Skip to content
Niezależne AI
PLEN

Guide · Updated October 2026

Which AI hardware for your business in 2026?

A comparison of the machines that run private AI inside a company: how many people they can serve, how quickly they answer when working with documents, what they cost and when to choose each one. Based on public benchmarks, specifications and prices, with sources. We don’t sell hardware, so we have no reason to push you towards the more expensive option.

The three numbers that matter

  1. Memory01 · which model fits

    The whole model has to fit in memory, with room left for several people’s conversations at once. With 128 GB you can run models of around 120 billion parameters (gpt‑oss‑120b takes about 63 GB). A 32 GB card such as the RTX 5090 tops out at models of around 40 billion parameters.

  2. Memory bandwidth02 · how fast it generates text

    Every word of an answer requires reading the active part of the model from memory. That’s why a Ryzen AI Max+ mini-PC (256 GB/s) and an NVIDIA GB10 (273 GB/s) generate text at a similar speed, while an RTX PRO 6000 card (about 1.6–⁠1.8 TB/s) is several times faster.

  3. Prompt processing03 · how soon it starts answering

    When you work with documents, the model receives your question together with excerpts from the documents, often around 4,000 tokens. With a prompt like that, a Ryzen AI Max+ mini-PC starts answering after about 9–⁠12 seconds and a GB10 after about 2 seconds, even though both then generate text at a similar speed. This is the figure most comparisons leave out.

How many people at once? In a typical company only a small share of the team is waiting for an answer at any given moment. We assume 5–10% of staff at peak. That is our assumption, not a measurement, so we check it at sign-off against your own tasks.

Hardware compared

Output speed is in tokens per second (a token is about ¾ of a word) for a single user. “Concurrent users” means simultaneous answers while working with documents, not headcount. Prices are indicative, net of VAT unless stated otherwise, based on Polish retailers in October 2026. Memory and graphics card prices change quickly, so every recommendation we send includes the current price.

  • Mini-PC with AMD Ryzen AI Max+ 395

    Memory · bandwidth
    128 GB · 256 GB/s
    Output, 1 user
    34–⁠55 tok/s (gpt‑oss‑120b)
    Time to first word*
    approx. 9–⁠12 s
    Concurrent users
    1–⁠3
    Team size
    up to about 10 people or one department
    Hardware price
    PLN 11–⁠15k
    Power, noise
    145–⁠186 W, quiet

  • NVIDIA GB10, 128 GB (DGX Spark, Dell Pro Max, HP ZGX Nano, Lenovo PGX, ASUS GX10)

    Memory · bandwidth
    128 GB · 273 GB/s
    Output, 1 user
    38–⁠59 tok/s (gpt‑oss‑120b)
    Time to first word*
    approx. 2 s
    Concurrent users
    5–⁠8 on gpt‑oss‑120b, 15–⁠30 on smaller MoE models
    Team size
    10–⁠50 people
    Hardware price
    PLN 22–⁠29k
    Power, noise
    60–⁠90 W, below 40 dBA

  • Mac Studio M5 Max (up to 128 GB)

    Memory · bandwidth
    128 GB · 614 GB/s
    Output, 1 user
    approx. 88 tok/s (gpt‑oss‑120b)
    Time to first word*
    approx. 3 s
    Concurrent users
    no public benchmarks
    Team size
    one person or a small team
    Hardware price
    from PLN 12,499 incl. VAT (36 GB); 128 GB costs considerably more
    Power, noise
    not measured

  • Workstation with RTX 5090

    Memory · bandwidth
    32 GB · approx. 1.8 TB/s
    Output, 1 user
    approx. 205 tok/s (gpt‑oss‑20b)
    Time to first word*
    under 1 s
    Concurrent users
    10–⁠30 (estimate)
    Team size
    20–⁠60 people, smaller models
    Hardware price
    card alone PLN 36–⁠49k incl. VAT
    Power, noise
    high power draw, louder

  • Workstation with RTX PRO 6000 (96 GB)

    Memory · bandwidth
    96 GB · 1.6–⁠1.8 TB/s
    Output, 1 user
    approx. 215 tok/s (gpt‑oss‑20b)
    Time to first word*
    approx. 0.3 s with 50 users
    Concurrent users
    about 50 on gpt‑oss‑120b, at about 31 tok/s each
    Team size
    50–⁠300 people
    Hardware price
    PLN 75–⁠95k (card alone approx. PLN 60–⁠72k)
    Power, noise
    600 W (Max-Q version 300 W)

  • Server with 2–⁠4 RTX PRO 6000 cards

    Memory · bandwidth
    192–⁠384 GB
    Output, 1 user
    as above
    Time to first word*
    as above
    Concurrent users
    hundreds
    Team size
    300+ people, multiple sites
    Hardware price
    PLN 190–⁠400k (estimate)
    Power, noise
    1.2–⁠2.4 kW, server room

* Time from sending a prompt of about 4,000 tokens (a question plus document excerpts) to the first word of the answer, calculated from public prompt-processing benchmarks. The measurements come from different software versions, so treat them as an order of magnitude.

The short answer

Office of up to 10 people

A Ryzen AI Max+ mini-PC. The cheapest way to get 128 GB of memory; quiet and frugal with power. It is slower to start answering long prompts, which rarely matters with only a few users.

Company of 10–50 people

An NVIDIA GB10 (DGX Spark or the Dell, HP, Lenovo or ASUS equivalent). The GB10 has 128 GB of memory and reads documents several times faster, handles several people at once well, and is quiet enough for an office.

50–300 people or heavier workloads

A workstation with an RTX PRO 6000 card. Dozens of people at once with fast answers. It needs better cooling; the Max-Q version is quieter.

Several sites

One device per site with the same configuration, or one more powerful server at head office. The choice depends on the links between sites and whether each site must keep working when the internet goes down.

What we don’t recommend

  • Large “dense” 70B models on a mini-PC or GB10They generate 3–⁠4 tokens per second, slower than you can read. On this hardware, MoE models, which use only part of their parameters for each word, work much better.
  • Ollama exposed to the network without sign-inIts API has no built-in authentication. In 2026 the CVE‑2026‑7482 vulnerability (rated 9.1/10) affected about 300,000 servers exposed to the internet. We always put the model behind a gateway with sign-in.
  • Buying the most expensive hardware “just in case”It’s better to buy a device sized for your current team and add a second one with the same configuration when you need it. Models and prices change every few months.
  • A Mac Studio as the server for a larger teamIt generates text quickly for a single user, but multi-user serving tools are less mature on the Mac and there are no public benchmarks under that kind of load.
  • Intel Arc Pro B60 cards in productionThey are cheap, but in public tests the software was still at an early stage and only some models worked correctly.

Models in 2026 and their limits

The best cloud models (Claude Opus 5.5, GPT‑6, Gemini 4) score 53–⁠58 on the Artificial Analysis Intelligence Index. Models that fit on company hardware score 12–⁠34. On the Polish language and culture benchmark (PLCC), the best open Polish models score about 55–⁠72 and the best cloud models 90–⁠97. For working with documents and citing sources that is enough, because every answer rests on an excerpt you can check. For free-form writing and coding, the cloud is clearly better today.

ModelLicence
Qwen3.8-27Bthe strongest model that fits on company hardware (34 on AA), long documentsApache 2.0
gpt-oss-120bfast MoE model, good for many simultaneous usersApache 2.0
Gemma 4 (26B MoE, 31B)also understands images and scansApache 2.0
Bielik 11B v3Polish language and letter style, small and fastApache 2.0
PLLuM 12Bofficial Polish and public administrationApache 2.0
Llama-PLLuM 8B and 70Badditional licence terms; 70B only on powerful cardsLlama 3.1 licence

We often run two models on one device: a stronger general model (such as Qwen3.8-27B or gpt-oss-120b) to understand the documents, and Bielik for natural Polish in letters and documents.

Licences few people mention

Open WebUI
Above 50 users you must keep its branding or buy an enterprise licence.
n8n
You may use it internally in your own company. Reselling it or offering it to others as a service is restricted.
Llama-PLLuM 8B and 70B
Llama 3.1 licence with additional terms, unlike PLLuM 4B and 12B, which are Apache 2.0.
NVIDIA AI Enterprise
Not needed for a typical deployment. It costs USD 4,500 per card per year.

Why prices are rising

The NVIDIA DGX Spark launched in 2025 at USD 3,999; since 2 October 2026 it has cost USD 6,950 (128 GB). The RTX PRO 6000 card went up by about 55% in a year. The cause is a memory shortage that TrendForce expects to last until at least 2027. That is why our recommendations quote prices on the day they are sent and are valid for a short time.

Which one is right for your company?

Answer three questions and you’ll see a recommendation straight away, with the cost of the hardware and of our work. Free and with no obligation.

Get free hardware advice Book a call

Sources

  1. LMSYS: NVIDIA DGX Spark review (SGLang and Ollama benchmarks)  (opens in a new tab)
  2. Hardware Corner: DGX Spark and Ryzen AI Max+ in llama.cpp  (opens in a new tab)
  3. Computingforgeeks: DGX Spark vs Strix Halo vs Mac Studio (October 2026)  (opens in a new tab)
  4. StorageReview: DGX Spark under multi-user load  (opens in a new tab)
  5. Database Mart: vLLM on the RTX PRO 6000  (opens in a new tab)
  6. ServeTheHome: DGX Spark power draw and noise  (opens in a new tab)
  7. NVIDIA: the RTX PRO 6000 family  (opens in a new tab)
  8. Epoch AI: which models fit on consumer GPUs  (opens in a new tab)
  9. Artificial Analysis: model leaderboard  (opens in a new tab)
  10. Codesota: Polish models and the PLCC benchmark  (opens in a new tab)
  11. Cloud Security Alliance: CVE-2026-7482 in Ollama  (opens in a new tab)
  12. Open WebUI: licence  (opens in a new tab)
  13. NVIDIA AI Enterprise: licence pricing  (opens in a new tab)
  14. AceCloud: DGX Spark prices (October 2026)  (opens in a new tab)
  15. Tom’s Hardware: RTX PRO 6000 price increase  (opens in a new tab)
  16. iMagazine: Polish prices for the Mac Studio M5 (in Polish)  (opens in a new tab)
  17. TrendForce: DRAM price forecast  (opens in a new tab)
Get free hardware advice