GPU clusters,built in LatAm.
A GPU neocloud in Santiago, Chile. We run NVIDIA hardware that is ours or yours: we buy it and you reserve the capacity, or you bring the fleet and we import, host and operate it. The hall, the builder, the hardware quotations and the import route are in place. What we are looking for is the offtaker.
HGX B200 and B300 nodes live from December 2026. The GB300 hall in Q1 2027.
What Fulmis does
Fulmis is a GPU neocloud for Latin America, based in Santiago, Chile. We operate NVIDIA clusters in a Santiago data centre, and the hardware can be ours or yours. If it is ours, we buy it, bring it into Chile as importer of record, commission it and run it, and you reserve the capacity for three to five years at a fixed price, take-or-pay. If it is yours, you bring the fleet, we bring it into Chile, host it and run it for you, on the same term.
Either way the line between us is drawn once and written into the contract, the workload, the models, the data and the keys are yours, and Slurm or Kubernetes is the one interface across it.
Our hardware
- We buy the GPUs and the rack
- We place the order and finance it
- We import, commission and accept
- We specify fabric, storage and software
- We operate it, on call, for the term
- You reserve the capacity, take-or-pay
Your hardware
- You own the GPUs and the rack
- We bring them into Chile
- We give them a position, power and cooling
- We commission them against the vendor baseline
- We operate them, on call, for the term
- You pay a fixed monthly fee for the term
Ready first
Blackwell nodes, live from December 2026
NVIDIA HGX B200 and HGX B300, eight GPUs per node, air cooled, in a Santiago data centre. From one node up, joined by InfiniBand into one cluster when you take more than one. Reserve now for a December 2026 start, while the GB300 hall follows in Q1 2027.
NVIDIA HGX B300
8 Blackwell Ultra GPUs, 2.3 TB HBM3e per node, ConnectX-8 at 800 Gb/s per GPU. Training and inference.
NVIDIA HGX B200
8 Blackwell GPUs, 1.4 TB HBM3e per node, ConnectX-7 at 400 Gb/s per GPU. Training and inference.
- Live from
- December 2026
- Smallest block
- One node, 8 GPUs
- Cooling
- Air cooled
- GPU fabric
- InfiniBand, sized to the cluster
- Software
- Slurm and managed Kubernetes
- Location
- A data centre in the Santiago Metropolitan Region
We are looking for the offtaker
The first hall is built against one signed contract. The data centre, the builder, the hardware quotations, the design, the import route and the companies are in place, and the financing is in discussion with lenders and a lessor and closes on the same signature. What is missing is the customer who takes the capacity for three to five years.
That can be a buyer of hours, an aggregator with a book, a cloud that wants a Latin American region, or an owner of GPUs that need a home in the region. Twelve GB300 NVL72 racks are the first hall. A smaller block or a different platform is a different hall, and we will price it.
In place
- The hall
- An existing data centre in the Santiago Metropolitan Region, identified, named under a mutual NDA
- The builder
- An EPC contractor, named under NDA, engaged since August 2026
- The hardware
- OEM quotations in hand for twelve GB300 NVL72 racks, fabric and storage specified
- The design
- InfiniBand fabric decided; the cluster follows the vendor reference architecture
- The import route
- Duty, IVA and the export qualification worked through with customs and tax advisers
- The companies
- Fulmis Inc. in Delaware; Fulmis SpA in formation in Chile
- The financing
- Hardware financing in discussion with lenders and a lessor, sized to the contract
- After the hall
- A permitted 20 MW campus site in the Santiago Metropolitan Region
Hardware
Any NVIDIA platform, rack scale or node by node, from one unit up, in a Santiago data centre. Bought by us against your contract, or brought by you. Every row has its own page and every page ends in the reservation form.
| Platform | GPUs | GPU memory | NVLink | Power | Cooling | Smallest block | Available | |
|---|---|---|---|---|---|---|---|---|
| GB300 NVL72Rack scale | 72per rack | 20 TB288 GB per GPU | 130 TB/s1.8 TB/s per GPU | 132 to 140 kW155 kW peak | Liquid | 1 rack, 72 GPUs | Q1 2027 | Reserve |
| GB200 NVL72Rack scale | 72per rack | 13.4 TB186 GB per GPU | 130 TB/s1.8 TB/s per GPU | 120 to 132 kWper rack | Liquid | 1 rack, 72 GPUs | Q1 2027 | Reserve |
| HGX B3008-GPU node | 8per node | 2.3 TB288 GB per GPU | 14.4 TB/s1.8 TB/s per GPU | About 14 kWper node | Air or liquid | 1 node, 8 GPUs | December 2026 | Reserve |
| HGX B2008-GPU node | 8per node | 1.4 TB180 GB per GPU | 14.4 TB/s1.8 TB/s per GPU | About 14.3 kWper node | Air or liquid | 1 node, 8 GPUs | December 2026 | Reserve |
| HGX H2008-GPU node | 8per node | 1.1 TB141 GB per GPU | 7.2 TB/s900 GB/s per GPU | About 10.2 kWper node | Air | 1 node, 8 GPUs | Q1 2027 | Reserve |
| HGX H1008-GPU node | 8per node | 640 GB80 GB per GPU | 7.2 TB/s900 GB/s per GPU | About 10.2 kWper node | Air | 1 node, 8 GPUs | Q1 2027 | Reserve |
| RTX PRO 6000 Blackwell8-GPU node | 8per node | 768 GB96 GB per GPU | PCIe Gen 5no NVLink | 5.4 to 6.7 kWper node | Air | 1 node, 8 GPUs | Q1 2027 | Reserve |
| Vera RubinRack scale | 72per rack | HBM4not yet published | Not published | Not published | Liquid | 1 rack | On request | Reserve |
NVIDIA's published figures. Tensor Core numbers carry sparsity unless the row says dense; H100, H200 and RTX PRO figures are per GPU because NVIDIA publishes no node total. Power figures are the ones we reserve and protect to, with NVIDIA's own where it publishes one.
Who we build for
Seven kinds of customer, two ways in: we buy the hardware and you reserve it, or you bring it and we run it. The hardware changes with the workload; the terms do not.
AI labs and model companies
Training capacity on a multi-year term: GB300 NVL72 racks joined by InfiniBand, Slurm on top, storage sized for checkpoints.
Compute aggregators and desks
Blocks of GPUs on term, in racks or nodes, under managed Kubernetes, resold to your customers under your name.
Inference platforms
B300, B200 or H200 nodes under managed Kubernetes, with the network edge and the model-load path your platform needs.
Enterprises whose data cannot leave Latin America
In-country GPUs from one node up, air cooled where that is enough, billed in dollars by a US company with a Chilean subsidiary.
Research and universities
From one node to a rack, Slurm, on a term that fits a grant cycle, in the region.
Clouds that want a Latin American region
Wholesale racks under your brand. Your customer, your margin, a Santiago hall and our operations.
Owners of GPUs that need a home in Latin America
A fleet you already own or are buying, rack scale or air cooled nodes. We bring it into Chile, give it a position in a Santiago data centre and operate it for you, for the term.
How it works
Five steps from the first message to a running cluster. The first one can start today.
- Mutual NDASigned within the week
- Technical callWorkload, storage, acceptance
- Term sheetTerm, price, prepayment, start
- ContractTake-or-pay, three to five years
- Cluster liveYours for the term
What we are building
The first hall: 864 GPUs
Twelve NVIDIA GB300 NVL72 racks in an existing Santiago data centre, joined by InfiniBand into one training cluster. First rack live in Q1 2027. The offtaker seat is open. The projects in detail.
- GPUs
- 864 Blackwell Ultra
- Platform
- 12 × NVIDIA GB300 NVL72
- GPU fabric
- InfiniBand, Quantum-X800, rail optimised
- Built for
- Training
- Rack power
- 132 to 140 kW rated, 155 kW peak
- Cooling
- Liquid cooled racks
- Software
- Slurm and managed Kubernetes
- Smallest block
- One rack, 72 GPUs
What comes next
The campus: 20 MW
A permitted greenfield site in the Santiago Metropolitan Region, planned as four 5 MW halls. Every permit the 20 MW campus needs is held, including the building permit. The builder is engaged and named under a mutual NDA.
- 20 MW IT
- 25 to 27 MW at the meter
- 8,640 GPUs
- 120 racks in four halls
- 18 months
- To all four halls energized, per the contractor
Why Chile
The region has almost no GPU capacity
The two largest declared AI clusters in Latin America are 96 GPUs each, and both are academic.
Data that stays in the region
In-country capacity for teams whose data has to stay in Chile, and in-region capacity for the rest of Latin America.
Two modern cables to the United States
Curie lands on the US west coast. Mistral runs up the Pacific side of the continent.
Round trip from Santiago
Questions buyers ask
- Who owns the hardware?
- Fulmis. We place the purchase order, import it into Chile and operate it.
- Which hardware can I get?
- Any NVIDIA platform, from one node or one rack up. Every platform is here; name it and the quantity in the reservation form and we quote it.
- How small can I start?
- One node of eight GPUs on the HGX platforms. One rack of 72 GPUs on GB300 NVL72.
- When can I start?
- HGX B200 and B300 nodes go live from December 2026. The first GB300 NVL72 hall goes live in Q1 2027.
- Where is the hardware?
- In a data centre in the Santiago Metropolitan Region, named under a mutual NDA.
- Who is behind Fulmis?
- Mateo Pirela, founder and CEO. Fulmis Inc. is a Delaware corporation. About Fulmis, or write to mateo@fulmis.ai.