The basics
What is sovereign AI? Under 243-FZ, Russia’s AI law, from 1 March 2027 it is a model that a Russian legal entity develops and fully reproduces, including training, with answers and data processed in Russian data centers; in everyday terms, it is a model that works at your site and sends no data out. Can you run foreign cloud models yourself? The flagship cloud models, no: their weights are closed; you can run open models: Russian ones with 32 and 8 billion parameters, and the world’s open families. Do you need internet? Only to download the model. How much worse are local models? On short chains of up to 3 steps, a 9-billion-parameter model did not fall behind cloud models in a test for a retail chain; on a 15-step task, no model gave a fully verifiable result.
Hardware
What runs on an ordinary laptop with 16 GB? Models of 4–8 billion parameters in 4 bits, 6.3 GB with a context of 8,000 tokens. Which Mac to buy? One with 128 GB for 70-billion models: a MacBook Pro for $8,100 or a Mac Studio for $11,000. Which graphics card? An RTX 5090 with 32 GB for $9,000 on 30 September 2026 for models up to 32 billion, and an RTX PRO 6000 with 96 GB for $13,000 at the February 2026 price for 70 billion and a department. What does a server cost? About $30,000 with two 96 GB cards, from $120,000 with four H100s. How much power? One RTX 5090 around the clock is about $39 a month, a server with four H100s about $250. How often does it break? By our calculation from Meta’s report on training on 16,384 H100 cards (419 failures in 54 days), one serious failure per about 2,100 days of one card’s work under full load.
Money
What does local AI cost per month? From $300 for a laptop to $6,100 for a server with four H100s, with power, hosting and an engineer. What does the cloud cost? At 60 million tokens a month, from $12 to $1,200 at prices on 29 September 2026. When does the local option pay back? When the cloud bill is clearly above the cost of your own hardware, or when data must not go out; by our calculation, a Mac Studio pays back in 10 months against Fable 5.1 and in 71 against Sonnet 5.5. What about rental? H100 from $3 an hour, from $1,700 a month at Russian providers.
Law
Is it legal to use foreign APIs from Russia? Russia is not on the supported country lists of Claude and OpenAI on 29 September 2026; OpenAI has blocked API traffic from such countries since 9 July 2024 and warns that access from outside the list may lead to the account being blocked. Most foreign suppliers have no data processing agreement under 152-FZ, Russia’s personal data law. Must a business use Russian models? No, there is no general ban; the criteria of 243-FZ take effect on 1 March 2027, and from the same date the government may define areas where only sovereign or national models are allowed; for systems created before that date, such limits do not apply until 1 September 2032 if data is processed and stored in Russia. Can you give personal data to a local model? Yes, if you meet your duties as a data operator: a local model removes cross-border transfer and third parties. What goes in the policy? Three classes of data: may go out, Russian cloud, inside only.
Adoption
Where do you start? With a data policy and 50 test requests. Do you need an engineer? Not for one computer; for a department server, 10–20% of a full-time role by our estimate. How long does adoption take? A pilot in 14 days from $1,499, a department in 4–6 weeks from $4,999. Can you move to the cloud later, or out of it? Yes, if the agents are built through a single model interface.