GPU server for AI weather models: running ECMWF AIFS and NVIDIA Earth-2 models in-house

Who runs AI weather models in-house

A GPU server for AI weather models runs data-driven forecasts such as ECMWF's AIFS, NVIDIA's Earth-2 models and the other models in NVIDIA's Earth2Studio framework on hardware the organisation controls. ECMWF's model card for AIFS ENS 1.0 states that it needs 38 GB of GPU memory per run by default and 24 GB with a chunking setting, so one NVIDIA RTX PRO 6000 Blackwell Server Edition with 96 GB holds the model with ample room, and a server with two or four cards runs ensemble members side by side. Earth2Studio asks for at least 40 GB on a card of compute capability 8.9 or higher.

The buyers are national and regional weather services, hydrology agencies, energy traders and grid operators, insurers and reinsurers, and research institutes in climate and atmospheric science. Their reasons for running models themselves differ. ECMWF already publishes AIFS ENS forecasts under its open data policy, so an organisation that only needs ECMWF's run can download it. A server pays off when you start your own ensembles from your own initial conditions, run extra members or longer lead times, downscale to a region at higher resolution, combine several models, or fine-tune a model on regional data. For a director, the hardware question is how many ensemble members must finish per forecast cycle and whether training is part of the plan.

Our guide on AI servers for energy and utilities places weather models next to load and generation forecasting; this guide sizes the weather workload itself. For classical numerical simulation on GPUs, our comparison of GPUs for CFD simulation covers the double-precision side.

AIFS: what ECMWF publishes about memory and hardware

ECMWF releases its AIFS models on Hugging Face. We read the model cards and file lists on 24 October 2026.

Hugging Face model cards and file lists of ecmwf/aifs-single-1.0, ecmwf/aifs-ens-1.0 and ecmwf/aifs-ens-2.0, read on 24 October 2026. Weights under CC BY 4.0, notebooks and scripts under Apache 2.0.

All three are built the same way: in ECMWF's words, AIFS "is based on a graph neural network (GNN) encoder and decoder, and a sliding window transformer processor". The checkpoints are small, about 1 to 2.6 GB, so memory goes to activations, not weights. That is why ECMWF can cut the footprint of AIFS ENS 1.0 by processing in chunks: the model card states that setting ANEMOI_INFERENCE_NUM_CHUNKS to a multiple of two lowers the 38 GB, and that 16 "reduces the memory requirements to 24GB at the cost of throughput".

AIFS ENS 2 is the version ECMWF runs today. Its model card states that it "was implemented on 12 May 2026 and supersedes version 1" and that ECMWF runs it to generate "a 51-member 15-day (6-hourly) global forecast four times per day". It adds a 10 hPa level for the stratosphere, wave variables and tropical cyclone tracks. It gives no memory figure for inference, and the checkpoint is almost three times the size of version 1. Until a measured figure exists, we plan with version 1's 38 GB as a starting figure and test version 2 on the target card.

Two software details matter for the hardware choice. The version 1 model card pins anemoi-inference 0.6.0 and PyTorch 2.5.0 in its example, and states that the use of the FlashAttention package "imposes certain requirements in terms of software and hardware". The FlashAttention README lists "Ampere, Ada, or Hopper GPUs" for FlashAttention-2 and names Hopper and Blackwell GPUs, with the H100 and B200 as examples, for FlashAttention-4, but does not name the RTX PRO 6000 or its compute capability 12.0. The H200 NVL (Hopper) and the L40S (Ada) are inside the families FlashAttention-2 names; for the RTX PRO 6000 we confirm the stack runs before the order is fixed. Hugging Face marks the version 2 checkpoint as unsafe because it uses pickle serialisation, which can execute code when a file is loaded, so load it only from ECMWF's own repository.

Earth2Studio and FourCastNet 3

Earth2Studio is NVIDIA's "Open-source deep-learning framework for exploring, building and deploying AI weather and climate workflows". Version 0.19.0 lists 45 prognostic model classes, among them ECMWF's AIFS, AIFS ENS 1.0 and both version 2 models, NVIDIA's FourCastNet 3, GraphCast, Pangu-Weather, Aurora and the regional StormCast, plus CorrDiff for diffusion downscaling. Its installation page recommends Ubuntu 24.04 LTS, Python 3.13 and CUDA 13.0, and its hardware table lists:

Earth2Studio installation page, read on 24 October 2026. The page states that Earth2Studio itself has no specific hardware requirements but "Most models do require a GPU with sufficient memory and compute capability".

The RTX PRO 6000 Server Edition (compute capability 12.0, 96 GB), the H200 NVL (9.0, 141 GB) and the L40S (8.9, 48 GB) all meet that row. The RTX PRO 4500 Server Edition, with 32 GB, falls below the 40 GB line and is a card for smaller models or AIFS ENS 1.0 in chunked mode.

NVIDIA's FourCastNet 3 model card lists 710,867,670 parameters on a 0.25-degree grid of 721 × 1440 points, an Apache 2.0 licence and "ready for commercial/non-commercial use". It states that the model produces "a 60-day global forecast at 0.25°, 6-hourly resolution in under 4 minutes" on a single GPU, without naming the GPU, lists the A100, H100 and L40S as test hardware and names the Ampere, Hopper and Turing architectures as supported; Blackwell is not among them, so FourCastNet 3 is part of the test on the RTX PRO 6000 as well. FourCastNet 3 is a probabilistic ensemble model, which NVIDIA publishes in the Earth-2 section of its NGC catalogue, and its model card names large ensemble predictions as a use.

Ensemble members per card and per server

An ensemble is many runs of the same model from perturbed starting points. Each run is independent, so members spread over cards and servers without any communication between them. Memory sets how many runs fit on one card at once; how fast each runs is not published for the cards in this article, so throughput has to be measured.

Our arithmetic from ECMWF's 38 GB and 24 GB figures and the listed card capacity, with a few GB left for the CUDA context; a memory fit, not a measured throughput. Whether several processes per card are faster than one run at a time is not documented.

For a 51-member ensemble, a server with four RTX PRO 6000 Server Edition cards running two members per card has eight in flight, so the ensemble finishes in seven rounds by our arithmetic; a two-card server needs thirteen. The length of each round depends on the measured time per member. A single member per card in default mode leaves 58 GB free on each RTX PRO 6000 Server Edition by our arithmetic for larger batch settings, a second model or downscaling with CorrDiff. The card also divides into MIG instances of 24 GB, as our guide to RTX PRO 6000 MIG profiles shows, but a 24 GB instance sits exactly at AIFS ENS 1.0's chunked figure and leaves no margin, so we do not plan members on MIG without a test.

We supply servers with two or four RTX PRO 6000 Server Edition cards for ensemble work; send us the models, members per cycle and cycles per day, and our team replies with a configuration and the test we propose before ordering.

Training and fine-tuning: where the H200 NVL fits

Running a model and training one differ by orders of magnitude. ECMWF's model card for AIFS Single 1.0 states that training used 64 A100 GPUs of 40 GB, that "the entire process takes about one week", and that each model instance spanned four 40 GB cards in one node. AIFS ENS 2 "was trained on 32 GH200 GPUs (120GB)", with "One model instance" split across two GH200. By our arithmetic one instance therefore needed about 160 GB and 240 GB of GPU memory, more than any single card holds.

Training a global model from scratch is therefore a job for a cluster or a supercomputer allocation, not for one server. What an in-house server can take on is fine-tuning an existing model on regional data or training smaller regional models. Here the H200 NVL fits: four cards joined by a four-way NVLink bridge offer 564 GB with 900 GB/s per GPU between them, and Hopper is a family the FlashAttention-2 README names. The sources we read do not document how efficiently Anemoi's model sharding runs over PCIe on cards without NVLink, such as the RTX PRO 6000, so for sharded training we recommend the H200 NVL. Anemoi, the framework behind AIFS, is open source under the Apache licence and, in its own words, development is "primarily driven by a number of European Meterological Organisations". Our guide to the four-card H200 NVL server shows that NVLink domain in detail.

Three configurations for a weather workload

Reference configurations of our AI servers page and NVIDIA's configuration guide for certified systems (2× GPU memory, at least six cores per GPU). Wall power is our estimate: cards at 600 W plus processors plus 15 per cent, as in our sibling guides; a measurement on the delivered system confirms it.

The RTX PRO 6000 Server Edition is the default for inference: 96 GB per card holds AIFS ENS 1.0 twice by memory and meets Earth2Studio's 40 GB line with room for downscaling. The four-card server fits an operational cycle with a full ensemble; the two-card server fits a research institute or a company that runs its own members on top of ECMWF's open forecasts. The H200 NVL server is the choice when fine-tuning is part of the plan, and also the card to pick if a test shows the FlashAttention build will not run on the RTX PRO 6000. An L40S server is a lower-power alternative for AIFS ENS 1.0, one member per 48 GB card, which our guide to buying an L40S server describes.

Storage and network are modest next to the cards. Earth2Studio lists 128 GB of disk, and the checkpoints are a few GB each, but initial conditions, outputs and verification data accumulate with every cycle, so we fit several NVMe drives and plan the link to the archive. Two cards fit a 230 V 16 A feed planned at 80 per cent, 2.94 kW; four need a 32 A or three-phase feed, as our guide to GPU server power on 16 A, 32 A and three-phase explains.

Open meteorological data in the EU

The model inputs and outputs connect to EU open-data rules. Commission Implementing Regulation (EU) 2023/138 of 21 December 2022 lists high-value datasets, and its meteorological category covers observations from weather stations, validated climate observations, weather alerts, radar data and numerical weather prediction (NWP) model data. Public bodies in scope make them available "under the conditions of the Creative Commons BY 4.0 licence or any equivalent or less restrictive open licence", through APIs and bulk download, except NWP model data, which "shall be made available through APIs only". For NWP data the annex sets a granularity of "Minimum 48 hrs ahead in 1hr steps, national, at 2.5km/best available grid". The regulation applies 16 months after its entry into force; with publication in the Official Journal on 20 January 2023 and entry into force on the twentieth day after, that is 9 June 2024 by our reading. Which datasets a particular body must publish is for its legal department to confirm; this is general information as of October 2026.

For a server buyer, this means national observations and forecasts are often available as inputs under open licences, and the AIFS weights themselves are under CC BY 4.0, which requires attribution. Forecasts generated on your own server can stay internal.

How we deliver the weather server

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. Sizing is free. We assemble the server, test it before delivery and deliver it anywhere in the EU on one contract and invoice, with manufacturer warranty on every component and the assembly covered by us; other destinations go through our logistics partner. On request we install the operating system, the NVIDIA driver, CUDA and a container runtime, and commission the server on site. Quantities and lead time are confirmed in writing with the quote.

To quote, we need the models and versions you plan to run, the members per cycle, cycles per day and lead time, the deadline for each cycle, any fine-tuning plans, the data sources for initial conditions, and the power feed at the rack position. Write to us with these details, and we reply with a configuration and a written quote.

FAQ