Oracle and AWS shipped LiteLLM gateway blueprints on Nov. 12, with budgets that don't enforce by default

Oracle▼−2.86% — Oracle, down 2.86 percent today and AWS▼−1.16% — AWS, down 1.16 percent today both published production blueprints on November 12 that put LiteLLM's open-source AI gateway between enterprise apps and the models behind them ¹². Read the fine print, and the layer they are handing their customers does not yet meter the money. AWS's own guide↗ warns that with Bedrock inference-profile ARNs, LiteLLM records spend as $0, so per-key budgets "NEVER enforce" until an operator sets per-token prices by hand ². The self-service key portal that pattern needs is a paid Enterprise feature; the open-source gateway is free for up to five users ².

Oracle's reference↗ ships a router trained on 3,590 graded answers across 20 OCI-hosted models in eight routes, but when every answer was labeled three times, the same cheapest passing route was chosen for only 66% of requests ¹.

Routing saves money, and that deflates vendor revenue

The router's economics run both ways. It sends a simple question to a model that costs about a thousandth of a dollar, while a four-turn code fix lands on a coding model that costs about $0.07, some twenty times as much ¹. That is routing working as designed, and it is revenue deflation for model vendors if enterprises adopt it. The chart below shows the same trade-off on Oracle's held-out set: routing policies that always buy quality pay several times what the trained router pays, and the router stays close to the cheap end ¹.

OpenAI moved the meter inside its own subscription

OpenAIPVT$1.18T — OpenAI, private, latest valuation $1.18T answered on November 10 by becoming its own meter: ChatGPT Business allowances now pay for AI inside 11 outside apps through Sign in with ChatGPT, with LovablePVT~$1.8B — Lovable, private, latest valuation ~$1.8B to follow; Enterprise and Edu are not eligible ³. Metering inside the subscription, outside the gateway, is the vendors' hedge against exactly this deflation.

AnthropicPVT$350B — Anthropic, private, latest valuation $350B prices its IPO on November 18, a $70 to $78 range implying about a $1.3 trillion valuation ⁷; what an agent task is worth after routing is part of the revenue argument that price will be argued on. Mercor'sPVT~$2B — Mercor, private, latest valuation ~$2B new $20 billion round funds evaluation and post-training, the business of measuring whether models complete tasks, and the startup was itself bitten by LiteLLM in April's hack ⁶. The market is already pricing the measurement layer.

The free layer carries a breach history

The layer both clouds handed out has a record we covered here in October, when Check Point▲+0.01% — Check Point, up 0.01 percent today found LiteLLM let any signed-in user seize admin (that piece, and our benchmark/OpenRouter piece from November 12).

What is new since: an admin key exposed in May sat unrevoked for over four months before an attacker used it to hijack the release pipeline, and the malicious version 1.107.0 shipped on PyPI and official Docker tags↗ for about two hours on October 21 ⁴⁵. March's earlier compromise produced 118,829 runner dumps from 2,488 organizations in a 153GB archive↗, many with live credentials nobody can attribute ⁴. The consequence for adopters: revocation, not deletion, is what closes a leaked key, and per-token prices set explicitly are what make budgets enforce at all.

The next contract decides whose meter counts

The chain ends at whose money moves. Oracle and Amazon win either way: they sell the capacity the gateway routes, and the routing saves the customer money ¹². The contested link is the meter. Enterprises need capped, attributed spend; vendors need usage they meter themselves. The concrete test: whether the next enterprise inference contract prices per routed model call, inside a gateway the customer controls, or rides a subscription allowance the vendor meters. The first deflates revenue per agent task; the second protects it. Anthropic's IPO price, set on November 18, will be argued partly on that question ⁷.