Former Vercel executive says voice agents will be won on infrastructure, not models

UK voice AI startup SLNG says the voice labs that have dominated the market have been systematically overcharging enterprises — and that the industry is heading for a price correction.

"The market has been shaped by voice labs whose business model depends on maximising compute at every step of every call," says Luke Miller, SLNG's CEO. "Every syllable through the most expensive TTS, every pause analysed by a full LLM call, every transcription through the highest-cost engine. We want to reprice the entire voice agent market."

SLNG, which raised €3.3M in pre-seed funding from Earlybird, StepFunction and a16z scouts in late 2025, is building what Miller calls "the Vercel for voice agents" — an execution layer that sits between a team's existing voice agent orchestrator and the underlying AI models. Teams bring their agent — built on LiveKit, Pipecat, or whichever framework they've chosen — plug into SLNG, and the platform handles model selection, in-region routing, failover and compliance across 11 sovereign regions.

Miller was previously a venture partner at Earlybird VC and the first seller at Vercel, where he built the company's international business.

The execution layer — a new category for voice agents

Miller argues that voice agents are following the same infrastructure pattern he saw at Vercel.

"Before Vercel, deploying a web app at scale meant stitching together CDNs, build pipelines, edge functions and caching layers yourself," he says. "Every team reinvented the same infrastructure. Voice agents are at exactly that inflection point — the models are powerful, the orchestrators work, but there's no platform layer between them and production."

Frameworks like LiveKit and Pipecat have made it possible for almost anyone to build a voice agent. But the prevailing approach — what Miller calls "token-maxing" — is to route every step of every call turn through the most expensive frontier models available. Every transcription, every response, every utterance gets the full treatment. Voice labs' revenue scales with compute consumed, so the incentive is to maximise model calls per conversation, not to optimise them.

The problem, Miller argues, is not just cost. It is outcomes. Repeatability and reliability matter far more at scale than raw model power on any single turn. A voice agent handling thousands of calls a day needs consistency — and frontier models called unnecessarily introduce latency, variability and cost that actively work against that.

SLNG sits in that gap, coordinating speech-to-text, text-to-speech and LLMs in real time, intelligently routing at every step of every call. The platform still calls the best models when they are needed — a complex financial advisory question gets different treatment to a simple appointment confirmation. But the gains, Miller claims, come from knowing when a deterministic response or a lighter model will deliver a better customer outcome than a frontier LLM.

The results are significant. Teams plugging into SLNG are seeing model costs across their voice agents drop by over 50%, latency per call turn cut by more than half, and target outcomes — whether that's appointment bookings, resolution rates or conversion — actually increasing as a result.

For teams already building on orchestrators like LiveKit or Pipecat, the integration is simple: keep your agent, plug into SLNG, and the execution layer handles model routing, failover, compliance and cost optimisation across every call turn.

Global by design, not by expansion

Much of SLNG's growth has been driven by financial services, banking, insurance and healthcare — sectors where data sovereignty is a legal requirement, not a preference.

Miller says the company's approach could not have been built in Silicon Valley, where abundant compute encourages teams to throw GPUs at every problem. In the markets where much of the world's enterprise demand sits — Southeast Asia, the Middle East, Latin America, India — GPU supply is limited and Northern Virginia's cost structures don't apply.

That constraint forced a different discipline. Co-founder and CPO Ismael Ordaz says the team developed approaches using CPU and memory to handle workloads that competitors route through expensive GPU accelerators. "When you don't have abundant GPUs to fall back on, you have to be creative," says Ordaz. "That discipline now runs through everything we build."

"A voice agent handling a mortgage application in Australia can't have its audio processed in Virginia," says Miller. "A patient triage system in Switzerland can't send recordings to a US-hosted model. These aren't edge cases — this is where the enterprise demand is."

One early example is Ixigo, India's second-largest online travel agent, which has shifted a significant portion of its customer support to SLNG. As Ixigo scaled its voice agent operations, the company found itself managing multiple vendor contracts, building in-house voice infrastructure expertise that wasn't a core competency, and absorbing the overhead of coordinating it all.

Since moving to SLNG, Ixigo accesses whichever models it needs through a single platform, has eliminated the vendor management burden, and redirected engineering time toward customer outcomes rather than infrastructure.

That demand has pushed SLNG to build sovereign infrastructure across 11 regions: Australia, Singapore, India, Indonesia, the UAE, UK, EU, Switzerland, Canada, the US and Brazil.

Miller says Europe's regulatory environment, often framed as a burden, is creating a structural advantage for SLNG. While US voice labs optimise for a single market, SLNG is building distributed infrastructure designed for global compliance from day one.

"We're not competing with the US on model size or fundraising headlines," he says. "We're building the infrastructure layer that makes voice agents actually work in production, globally. AWS democratised compute, Stripe democratised payments, Vercel democratised web deployment. The execution layer will do the same for voice agents."

Looking ahead, Miller sees SLNG's role expanding with the emerging agentic economy — where AI agents increasingly build and deploy other agents, and voice becomes the default human interface.

"We're positioning SLNG as the default environment for creating voice agents — whether it's a human building one or an agent spinning one up on the fly," says Miller. "Voice is how humans interact, and the execution layer needs to be ready for agents to create that interface on demand."