How many GPUs do I need to fine-tune a 70B parameter model?

Roughly 4-8× H200 (141GB) or B200 (192GB) for LoRA fine-tuning; 16-64× H100 (80GB) for full fine-tuning depending on batch size + context length. Servnet runs workload sizing against your specific model + technique.

Can I start with H100 and upgrade to B200 later?

Yes for fabric / network / storage. NO for chassis — H100 SXM and B200 SXM use different baseboards (HGX H100 vs HGX B200). Plan chassis around your 3-year horizon. Air-cooled H100 chassis cannot be converted to liquid-cooled B200.

Is NVIDIA DGX better than Supermicro?

For turnkey simplicity + NVIDIA AI Enterprise software bundle + NVIDIA support: DGX. For lower cost + identical compute capability + your choice of OS / cluster manager: Supermicro SYS-821GE. Both use the same HGX H100/H200/B200 baseboards.

What's the lead time on a 10-GPU server?

H100 8-GPU: 4-8 weeks. H200 8-GPU: 8-12 weeks. B100/B200 8-GPU: 12-20 weeks. Q4 / H1 demand spikes can extend. Servnet quotes honest lead time at the point of bid.

NVIDIA H200 vs H100 vs B100 vs B200: UK AI GPU buyer's guide 2026

NVIDIA H100, H200, B100, B200, and GB200 NVL72 are the GPUs UK enterprise AI buyers shortlist in 2026. H100 + H200 (Hopper architecture) are mainstream + available; B100 / B200 (Blackwell) are the latest generation, with GB200 NVL72 as the rack-scale superchip. This guide explains which to pick when — and the UK supply + power reality.

H100 · H200 · B200 — UK supply summary

The current NVIDIA AI GPU lineup

H100 (Hopper, 80GB HBM3) — the workhorse. Mainstream availability since 2023. Still the most-deployed AI GPU in UK enterprise. 700W TDP. SXM5 + PCIe form factors.

H200 (Hopper, 141GB HBM3e) — H100 successor. 141GB memory (vs 80GB H100) makes it the right choice for large-model training + inference. 700W TDP. Available in UK from H2 2024.

B100 (Blackwell, 192GB HBM3e, 700W) — successor to H100 on Blackwell architecture. Roughly 2.5× training perf vs H100. Mainstream availability H2 2025.

B200 (Blackwell, 192GB HBM3e, 1000W) — higher-power Blackwell variant. 4× H100 training perf. Requires liquid cooling at rack density. Available as DGX B200.

GB200 NVL72 — rack-scale superchip. 36 Grace CPUs + 72 B200 GPUs in one liquid-cooled rack. Single coherent memory pool. For frontier model training only. Available as NVIDIA GB200 NVL72.

Which GPU for which workload

Inference (LLM serving, RAG, embeddings): H100 80GB is still the practical sweet spot. Plenty of memory, mature software, lowest cost per inference. H200 only if your model exceeds 80GB single-GPU.

Fine-tuning (smaller models, LoRA, adaptation work): H100 or H200. H200's extra memory becomes valuable for larger context windows.

Pre-training (medium-scale, under 100B parameters): H200 or B100. The B100 power efficiency improvement matters at multi-rack scale.

Pre-training (frontier scale, 100B+ parameters): B200 or GB200 NVL72. Memory + interconnect bandwidth + power efficiency are all step-changes vs Hopper generation.

UK availability + lead times (Q2 2026)

H100 80GB: widely available, 4-8 week lead time. Servnet stocks Supermicro SYS-821GE 8U 8× H100/H200 servers.

H200: improving availability through 2026, 8-12 week lead time typical. Same Supermicro chassis (SYS-821GE) supports H100 or H200 SXM modules.

B100: mainstream availability through 2026, 12-16 week lead time. Higher demand from hyperscalers competes for supply.

B200: 16-20 week lead time typical. Requires liquid-cooled 4U chassis at high density.

GB200 NVL72: order-to-deliver typically 9-12 months. Frontier deployments only.

TCO crossover — training a 70B model on H100 vs B200

Power + cooling reality for UK data halls

8× H100 SXM in SYS-821GE: ~10.5 kW per server. Air-cooled. Fits in standard 42U rack with thermal headroom.

8× B200 SXM in liquid-cooled 4U: ~14 kW per server. Requires direct-to-chip liquid cooling + CDU. Most UK colos can support but pre-survey is essential.

GB200 NVL72: ~120 kW per rack. Requires purpose-designed liquid-cooled facility — not all UK colos support this density yet. Talk to Servnet about UK colo options that do.

What Servnet does

Servnet is an authorised UK partner of NVIDIA + Supermicro. We quote NVIDIA DGX (turnkey) and Supermicro GPU SuperServer (component-built) configurations.

Typical AI infrastructure engagement: 1) workload sizing (model architecture + training vs inference + concurrency), 2) sized commercial bid across DGX + Supermicro options, 3) data-hall pre-survey (power + cooling + space), 4) deployment + commissioning, 5) optional ongoing managed AI infrastructure service.