Self-hosted AI vs API costs: the 2026 break-even

For agent workloads in 2026, self-hosted AI breaks even startlingly early. A £4,899 desk-side box such as NVIDIA's DGX Spark, at moderate utilisation, works out around £4–4.50 per million generated tokens — small-model API territory, roughly a third of mid-tier output rates — and input tokens stop being metered altogether. Against a mid-tier API bill the box can pay for itself in about six months. Against small-model pricing or spiky volume, the API still wins.

What are the three ways to buy AI inference in 2026?

There are three ways to pay for the same fundamental thing — a model turning input tokens into output tokens — and they have almost nothing in common commercially.

Frontier API. You rent intelligence per token from a closed-model provider. As of July 2026, public rate cards put frontier pricing at roughly $5 per million input tokens and $25 per million output for Opus-class models, about $3 and $15 for Sonnet-class mid-tier models, and $1 and $5 for small models. Zero upfront, zero ops, always the newest model, elastic to any burst. The catch is structural: the bill scales linearly with volume, forever. There is no volume at which the API is paid off.

Hosted open-weight. A cloud provider serves an open model — GLM-5.2, DeepSeek V4, Qwen3.5 — from their GPUs, priced per token or per GPU-hour. Cheaper per token than frontier, no capex, and you inherit the open-weight capability ceiling. But your data still transits someone else's infrastructure, and the linear-scaling problem is untouched. You have swapped landlord, not tenancy.

Your own hardware. You buy the box once; tokens become a by-product of electricity. In 2026 this no longer implies a server room. NVIDIA's DGX Spark Founders Edition costs £4,899 including VAT with 128 GB of unified memory; Lenovo's ThinkStation PGX puts the same silicon at about £3,211; and the range runs up through a twin RTX PRO 6000 workstation at £36,999.98 — currently the cheapest box that serves DeepSeek V4 Flash natively — to a 748 GB DGX Station GB300-class machine at £117,599.99. We maintain a live catalogue with UK pricing at /self-hosted/hardware. The marginal token is close to free; the open questions are utilisation and ops, and the rest of this article is about exactly those.

What does an AI agent's token bill actually look like?

Most published cost models assume chat traffic: a person types a short question, the model writes a long answer, and input and output roughly balance. Agent workloads are the opposite shape, and the shape decides the economics.

Look at what one turn of a working agent actually contains. Usermode's named agents — Sarah in credit control, Henry in property compliance — run on standing daily schedules inside the customer's own Microsoft 365 tenant. A single turn carries the system prompt, the role's operating instructions, the schema of every tool the agent may call, the accumulated history of the run, and the full result of the last tool call: a ledger extract, an email thread, a certificate register. All of it is input, and much of it is re-sent every turn. The output is a decision — a tool call, a short email, a line in the audit ledger. A few hundred tokens out against tens of thousands in.

Three properties follow, and each one moves the break-even:

• •Agent bills are input-dominated. When every turn re-sends history plus tool results, input can outweigh output by an order of magnitude. The rate-card line that matters is the input price multiplied relentlessly, not the headline output price.
• •Agent traffic is heavily cacheable. The system prompt, role instructions and tool schemas are byte-identical across thousands of runs. API providers discount cached input for precisely this reason; on your own hardware the same repetition becomes a KV-cache engineering win instead of a billing concession.
• •Agent traffic is bursty and latency-tolerant. A fleet on morning crons hammers inference from 06:00 and idles overnight. Nobody watches the tokens stream. A scheduled run that takes four minutes instead of ninety seconds is invisible — what matters is that the chaser went out and was logged, not how quickly the prose rendered.

Input-heavy, cacheable, bursty, latency-tolerant: that is close to a specification for owned hardware. It is also why generic break-even figures deserve suspicion — most were computed for a different workload.

Where is the break-even on desk-side hardware?

Published break-even figures for self-hosted AI disagree to an almost comic degree: 2026 TCO analyses variously place the break-even anywhere from a few million tokens a month to tens of millions a day — the spread tells you the modelling assumptions matter more than the headline. Open those spreadsheets and the same two assumptions appear: datacentre GPUs rented by the hour, and a dedicated ops engineer on a six-figure salary to mind them. Price in a mostly-idle H100 and half a DevOps hire and of course self-hosting looks brutal. Neither assumption survives contact with a £4,899 box under a desk.

So here is the arithmetic for 2026 desk-side hardware, laid out so you can rerun it with your own numbers. This is illustrative maths built from published specifications, not measured telemetry.

Take the DGX Spark Founders Edition at £4,899, running OpenAI's gpt-oss-120b at 35–40 tokens per second. Assume UK electricity at 27p/kWh, a draw of roughly 240W under load, and 45% utilisation — the box working a little under half the time, which has to cover prompt ingestion, bursty demand and quiet nights.

• •A month contains about 2.6 million seconds; at 45% utilisation the box is generating for roughly 1.17 million of them.
• •At 35–40 tokens per second, that is about 41–47 million generated tokens a month.
• •Electricity, costed pessimistically at full load around the clock: 240W × 720 hours ≈ 173 kWh ≈ £47 a month.
• •Hardware amortised straight-line over three years: £136 a month.
• •Total: about £183 a month for 40-odd million generated tokens — roughly £4–4.50 per million, or $5–6 at a round $1.30 to the pound.

Set that against the rate cards. Per generated token, the box lands at small-model API prices while running a 120-billion-parameter model. It is roughly three times cheaper than mid-tier output and four to five times cheaper than Opus-class output. And input stops being a metered line item entirely: prompt ingestion just consumes duty cycle you have already paid for — which, for the input-dominated agent traffic described above, is where the real money was.

The break-even is not a token count; it is a workload shape.

Payback makes it concrete. Take an illustrative fleet burning 300 million input and 15 million output tokens a month — comfortably within one Spark's capacity. On mid-tier API pricing that is about $1,125 a month, call it £865; against the box's £47 of running costs, the £4,899 purchase pays back in about six months. Against Opus-class pricing ($1,875, roughly £1,442), three and a half months. Against small-model pricing ($375, roughly £288), payback stretches to about 20 months — before counting one hour of anyone's time looking after the machine.

One honesty note before the comparison table: gpt-oss-120b is not an Opus-class model, and nothing here pretends otherwise. The like-for-like fight is the small-to-mid tier — which happens to be where most agent volume actually lives: extraction, classification, drafting, summarising private documents that never needed a frontier model. Where the open-weight ceiling sits this year is its own subject, covered in our 2026 open-weight landscape.

What are the hidden costs the TCO spreadsheets skip?

Amortisation plus electricity is honest arithmetic, but it is not the whole invoice of owning inference hardware. Six costs turn up after purchase, and a vendor who never mentions them — and Usermode sells this hardware — is selling, not advising.

• •Model churn. The open-weight leaderboard turns over in months: GLM-5.2 leads today, Meta has exited open weights entirely, and the leader two quarters from now may want a different quantisation or more memory than you bought. Every swap means downloads, re-quantisation and re-running your evals — an afternoon to a few days, several times a year.
• •Someone must own the box. Driver updates, inference-server releases, monitoring, the occasional restart. Not a hire for one machine — but a named owner, or the box decays into the office appliance nobody quite trusts.
• •Power, cooling and noise are class-dependent. A 240W Spark is office furniture. A twin RTX PRO 6000 workstation or a GB300-class station is kilowatt-class with acoustics to match; that machine lives in a comms room, and the comms room needs the cooling budget.
• •There is no SLA. When the box fails, inference stops until someone fixes it. The boring mitigation: keep API keys warm as failover, and accept that an outage week shows up on the API bill instead.
• •The capability gap is real. Open weights trail the closed frontier by roughly three to six months. For bulk work this is irrelevant; for the judgement-heavy step in a workflow it can be the entire answer.
• •Utilisation is the whole game. Everything above assumed 45%. A box bought for one workflow that runs 5% of the time carries an effective token cost nine times worse — worse than the API it was meant to beat. Buy against measured volume, not ambition.

When is the API still the right answer?

Often — and a piece written by a hardware vendor owes you this section more than any other.

The API remains the right call when volume is low, spiky or simply unmeasured, because capex that idles is the most expensive compute there is. When the task genuinely needs frontier capability this quarter — contract nuance, novel judgement, the hard ten percent — rather than in three to six months. When humans are watching the response render, because 35–40 tokens per second is fine for an overnight batch and noticeable in live chat. When nobody in the organisation will own a physical machine. And when you need this week's model the week it ships.

In practice the estates worth running are rarely pure. The pragmatic posture is a split: bulk, repetitive, privacy-sensitive agent traffic on the box; the occasional frontier-grade judgement call over the API, with the same approval gates and audit trail either way. And if your real driver is sovereignty rather than cost — data that must not leave the building, whatever the spreadsheet says — then the arithmetic above is the board memo, not the reason. We wrote the sovereignty case separately in GDPR and data sovereignty for AI agents.

A note on our own incentives, since candour is cheaper than trust recovered later: Usermode bills AI usage at cost — there is no margin for us on your token bill on either side of this trade (see pricing). The recommendation follows your workload, not our revenue line.

How should you decide between self-hosted and API?

Three questions — volume, sensitivity, latency — decide nearly every case, with capability and ops as the tie-breakers.

If the left column dominates, stay on the API with a clear conscience; the arithmetic is on your side. If the right column dominates, 2026 desk-side hardware has quietly moved the numbers in your favour — and you should not take our worked example's word for it. The self-hosted page has an interactive TCO calculator with frontier-model presets: put in your own volumes, electricity price and utilisation, and see where your break-even actually lands. Usermode supplies the hardware, stands it up and supports it — current UK pricing lives at /self-hosted/hardware — and runs governed agent workforces on both sides of the trade.

If you want to walk your actual token bill through this arithmetic — and see the agents that generate it running live — book a demo.