulo, may simply never get to try.
Not because they can't.
Because they can't afford the room.
Why this matters even if you'll never train a model
When expensive AI training is accessible only to well-funded teams, the people able to attempt that work are a narrower group.
And it inherits that slice's blind spots.
The tools everyone ends up depending on get shaped by whoever could afford to make them.
That should bother you even if you never plan to touch a model yourself.
What we're trying to do about it
We can't fix the cost of GPUs.
But we can refuse to hoard what we learn paying it.
Editorial update, October 11, 2026: The public CBLRE repository provides a harness and methods. Its gated evaluation data require separate access; public code is not proof of a public item bank or independently reproduced scores. Readers should inspect each release’s actual files, license and access conditions.
Release availability varies by artifact. Public model weights, source code, datasets and methods have separate licenses and access conditions; this archive does not establish that every artifact is openly released.
The goal is simple: lower the entry fee for the next team by as much as we possibly can.
We made it into the room the slow way.
The least we can do is leave the door open behind us.
The line I keep coming back to
For our project, access to compute became a barrier alongside the work itself.
Our funding cushion gave us room to attempt that work.
Until that changes, the most useful thing the people already inside can do is hand the next person a key.