Open-weight models are good enough now that self-hosting is a real option rather than an ideological one, and the obvious question is at what volume it starts paying. The published answers are not close to each other.
One analysis puts the break-even at roughly 500,000 tokens a day. Others put it above 2 million. Others argue the maths only genuinely flips past 100 million tokens a day or under a hard privacy requirement. That is a spread of more than two hundred times, between sources published in the same year about the same decision.
Treat all of these as weakly sourced — most are vendor-adjacent or SEO content rather than audited studies. But the disagreement itself is the useful finding, because the reason for it is knowable.
Why the estimates diverge
The low numbers compare a GPU rental invoice against an API invoice. The high numbers include the things that surround the GPU. That is the entire gap.
Utilisation is the big one. A rented GPU costs the same whether you send it traffic or not. Real product traffic is bursty — busy for a few hours, near-idle overnight — so effective utilisation is often a fraction of what a back-of-envelope calculation assumes. If you use 20% of the capacity you rent, your true cost per token is five times the number you started with. Everything else is rounding compared to this.
People are the other one. Running inference in production is a job: deployment, model updates, scaling, monitoring, incidents at inconvenient hours. Estimates for a minimum viable team run to one and a half to two engineers, which any founder can price against their own payroll and get a number far larger than their API bill.
The low estimates are not lying. They are answering "what does the hardware cost" while the decision is "what does the capability cost".
The ground moved this summer
Any break-even calculation from earlier this year is now stale in the API's favour. Average inference pricing fell roughly 43% over ten weeks to around $1.16 to $1.18 per million tokens, driven by a price war rather than by anything structural.
A falling price on one side of a comparison pushes the crossover further away. Self-hosting is a fixed commitment; API pricing has been dropping. Committing capital and headcount to escape a cost that keeps falling is a bet that the price war ends, which it may — but that is the bet, and it is worth naming.
Work out your own number
Four inputs, all of which you can get today.
Your real token volume. Not your projection — last month's actual usage from your provider dashboard, split by model.
Your real blended rate. Divide last month's bill by last month's tokens. This is almost never the headline price, because it mixes models, caching and retries.
Realistic hardware cost. Take the hourly GPU rate for the instance that fits your model, multiply by 730 hours, then divide by the fraction of capacity you will genuinely use. Guessing 30% is more honest than assuming 100%.
The engineer. Even a conservative fraction of one salary usually dominates everything above at small scale. Include it, or the exercise is theatre.
If the answer is not obvious in favour of self-hosting after that, it is a no. Marginal wins do not survive contact with an incident at 3am.
When self-hosting genuinely wins
Three cases, and they are mostly not about cost.
Data cannot leave. Health, legal, defence, or a customer contract that forbids third-party processing. This decides the question by itself and the cost comparison is irrelevant.
High, steady, predictable volume. Not bursty consumer traffic but constant batch work — enrichment, classification, transcription — that keeps a GPU genuinely busy around the clock.
A small fine-tuned model doing one narrow job. This is the underrated case. A small open-weight model tuned for a single task can beat a frontier model on that task while running on hardware you can afford, and it does not need frontier capability because it is not doing a frontier job.
Notice that none of these describes a typical early SaaS calling a general model for a general feature.
The bottom line
For most small teams in 2026 the API wins, and it won more decisively this summer than it had in the spring. Self-host when data residency forces it, when your volume is high and genuinely steady, or when a small tuned model does a narrow job well. Otherwise, spend the engineering time on the product and let the price war work for you — but run the four-input calculation yourself, because the published break-evens disagree far too much to borrow.