What the study examines
The authors set up a cost model: one-off hardware costs plus monthly electricity costs against the API costs for the same token volume. They assume operation of eight hours a day and 20 days a month at 0.15 US dollars per kilowatt hour. API costs are normalized to the token volume the local hardware can produce in that time. From this follows after how many months own hardware is cheaper than the API.
Nine open models (among others Qwen3, Llama-3.3-70B, gpt-oss-120B, GLM-4.5, Kimi-K2) are compared with six commercial models (GPT-5, Claude 4 Opus and Sonnet, Grok-4, Gemini 2.5 Pro). The authors use as a criterion that open models should be within 20 percent of the top models in benchmarks. The benchmark and price data come from Artificial Analysis and are as of 2025.
The results
Small models up to about 30 billion parameters run on a single graphics card (in the study an RTX 5090 for about 2,000 US dollars). Break-even is at 0.3 to 3 months depending on the comparison. Medium models such as gpt-oss-120B, GLM-4.5-Air or Llama-3.3-70B need one to two A100 GPUs (15,000 to 30,000 US dollars) and reach 2.3 to 34 months. Large models such as Kimi-K2 with hardware costs of up to 240,000 US dollars lie between 4.3 and 69.3 months.
The study names three types. Small companies with fewer than 10 million tokens a month should start with small models. Medium companies with 10 to 50 million tokens are the sweet spot for medium models, and especially regulated industries such as healthcare and finance prefer medium setups. Large companies with more than 50 million tokens a month could run large models economically. Against very cheap APIs such as Gemini 2.5 Pro the study sees break-even times of 5 to 9 years there, so that non-financial reasons such as data protection, sovereignty or vendor lock-in often weigh more than price.
What the study doesn't include
Three limitations matter. First, the cost model counts only hardware and electricity. Staff, maintenance, updates and outages are not included; the authors name this as a task for future research. Second, it is a preprint, and the authors themselves write that the estimates still need empirical validation. Third, open models partly remain behind the top models in benchmarks. Qwen3-30B reaches 70.7 percent on GPQA per the study, GPT-5 85.4 percent. Whether that suffices for your task you have to test yourself.
There is also a point the study doesn't highlight but which we derive from its model: the break-even times only apply if the hardware is utilized. A single graphics card produces about 114 million tokens a month per the study. That is far more than many small companies consume. The study itself names fewer than 10 million tokens a month as the order of magnitude for small companies.
An example calculation with current prices
We recalculated with the study's formula how things look at 10 million tokens a month. Assumptions: hardware 2,000 US dollars as in the study, without electricity and operation, ratio of input to output tokens as in the study (one third input, two thirds output). As API comparison we take the prices VentureBeat names in October 2026: Claude Sonnet 5.5 at 2 and 10 US dollars and Claude Opus 5.5 at 4 and 20 US dollars per million tokens. Against Sonnet this gives about 73 US dollars of API cost a month, so about 27 months to break-even. Against Opus it is about 147 US dollars a month and about 14 months. At 100 million tokens a month both shorten by roughly a factor of ten.
This is a rough calculation. It ignores electricity, staff and the quality differences between a 30-billion model and a top model. We haven't checked the comparison prices with the vendors. But it shows that the study's figures of 0.3 to 3 months occur in practice only with high utilization and against expensive APIs. Since 2025 premium prices have also fallen: Claude Opus 4 then cost 15 and 75 US dollars; per the figures named, Opus 5.5 is now at 4 and 20 US dollars.
Cost control: predictability, not just price
From the perspective of IT managers, besides the average price it matters how predictable costs are. Own hardware causes fixed costs for purchase and electricity, regardless of how many requests come in. API prices, by contrast, are variable and change. Google launched Gemini 4 Argon with an introductory price of 2 and 10 US dollars which, per the pricing page, ends on 30 November and doubles to 4 and 20 US dollars from 1 December. Gemini 3.8 Flash doubles on 1 January 2027. Anyone building on API prices must plan for such jumps; anyone running their own hardware carries other risks instead, such as utilization.
Data protection: what local changes and what it doesn't
The study names local operation as a solution for sensitive areas such as healthcare, finance and law. With operation in your own house, requests and documents don't go to an external provider. That removes, for example, questions about transfer to a third country and about data processing for inference. Our assessment: GDPR duties don't disappear with it, they lie solely with the operator. These include access rights, logging, deletion concepts and securing the appliance. Whether a concrete application is permissible depends on the data and purpose; you should clarify that with your data protection officer.
For whom an own appliance is more likely to pay off
- When the load is constant, such as document processing, internal knowledge search or customer service automation with many requests. Sporadic use doesn't utilize the hardware.
- When the data is sensitive or sector rules limit data residency (per the study, for example healthcare, finance, law).
- When a medium or small model suffices for the task. For tasks that need a top model the API is often the better choice.
- When predictability matters more than the lowest price per token.
- When someone can take over operation, internally or as a service. Without operation, hardware and model are only half the solution.
- Hybrid: sensitive and regular tasks locally, rare peaks and especially demanding tasks via the API. The study explicitly recommends this mix for medium setups.
What remains open
The study is a 2025 preprint with prices and models of that time. Since then new open and commercial models have appeared, and API prices have changed. Costs for operation and staff are not modeled, and the hardware prices are the authors' assumptions. For a decision you need your own calculation with your volume, your model and your operating costs.