SwitchGate for Enterprise
Run your own models on your own hardware behind the same governed gateway your teams already use for cloud models. Private inference, public-cloud breadth, one ledger, and controls your auditors will actually like.
99.95%
uptime SLA target, backed by cross-provider failover
10×
rate limits vs Team, with dedicated capacity options
<4h
P1 response from a named solutions engineer
CA · EU
data residency options under a DPA
Targets shown are placeholder data pending the console launch.
Built for organizations
Everything the free tier has, plus the parts legal asked about
Private model hosting
Serve your fine-tuned or open-weight models from your own GPUs through the SwitchGate Model Tunnel, in your catalog next to the cloud frontier.
SSO & SCIM
SAML and OIDC single sign-on with directory sync, so keys assigned to users follow your identity provider, joiners and leavers included.
DPA & GDPR
Canadian hosting, a signed data-processing agreement, and pure-routing guarantees: no training on your data, no prompt retention by the gateway.
Volume fee tiers
The 4.9% top-up fee steps down with committed volume; model usage stays at provider list price at every tier, and private inference bills 0 credits.
Org-wide governance
Budgets, model locks and the kill switch applied per user, per team and per environment, with keys assigned to named people and audit-log export.
Priority support & SLA
A named contact, an uptime SLA on the gateway, and migration help from whatever you route through today.
The enterprise platform
What changes at enterprise scale
Dedicated capacity
Provisioned throughput on the models you depend on, so a provider's public queue is never your incident.
Higher rate limits
Custom ceilings across the catalog - burst for batch jobs, sustained for products - raised without a support ticket.
No-logging mode
Metadata-minimized deployment for regulated workloads: the ledger keeps cost math and drops everything else.
Provider & region allowlists
Org-level policy over where requests may route: pin to Canadian or EU regions, exclude providers wholesale, per team.
Data residency & DPA
Canadian hosting standard, EU residency available, subprocessor list and pen-test summaries under NDA.
Volume fee tiers
The 4.9% top-up fee steps down with annual commitment. Model usage stays list price - we discount our fee, not your tokens.
Procurement, solved
POs, wire/ACH, annual invoicing, security questionnaires answered by people who wrote the systems.
Identity at scale
SAML SSO and SCIM provisioning: joiners get governed keys automatically; leavers lose them the same minute.
Early model access
New frontier checkpoints appear in enterprise catalogs first, with GateBench scores attached before general rollout.
Full plan details on the Enterprise plan page, or line-by-line on the comparison board.
The Model Tunnel, in detail Enterprise exclusive
Your own model, behind the same API
Full product page: architecture, install, security model and FAQ →
The Model Tunnel is not an AI: it's a single static binary or container you install on your own server next to your inference runtime. It solves the hard part of private AI by opening a secure tunnel from your model to the gateway, so the model becomes usable through our API, on your website, by your teams, or by paying customers, without punching a hole in your firewall.
How it works
Install next to your runtime
One binary via curl, a Docker container, or a Helm chart into the same cluster as your vLLM or Ollama deployment. It needs outbound HTTPS and nothing else.
The Agent dials out, never in
It opens a persistent outbound mTLS tunnel to the gateway. No inbound ports, no public IP, no VPN appliance; it works from a datacenter, an office, or a workstation.
Your models join the catalog
Whatever the runtime serves is registered under
private/your-org/<model>, visible only to your organization, callable through the standard endpoint.Governance applies automatically
Per-user budgets, model locks, user-assigned keys and the kill switch treat private models exactly like cloud ones; the ledger records every call at $0.
Route, mix and fail over
Send sensitive traffic to private models and overflow to cloud, or seat your fine-tune on a review council. And once the marketplace opens, the same tunnel will let you charge for access (see below).
Deploy it your way
helm repo add switchgate https://charts.switchgate.ai
helm install sg-tunnel switchgate/tunnel \
--set org=acme \
--set runtime.url=http://vllm.internal:8000 \
--set replicas=2 # HA: agents load-balance the tunnelorg: acme
runtime:
kind: vllm # vllm | ollama | llamacpp | openai-compatible
url: http://vllm.internal:8000
publish:
- "private/acme/llama-4-maverick-ft"
- "private/acme/legal-qwen-72b"
limits:
max_concurrency: 32
allowed_teams: ["platform", "legal-ai"]Security posture
- Data plane stays home. Prompts and completions for private models are processed inside your network; the control plane sees metadata only.
- Outbound-only mTLS with pinned certificates and per-agent identities you can revoke individually.
- High availability: run multiple agents; the gateway spreads traffic and routes around a lost tunnel in seconds.
- Least privilege: publish only the models you list, to only the teams you name.
- Full audit trail: every private call lands in the same exportable ledger as cloud traffic.
Model marketplace · on the roadmap
You'll be able to resell your model on SwitchGate
The same tunnel that serves your teams is built to serve paying customers next. When the marketplace opens, you'll flip a listing from private to public, set a per-token price, and your model will join the catalog next to the frontier: running on your hardware, billed by us, paid out to you. Here's how it will work.
Publish the listing
You'll promote
private/acme/legal-qwen-72bto a public catalog entry with a name, description and your own model card.Set your price
You'll choose the $/1M input and output token rates. Change them any time; buyers will always see the current list price, pass-through, like every other model here.
We meter and bill
Every buyer call will be metered by the gateway with the same exact-cost tracing you already use. Buyers will pay with their SwitchGate credits; you never handle billing.
You get paid
Monthly payouts of your earnings minus a marketplace share (rate to be announced; enterprise agreements will be able to negotiate it), with a revenue dashboard alongside.
You stay in control
You'll cap concurrency so buyers never starve your own teams, restrict listings by region, and unpublish instantly: the tunnel is yours, so the off switch is too.
curl https://api.switchgate.ai/v1/marketplace/listings \
-H "Authorization: Bearer sg-admin-…" \
-d '{
"model": "private/acme/legal-qwen-72b",
"public_id": "acme/legal-qwen-72b",
"price_per_m_input_usd": 0.40,
"price_per_m_output_usd": 1.60,
"max_buyer_concurrency": 16,
"regions": ["ca", "eu"]
}'- Buyers will never see your infrastructure. They'll call the public model ID; the gateway routes through your tunnel and returns the response.
- Your weights will never move. Only tokens travel; the model stays on your server, under your control.
- Same governance for buyers. Their budgets, locks and user keys will apply to your model like any other, which makes it easier to buy.
FAQ
Enterprise questions, answered
Will we be able to sell access to our model through SwitchGate?
Do prompts sent to our private models ever leave our network?
Which runtimes does the SwitchGate Model Tunnel support?
Can we mix private and cloud models in one workflow?
What does enterprise pricing look like?
Bring your model. Keep your perimeter.
Tell us about your runtime and your compliance constraints; a pilot Agent is usually live in a day.