AI hardware budget for 2027: the order of decisions and what to order first

AI hardware budget for 2027: what to order first

In an AI hardware budget for 2027, the first enquiries belong to the items with the most dependencies: systems that pass a project approval before they can be quoted, servers with a large amount of host memory, and any work on power and cooling. Workstation cards, licences and the network can follow. Before any of them, three decisions fix what is being bought: the workload with its peak number of concurrent users, the memory each model needs, and the system class.

This article contains no prices and no budget figures. It describes each level by system class and card count, and a level becomes a figure through a written quote for one configuration on one date.

The order of decisions: workload, memory per model, system class

The workload comes first, because inference, retrieval and fine-tuning load a system differently, and with it the peak number of requests in flight. A hundred employees with an assistant open produce far fewer simultaneous generations, so the figure that sizes a system is the busy-hour peak. Our guide on what to specify when buying an AI server shows how to arrive at it.

The memory per model comes second. By our sizing rule a card has a budget for weights and KV cache of 0.9 times the memory the driver reports, less 3 GiB: 25.7 GiB on a 32 GB card, 40.0 GiB on a 48 GB card, 83.0 GiB on a 96 GB card and 123.4 GiB on an H200 NVL. For the 32 GB and 48 GB cards we scale the reported memory from the 95.6 GiB the driver shows on the 96 GB card, which is an assumption; the driver on the delivered card decides. One example shows how the model sets the class. Llama 3.3 70B in FP8 has 67.7 GiB of weights. That is above the budget of a 48 GB card and inside that of a 96 GB card, where 15.3 GiB remain for the cache, enough for 12 concurrent sessions of 8,192 tokens with an FP8 cache.

The system class comes third and follows from the first two.

memory for models by our sizing rule (0.9 × driver-reported memory, less 3 GiB per card); the 2.1 TB is NVIDIA's figure for the eight GPUs of a DGX B300; host memory from the reference configurations on our AI servers page, text of 12 November 2026.

The cards behind the rows are the NVIDIA RTX PRO 5000 Blackwell with 48 GB (buy page, 48 GB and 72 GB compared), the RTX PRO 6000 Blackwell Server Edition with 96 GB and the NVIDIA RTX PRO 4500 Blackwell Server Edition with 32 GB (buy page, specs). Which starting point suits a first project is compared in our article on which GPU to start a first AI project on.

What belongs at the front of the plan

Three items have more steps between enquiry and operation than the rest, so their enquiries come first.

Systems supplied per project. We supply the DGX B300, HGX B300 servers and manufacturers' versions of the DGX Station GB300 per project. Before we can quote, the distribution asks for the company, the project and the intended use, and an approval is not certain until it is given. The quote that turns this budget line into a figure exists only after the approval, so the project data should go out at the start of the planning period.

Configurations with a large amount of host memory. Two server makers named memory first among their supply constraints in September 2026: Dell with "DRAM, DRAM, DRAM followed by NAND, NAND, NAND" on 1 September, HPE with DDR5 as "a great example" on 2 September. TrendForce wrote on 30 September 2026: "Server DRAM is therefore expected to remain undersupplied in 4Q26." A training node with 2.3 TB, or an inference node whose memory is raised for KV cache offloading, depends on that component more than an entry server does. Our articles on how much RAM a GPU server needs and on server DDR5 supply and what it does to a quote give the rule for the amount and the effect on a quote.

Power and cooling work. By our estimate, the rated power of cards and processors × 1.15 for the rest of the system, a server with two 600 W cards draws about 1.8 kW at the wall and fits a 16 A single-phase feed. One with four draws about 3.4 kW, above the 2.94 kW we plan for such a feed under continuous load (0.8 × 16 A × 230 V). One with eight draws about 6.3 kW and needs three-phase feeds. A DGX B300 is rated at 15.1 kW in NVIDIA's datasheet for the version with power supplies. A new feed or a cooling upgrade is ordered from other trades than the server and has to be finished on the day of delivery. NVIDIA's Form 10-Q for the quarter ended 26 July 2026 names the same dependency from the seller's side: among the reasons why customers may postpone purchasing new architectures, it lists missing data-centre infrastructure to deploy the products. The checks are in our article on the site survey for GPU servers.

We supply the DGX B300 and HGX B300 servers per project; send us the company, the project and the intended use at the start of your planning and we check price and availability for your project. The steps every server order passes are set out in our article on what decides GPU server lead times.

Workstation cards, licences and the network can follow

An order for workstation cards has fewer steps than one for a server: the distribution asks for no project approval, and a card supplied without a system is not assembled. Its line in the plan needs the workstation model it goes into and the driver branch the fleet runs.

Licences follow the hardware they are counted on. NVIDIA AI Enterprise is licensed per GPU, as a subscription of one to five years or as a perpetual licence with five years of support, and each H200 NVL includes a five-year subscription. An open-source engine such as vLLM on bare metal needs no NVIDIA licence, and NVIDIA's product page of 17 September 2026 offers "a 90-day trial license for NVIDIA AI Enterprise", so a pilot can run before the licence line is spent. What NVIDIA's own pricing page states, with its date, is in our article on NVIDIA AI Enterprise licensing; this page prints no figures.

The network of one inference server is a small line, because streamed text needs little bandwidth. It grows when several servers work on one model: NVIDIA's configuration guide for NVIDIA-Certified Systems asks for at least 200 Gbps for multi-node inference.

What to budget besides the hardware

licence metric from NVIDIA's licensing guide (pricing page last updated 2 September 2026) as described in our licensing article; the other lines from our AI servers page, text of 12 November 2026, and our own summary.

No primary source read for this article gives a share of an AI budget for these lines, so we print none. What can be said is how each is counted. The licence line multiplies with the GPU count. The energy line multiplies the wall power above with the hours of operation and your own tariff. The support line has two parts: the manufacturer warranty covers every component and is handled through us, while response on site is a service level of the server maker, as our article on GPU server warranty in the EU explains.

You can order the server alone, the setup of the software alone, or both together: the server delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

NVIDIA's platforms and support dates that touch a 2027 plan

A plan can only budget what can be specified and quoted today. NVIDIA's Form 10-Q states that its next data-centre architecture, Vera Rubin, "began production shipments in the third quarter of fiscal year 2027", the quarter that ended on 25 October 2026. Vera Rubin NVL72 is a rack-scale system with 72 Rubin GPUs and 36 Vera CPUs. NVIDIA's HGX page of 30 September 2026 lists four eight-GPU platforms: HGX Vera Rubin NVL8, HGX Rubin NVL8, HGX B300 and HGX B200. The eight-GPU systems we supply per project are the Blackwell Ultra systems, the DGX B300 and HGX B300 servers. Beyond Rubin, NVIDIA's GTC 2026 roundup of 19 March 2026 names the next architecture, Feynman, without a year. For workstation cards, NVIDIA's desktop line-up page of 8 September 2026 lists the RTX PRO Blackwell range, and the NVIDIA documents read for this article announce no new RTX PRO generation for 2027.

Several support dates fall inside the budget year.

the documents named in the third column; Regulation (EU) 2026/1744 has been in force since 27 July 2026. General information, not legal advice.

A system delivered in 2027 should start on a long-term support branch that outlives the year. NVIDIA's lifecycle table lists two: R580, in the table above, and the newer R620, whose end date should be taken from NVIDIA's table on the day of planning. The cards decide between them: the RTX PRO 4500 Server Edition is not in the newest R580 driver lists and needs R595 or R620. A GPU host planned for vSphere should be planned on a release that is in general support for the life of the plan; the dates are in our article on the vSphere 8 end of general support.

A plan in three stages, from the pilot system to the second node

Stage one, the pilot system. Its job is to replace estimates with measurements: the peak number of requests in flight, the model, its precision and the context length. Its hardware is the lowest level of the table: a workstation with one or two 48 GB cards, or the entry server with two 96 GB cards if the pilot has to run in the rack. What to measure is in our guide to the private AI pilot, and the step to production hardware in our article on scaling private AI hardware from pilot to production.

Stage two, the production node. It is sized from the pilot's figures: a four-card node, an eight-card server or, where one model has to be split across eight GPUs in one NVLink domain, a system supplied per project. Its memory amount and its power feed are the two items to fix first.

Stage three, the second node. Where a model fits one card, more cards run more copies behind a load balancer, and copies in a second host remove the single point of failure. Budget the second node as a line of its own with the same configuration, so that it can be quoted again when it is released.

We size each stage free of charge; tell us the models, the peak number of users and the site and we reply with a configuration and a written quote within one business day.

What our written quote contains and why it carries a date

A request is answered within one business day with a configuration and a written quote. The quote names the price, the lead time and the warranty terms; quantities and lead time are confirmed in writing with the quote, and the invoice comes from Austria on one European contract.

The quote carries a validity date because the prices of its components move. TrendForce wrote on 9 July 2026 that memory suppliers "may continue revising quotations upward throughout the remainder of this quarter", and Dell said on 1 September 2026 that its input costs continue to rise. For a budget this has two consequences. The figure in a budget line is the figure of a quote on its date, and it is renewed before the order is approved. And the plan should keep the configuration as separate lines for cards, memory, drives and licences, so that each can be renewed on its own. What a sizing request should contain is described in our article on the Private AI sizing request.

FAQ