The four generations you will see in on-prem AI quotations from 2024 through 2027. What each one buys you, which EMARQUE and DGX systems carry it, and how to plan a multi-year refresh path.
Transformer Engine with FP8. First generation tuned for LLM training and inference at scale. H200 adds HBM3e for larger KV caches.
Still a strong on-prem choice for 70B-class production inference. Cheaper per node than Blackwell and shipping in volume.
FP4 compute, 5th-gen NVLink, much larger HBM3e per GPU. Step change for training throughput and long-context inference.
The current production-volume generation. Right answer for most enterprises refreshing in 2025–2026.
Same architecture as Blackwell with denser HBM3e per GPU and higher dense FP4 throughput. Built for reasoning workloads and very long context.
Right answer when reasoning workloads need the per-GPU memory headroom, or when you want the NVL72 rack-scale fabric.
HBM4 memory and next-gen NVLink Switch fabric. NVIDIA targets ~3.3 EFLOPS dense FP4 per NVL72 rack.
Multi-year planners only. Allocation conversations now for late-2026 / 2027 deployment windows.
Cheapest tokens per MYR for 70B-class production inference. Use it when budget dominates and the workload isn't HBM-pressure-bound. Plan a Blackwell refresh in 18–24 months.
Current refresh sweet spot. FP4 economics, mature shipping volume, easy DGX SuperPOD scale-out. Step up to B300 in-place when the workload demands more HBM per GPU.
Long-context reasoning, rack-scale fabric, or a roadmap that lands in 2026–2027. Have the allocation conversation now — supply is the gating factor, not technology.
Architecture consult to map current Hopper / Blackwell estate onto a Blackwell Ultra or Rubin roadmap — without throwing out what already works.
Tell us about your workload. We reply within one business day with a quote sized to fit.