Carthage
We are building the control plane for enterprise AI, enabling them to (1) reduce AI costs, (2) continuously find pre-qualified best models and execution paths for their specific workloads, (3) switch automatically as performance, price, or availability changes, and (4) maintain reliable alternatives so they are never dependent on a single model or provider.
Mission
Our goal is to make enterprise AI cheaper and perfectly reliable.
Companies already increasingly depend on AI for real workloads, but choosing which model runs those workloads is still driven, largely, by general benchmarks, reputation, and task-level routing. The result is concentrated usage: Menlo Ventures, for example, estimates that Anthropic, OpenAI, and Google account for 88% of enterprise LLM API usage. Our math says this is bad.
Carthage is building infra that learns what actually works for each institution.
We evaluate execution paths against the institution’s own workloads: the model, prompts, retrieval, tools, context, guardrails, verification, latency, reliability, and cost. Continuously qualify alternatives and move production toward whichever configuration delivers the required outcome best.
This means a model does not have to be the best general-purpose model in the world to win a workload. A smaller model, open model, local model, or specialist model can win if, inside the right harness, it is better for that institution’s particular job.
We believe this creates a better form of model competition.
Most routing systems optimize among models on already won distribution. Carthage makes that competition continuous. Models earn workloads based on measured usefulness to the institution, rather than provider reputation or benchmark position alone.
We believe this is also one of the strongest approaches to AI safety and alignment: democratize access to models, make them substitutable, and allocate usage meritocratically and provider-agnostically.
That changes the incentive structure. If a provider raises prices, degrades reliability, changes behaviour, or introduces data terms an institution will not accept, the institution can move the workload elsewhere. For example, Anthropic currently requires 30-day retention for certain covered models even for customers that would otherwise use zero-data-retention arrangements. In a genuinely substitutable system, the institution can make its own privacy requirements part of qualification and simply choose another execution path.
Providers therefore compete to align themselves with users, rather than users having to continually adapt themselves to providers.
The consequences extend beyond individual companies.
As AI becomes infrastructure, economies that consume intelligence primarily through a handful of foreign providers become exposed to their prices, capabilities, policies, outages, and strategic decisions. These are First-Meter problems: early technological dependencies that can compound into much larger economic dependencies as AI becomes embedded across production.
Carthage gives institutions the opposite trajectory. Everyday AI usage continuously creates tested alternatives: cheaper models, specialist models, local models, open models, different providers, and eventually models developed closer to the institutions and countries using them.
If that works at scale, the current concentration of AI expenditure can fall substantially. Competition moves from “which frontier lab has the strongest general model?” toward “which system can deliver this outcome, for this institution, under its constraints, at the best cost and reliability?”
We believe that is how AI gets both cheaper and more aligned with the people using it.
Not by choosing one model or one company that everyone must trust, but by making models compete continuously for the right to be used.
A race to the top, enforced by the ability to switch.
