1. Few tokens
At 60 million tokens a month, cheap cloud models cost $12–$100 and mid-range ones $250. At this volume a Mac Studio for $11,000 pays back against Sonnet 5.5 in 71 months, and against GPT-6 Luna never. By our calculation, your own hardware starts to pay off at hundreds of millions of tokens a month, or at a cloud bill of $800–$1,200 a month.
2. Long chains of steps
On chains of 15 steps in a test by a retail chain, no model gave a fully verifiable result. A large cloud model with 235 billion parameters partly succeeded twice as often as a local one with 9 billion. The longer the chain, the more model size matters. By our calculation, a 235-billion model like the one in this test needs two 96 GB cards on site, which is a server from $30,000. While no model solves long chains reliably, it is cheaper to pay the cloud by the bill, even $51 per million output tokens for flagship models, than to buy a server for such a task.
3. No engineer
By our estimate, a department server needs 10–20% of a DevOps engineer’s time, at a median salary of $2,700–$3,000, and up to $3,800 in Moscow for a mid-level engineer. Without a person, a server degrades: an update breaks compatibility, a card overheats, nobody reads the logs. If you have nobody to hire, choose the cloud or a rental with provider support.
4. Seasonal or peak load
Hardware costs the same in season and out of season, while the cloud charges only for tokens. Renting an H100 for two months of the season costs from $3,500 at the monthly rate, and buying one costs $28,000 plus a server: with a peak of two months a year, buying loses over 5 years.
5. You need vision, speech and text at once
Locally this usually means separate models for vision, speech and text, memory for each and their own engines. In the cloud one model takes a picture, a document and text. While the task is not yet defined, the cloud is cheaper at the stage of searching for a solution.
6. Data may leave the company
If your requests hold no personal data, code, finance or medical data, or you have a data processing agreement with a Russian provider under 152-FZ, Russia’s personal data law, the main argument for your own hardware is gone. Russian clouds give a contract with a Russian legal entity and keep the data in Russia.
7. You are only starting
A cloud pilot in 14 days will show the token volume, the length of chains and the quality on your data. Only after that do the numbers in the calculators become yours, not averages. Buying hardware before a pilot is buying blind.
How to test it in a week
Run 50 of your own requests on a mid-range cloud model and on an open 8–32 billion model on a rented card for about $0.50–$7 an hour, compare the quality and count the tokens. Put the volume into the calculators. If the local total is higher than the cloud one and the data may leave the company, the answer is the cloud, and that is a good answer.