Proximal Agentic Compute × Open Models

Private AI that runs your way.

Open models and an agent marketplace, running on Proximal Agentic Compute (PAC). Try the same bundle in Google Cloud, burst onto a partner NeoCloud, then run the identical stack in your own data center. Hybrid by design, open by default.

Any open model — no model lock-in Runs on x86 you already own No GPU lock-in
The Bundle⇄ portable
Open Models + Agent Marketplace
Llama · Mistral · Qwen · DeepSeek · your own
Model serving & connectors
RAG · Jira · Confluence · SharePoint · OneDrive
PAC — Proximal Agentic Compute
Security, IAM, ACLs & Proximal AI Eyes — reused
x86 agentic compute
CPU-optimized · cost, reliability, availability
deploy once · run in Google Cloud, NeoCloud, or your DC
Hybrid cloud

One bundle. Three homes. Zero rewrites.

The exact PAC + open-models stack runs wherever your data needs to live. Evaluate in Google Cloud, scale on a partner NeoCloud, then graduate the identical bundle into your own data center for residency, latency, and TCO — with no re-architecting.

OPTION 1 us-central1

Google Cloud

Try & evaluate instantly
Open ModelsAgent MarketplacePAC

Provision a managed PAC sandbox in minutes. Validate models, agents and policy against real workloads before you commit hardware.

OPTION 2 partner region

Partner NeoCloud

Burst & scale on demand
Open ModelsElastic capacityPAC

Spill over to a Proximal partner NeoCloud for elastic agentic capacity — same bundle, pay-as-you-go, no long-term hardware commitment.

SAME
BUNDLE
on-prem

Your Data Center

Your private cloud, your control
PACOpen Modelsx86 fleet

Run the identical bundle on the infrastructure you already operate. Your data never leaves the building — and the experience is byte-for-byte the same as the cloud.

evaluate in Google Cloud  →  burst on NeoCloud  →  graduate to your DC  — same models, same agents, same policy
Tightly integrated

Open models, built into PAC — not bolted on.

Any open model deploys as a native workload on Proximal Agentic Compute — and inherits the platform your team already trusts. With it comes a full marketplace and ecosystem of agent partners.

OPEN

Any open model

Llama, Mistral, Qwen, DeepSeek or your own fine-tunes — deploy as a first-class PAC workload. No model lock-in, ever.

MARKETPLACE

Agent Marketplace

Thousands of partner-built agents, one click to deploy on-prem — the same catalog across every environment.

SECURITY

Reuse your security

Existing ACLs, IAM and encryption carry straight over. No new identity plane, no new policy model to learn.

VISIBILITY

Proximal AI Eyes

A built-in console shows every agent users deploy or create — plus agent-to-data and agent-to-agent activity.

DATA

Your data, connected

Built-in connectors to Jira, Confluence, SharePoint, OneDrive and more — grounded in your enterprise sources.

PARITY

Same workflow everywhere

Identical across Google Cloud, partner NeoCloud and your DC — same app, same agent setup, same experience.

PAC · Compute AI

Compute that thinks about cost.

The core of Proximal Agentic Compute isn’t just running models — it’s Compute AI: a runtime that picks the right model for every request. Local open models handle the work cheaply on x86; when quality demands more, PAC transparently escalates to a public API model. You get the best answer at the lowest defensible cost — automatically.

01

Compute AI thesis

PAC treats inference as an optimization problem, not a fixed pipeline — matching each request to the cheapest model that clears the bar.

02

Runtime model selection

Per-request routing across your model fleet. PAC chooses the model at run time based on the prompt, latency budget and live capacity.

03

Cost-based optimization

Local-first. Open models on your x86 run at near-zero marginal cost — PAC defaults to them whenever they can do the job.

04

Quality-gated fallback

Every local answer is scored. When it falls short of your quality threshold, PAC escalates to a public API model — only paying for premium when it matters.

PAC Router live routing
req ›Summarize this 8-page vendor contract…
served ✔
Local · open model
Llama 3.1 70B
x86 · ~$0.0008
escalated ⚡
Public API · premium
Frontier API
on-demand · ~$0.0210
Local quality score0.00
threshold 0.78
Running cost saved vs. always-premium 0%
Compute.AI microkernel

Solving the real enterprise query.

Business questions don’t map to a single model — they span all of an enterprise’s data. The Compute.AI microkernel decomposes each query into subqueries and dispatches every one to the silicon built for it.

“How do the latest tariffs and steel prices affect my quarterly results, my ability to deliver, my revenue and my cost?”
Relational SQL
transactions, financials, ERP
■ CPU
Graph traversal
supply-chain relationships
■ CPU
Vector search
semantic recall, embeddings
■ CPU / GPU
LLM reasoning
synthesis & the final answer
■ GPU
Memory-first node · HBM at the center
Every subquery reaches shared high-bandwidth memory in one hop — no rack-crossing, no copy tax. The microkernel keeps CPU, GPU and xPU working the same resident data.

Why memory-first wins: most subqueries are memory-bound lookups, not raw FLOPs. Co-locating SQL, Graph, Vector and LLM next to one pool of HBM is what makes running all four together — fast and private — actually possible.

Explore the Unity™ hardware platform
Ecosystem & marketplace

A marketplace of agents — tuned for your vertical.

The Proximal partner ecosystem brings purpose-built agents for the industries you run. Deploy them privately, on open models, on your own infrastructure.

🏦

Financial Services

Underwriting, fraud, compliance & complex transaction workflows.

Risk & underwriting agentsFraud detection & AMLInvestment research copilots
🩹

Healthcare

Clinical documentation, imaging & administrative automation.

Clinical documentation agentsImage analysis & triageCare-coordination workflows
🏛

Public Sector / SLED

Citizen services and case management with data sovereignty.

Constituent service agentsRecords & case automationPolicy & document summarization
🛒

Retail & CPG

Customer service, merchandising and supply-chain copilots.

Customer service agentsCatalog & content generationDemand & inventory assistants

Legal & Professional

Contract review, research and knowledge retrieval.

Contract analysis agentsLegal research copilotsKnowledge-base retrieval

IT & Engineering

Coding, IT automation and digital workflow assistants.

Coding agentsIT automation agentsWorkflow orchestration
From the Agent Marketplace
Coding agentsCursorn8n AILindyMoveworksServiceNow AI+ thousands more
x86 · CPU-optimized agentic compute

Why Proximal for the agentic era.

Agents are workflows, not training runs — they run beautifully on x86 CPUs. Proximal turns the infrastructure you already own into the most cost-effective, dependable home for private AI.

×

Proximal is partnered with VMware to deliver agentic compute — PAC builds on VMware’s proven virtualization, HA/DR and operations to run open-model agents reliably on x86, in any cloud or your own data center.

01 · COST

Lower TCO

x86 not GPU

Agents run on the CPUs already in your racks. Skip the GPU premium and scarce-silicon waitlists.

02 · RELIABILITY

Battle-tested

99.9%+

Built on VMware — the platform that runs the world’s data centers. That hardening, brought to AI.

03 · AVAILABILITY

Always on

HA · DR

High availability, disaster recovery and live migration — inherited from VMware, built in.

04 · OPERATIONS

Familiar tools

vSphere

Govern AI with the same VMware automation & vSphere tooling your team already masters.

Try it in 60 seconds

Launch a live PAC + Open Models instance.

Spin up in Google Cloud, burst onto a partner NeoCloud, or deploy to your own infrastructure — the bundle is identical.

proximal-pac · provisioning console

Choose where it runs

The bundle is identical in all three. Pick a target and launch.

BundlePAC + Open Models
ModelLlama 3.1 70B
Computex86 · 64 vCPU
Regionus-central1
pac@console:~$ ready to provision — press the launch button
✔ Instance ready
https://console.proximal.cloud/us-central1/pac-7f3a