How much does an AI server cost? Price by class, running costs and rental

Run local AI with predictable costs? The AI Cube is the turnkey AI server for your own network, the Managed GPU servers run in a German data centre. Neither is billed per token. Support, consulting and implementation by WZ-IT. Book a call

In this article, an AI server is a server with one or more GPUs, or with a GB10 chip and unified memory, that runs language models for a company's own use. What it costs depends on the performance class, on who operates it and on how long it runs. This article sets the prices side by side: purchase by class, running costs, rental over 12, 24 and 36 months and the break-even against a model API. Third-party figures carry their source and date, calculations show their formula, and WZ-IT prices are the current prices of the offers. As of 11 November 2026.

What does an AI server cost?

The price is set by the performance class. As of 11 November 2026:

• 24 GB class: a workstation GPU card with 24 GB (NVIDIA RTX PRO 4000 Blackwell SFF Edition) from €3,099.90 including VAT (Geizhals). The card alone; the server around it comes on top.
• GB10 class with 128 GB unified memory: the NVIDIA DGX Spark at a list price of USD 4,699 (NVIDIA). As a system set up for use, the AI Cube costs from €7,999 excl. VAT one-time plus AI Cube Care from €349.90 excl. VAT per month.
• 96 GB class and servers with eight GPUs: a single RTX PRO 6000 Blackwell card from €16,798.94 including VAT (Geizhals); a server with eight RTX PRO 6000 GPUs €170,000 to €240,000 before VAT, a server with eight B200 GPUs USD 360,000 to 470,000 as an indicative market range (CHG-MERIDIAN).
• Rental: dedicated GPU servers from €214 per month plus €209 setup, net (Hetzner); the Managed GPU Server 24 at €699 per month plus €499 setup and the Managed GPU Server 96 at €2,199 per month plus €1,999 setup, operated by WZ-IT.
• Running costs: power, for example 150 W of continuous load equals 109.5 kWh or €19.60 per month at 17.9 cents per kWh (BDEW), plus administration, updates, monitoring and backup.

What the price depends on:

1. Model size and context length. They determine how much GPU or unified memory is needed, see Sizing GPU and VRAM for LLMs.
2. Concurrent users. Not the number of accounts but the number of simultaneously generated answers decides, see Size a local AI server by users.
3. Hours of use. They decide how well the capacity is used. A rented server is billed as long as it exists, so idle time costs the same as use.
4. Operating scope. Patching, monitoring, backup and support are either done internally or bought as a service.
5. Term. One-time and monthly amounts only become comparable over the same period.

Table of contents

• Purchase price by performance class
• AI Cube and Managed GPU servers: WZ-IT prices
• Running costs: power, operations and support
• Buy or rent: total cost after 12, 24 and 36 months
• Four cost blocks
• Break-even per token: the assumptions
• Cost per million tokens by utilisation
• What the break-even volumes mean
• When the token price does not decide
• Compare three years instead of purchase price
• When standard hardware is not enough
• As of and methodology
• Our approach at WZ-IT
• Sources and further reading

Purchase price by performance class

Three points for reading the table:

• A card is not a server. A GPU card needs a host with processor, memory, storage, power supply and cooling; these parts are not in the price. The DGX Spark and the AI Cube are complete devices.
• Prices move. NVIDIA raised the list price of the DGX Spark Founders Edition from USD 3,999 to USD 4,699 in February 2026 and named worldwide constraints in memory supply as the reason (NVIDIA). Prices of OEM systems with the same chip are set by the OEMs. CHG-MERIDIAN also names sharply higher memory costs as the reason for rising AI server prices in 2026.
• The classes serve different loads. How much memory a model needs and how many users a device carries is covered in Sizing GPU and VRAM for LLMs and Size a local AI server by users. According to the product page, one AI Cube is generally enough for teams of five to twenty people; the larger classes are used for larger models, longer contexts or more concurrent requests.

AI Cube and Managed GPU servers: WZ-IT prices

With the AI Cube the device belongs to the customer; operations run via AI Cube Care. The GPU servers are an ongoing service; host system, term and provisioning are confirmed in the quote. Neither option bills per token or per user. Whether ownership or a service model fits better is covered in Buy or rent an AI server?.

Running costs: power, operations and support

Power. The product page of the AI Cube states about 150 W under load and a 240-watt power supply; the table uses 150 W as a calculation example. The table calculates a constant load over 730 hours per month (8,760 hours per year divided by 12); actual consumption depends on the load profile and the idle draw. Two reference prices: the average for small and medium-sized industrial customers was 17.9 cents per kWh in 2026 so far, the household average 37.3 cents, both as published by BDEW (BDEW, as of October 2026).

For the AI Cube at 150 W this adds up to 1,314 kWh per year, or €235.21 at 17.9 cents. Cooling of the room is not included. A server in a data centre, such as the Managed GPU servers, causes no server power bill at your site.

Operations. Updates, security review, monitoring, backup, recovery, model changes and support arise over the entire service life. With the AI Cube, AI Cube Care from €349.90 excl. VAT per month covers patching, backup, monitoring and support. With the Managed GPU servers, monitoring, updates and technical operations are included in the monthly price. A rented dedicated server without platform operations leaves all of this with the customer; internal administration time belongs in the calculation.

Not included in the prices of the basic configurations above: integrations with Nextcloud, SharePoint, DMS or business software, internal staff time and the replacement of hardware after its service life.

Buy or rent: total cost after 12, 24 and 36 months

Cumulative payments in euro excl. VAT for the base configurations: one-time amount plus months multiplied by the monthly price. Setup fees are included; power, internal work, integration, replacement and residual value are not.

Prices of the rented dedicated servers: Hetzner models GEX45, GEX63 and GEX430, net, as of 11 November 2026. The AI Cube figure assumes that Care runs throughout, although it is cancellable monthly after the first twelve months. The Managed GPU server figures assume continuous service for the stated period.

How to read it:

• The break-even between purchase and service. The AI Cube with Care costs less in total than the Managed GPU Server 24 from the 22nd month (€15,687.80 against €15,877 after 22 months) and less than the Managed GPU Server 96 from the 4th month. The AI Cube figure includes operations, the power at your site is extra (see above).
• Within a GPU class, rental without platform operations is the lower line: €7,913 against €25,663 after 36 months in the 24 GB class, €36,963 against €81,163 in the 96 GB class. It covers the server in a data centre with its network connection. Operating system, drivers, model server, user interface, updates, monitoring and backup stay with the customer, while the managed offers cover monitoring, updates and technical operations and AI Cube Care covers patching, backup, monitoring and support.
• The classes are not equal in performance. The table compares cost, not speed. Which class carries a model and a team is shown by the sizing guides linked above and by the Token Simulator with GB10 measurements.
• Hourly billing is capped by the monthly price. Hetzner bills rented GPU servers by the hour, and the total does not exceed the monthly price (billing documentation). For the 96 GB model GEX63 at €1.6009 per hour, the monthly price of €999 is reached after about 624 hours. A server is billed as long as it exists in the account; to stop paying, it has to be cancelled, and a new order incurs the setup fee again. The models GEX45, GEX63 and GEX131 as of 13 October 2026 are compared in Hetzner GPU servers compared.

Four cost blocks

1. Hardware and base system. For local language models, available memory, memory bandwidth and desired throughput determine the hardware class. How much memory model and context occupy is calculated in Sizing GPU and VRAM for LLMs.

2. Setting up the AI platform. Operating system, drivers, model runtime, user interface, model, user management, network, hardening and function test. Anyone doing this internally has to account for the time and the knowledge needed in the team later.

3. Integration into workflows. Using your own documents manually in a knowledge interface is different from a permanently synchronised RAG pipeline. Interfaces, source permissions and automations depend on the project.

4. Ongoing operations. Updates, security review, monitoring, backup, recovery, model changes and support arise over the entire service life, either internally or as a managed service.

Break-even per token: the assumptions

A model API bills per million tokens, own hardware costs a fixed amount per month. The break-even is the monthly token volume at which both cost the same. The comparison uses the list prices of the IONOS AI Model Hub, a German provider processing in Germany, for two open-weight models that also run on WZ-IT hardware.

gpt-oss-120b is a mixture-of-experts model with 117 billion parameters that, according to its model card, fits on a single 80 GB GPU; it therefore runs on GPU Server 96 and on the 128 GB of the AI Cube. Mistral Small 24B fits on 24 GB with 4-bit quantization. All prices exclude VAT.

Cost per million tokens by utilisation

Dividing the monthly fixed cost by the tokens actually processed gives the cost per million tokens. It falls with every additional request processed.

Above the break-even, own hardware is cheaper per token; below it, the API is. Whether the hardware can actually process that volume is a second question that depends on model, context and concurrency.

What the break-even volumes mean

A request with 4,000 tokens of context and a 1,000-token answer uses 5,000 tokens. This gives:

For context: a publicly documented measurement on a single GB10 with the smaller model Qwen3.6-35B-A3B (NVFP4, vLLM, 1,024 input tokens, 256 output tokens) reached around 320 output tokens per second with eight and around 432 with 16 concurrent chats. This is not a measurement for gpt-oss-120b, but it shows the order of magnitude.

Three factual conclusions follow:

• With use during office hours only, the pure token price of European APIs for open-weight models is usually below the cost of own hardware. The break-even requires a utilisation that an interactive assistant for a small team does not reach.
• Under continuous load the picture shifts. Round-the-clock batch processing, such as classification, document extraction or embedding large collections, uses hardware evenly and can reach the break-even.
• API prices change. List prices, model catalogues and retirement dates are set by the provider. The fixed cost of own hardware is known over the term.

When the token price does not decide

In many projects the token price is not the decisive criterion:

Which European APIs exist and how they differ in location, contract and models is compared in European LLM APIs compared. The general trade-off is covered in Cloud AI vs. self-hosted. A mixed setup is also possible: embedding and reranker locally, the language model locally or via an API depending on protection needs, bundled behind a gateway such as LiteLLM.

Compare three years instead of purchase price

A robust calculation uses the same assumptions for both options:

Local platform: purchase price or monthly rate + setup + integration + power and operations + risk buffer for replacement.

Cloud platform: 36 months of user licences or API consumption + integrations + administration + potential add-on features and data regions.

Anyone who already knows their API consumption can enter it directly into the table above. Anyone who does not measures it in a pilot phase, for example via the usage logs of a gateway. Non-financial factors also belong in the decision: permitted data types, control over model and updates, offline capability and dependence on the provider.

When standard hardware is not enough

With the standard configuration an AI Cube serves up to ten simultaneously generated answers, depending on model and context length; for teams of five to twenty people one device is generally enough. A larger setup makes sense when several large models run in parallel, very long contexts are needed, or availability and redundancy have to be contractually assured. We then plan multi-GPU systems or clusters based on measured load; these architectures are quoted separately.

How account count, concurrency and answer length become a load profile is shown in Size a local AI server by users. Which models fit on 128 GB is covered in LLM models on 128 GB unified memory.

As of and methodology

Prices retrieved on 11 November 2026; the date of third-party statements is given per line.

• Purchase prices of GPU cards: lowest listing on the price comparison Geizhals on that day (product pages linked in the table), in euro including VAT and excluding shipping. Listings change daily, and a card is not a server.
• DGX Spark: list price of the Founders Edition according to the NVIDIA announcement of February 2026 in the NVIDIA Developer Forums. NVIDIA refers questions on OEM systems to the OEMs.
• Servers with eight GPUs: price ranges as of October 2026 as stated in the CHG-MERIDIAN article on AI server cost, which does not publish the underlying offers; the range for RTX PRO 6000 servers is based on European reseller listings before VAT, the HGX figures on a market report of October 2026 named there.
• Rented servers: the Hetzner price list for dedicated GPU servers, net, monthly price plus one-time setup fee, on 11 November 2026. The website shows prices with VAT depending on the selected country.
• Power: BDEW electricity price analysis, October 2026 (2026 year to date). Consumption is calculated as watts times 730 hours per month divided by 1,000, which gives kWh.
• API prices: IONOS price list, retrieved on 11 November 2026, models gpt-oss-120b and Mistral Small 24B Instruct.
• WZ-IT prices: the current prices of the offers, excl. VAT. Totals are one-time amount plus months times monthly price, without discounting, without residual value and without VAT.
• Not covered: clusters beyond eight GPUs, network and shared storage, internal staff time, residual value and resale.

Our tables and calculations may be reused with attribution and a link to this article.

Our approach at WZ-IT

Which class fits depends on model, context length and the number of concurrent requests. We check that before the quote. For larger setups we plan multi-GPU systems or clusters based on measured load. Options are the AI Cube and the Managed GPU servers.

Sources and further reading

• Geizhals: NVIDIA RTX PRO 4000 Blackwell SFF Edition
• Geizhals: NVIDIA RTX PRO 6000 Blackwell
• NVIDIA Developer Forums: DGX Spark price change announcement
• CHG-MERIDIAN: AI server cost from purchase to retirement
• Hetzner: GPU servers, price list
• Hetzner: billing documentation
• BDEW: electricity price analysis
• IONOS Cloud, price list including AI Model Hub
• OpenAI, gpt-oss-120b model card
• Public GB10 measurement with Qwen3.6-35B-A3B NVFP4 (dgx-spark-bench)
• NVIDIA: DGX Spark hardware with 128 GB unified memory
• WZ-IT: Sizing GPU and VRAM for LLMs
• WZ-IT: Buy or rent an AI server?
• WZ-IT: Hetzner GPU servers compared
• WZ-IT: Cloud AI vs. self-hosted
• WZ-IT: Token Simulator with GB10 measurements

Which class fits your use? The AI Cube is the turnkey AI server for your own network; for larger models there are the Managed GPU servers. Which class fits depends on context length and concurrent requests; we check that before the quote. Book a call