Table of Contents
For a private document assistant serving about 300 employees, a two-GPU on-premise server costs roughly $76,200 over three years: about the same as an upper-tier cloud API ($80,800), more than mid-tier and small models, and less than half a premium model. Above a few hundred regular users, on-premise is cheaper; below that, APIs are.
Key takeaways
- At the prices we checked on 10 November 2026, cloud API costs for the same workload range from about $4,000 to $161,600 over three years, depending on the model tier.
- A two-GPU on-premise server running an open-weight model costs about $76,200 over three years in our model. Hardware is 68% of that; electricity at the official Algerian high-voltage tariff is about 2%.
- The break-even point against a $2 input / $10 output per million token API is about 280 regular users. Against a $1/$5 API it is about 570 users.
- Renting the same two GPUs in a foreign cloud around the clock costs about $159,300 over three years, roughly twice the cost of owning them.
- For many Algerian industrial companies the decision is set by data rules before cost: Law 18-07 requires authorisation before personal data is transferred abroad.
What exactly are we comparing?
Three ways to run the same private assistant over internal manuals, procedures and maintenance reports:
- Pay-per-token cloud API. Each question and its retrieved passages go to a provider such as OpenAI, Anthropic or Mistral, billed per million tokens.
- On-premise server. You buy a GPU server and run an open-weight model on your site with an open-source engine such as vLLM.
- Rented cloud GPUs. You rent the same GPUs by the hour and run the open-weight model yourself.
The retrieval application, connectors and interface cost about the same in all three, so we compare only the model-serving cost. For the wider architecture and governance questions, see our guide to sovereign, on-premise AI.
What workload did we price?
These are assumptions, not measurements. Replace them with your own figures.
- Users: 300 regular users.
- Usage: 20 questions per user per working day, 22 working days per month. That gives 300 × 20 × 22 = 132,000 questions per month.
- Tokens per question: 6,000 input (system prompt, retrieved passages, question) and 500 output.
- Monthly volume: 132,000 × 6,000 = 792 million input tokens; 132,000 × 500 = 66 million output tokens.
- Three-year volume: 28,512 million input and 2,376 million output tokens.
- No discounts: standard list prices, no prompt caching or batch discounts, since retrieved passages change with every question.
What does a cloud LLM API cost over three years?
All prices below were checked on the providers' pricing pages on 10 November 2026. The formula for each tier is: (28,512 × input price) + (2,376 × output price), with prices in USD per million tokens.
- OpenAI gpt-6-luna or Claude Haiku 5.5 (prompts up to 100,000 tokens), both $0.10 / $0.50 (OpenAI pricing, Anthropic pricing): 2,851 + 1,188 = $4,039.
- Mistral Large, $0.50 / $1.50 (Mistral pricing): 14,256 + 3,564 = $17,820.
- OpenAI gpt-6.1-terra, $1 / $5 (OpenAI pricing): 28,512 + 11,880 = $40,392.
- Claude Sonnet 5.5 or OpenAI gpt-6.1-sol, both $2 / $10 (Anthropic pricing, OpenAI pricing): 57,024 + 23,760 = $80,784.
- Claude Opus 5.5, $4 / $20 (Anthropic pricing): 114,048 + 47,520 = $161,568.
Keeping data in a given region costs extra. OpenAI charges a 10% uplift on regional processing (data residency) endpoints for models released on or after 5 March 2026, and Anthropic charges 1.1 times the standard price for US-only inference. Neither option processes data in Algeria.
What does an on-premise LLM server cost over three years?
Model. OpenAI's open-weight gpt-oss-120b is published under the Apache 2.0 licence, has 117 billion parameters with 5.1 billion active, and is documented to fit on a single 80 GB GPU. Google Cloud describes it as reaching near-parity with OpenAI o4-mini on core reasoning benchmarks. Licence cost: $0.
GPUs. Two NVIDIA RTX PRO 6000 Blackwell cards, 96 GB each. NVIDIA's US marketplace listed the Workstation Edition at $16,000 in August 2026, and Thunder Compute reports the same $16,000 marketplace price in November 2026. A rack server uses the passively cooled Server Edition, for which we found no public list price; we assume it costs the same (assumption). Two cards: $32,000.
Host server (assumption). As a public reference, Exxact lists a 4U rack-mountable system with one 16-core AMD EPYC 9124 processor, 96 GB of RAM, two redundant 3,000 W power supplies and a three-year parts and labour warranty at $16,005 without GPUs. We budget $20,000 to allow for more RAM and storage. That configurator does not offer the RTX PRO 6000, so get two vendor quotes for the exact configuration.
Electricity. The Server Edition card is rated at 400 to 600 W (Thunder Compute) and the reference processor at 200 W (Exxact). We assume a pessimistic 1.5 kW for the whole server (2 × 600 W + 200 W + 100 W for memory, storage and fans), 24 hours a day. We multiply by a power usage effectiveness (PUE) of 1.56 for cooling, the industry average reported by Uptime Institute in its 2024 survey: 1.5 × 1.56 × 8,760 hours = 20,498 kWh per year, or 61,495 kWh over three years.
The CREG tariff sheet for high voltage type A (sites supplied at up to 30 kV) gives a single-rate energy price, tariff 44, of 375.62 centimes, or DZD 3.7562 per kWh, excluding taxes. 61,495 × 3.7562 = about DZD 231,000. At DZD 136.02 per US dollar (ValutaFX, 9 November 2026), that is $1,698. The sheet refers to a CREG decision of 29 December 2015, so confirm the rate on your current Sonelgaz invoice; fixed and power charges, which your site already pays, are excluded. As a cross-check, the average business price of DZD 4.68 per kWh reported by GlobalPetrolPrices for March 2026 would give about $2,116, still under 3% of the total.
Operations staff (assumption). A quarter of an engineer's time, at an assumed fully loaded cost of $30,000 per year: 0.25 × 30,000 × 3 = $22,500.
Total: 32,000 + 20,000 + 1,698 + 22,500 = $76,198 over three years, or about $2,117 per month.
Not included: customs duties and VAT on imported hardware, shipping, and the rack space and network work at your site. Ask your customs broker and add them.
Three-year comparison: all options side by side
| Option | Price basis | 3-year cost (USD) | Where your data is processed |
|---|---|---|---|
| Cloud API, OpenAI gpt-6-luna or Claude Haiku 5.5 | $0.10 in / $0.50 out per M tokens | 4,039 | Provider's servers abroad |
| Cloud API, Mistral Large | $0.50 / $1.50 | 17,820 | Provider's servers abroad |
| Cloud API, OpenAI gpt-6.1-terra | $1 / $5 | 40,392 | Provider's servers abroad |
| Cloud API, Claude Sonnet 5.5 or gpt-6.1-sol | $2 / $10 | 80,784 | Provider's servers abroad |
| Cloud API, Claude Opus 5.5 | $4 / $20 | 161,568 | Provider's servers abroad |
| On-premise, 2 × RTX PRO 6000, gpt-oss-120b | Hardware, power, 0.25 FTE | 76,198 | Your site |
| Rented cloud GPUs, 2 × RTX PRO 6000, 24/7 | $3.03 per GPU-hour | 159,257 | Provider's servers abroad |
The rented-GPU row uses the $3.03 median hourly price across nine providers reported by Thunder Compute in November 2026: 2 GPUs × 8,760 hours × 3 years × $3.03 = $159,257.
Model quality is not equal across rows: premium models are likely to answer hard questions better than gpt-oss-120b. Test each candidate on your own Arabic and French documents before comparing prices.
Where is the break-even point?
The on-premise cost is mostly fixed: about $2,117 per month whether you ask 1,000 questions or 100,000. The API cost grows with every question. Cost per question is (6,000 × input price + 500 × output price) ÷ 1,000,000.
| API tier | Cost per question | Break-even questions per month | Equivalent regular users |
|---|---|---|---|
| gpt-6-luna or Haiku 5.5 ($0.10 / $0.50) | $0.00085 | about 2.49 million | about 5,660 |
| Mistral Large ($0.50 / $1.50) | $0.00375 | about 564,000 | about 1,280 |
| gpt-6.1-terra ($1 / $5) | $0.0085 | about 249,000 | about 570 |
| Sonnet 5.5 or gpt-6.1-sol ($2 / $10) | $0.017 | about 124,500 | about 280 |
| Opus 5.5 ($4 / $20) | $0.034 | about 62,300 | about 140 |
Break-even questions = 2,117 ÷ cost per question; users = questions ÷ 440 (20 questions × 22 days). At 1,000 users, the $2/$10 tier would cost $269,280 over three years, while the on-premise cost stays near $76,200 if one server still covers the load. The two cheapest APIs are rarely beaten on cash alone: the server would need expanding well before 5,660 users.
Can one server actually handle 300 users?
Assume the busiest hour carries 15% of a day's 6,000 questions (assumption): 900 questions, or 0.25 per second. That needs 1,500 input tokens per second and 125 output tokens per second.
Published measurements on a single RTX PRO 6000 with gpt-oss-120b:
- Hardware Corner measured 3,527 tokens per second of prompt processing and 148 tokens per second of generation with a 12,000-token context (llama.cpp, Flash Attention on). The test does not report concurrent load.
- Database Mart measured 1,525 output tokens per second in total with vLLM and 50 concurrent requests, using short prompts of 100 input tokens and 600 output tokens.
On paper one card covers the peak and the second adds headroom and redundancy. Neither test matches 6,000-token prompts under concurrent load, so run your own load test before buying.
Which costs never show up in the token price?
- Data transfer rules. Under Law 18-07, a controller may transfer personal data to a foreign state only with authorisation from the national authority and only if that state ensures an adequate level of protection, with exceptions in Article 45 such as the data subject's express consent. The law was amended by Law 25-11 of 24 July 2025 (DLA Piper summary). Maintenance reports and HSE incident files often name employees. Confirm your position with counsel.
- Hardware price risk. The same card was listed at $8,565 at launch in March 2025, then $16,000 in August 2026, after $13,250 in June 2026. In November 2026 one US retailer listed the Workstation Edition at $19,999 (Thunder Compute). Get a firm quote before you build the business case.
- Usage growth. API bills grow with adoption; the server cost does not.
- Foreign-currency billing. API providers bill in foreign currency. Check with finance how recurring foreign payments work in your company.
- Connectivity. An API assistant stops when the site loses its internet link; an on-premise one keeps working.
For the regulatory background, see our guide to sovereign AI for Algerian industry.
How should an IT director decide?
- Measure real usage on a pilot: users, questions per day, tokens per question.
- Check whether your documents contain personal or classified data that cannot leave the site.
- Price the API tier whose answers are good enough on your documents, not the cheapest one.
- If you are above the break-even line, or data cannot leave, cost the on-premise option with firm quotes.
DALILE is KABAS AI's on-premise assistant for industrial documents in Arabic, French and English; the DALILE deployment guide explains how it is installed. Our consulting team can run this cost model with your own figures.
Sources
Prices and exchange rates checked on 10 November 2026. They change often; recheck them before you decide.
- OpenAI API pricing
- Anthropic Claude pricing
- Mistral AI pricing
- Let's Data Science: Nvidia lists RTX PRO 6000 Blackwell at $16,000 (12 August 2026)
- Tom's Hardware: Nvidia raises RTX Pro 6000 Blackwell pricing to $13,250
- Thunder Compute: NVIDIA RTX PRO 6000 Blackwell pricing (updated 9 November 2026)
- Exxact: TensorEX 4U configurator, TWS-173594376
- Hugging Face: openai/gpt-oss-120b model card
- Google Cloud: gpt-oss-120b on Vertex AI
- vLLM on GitHub
- Hardware Corner: Testing GPT-OSS 120B on RTX Pro 6000 Blackwell
- Database Mart: vLLM benchmark on RTX Pro 6000
- CREG: Tarification de l'électricité haute tension type A (PDF)
- GlobalPetrolPrices: Algeria electricity prices
- ValutaFX: USD to DZD historical rates, 2026
- Data Center Knowledge: Uptime Institute 2024 survey, average PUE 1.56
- DLA Piper: Data protection laws of the world, Algeria
Frequently asked questions
Is an on-premise LLM cheaper than using the ChatGPT or Claude API?
It depends on volume and on which API tier you compare against. In our model, a two-GPU server costs about $76,200 over three years. For 300 users that roughly equals Claude Sonnet 5.5 or OpenAI gpt-6.1-sol ($80,800), costs more than mid-tier and small API models, and costs less than half as much as Claude Opus 5.5 ($161,600).
How many users justify an on-premise LLM server?
With our assumptions of 20 questions per user per working day and 6,500 tokens per question, the server breaks even at about 280 regular users against a $2/$10 per million token API, about 570 users against a $1/$5 API, and about 140 users against a $4/$20 API. Rerun the calculation with your own usage logs.
What hardware do I need to run a private LLM for internal documents?
An open-weight model such as gpt-oss-120b is documented to fit on a single 80 GB GPU. A server with two 96 GB NVIDIA RTX PRO 6000 Blackwell cards covers one model with headroom and redundancy for a few hundred users, but you should confirm the sizing with a load test on your own prompts.
Does Algerian law allow sending company documents to a foreign AI API?
Law 18-07, amended by Law 25-11 of 24 July 2025, allows transfers of personal data to a foreign state only with authorisation from the national data protection authority, subject to exceptions. Technical documents often contain personal data such as names in incident reports, so check your case with legal counsel before using a foreign API.
What costs do companies forget when comparing on-premise and cloud LLMs?
On the on-premise side: staff time, cooling, customs and VAT on imported hardware, and GPU price changes before purchase. On the cloud side: token volume growth, foreign-currency billing, data residency surcharges, and the compliance work needed to send documents abroad.



