When does private AI become cheaper than a global API?

Two months ago, in a meeting with a bank's CFO, we got a direct question: "We now pay $6,200 per month to a global API. Will a private build save us money, or cost more?" We asked for a day. We came back with a curve. The answer was neither "yes" nor "no," but: "Savings begin at 500M tokens per month. You are currently at 620M."

That is the correct way to think about AI economics. Not in choices, but in curves. Not in prices, but in functions. This piece simplifies the math, and shows when private AI wins — and when it loses.

The visible cost: token price.

What appears on the invoice is what accountants call the "direct variable cost": how many dollars per million tokens. Global API vendors compete aggressively here: modern models range from $3 to $20 per million output tokens, and $0.50 to $5 per million input. For light use, this is cheap. For heavy use, it compounds fast.

Private AI, on the other hand, runs on a different equation: a near-fixed cost (server, power, maintenance) plus a minimal variable cost (inference runtime). Different equation, different curve.

The hidden cost: four line items not in the quote.

If you stop at comparing token prices, you are comparing one-third of the equation. The other two-thirds are cleverly buried in the terms of service, or not mentioned at all:

• Data egress fees: some providers charge for data leaving their network. At scale, this can be 8% to 15% of the bill.
• Compliance cost: every time a regulator changes something (a new data-protection clause, for example), you need legal review and sometimes contract updates. That is time and money.
• Lock-in cost: after six months of building on a specific API, migrating costs at least three team-months. The price is low today? Look at the price in two years.
• Latency cost: a geographically distant API adds 200–500 ms to every call. In a chat bot, that is the difference between "sharp" and "sluggish." The difference translates to customers who leave, and that loss never hits the AI line item.

Break-even: the simple math.

Assume the global API costs you p dollars per million tokens. Private AI costs you F fixed per month, plus q dollars per million tokens. Your monthly volume is V million tokens.

Private AI is cheaper when: F + q·V < p·V, i.e. V > F / (p − q).

In the bank case: F = 5,000; p = 10; q ≈ 0. Break-even = 5,000 / 10 = 500M tokens per month. The bank consumes 620M. Private AI saves, per month: (10 − 0) × 620 − 5,000 = $1,200. Not a huge saving, but enough to justify the move — and with performance gains, savings compound over time.

The advertised token price is as tempting as a hotel's first invoice. The real bill comes at checkout.

Inside the invoice: anatomy of total ownership cost.

The figure below compares a real case for one of our clients who consumes 800M tokens per month. Notice not only the total, but its composition — line by line:

When is it worth it? Three conditions.

Private AI is not the default answer. It is the answer when three conditions coincide:

• Volume exceeds break-even with a 20% margin (to absorb fluctuation).
• Data sensitivity that justifies sovereignty: bank, health, government, law, education.
• Time horizon of at least 18 months — infrastructure takes two months to build and six to tune.

When is it not worth it?

Conversely, three cases in which we refuse to sell private AI, even if the client asks:

• Volume below 200M tokens per month — the math does not work, and the ops overhead is a drag.
• Proof-of-concept stage: do not build a factory to test an idea. Use an API first, migrate later.
• A technical team not ready to operate infrastructure. Private AI needs operations, not just developers.

A real (anonymized) example.

An insurance client consuming 1.2B tokens per month on a global API. Visible monthly bill: $12,000. True bill with hidden costs: $14,200. We built them a private AI stack in a local data center, with a one-time build cost of $35,000 and a monthly run cost of $6,500.

Monthly gap: $7,700. Payback period: $35,000 / $7,700 = 4.5 months. After that, every dollar of saving is net. Year one net saving: $57,400. Year two, after the build is amortized: $92,400.

This does not include what the math cannot count: data control, a 300 ms latency improvement, and the ability to fine-tune the model on private insurance records. The last alone raised resolution rate on their customer bot from 52% to 74%.

Closing.

AI economics is not an ideological question ("global cloud" vs. "digital sovereignty"). It is a calculation. Before signing a long contract with any provider, compute: What is your current volume? What is your expected growth? What is the break-even of your private option? And how long until the build cost is recovered?

At Nuqta, we run this calculation free for any company that asks. If the math favors a global API, we say so. If it favors private AI, we say so too. Numbers, not enthusiasm, are what sell a solution that lasts five years.