Thoughts on the new 'bring your own model' option? Pricey gateway.

I've been evaluating the recently announced 'bring your own model' (BYOM) option from several major observability and APM vendors over the last fortnight. The premise is undoubtedly compelling for teams with significant existing investments in proprietary or fine-tuned machine learning models for anomaly detection, log pattern classification, or predictive alerting. However, after a detailed review of the technical implementation and, crucially, the pricing schema, I am arriving at the conclusion that the current gateway fee structure creates a significant barrier to entry that may negate the purported flexibility for all but the largest enterprises.

The core value proposition is clear: escape vendor lock-in for inference and leverage your specialized models within a managed data pipeline. In practice, this involves their platform providing a standardized gateway interface—often a gRPC or HTTP endpoint—where you host your model container, and they handle the data ingestion, feature extraction, and subsequent routing of the model's output back into their dashboards and alerting rules. The technical execution is generally sound. The primary cost, however, is not in the model execution itself but in the 'gateway' fee, which is typically a per-hour charge regardless of inference volume, coupled with a data processing fee for the traffic routed to and from your model instance.

Let's break down the cost components as I've observed them across three providers:

* **Base Gateway Fee:** This is a fixed hourly or daily rate to simply have the gateway provisioned and available. It covers the management overhead and integration points. In one case, this equates to nearly $500 per month before a single inference is processed.
* **Data Processing Unit (DPU) Fees:** After the base fee, you are charged for the volume of data shuttled to your model. This is calculated in DPUs, which are aggregates of compute and network. This cost is non-trivial when dealing with high-cardinality metric streams or raw log events.
* **Platform Consumption Costs:** The outputs from your model are then treated as new data sources within the platform, consuming your existing allotment for custom metrics, events, or logs. This effectively double-dips, as you pay to process the data into your model and then pay again to store and alert on its results.

When I modeled this for a hypothetical scenario involving a custom anomaly detection model on ~500 critical business metrics, the BYOM gateway solution was approximately 3.2 times more expensive over a 12-month period than using the vendor's native, albeit less specialized, machine learning features. The break-even point where the custom model's precision would have to drive enough incident reduction to justify the overhead was extraordinarily high. This pricing model feels less like enabling customization and more like monetizing a niche enterprise request.

My question to the community is this: have others performed similar cost-benefit analyses, and what were your findings? Are there specific use cases—perhaps in highly regulated industries with non-negotiable model requirements—where the gateway cost is genuinely justifiable despite the premium? Furthermore, has anyone successfully negotiated alternative pricing structures, such as a cap on the base gateway fee after a certain DPU volume is reached? The feature is technologically impressive, but its economic viability for mainstream SRE or platform engineering teams seems questionable at present.