AI Model Costs: When Does Self-Hosting Break Even for Business?

01 / The story

The business decision is a deployment decision

Open-weight AI can cost less than a hosted API, but only when the whole service—not just the model download—works at a lower cost for acceptable results. Volume, engineering, hardware use, quality and uptime decide the answer. A cheap API can also make buying servers harder to justify.

This is a business deployment comparison, not a laptop buying guide or a comparison with personal chat subscriptions. We examine a service that processes work for an organization. Hosted API pricing is a dated fact; every workload, pass rate and self-host budget below is explicitly hypothetical. Qlauson has not measured model performance.

The useful question is: what will it cost to deliver an accepted result, under our data rules and deadline? That question prevents a finance team from comparing a token bill on one side with an incomplete hardware budget on the other.

02 / The story

Hosted and open-weight are not opposites

An API provider runs the model and bills your application for use. With self-hosting, your team runs downloaded model weights—the learned numerical values—on owned equipment or rented infrastructure. Open weights describe access to a model; hosted describes how it is served. A business can buy hosted access to an open-weight model.

A hosted service reduces some infrastructure work, but your team still owns integration, access controls, evaluation and fallback design. Self-hosting adds model serving, capacity planning, patching and incident response. Neither route removes the need to protect logs, tools and stored documents.

For procurement, compare a complete hosted service with a complete self-hosted service. Do not compare a managed API with one unprotected development server, or assume that paying for a cloud account automatically buys an uptime commitment.

03 / The story

Three official hosted price points, checked October 19, 2026

The table uses USD per 1,000,000 tokens for text inference. Input is what you send; output is what the model generates. These examples show different billing structures, not equal capability. Use each provider’s billable token counts rather than assuming identical text produces identical counts. [1] [2] [3]

DeepSeek’s row deliberately shows both time bands. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday, excluding Chinese public holidays. Other hours, including weekends, are off-peak. Its pricing page lists the offered API name deepseek-flash; no release story or model-performance claim is needed for this cost analysis. [1]

Google lists the displayed Gemini rates through December 31, 2026. From January 1, 2027, its announced Standard rates are $1.50 input, $7.50 output and $0.15 cached input per million tokens, with cache storage at $1.00 per million token-hours. This is a published schedule to recheck, not a promise about your future invoice. [3]

04 / The story

Cache discounts are conditional, not a default budget

Anthropic documents a 512-token minimum cacheable prompt for Sonnet 5.5 on its first-party platform. Hits require identical prompt segments. The default cache lifetime is five minutes; a one-hour option costs more to write. A workload with changing prefixes or few repeats may not recover its write cost. [4]

DeepSeek enables context caching by default, but a hit requires fully matching an already persisted cache prefix unit. Its documentation calls this best effort, not a guaranteed hit rate. Use reported hit and miss token counts when estimating the bill. Google’s cache storage also makes retention time part of the calculation. [5] [3]

Tool use can sit outside token charges. DeepSeek’s page lists standalone Search and Fetch at $0.75 per 1,000 successful requests each, and says specified server-side calls through DeepSeek Harness, Claude Code or Codex are currently free, with per-request billing starting November 1, 2026 at 00:00 Beijing time. Do not turn that temporary condition into a permanent saving. [1]

05 / The story

Build a self-hosting budget before estimating savings

A downloaded model does not arrive with a free production service. Include equipment purchase or rental, power, cooling, storage, network transfer and replacement parts. Spread a purchase across a chosen useful life for monthly TCO, but keep the upfront cash payment visible in a separate cash-flow plan.

Add engineering for deployment, load testing, monitoring, upgrades and security. Include initial integration work spread over the evaluation period and the recurring cost of people maintaining the system. Existing staff time is not free: assigning it here may delay other work.

Budget for spare capacity, backups, failover tests and on-call coverage. A target for service availability changes the design and the cost. Redundant hardware only helps if traffic can switch safely and enough capacity remains during failure or maintenance.

Licensing needs its own check. The inspected DeepSeek-V3 LICENSE-MODEL grants rights subject to conditions and use restrictions, including obligations for redistribution and hosted access. That is one identified license, not a claim about newer DeepSeek models. Check the exact weights, version, code and dependencies with your legal team. No model royalty should be assumed either zero or payable without reading its terms. [6]

No cloud GPU hourly price or hardware sizing claim is used here. We have not verified a matching configuration, region and capacity quote. Memory needs depend on the chosen weights, precision, context and concurrent requests; a purchase recommendation would require those details and a load test.

06 / The story

Worked example: cost per successful business task

Every number in this scenario except the cited Sonnet rates is hypothetical. Suppose a business submits 100,000 document-processing tasks per month, one model attempt each, with 4,000 input tokens and 800 output tokens per attempt. Assume no cache, tools, extra reasoning usage, retries or tax. These are total billed token assumptions, not visible-answer estimates.

At Sonnet 5.5 base rates, one attempt costs (4,000 × $2 + 800 × $10) ÷ 1,000,000 = $0.016. For 100,000 attempts, the token subtotal is $1,600. Assume another $1,000 monthly for application operation and engineering: hosted TCO is $2,600. [2]

Now assume a 90% accepted-result rate, defined by a fixed business review test. That yields 90,000 successful tasks. Hosted cost per success is $2,600 ÷ 90,000 = about $0.0289. The remaining 10,000 tasks fail; this scenario does not silently retry them or call them completed.

For a hypothetical self-hosted alternative, assume $1,200 monthly for equipment allocation, power and hosting; $1,300 for engineering; and $500 for monitoring, security and availability provisions. Those are invented budget allowances, not supplier quotes. Total fixed cost is $3,000. Add an assumed $0.002 variable cost per attempt: 100,000 × $0.002 = $200, making TCO $3,200.

Assume its acceptance rate is 80%, giving 80,000 successes. Self-host cost per success is $3,200 ÷ 80,000 = $0.0400. In this invented case, self-hosting loses despite its lower variable cost. Neither pass rate belongs to an actual model; no quality ranking follows from the example.

07 / The story

Solve for break-even—and check whether capacity permits it

Let N be monthly attempts. Keeping the invented budgets and acceptance rates unchanged, hosted cost per success is ($1,000 + $0.016N) ÷ (0.90N). Self-host cost per success is ($3,000 + $0.002N) ÷ (0.80N). Equalize these two expressions to compare the same outcome unit.

The arithmetic is 0.80 × (1,000 + 0.016N) = 0.90 × (3,000 + 0.002N). Therefore 800 + 0.0128N = 2,700 + 0.0018N; 0.011N = 1,900; N = about 172,727 attempts per month.

This is a conditional crossover, not a buying threshold. It assumes both systems can serve that volume at those costs and maintain their invented acceptance rates. If another server, engineer or support tier is needed, the cost curve changes. Quality-adjusted crossover is not the same as equal total monthly bills, because the successful output counts differ.

At 200,000 attempts, the same assumptions give hosted TCO of $4,200 and 180,000 successes, or $0.0233 each. Self-host TCO is $3,400 for 160,000 successes, or $0.02125 each. Review labor could reverse that narrow advantage. For example, an extra hypothetical 30 seconds of repair at $30 per hour costs $0.25 per affected task—far more than these inference amounts.

08 / The story

Which business conditions favor each route?

A developer testing demand often benefits from hosted usage because a small experiment does not need dedicated serving infrastructure. A startup should stress-test peak traffic and credit expiry before turning a trial bill into a production forecast.

A company with steady, high-volume routine processing has a stronger reason to investigate self-hosting, provided the model passes its tests and equipment stays usefully occupied. Large daily spikes can leave owned capacity idle for most of the day. Non-urgent queues may improve utilization, but only if deadlines permit.

A business with strict data rules should first identify which deployment is allowed. Self-hosting may offer control, but brings access, retention and audit duties. A permitted hosted endpoint with suitable terms may also be viable. For coding or research, include tool calls, long outputs, review and failed multi-step runs; a low token rate does not establish lower project cost.

09 / The story

Make the decision with a representative pilot

Choose a fixed set of real, permitted tasks and define acceptance before testing. Include ordinary cases, long inputs, peak concurrency and failures. Record the exact model, deployment settings, billed tokens, cache hits, completion time, retries and human review minutes. Report costs per attempt and per accepted outcome separately.

Build low, expected and high volume budgets, including the traffic pattern rather than only a monthly average. Test failover and account quotas. Then compare equivalent privacy rules, deadlines and quality. A cheap deployment that misses the required service level is not a substitute.

Watch official price changes, license revisions and evidence from your own workload. Open weights can give businesses useful control and bargaining options; low hosted prices can remove the economic case for running servers. Choose the lowest sustainable cost per accepted task, not the cheapest-looking token row.