Conscious this short article will very quickly age given the velocity of AI product innovation. However, a quick side by side of vendors shows where product offerings sit against a target set of consumption layer features. Most of what it needs is already on sale, though very little of it comes from one supplier. Have also referenced some internal solutions.

Kong, Portkey, LiteLLM, Cloudflare and TrueFoundry sell the plumbing: a single endpoint across many providers, with quotas, caching, fallbacks and per-key spend tracking. Databricks positions its Unity AI Gateway explicitly as a governance and cost-control point for agents, with dollar-level cost attribution and governance over which agents may reach which external systems.[13] All of it meters tokens and not pricing work.

Model choice has its own market. Beyond the Bedrock and Foundry routers mentioned in main Article 6, Ramp (a US fintech — corporate cards and spend management) has released to the market the router it runs internally, which it says handles more than 2.75 trillion tokens a month and has cut Ramp's own AI costs by 30 per cent.[14] Cursor (aka Anysphere, now SpaceX, anAI-native development tool company), reports frontier-quality coding at roughly 60 per cent lower cost, and publishes its unit economics as cost per commit: $6.76 in its routed intelligence mode against $12.69 for a premium model used throughout.[15]. Note caution, figures are vendor-reported. The unit is the more interesting disclosure. Two companies no longer counting tokens but counting completed work.

Vantage, Amberflo, CloudZero and Finout will now attribute AI spend by model, team, customer and feature.[16] Feature sits closer to the business than token does, and still some short of outcome.

The most complete example so far was built rather than bought. Walmart's Element platform routes across models on cost and performance, serves around three million queries a day, and allows new models to be tested without architectural change. Its leadership describes AI as a dynamic optimisation problem rather than a fixed cost.[17]

Set the categories against what the layer actually has to do and the pattern is hard to miss. Every column has strengths. No column is the layer.

Consumption-layer capabilities, benchmarked against what is on the market

● provided  ◐ partial, or available only at a coarser level  ○ not offered  n/d not disclosed. Assessment based on published vendor documentation and company statements as at August 2026; capabilities in this market change quickly.

Nobody yet sells a single place that governs models, data, tools and human review together, budgets an agent by task rather than by application, treats sovereignty as a routing decision rather than a deployment choice made once, and reports what a finished outcome cost. An enterprise that wants that today must assemble it from three markets or build it!

Additional sources

[13] Databricks, Expanding agent governance with Unity AI Gateway. https://www.databricks.com/blog/ai-gateway-governance-layer-agentic-ai

[14] Ramp, Ramp Router. https://router.com/ (accessed 21 August 2026)

[15] Cursor, Introducing Cursor Router, 22 July 2026. https://cursor.com/blog/router

[16] Vantage, AI Cost Observability: Measuring and Justifying Token Spend in 2026. https://www.vantage.sh/blog/finops-for-ai-token-costs

[17] VentureBeat, How Walmart built an AI platform that makes it beholden to no one. https://venturebeat.com/ai/walmart-ai-foundry-ships-first-apps-3m-daily-queries-67-faster-planning