packetaia hosted·ai project
Start building
GPUs/NVIDIA H200

H200
coming soon.

141 GB of HBM3e at 4.8 TB/s: the H100’s compute with 76% more memory, for long-context inference and large-batch training.

Get notified Compare GPUs
HopperComing soon
141GB HBM3e
Not on sale yetComing soonTell us what you would run on itNotify me →

Not available to rent yet

Memory141GB
Bandwidth4,800GB/s
ArchitectureHopper
Form factorSXM5
Good for

What the H200
is built for.

Tensor-core precisions: FP8, BF16, FP16, TF32 and FP64.

Best at

  • 70B models on one GPU with long contexts
  • Large-batch training
  • Vision-language fine-tuning
  • Memory-bound HPC

Coming soon

The H200 is not on sale yet. Tell us what you would run on it and we will let you know when it launches.

Get notified
Sizing

Will my model fit
on 141 GB?

Rules of thumb for weights plus a modest KV cache or optimizer state. Longer contexts and bigger batches need more.

WorkloadNeedsOn an H200
Llama 3.1 8B · FP16 inference~18 GBFits
Qwen 32B · FP8 inference~36 GBFits
Llama 70B · FP8 inference~78 GBFits
Llama 70B · FP16 inference~140 GBTight
Llama 8B · full fine-tune (BF16)~70 GBFits
70B · QLoRA fine-tune~48 GBFits
Flux.1 [dev] · BF16~33 GBFits
Llama 405B · FP8 inference~410 GBToo big
Other GPUs

Not the right fit?
Compare the fleet.

These are on sale now, billed hourly with no contract.

Blackwell

B200

180 GB HBM3e · from $4.25/hr

Blackwell

RTX PRO 6000

96 GB GDDR7 · from $0.66/hr

Ampere

A100 80GB

80 GB HBM2e · from $0.89/hr

Ada Lovelace

RTX 4090

24 GB GDDR6X · from $0.39/hr

The H200
is coming soon.

Tell us what you would run on it and we'll let you know when it launches.

Get notified
No contractTalk to a human any time

Black Friday special - Top up $100, get $50 free with code BLACKFRIDAYSPCL

AVAIL NOW
Start Building →
packet.ai/GPUs/NVIDIA H200
Coming soon, join waitlist
NVIDIA Hopper141 GB HBM3eSXM5

NVIDIA H200 GPU

Hopper, maxed out.

The NVIDIA H200 is the peak of the Hopper architecture, 141 GB of HBM3e at 4.8 TB/s, built for long-context LLM inference and large-batch training. Where the H100 runs out of memory, the H200 keeps going. Available on packet.ai from $2.49/GPU-hour.

from $2.49/GPU-hr· Launching soon
≈ 56% below market median
See all pricing
No contractsHourly billingSSH in <5 min
NVIDIA H200 GPU
141GB
HBM3e memory
4.8TB/s
Memory bandwidth
990TFLOPS
FP16 Tensor compute
900GB/s
4th-gen NVLink
Specifications

NVIDIA H200 specifications.

SpecificationValueGreat for
GPU architecture
NVIDIA HopperGH100 · 80B transistors
Same proven Hopper compute die as H100, battle-tested in production.
GPU memory
141 GBHBM3e
76% more than H100.
Memory bandwidth
4.8 TB/s
1.4× faster than H100.
FP8 Tensor compute
~1.98 PFLOPS
High-precision FP8 training.
FP16 Tensor compute
~990 TFLOPS
Standard mixed-precision training.
Transformer Engine
4th generation
Up to 4× FP8 speedup vs FP16.
NVLink
900 GB/s4th generation
Scales across NVSwitch-connected nodes.
MIG support
Up to 7 instances
Split a single H200 into isolated GPU instances.
Availability
On-demandUS & EU regions
Contact us for H200 capacity.
Architecture

Built on NVIDIA Hopper, with 1.4× more memory.

The H200 uses the same Hopper GH100 die as the H100 but swaps HBM3 for HBM3e, adding 76% more memory capacity and 1.4× more bandwidth.

GH100 Hopper die

Same 80-billion-transistor die as the H100. 4th-gen Tensor Cores, MIG support, NVLink 4.0, with HBM3e delivering 1.4× the bandwidth.

141 GB HBM3e at 4.8 TB/s

76% more memory than H100. Long-context LLMs, multi-modal pipelines, and large-batch training fit without sharding.

4th-gen Transformer Engine

FP8 training and inference with per-tensor scaling for up to 4× speedup over FP16 on Hopper.

NVLink 4.0 at 900 GB/s

Scale across NVSwitch-connected nodes for large training runs where gradient communication is the bottleneck.

Use Cases

What the H200 is built for.

Long-context LLM inference

141 GB fits 70B models with large KV caches, serve 128K+ context windows without degrading throughput.

  • 128K+ context windows
  • 70B on a single GPU
  • No KV cache starvation

Large-batch training

More memory means larger batches and fewer gradient accumulation steps for faster convergence.

  • Larger batch sizes
  • 1.4× more bandwidth than H100
  • Multi-node NVLink

Multi-modal fine-tuning

Vision-language and audio-language models need more VRAM, the H200 provides the headroom.

  • Vision + language models
  • Full 141 GB VRAM
  • Hourly billing

Scientific HPC

Memory-bound genome sequencing, protein folding, and climate modelling that exhaust H100 capacity.

  • 4.8 TB/s bandwidth
  • FP64 capable
  • MIG for multi-tenant
Pricing

H200 pricing.

H200 is launching soon on packet.ai. Join the waitlist to be notified.

FAQ

NVIDIA H200, answered.

What is the NVIDIA H200?

The H200 is the memory-upgraded H100, 141 GB HBM3e at 4.8 TB/s, 76% more memory and 1.4× more bandwidth.

How much does H200 cost?

H200 starts at $2.49/GPU-hour dynamic. See pricing →

Difference between H100 and H200?

Same compute die. H200 replaces HBM3 with HBM3e: 141 GB vs 80 GB, 4.8 TB/s vs 3.35 TB/s.

Does H200 support MIG?

Yes, up to 7 isolated MIG instances per GPU.

How do I get an H200?

Available on-demand from $2.49/hr. Deploy now →

Coming soon

H200: Launching soon. Get notified.

Join the notify list and be first to know when H200 capacity opens on packet.ai.

Talk to a human

No commitment · we'll notify you by email

NVIDIA H200from $2.49/GPU-hr