At Least 3 of the 5 Major Cloud AI Gateways Will Ship Per-Prompt Model Routing as a GA Feature by Q3 2027

The Prediction

By 30 September 2027, at least three of the five major cloud AI gateways — AWS Bedrock, Azure AI Foundry, Google Vertex AI, Cloudflare AI Gateway, and Vercel AI Gateway — will ship per-prompt automatic model routing as a documented, generally available feature. The feature must:

1. Accept a prompt and choose a target model from a configurable pool based on declared capability, latency, or cost preference.
2. Be invoked via a single API call (not require the customer to build their own routing logic).
3. Be marked GA in the provider's official documentation — not beta, not preview, not "labs."

As of May 2026, one of the five meets this bar (Cloudflare AI Gateway, shipped in late April). For the prediction to land, two more must follow within sixteen months.

Why This Will Happen

The economic forcing function is already in place. The 96x price spread between commodity and frontier inference makes routing a measurable operating expense for any team running material AI traffic. Customers are asking for it. Sales engineers are getting requirements docs that mention it by name. Once one cloud ships routing as a documented feature — which Cloudflare has done — the others move within their normal release cadence because their largest customers ask.

The same pattern played out with API gateways (2014-2016), with managed Kubernetes (2016-2018), and with managed object storage (2008-2011). The gap between "first cloud ships it" and "three of five clouds ship it" was roughly twelve to twenty-four months in each case. Sixteen months from May 2026 lands at September 2027 — the midpoint of that historical window.

The five-gateway list itself is partially forecast. AWS, Azure, and Google are stable bets — they will ship something. Cloudflare already shipped. Vercel is included on the strength of their AI SDK 5.0 multi-provider default, which strongly suggests a gateway-level routing offering is on the roadmap.

Why This Might Not Happen

Three failure modes could break the prediction.

Vendor strategy stays bundled. AWS Bedrock and Azure AI Foundry both have a strategic incentive to keep customers inside their own model ecosystems. Routing across providers — or even across all the models within a single provider's stable — disrupts the "land on our preferred model" sales motion. They may ship routing only for internal pools (Bedrock-to-Bedrock), which is a weaker form of the feature and could arguably miss the bar above.

Per-prompt routing collapses into the SDK layer. If Vercel AI SDK and LangChain succeed at making client-side routing the dominant pattern, gateway-level routing becomes redundant for new customers. Established clouds may not bother to ship it because the value already lives in the client.

Customer adoption proves smaller than expected. If the engineering teams currently building in-house routers do not extract enough value to justify a buy decision, the platform demand signal weakens and the GA timelines slip past September 2027 into 2028.

These are real risks, but the trajectory of the past six months suggests the gateway pattern wins. Cloudflare's edge-routing latency advantage is hard to replicate in client code, and enterprise procurement teams prefer gateway features over libraries. My confidence sits at 72% — high enough to publish, low enough to acknowledge the genuine uncertainty.

What Would Falsify This

The prediction is falsified if, on 30 September 2027:

• Two or fewer of the five gateways have GA per-prompt routing matching the three-part bar above; or
• The dominant pattern has clearly shifted to SDK-level routing with no GA gateway equivalents.

Beta or preview features do not count. Cross-region failover routing (which is load balancing) does not count. Per-prompt automatic model selection from a customer-configured pool is the specific bar.

Related Context

This prediction is the platform-level companion to my tutorial on building an in-house cost-aware router. The two trajectories — in-house and platform — coexist for now. The prediction is essentially that the platform side reaches feature parity with the in-house pattern within the next sixteen months.

The economic context — Gemini 3.1 Flash-Lite at $0.25 per million tokens versus Opus 4.7 at $75 per million output — is documented in the inference price floor analysis.