Is our data ever retained or used for training?

No. Every ChatVLM plan, including the free tier, ships with contractual zero data retention and a contractual commitment that customer data is never used to train models. Privacy is never a paid upgrade here. Enterprise and on-premises deployments add verifiable controls on top of that guarantee.

What exactly are Sovereign Presets?

Sovereign Presets run leading open-weight models on ChatVLM-operated private infrastructure with zero third-party subprocessors. Prompts routed to a Sovereign Preset never reach OpenAI, Anthropic, Google, or any other external provider, and the model lineup is refreshed continuously as the open-weight landscape evolves. Admins can set hard policies that force sensitive departments onto Sovereign Presets only.

How is Enterprise priced?

Enterprise SaaS is priced per seat with a 25-seat minimum, billed annually by invoice, using the same Pro and Ultra seat tiers as Teams. Sovereign Boxes are priced as one-time hardware plus an annual software license covering unlimited seats, or as a single monthly Hardware-as-a-Service figure on a 36-month term. Multi-node and multi-site deployments are priced as a single quote, and support is priced per managed node equivalent beyond the first three, counted across the fleet.

Can our developers use Claude Code and other coding agents?

Yes. Enterprise endpoints and Sovereign Boxes expose industry-standard APIs, so coding agents like Claude Code, IDE assistants, and in-house agents connect by changing the base URL. Combined with Sovereign Presets or an air-gapped Box, that gives an engineering org fully private agentic AI, and on a Box there is no per-token cost no matter how much those agents run.

How many people can a Sovereign Box cluster support?

Up to 1,000 people on a three-node Theta cluster and 4,000 on a three-node Omega, and both scale from there one node at a time. That is many more people than the stream counts suggest. The 150 and 600 figures describe simultaneous generation at full tilt rather than headcount, and real usage is bursty: a knowledge worker spends most of a session reading rather than waiting on tokens, so a few hundred of them add up to a small fraction of an appliance. Engineers running coding agents are the exception, and a heavy one counts as more than a full stream on their own. An agent holds a stream for minutes at a time with no human reading between turns, and the engineer keeps working while it runs rather than waiting on it. That mix is what decides sizing rather than headcount alone. The calculator above works it out for a specific org, and a pre-sale sizing review confirms it before purchase. Heavy agentic workloads consume far more capacity per person than everyday chat, so engineering-led deployments are sized differently, and every quote is based on an actual usage mix.

What happens if a Sovereign Box fails?

Nothing stops. Clusters start at three nodes and grow one node at a time, and every cluster holds two nodes worth of capacity in reserve, so losing any two nodes still serves the committed load. Traffic fails over automatically, updates roll node by node with no downtime, and the data tier keeps a warm spare that takes over without losing state. Every node beyond the first is priced at list hardware, and the software license covers the cluster as a whole.

What support comes with a Sovereign Box?

Support comes in three tiers, priced per managed node equivalent so it tracks the size of the fleet rather than the license. Standard is $10,000/MNE/yr, and the first three MNE are included, so a first cluster carries it at no extra cost: 24/7 automated monitoring, a 4-hour P1 response with next-business-day P2, quarterly offline model and software updates, and software upgrade assistance. The allowance counts once against the whole fleet, so a second site pays for all three of its nodes. Mission Critical is $30,000/MNE/yr with a $90,000/yr minimum, which is what a three-node cluster comes to. It adds a 30-minute acknowledgement and 1-hour P1 response, a named SRE pod on the account, an optional on-site spares cache, an annual deployment health review, and priority hardware advance replacement. Resident Engineer carries Mission Critical response times or better and adds a full-time engineer assigned to the deployment, with regular on-site presence, a direct escalation path, capacity planning and model tuning, and runbook and enablement work for the internal team. It is $420,000/yr per FTE, and required above 15 managed nodes.

What else do we need besides the GPUs?

ChatVLM needs a small amount of conventional infrastructure alongside the inference hardware: a database, cache, object storage for attachments, a vector store, observability, and a backup target. Sovereign Box Theta carries all of it on board and replicates it across the cluster, so a Theta deployment is genuinely three boxes on ordinary wall power and nothing else. For Sovereign Box Omega we recommend the Sovereign Control Plane: three redundant 1U servers with a warm database spare, 100 GbE switching, and a 50–100 TB NVMe storage array, integrated and supported by us. Organizations with existing Kubernetes, Postgres, and storage teams can instead supply their own to our published reference architecture, which we certify before deployment.

Do you support multi-site or fleet deployments?

Yes. Nodes are managed as a fleet through a local control plane that handles configuration consistency, canary updates, and capacity monitoring across every site. For air-gapped facilities, diagnostics are exported as a signed encrypted bundle that the customer hands to us, so we can debug a deployment we have no network path to. Sites beyond the first are $25,000/yr, and managed node equivalents are counted across the whole fleet against a single three-MNE allowance, so the nodes at a second site are all chargeable.

Can we avoid a large capital expenditure for a Sovereign Box?

Yes. Both appliances are available as Hardware-as-a-Service on a 36-month term, including the software license, Standard support, and hardware replacement: a three-node Theta cluster from $12,200/month, and a three-node Omega cluster from $65,500/month with the Sovereign Control Plane. Every node beyond the third bills the same flat rate, $2,400/month on Theta and $16,100/month on Omega, so capacity is an OpEx line that scales one node at a time instead of a CapEx approval.

What does installing a Sovereign Box Omega involve?

Omega is a datacenter system: each node draws roughly 14 kW on 3-phase power and requires liquid-integrated rack cooling, so a three-node cluster is a rack-scale install. A mandatory pre-sale site survey ($25,000, credited against purchase) validates the facility before we ship. Theta, by contrast, runs on standard office wall power with no special cooling.

How do updates reach an air-gapped deployment?

Quarterly model and software updates are delivered on encrypted physical media, by secure courier for Omega. License enforcement is fully offline, so a Sovereign Box never needs to phone home.

ChatVLM Enterprise

Frontier AI, from private cloud to air gap

Every ChatVLM plan carries contractual zero data retention and no training on customer data. Enterprise goes further, running open-weight frontier models on private infrastructure or entirely on-site.

Contact sales See the cost comparison

Private models without the hardware

The same Pro and Ultra seat tiers as Teams, plus the identity and governance controls a security review asks for.

Enterprise

Private cloud

Pro and Ultra seats for the whole organization, plus private models, identity controls, and compliance infrastructure.

Everything in Teams
Pro & Ultra seat tiers
Sovereign Presets on private infra
Zero third-party subprocessors
SAML SSO + SCIM provisioning
Workspace model allowlisting
Custom retention & legal holds
Immutable audit logs & SIEM export
SOC 2 Type II & HIPAA BAA package
Dedicated CSM & security review help

Custom

Contact sales

25-seat minimum · Annual, invoiced billing

Sovereign Box: the whole stack in one rack

Two appliances, one sized for an office and one for a datacenter. Both ship as clusters of at least three nodes, with unlimited seats, offline license enforcement, and quarterly model updates on encrypted physical media.

Sovereign Box Theta

The office-ready private AI server. High-concurrency local routing on standard wall power.

From $85,000

per node, one-time hardware+ $60,000/yr cluster license

4× RTX PRO 6000 Blackwell Server Edition, 384 GB VRAM
Multiple mid-size open models, routed intelligently
Up to 4 models served simultaneously
From 150 concurrent streams and 1,000 users
+50 streams per node beyond the third
Full capacity with two nodes down
Standard 110/220 V office power
Self-contained, with database, cache, and storage on board
Unlimited seats
Quarterly offline model updates

Talk to sales

Priced by CPU/RAM configuration · Or from $12,200/mo as HaaS for a three-node cluster (36-month term)

Sovereign Box Omega

Air-gap flagship

The datacenter deployment, for organizations standardizing on the largest open-weight models.

$580,000

per node, one-time hardware+ $150,000/yr cluster license + Sovereign Control Plane: $110,000 hardware + $20,000/yr Or bring existing infrastructure to our reference architecture

8× NVIDIA HGX B300, 2.3 TB HBM3e
Largest open-weight models resident in memory
Headroom for smaller routing models
From 600 concurrent streams and 4,000 users
+200 streams per node beyond the third
Full capacity with two nodes down
24/7 monitoring, 4-hour P1 response
Unlimited seats
Secure courier-delivered updates

Talk to sales

Or from $65,500/mo as HaaS for a three-node cluster with control plane · Pre-sale site survey required

One private endpoint for every AI tool

Enterprise endpoints and Sovereign Boxes speak the standard APIs these tools already use, so coding agents, IDE assistants, and internal automations all run on private models.

Coding agents on private models
Point Claude Code or any other AI assistant at the private endpoint. Source code, prompts, and diffs stay on infrastructure the company controls.
Standard, drop-in APIs
Industry-standard API compatibility means existing SDKs, IDE plugins, and internal tools work unchanged. Swap the base URL, keep the workflow.
Internal automations, same perimeter
Nightly extraction jobs, CI reviews, and in-house agents call the same endpoint, so prompts, source code, and documents never leave the deployment.

As many agents as you want, for the same bill

Teams arrive with an agentic workflow whose API bill grows every month it works well, and leave running the same workflow flat out on machines they own. On a Sovereign Box the marginal token is free; a private cloud is priced on capacity, not usage.

The meter is off
You pay for capacity: a box in your rack, or single-tenant infrastructure we operate. Whether it idles or runs all night, the invoice at the end of the month is the one you signed.
Agents without a quota
Coding agents, extraction pipelines, CI reviews, and overnight evaluations run at once, around the clock. The tenth agent costs what the first one did.
More tokens for the same money
Nobody trims context to save cents or thinks twice about a retry. The work gets the tokens it actually needs, and the number on the invoice stays where it was.

When the big providers go down, you don't

A public API is one shared system having one shared bad day. A Sovereign Box serves your people and nobody else's, from hardware in your own building, so an incident somewhere else is not an outage for your company.

Nobody else's traffic, nobody else's incident
The cluster serves your users and no one else, so another company's spike or a provider's control-plane failure has no path to it. An air-gapped Box does not even need the internet to be up.
Provider outages pass you by
Large providers go down for hours at a time. Work on a Box has carried on straight through those incidents, because nothing in the answer path belongs to a company you do not control.
Full capacity with two nodes down
Every cluster ships as at least three nodes, so a failed node, or one pulled for maintenance, does not take the service with it. Updates land in your change window, not when a vendor ships.

Total cost of ownership

Metered seats and agentic usage scale with every person added. An appliance doesn't. Set the headcount below to see where the lines cross.

Marketing, sales & operations

Drafting, research, summarizing, everyday chat

Analysts, legal & research

Long documents, deep research, council answers

Engineers

Coding assistants and agentic workloads

How hard do your engineers push it?

Agentic workloads are the line that dominates a metered bill.

Light

Completions and the occasional chat

Standard

Daily agentic coding in the IDE

Heavy

Agents running continuously, CI review, batch jobs

Sovereign Box support

Per managed node equivalent, with the first three included across the fleet.

Standard

24/7 monitoring, 4-hour P1 response. $10,000/MNE/yr, first three included.

Mission Critical

1-hour P1 response and a named SRE pod. $30,000/MNE/yr, minimum $90,000/yr.

Resident Engineer

Mission Critical response times or better, plus a dedicated engineer on the deployment. $420,000/yr per FTE, required above 15 nodes.

Saved per month

$12,800

Saved per year

$153,600

Running 100 people on your own hardware, against the same org on metered Enterprise SaaS seats.

Enterprise SaaS

$25,000/mo

Seats: $5,000
Agentic usage: $20,000

3 × Sovereign Box Theta

$12,200/mo

3-node HA cluster, automatic failover

3 nodes (hardware): $7,200
Software license (cluster): $5,000
Standard support: Included
Seats: Unlimited
Agentic usage: $0 (marginal token is free)

Serving 100 people and ~30 agents

22 of 150 streams

Sizing is set by how many people generate at the same instant rather than by headcount, because most of an org is reading rather than waiting on tokens. At this mix the configuration has room for about 560 people before you need more capacity.

The bar shows whichever ceiling sets the node count, streams or population. Committed capacity here is 150 streams across 3 nodes at 50 each, with 1,000 rated seats. Each node commits a third of what it can physically serve, so 2 nodes' worth stays in reserve and losing any two of them serves this load without interruption.

Estimates only, for comparison. Seats priced at the published Pro ($30) and Ultra ($80) rates; appliances at Hardware-as-a-Service monthly pricing on a 36-month term, software license included. Agentic figures assume $1,000/engineer/month of metered model usage at standard intensity, and about 1.5 concurrent agents per engineer. Clusters start at 3 nodes, hold 2 nodes' worth of capacity in reserve, and are sized to stay under 80% of committed streams; every node bills the same monthly rate under a single cluster-wide software license. A Theta cluster scales horizontally until metered spend passes $131,000/month, where Omega and the largest open-weight models become the better step. Support is shown at the tier selected above, priced per managed node equivalent with the first 3 MNE included once across the fleet.

Frequently asked questions

Add-ons and services

Listed up front rather than averaged into the quote.

Additional sites$25,000/yr beyond the first
Three-year prepayment10% discount
Hardware refresh20% trade-in credit at year three
Pre-sale site survey$25,000, credited against purchase
Sovereign model evaluation on in-house data$25,000, credited against purchase
Theta installation and commissioningFrom $10,000
Omega installation and commissioningFrom $50,000
On-site spares cache$50,000
RAG and data-source integrationFrom $50,000
Domain distillation and model tuningFrom $75,000

Put frontier AI where the sensitive work happens

Talk to us about Sovereign Presets, a proof of concept on a Sovereign Box, or a compliance review.

Contact sales Buying for a smaller team?