Kentino AI Node H200 — 8× NVIDIA H200 SXM5, 1,128 GB HBM3e
Enterprise-grade platform, assembled and tested in the EU.
Kentino AI Node H200
1,128 GB HBM3e in One Chassis
8x H200 SXM5 | NVSwitch | 15,832 INT8 TOPS
The tier where frontier models stop needing quantisation.
An 8U node carrying eight NVIDIA H200 SXM5 modules on a single NVLink/NVSwitch baseboard: 1,128 GB of HBM3e pooled at 4.8 TB/s per GPU, with full any-to-any 900 GB/s GPU-to-GPU bandwidth. Dual Intel Xeon Platinum processors, 2 TB of DDR5 system memory, U.2 NVMe boot and data tiers, and high-speed networking. This is a different class of machine from a PCIe multi-GPU build: because every GPU talks to every other GPU at full bandwidth, tensor parallelism scales close to linearly instead of degrading on dense models.
Hardware
| Component | Detail |
|---|---|
| GPUs | 8x NVIDIA H200 SXM5 141 GB HBM3e (Hopper, 700 W per module) |
| VRAM pool | 1,128 GB HBM3e, NVSwitch-pooled - usable as one address space |
| Memory bandwidth | 4.8 TB/s per GPU |
| GPU interconnect | NVLink 4 via 4x NVSwitch, 900 GB/s any-to-any (not PCIe-limited) |
| Compute | 7,912 TFLOPS bf16 dense | 15,832 TFLOPS fp8 dense | 15,832 INT8 TOPS |
| CPUs | 2x Intel Xeon Platinum, dual-socket |
| System memory | 2 TB DDR5 |
| Storage | 2x 1.92 TB U.2 NVMe boot + 2x 3.84 TB U.2 NVMe data. Expandable |
| Networking | 2x NVIDIA ConnectX-7. Up to 8x on request for multi-node InfiniBand fabric |
| Chassis | 8U rack-mount, front-to-back airflow, 24/7 rated |
| Power | Redundant, up to 10.2 kW peak system draw |
What it runs
1,128 GB is the point at which the largest open-weight models run at their native precision instead of being squeezed into a quantisation that costs you quality. Estimates below are for this 8-GPU configuration with tensor parallelism.
| Component | Detail |
|---|---|
| DeepSeek-V3.2 / R1 671B | Native fp8, full context. ~30-50 tok/s single stream, 2,000+ tok/s batched |
| Kimi-K2 1T class | fp8 with room for KV cache - no 1.58-bit compromise needed |
| Llama 3.3 70B / Qwen3-72B | fp8. ~90-130 tok/s single stream, 5,000+ tok/s batched |
| Qwen3-235B-A22B MoE | fp8 native with long-context headroom |
| Long-context serving | 128k+ context with KV cache resident, not spilled |
| Full fine-tune 70B | 8-way tensor parallel, no gradient offload |
Throughput figures are estimates derived from published vendor specifications and community benchmarks for this GPU generation, not measurements taken on this exact configuration. We will give you measured numbers for your model and context length before you commit.
Power and cooling
Up to 10.2 kW peak system draw, the bulk of it the 8x 700 W GPU complex. In practice that means a dedicated three-phase feed per node and a rack budget planned around roughly one node per 12 kW. Air-cooled at this density needs real front-to-back airflow with a contained cold aisle below 24 C - perforated doors alone will not do it. A liquid-cooled baseboard variant is available and is the better answer above two nodes per rack; it needs a CDU and secondary loop, which we will scope with you rather than assume.
Export control - read this before ordering
The H200 is classified as an Advanced Computing Item (ECCN 3A090.a). Shipment requires a completed KYC form and an End User Certificate naming the end user, the installation address and the intended use. This is not optional and it is not something we can waive.
Practically: budget several weeks for clearance on top of hardware lead time, and expect to name the legal entity that will own and operate the machine. If the end user cannot be named, this hardware cannot be supplied - by us or by anyone else in the EU.
What you get from Kentino
Kentino supplies, configures and supports this node as a complete deployment: rack and power design against your site's actual feed, front-to-back airflow sizing for 24/7 operation, OS and driver stack (Ubuntu with pinned NVIDIA driver, CUDA, container toolkit), the inference or training stack of your choice, burn-in with a written report, EU-based logistics and warranty.
We also handle the export paperwork with you rather than mailing you a form and hoping. That part takes longer than the hardware and is where most first-time buyers of this class of machine get stuck.
Price shown is EUR excluding VAT, EXW Prague (Incoterms 2020). Collection, freight, insurance and any destination import duties are for the buyer’s account. Lead time on request.
Channel pricing on 8-GPU baseboards moves, so we confirm the figure against documented acquisition cost when you order — if it has moved in your favour, so does the price. Export clearance is part of the delivery. Contact us with your model, context length and site power envelope and we will come back with a firm quote, a measured throughput estimate and a delivery date.
The questions buyers ask us most often before ordering a server.
How long does it take?
Machines built from components we hold ship quickly; anything requiring a specific GPU generation depends on supply. We give you a date before you pay, and if it moves we tell you rather than letting you find out.
Can the configuration be changed before you build it?
Almost always. GPUs, memory, storage and cooling are chosen per order, and the listed configuration is a starting point rather than a fixed package. If you need more VRAM, faster storage or a different cooling approach, say so before you order and we will quote the change.
Can I collect the server in person?
You can. Our warehouse is in Prague, and collection in person is welcome — most people who come use the visit to go through the machine with our engineer and ask the questions that are awkward over email. For orders within the Czech Republic we also try to deliver personally and walk you through the setup on site.
Can I talk to someone who actually understands the workload?
Yes. We have an engineer who works on AI systems specifically, not a general sales desk. If your question is about batch sizes, quantisation, interconnect or where your bottleneck will be, ask it — that conversation usually changes the configuration for the better.
Which model can I run on this configuration?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
How do you test a server before shipping?
We run it against actual AI workloads rather than synthetic scores: inference throughput, sustained load behaviour and thermals under continuous operation. You get the benchmark results with the machine, so the performance you were promised is the performance you can verify on day one.
Is the server ready to run when it arrives?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
Ships from our EU warehouse. Heavy items may require freight arrangement — contact us for a shipping quote and lead time. 2-year limited warranty with advanced RMA support; extended warranty available.
Not exactly what you need?
Tell us your workload and we'll spec this platform around it — GPUs, memory, storage and cooling matched to what you actually run.