Solution Brief
Every revenue cycle agent. PHI inside.

Every agent.One server.

A revenue cycle agent suite on AMD Instinct™ MI350P.Prior auth, denial appeals, eligibility and documentation on one LIQID UltraStack server, with PHI inside your walls.

Download PDFTalk to an Expert
20
GPUs
MI350P on one host, room for 30
2.88
TB
HBM3E on one server
20
models
one per GPU, in parallel
0
hops
network hops between GPUs

The Revenue Cycle Gap

Clinical AI is mainstream. Revenue cycle AI is still a stack of competing pilots.

Adoption Gap

Clinical AI is everywhere

75% of health systems run AI; ambient notes went from 42% to 68% in a year.

Revenue cycle lags

Only 28% use prior auth AI and 25% denial prediction.

Pilot Sprawl

Every agent is its own pilot

Each one competes for its own GPUs.

The ROI is already proven

59–71% of adopters report 2× or better ROI.

PHI Pressure

PHI can’t go to a cloud API

Not at the scale revenue cycle runs.

Staff own the submission

Every draft must cite the chart and the payer rule.

One Server. Five Agents.

One 20-GPU LIQID UltraStack server runs five revenue cycle agents side by side, one model per GPU, each agent on its own GPUs. Every draft cites the chart and payer rule.

  • Fixed capacity replaces per-token fees
  • A CFO budget: denial rates, days in A/R and vendor fees
  • Grow from 10 to 20 to 30 GPUs on the same host
GPU allocation for five revenue cycle agents with pinned models, one model per GPU: 4 GPUs each for prior auth, appeals, eligibility and CDI, claims triage, and denials, on a LIQID GPU chassis with 20 AMD Instinct MI350P and 2.88 TB HBM3E, connected over PCIe Gen 5 to one dual-socket AMD EPYC 9005 host with one OS image.

Prior Auth + Appeals · 8 GPUs

gpt-oss-120b drafts prior auths; Llama 3.3 70B drafts cited appeals.

1,256
tok/s, 72B-class, 8 GPUs
1 GPU
per 70B model at FP8

Eligibility + CDI · 4 GPUs

Mistral Small 3.2 and Llama 3.1 8B run eligibility and CDI checks.

9.5×
scaling, 1 to 10 GPUs
44,383
tok/s, 7B-class, 10 GPUs

Claims Triage + Denials · 8 GPUs

Llama 3.1 8B routes claims; Llama 3.3 70B extracts denial reasons for appeal.

~89,000
tok/s, 7B-class, 20 GPUs
~65 GB
at MXFP4, fits one card
Measured results: LIQID Phase-3 testing on 1 to 20 MI350P GPUs on one host. Model specs are published. Notes on page 2.
LIQID UltraStack with AMD MI350P for Healthcare

Proof: Measured on MI350P

Four agents don't need four pilots: one 20-GPU server runs them all, one model per GPU.

Result
What was measured
Why it matters here
1,256 tok/s
on 8 GPUs
72B-class model, FP8, under interactive limits: P99 TTFT ≤ 2 s, ≥ 30 tok/s per user
Appeal letters and prior auth drafts at interactive speed for staff
9.5× scaling
1 to 10 GPUs, 7B-class
4,653 tok/s on 1 GPU to 44,383 tok/s on 10, one replica per GPU
Eligibility and CDI checks scale with every GPU you add
~5B active
of 117B, gpt-oss-120b
Published model spec: only ~5B parameters active per token, ~65 GB at MXFP4
Reasoning-class model on one card, expected to decode at least as fast as the 72B result
What it means
Five agents, one server
No agent waits for another pilot's GPUs
7B-class: FP4, one replica per GPU, concurrency 128. 72B-class: FP8, one replica per GPU, P99 time-to-first-token ≤ 2 s, ≥ 30 tok/s per user. Different test conditions; don't compare them. Source: LIQID Phase-3 testing. gpt-oss-120b, Mistral Small 3.2 and Llama performance is projected from these results; 7B-class scaling is measured to 20 GPUs (~89,000 tok/s).

Grow on the Same Server: 10, 20 and 30 GPUs

Add GPUs and a new job, not a new cluster.

10
GPUs
1.44 TB HBM3E

Health-system-wide clinical documentation

Scale summarization from a pilot unit to every clinician, sized for shift-change peaks.

Model layout

1 × speech-to-text (Whisper-class, US) · 7 × gpt-oss-120b note drafting · 2 × Llama 3.3 70B coding suggestions

Who buys, why now

CMIO / CNIO. Ambient pilot succeeded; PHI can’t go to a cloud API at scale.

Start here
20
GPUs
2.88 TB HBM3E

Revenue cycle agent suite

Prior auth, appeals, eligibility, CDI and claims triage as separate agents on one host.

Model layout

4 × prior auth (gpt-oss-120b) · 4 × appeals (Llama 3.3 70B) · 4 × eligibility and CDI (Mistral Small 3.2, Llama 3.1 8B) · 4 × claims triage · 4 × denial extraction

Who buys, why now

CFO / VP Revenue Cycle. Denial rates, payer rule changes, RCM vendor fees.

30
GPUs
4.32 TB HBM3E

Health-system-wide private model hub

Clinical documentation, revenue cycle and patient access on one host.

Model layout

10 × clinical documentation · 14 × revenue cycle agents · 6 × patient access (routing and reply drafting)

Who buys, why now

CIO / Chief Digital Officer. Consolidating clinical and administrative AI; PHI stays on-prem.

Every GPU runs one model with a fixed job. Starting with one job? See the 4-GPU Clinical Summarization and Coding brief.

What UltraStack Adds

  • One model per GPU: 144 GB HBM3E holds a full 70B model at FP8, or gpt-oss-120b.
  • No network between GPUs: peer-to-peer over PCIe Gen 5 to one EPYC host.
  • Air-cooled: about 22 kW for the full 30-GPU system.

Start Fixed, Grow on One Host

  • One host: one OS image, one license footprint, one thing to patch.
  • Room to grow: 10 to 20 to 30 GPUs on the same host.

What Is Proven Today

  • P2P bandwidth saturates the link; RCCL collectives pass, 1 to 10 GPUs.
  • ROCm integration complete, developed jointly with AMD.
  • 200+ production deployments, 85+ patents.
The Efficiency Case · Targets, not measurements*
75%
Up to 75% fewer servers
2.1×
Up to 2.1× more tokens per dollar
1.8×
Up to 1.8× better tokens per watt
50%
Up to 50% lower power

Bring Us Three Numbers

Two questions settle the size of the opportunity: how many prior auths and denials does your team work each month, and what does each one cost?

Size Your Revenue CycleContact UsDownload PDF
1
Volume:

prior auths, denials and claims per month.

2
Case size:

notes, labs and payer policies per case.

3
Concurrency:

sessions and staff at peak, incl. month-end.

Built for

Revenue cycle · Health information management · CDI · Patient access

Proven on

AMD Instinct™ MI350P · AMD EPYC™ 9005 · ROCm™ · 200+ LIQID deployments

*Efficiency targets compare against multi-server deployments of the same GPUs; they are LIQID internal projections using AMD published specifications, and results vary by configuration and workload.

Performance figures for the AMD Instinct MI350P are based on AMD published specifications and LIQID Phase-3 testing from 1 to 20 GPUs, and are subject to change. 30-GPU figures are projected from measured 20-GPU scaling. Every layout runs one model per GPU; 20- and 30-GPU layouts are sizing arithmetic on published model sizes, not benchmark results. Models shown are US and EU open-weight models. Not a medical device; outputs require clinician, coder or revenue cycle staff review, and payer submissions remain the customer's decision. Adoption and ROI data: Eliciting Insights, Feb 2026 (n=120).

Speak With An Expert