TrustedRouter raises $6m to sell privacy-focused AI model routing

TrustedRouter has raised a $6 million seed round to build an AI model-routing service around a specific promise: keeping customers’ prompts and outputs out of reach of both the routing company and the data centre running the model. That could make switching between models less of a privacy trade-off, though the assurance is the company’s pitch, not an independently tested result.

Watch Desk analysis

What happened

The startup, which launched publicly six weeks ago, told Semafor the round was led by Sierra Ventures. It aims to help developers and businesses route requests between different AI models, a growing market for customers looking to manage costs and avoid dependence on a single provider.

Chief executive Joseph Perla says TrustedRouter uses end-to-end encryption and verifiable confidential computing with some open-weight models. He says those protections mean neither TrustedRouter nor the data centre can read customer prompts or outputs. The company says its approach differs from closed models from OpenAI and Anthropic, which do not currently offer the same guarantees.

Perla also told Semafor that Chinese models, including DeepSeek, Zhipu and Moonshot, are among the most popular on the platform. He described a longer-term ambition to create a marketplace where many models and hosting providers compete to handle requests.

Why it matters

Model routing can give developers more choice over which system handles a task. But sending prompts through another service creates a question that is less glamorous than model benchmarks and rather more important to anyone with sensitive data: who can see what was sent?

TrustedRouter is betting that privacy protections will help it stand out in a crowded field. The trade-off, as Perla presents it, is that customers may consider less advanced open-weight models for private or cheaper workloads while reserving closed models for tasks where they prefer the output. Those are reported claims and examples, not a guarantee that every workload will see the same results or savings.

Our read

This is a useful bet on the plumbing around AI, not just another entrant promising a cleverer model. If TrustedRouter can make its privacy claims verifiable in practice, it could give businesses a more credible way to balance cost, capability and confidentiality. The key word is “if”: security language is easy to put in a pitch deck; evidence is what makes it useful.

What to watch

• Which models and hosting providers support the promised protections.
• How TrustedRouter demonstrates that prompts and outputs remain inaccessible.
• Whether customers adopt routing for sensitive work, not just cheaper experiments.

Discussion spark: Would you route sensitive work through a service promising verifiable privacy if it meant using a less capable model, or keep that work with a closed provider?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.