When EU MSPs should route AI traffic open vs. closed

EU managed service providers routing AI workloads for customers face a recurring decision: when should bulk or low-stakes traffic go through an open, high-volume gateway, and when does a task warrant a closed, premium API from a specific vendor? The answer is a matter of posture and risk tolerance, not a ranking of vendors.

Open, high-volume traffic

Drafts, summarization, and internal tooling — work where a mistake is recoverable and easily corrected — are good candidates for an open or mixed gateway that an MSP can observe and rate-limit per customer tenant. The value here is visibility: a gateway that logs usage and lets an MSP set caps per tenant makes it easier to catch runaway costs or misuse before they become a customer-facing problem.

Closed, high-stakes traffic

Customer-facing decisions, regulated data paths, and evaluations an MSP already trusts from a specific vendor call for tighter control: explicit allowlists, a named model version, and a vendor relationship the MSP can hold accountable if something goes wrong. Here the priority shifts from cost and flexibility to predictability.

One thing to avoid

Do not infer a market-wide adoption percentage from a single gateway's own traffic sample — a vendor's dashboard shows its own customer mix, not the industry. Vercel's own AI Gateway Production Index is a useful illustration of why: it reported open-weight models reaching 56% of that gateway's token volume in August 2026 while accounting for only 14% of spend, a split driven by how Vercel's specific customer base happens to route cheap, high-volume work. That is a real pattern worth knowing about, but it describes one gateway's traffic, not the AI market as a whole. Any open-vs-closed split an MSP adopts should be justified by the risk of the specific workload, not by a borrowed statistic from someone else's dashboard.

Where the tenant boundary lives

Whichever split an MSP chooses, the boundary that actually protects a customer sits at the identity layer, not the routing layer: the control plane decides which route traffic takes, but a separate, tenant-aware authorization layer needs to decide who may invoke which route in the first place. Altovar positions Lodestar as the EU-based control plane for that routing decision, with Auris handling the tenant boundary underneath it — two different jobs that are easy to conflate into one product claim.

The practical test for any routing policy is whether an MSP can explain, for a specific customer workload, why it landed on the open or the closed side — and whether that explanation would survive a client audit.

Sources: Vercel — AI Gateway Production Index, September 2026.