NetApp Iterate.ai
Field Brief
Updated July 2026
Field Guide · The Inference Layer

What Actually Runs the Model

Between your question and the answer sits an engine. How good it is decides what private AI costs — and whether it runs on one small box or three hundred.

What you'll learn
Why agents break inference engines built for chatbots
What Lifeboat does differently, and what it measures
Why running well on small hardware does not mean only small hardware

The core argument

A chatbot asks the model one question and gets one answer. An agent does not. One user request can trigger five to ten model calls — planning, fetching, checking, retrying — and they run for minutes, holding the whole conversation in memory the whole time.

Most engines were built for the chatbot. Point agents at them and they buckle. Standard software can collapse at four or five long jobs on one GPU.

Lifeboat is Iterate.ai's proprietary inference engine — the layer that runs the model on your own hardware. It is built for the agent, and handles 25 or more long jobs on that same GPU.

Measured head to head against SGLang

Measured July 2026. One NVIDIA RTX PRO 6000 Blackwell GPU, Qwen 30B, long-context workloads simulating agents reading large documents in parallel. This work is continuous — the gap has widened at every round of testing so far.

Lifeboat
2,048 sessions
SGLang
1,024 sessions

Concurrent sessions at 100% request success.

Lifeboat
8,714 tokens/sec
SGLang
4,965 tokens/sec

Throughput at 2,048 concurrent sessions — 76% faster.

2–6×
More concurrent capacity on the same GPU
129×
Faster response under heavy agent load — 1.5s against 189s
40%
Less GPU memory for conversations and multi-turn work
NetApp | Iterate.aiWhat Actually Runs the Model · 01 / 02

Small hardware is a choice, not a ceiling

The story starts with a box you can carry, so people assume that is the limit. It is not. The same engine runs a departmental AIPod Mini and a full enterprise cluster. You change the hardware, not the software.

Departmental
AIPod Mini

RTX-class GPUs in one box, in an office or a closet.

Divisional
Multi-GPU server

Tensor parallelism across 1 to 8 GPUs in a single machine.

Enterprise
Clustered A100 / H100

300+ node clustering across a data center.

One Docker container · 168 model architectures · OpenAI-compatible API · air-gapped deployment.

Memory compression

Agents re-read the same conversation constantly. Compressing it frees 40% of the GPU.

Fair scheduling

Quick requests stay responsive while long agent jobs run, instead of queueing behind the slowest.

Model optimization

Tuned pipelines for mixture-of-experts models, so a capable model fits smaller hardware.

The same engine, on somebody else's models

Two production workloads, neither on our hardware. The model stayed as it was in both, and gains on this scale have been consistent.

A national beauty retailer
Site search and personalization
95%
lower processing cost
75%
less compute
20×
less memory
An outdoor equipment brand
Image generation, vs a major cloud
12s→2s
per image
faster
12×+
cheaper per image

We did not replace the models. We made them run on less hardware, in less memory, for less money.

An economics problem first

Private AI is an economics problem before it is a privacy one. Get twice the work from one GPU, and owning your inference stops being the expensive choice.

Further reading

Full paper, 12 pages — iterate.ai/partners/netapp/papers. Benchmarks: iterate.ai/lifeboat.

About this series

A joint educational series on private AI and the AIPod Mini. NetApp — the governed data-control layer. Iterate.ai — the private intelligence layer.

NetApp AIPod Mini
NetAppIterate.ai
v1.0 · July 2026
NetApp | Iterate.aiWhat Actually Runs the Model · 02 / 02