How Sipfront runs 50,000 concurrent test calls on custom bare metal without tripping abuse alarms

A typical Sipfront test run pours simulated calls out of a handful of servers toward a single destination: a flood of very small packets from three or four source IPs, converging on one target. That is an orchestrated-attack signature, and automated security systems treat it as one. Sipfront's previous providers couldn’t deal with it. One flagged legitimate load tests repeatedly, with alerts going upstream to other providers in the stack, where they damaged reputations at multiple levels at once. Another off-boarded the company for its load. Sipfront realized they weren’t looking for the cheapest bare metal provider, but a partner who would take the time to understand their workloads, recognize that their network traffic shouldn’t trigger security alerts, and tailor a cost-effective solution to fit.

The workload

Telephony traffic flips most hosting assumptions on their heads. The packets are tiny, the volumes are enormous, and per-packet processing rewards raw clock speed over data-center pedigree.

"Typically you might get a Dell or an IBM server with Intel Xeon processors, which is usually a good thing if you run databases or web servers," Granig says. "Not so much if you need to handle and process a lot of real-time traffic. There you want consumer-grade AMD Ryzen CPUs with five gigahertz of clock rate, because you can then generate and absorb a lot of packets."

Availability requirements also deviate from norms. Test agents are ephemeral by design: workloads spin up, run for days, and are torn down, and if a node drops, the orchestration layer reallocates to another server. Sipfront deliberately trades five-nines hardware for clock speed, and the architecture absorbs the difference. That trade is only safe because the fleet is custom-built for it, which is the difference between a provider that stocks standard hardware configurations and one that builds it to order.

The financial piece of their puzzle is similarly unique. Egress pricing is what pushed Sipfront off a pure hyperscaler footprint in the first place. "These hyperscalers, they are not cheap," Granig says. "The way they charge for network traffic, when you do call simulations, 50,000 concurrent calls, point to point, that's going to be a lot of traffic and it's going to be expensive."

He was, however, willing to pay a premium for bare-metal servers from Servers.com because they could provision systems in global regions, including EMEA, US and Southeast Asia, and because they were able to configure the network and security components to treat the traffic as a specification instead of a violation.

"What looks like a denial-of-service attack to 90 or 95% of the typical hosting customers is actually legitimate traffic for us. It's real voice traffic."

Andreas Granig

CEO and Co-Founder of Sipfront

The architecture

Sipfront runs a deliberate hybrid infrastructure. The control plane remains on AWS, managed through ECS Anywhere, so orchestration, APIs, and portals work exactly as they always have. The execution infrastructure is Servers.com bare metal, where every traffic-generating and traffic-ingesting workload lives. "You can still manage all of your workload through their APIs, their portals, but you actually plug your own infrastructure in there," Granig says. "That is a super efficient way to have the cake and eat it too." This bare-metal-behind-a-cloud-control-plane pattern allows infrastructure repatriation to be scoped to the workloads that need it, with the cloud operating model intact and the data layer untouched.

The results

The future

The arrival of voice AI added a new dimension to Sipfront's call quality tests. Before deploying AI-driven voice agents, Sipfront’s customers used more primitive systems that detected callers’ prompts and selected from a predetermined set of actions and responses. In that paradigm, Sipfront could measure quality by comparing a call to a reference recording. Their customers’ new voice AI systems respond unpredictably by design, and the most effective way to test them at scale is to grade recorded test calls with AI models. However, off-the-shelf LLMs aren’t suited for the task, so Sipfront needs to use specialized models that are trained on hundreds of thousands of minutes of recorded calls. And GPU-equipped Servers.com bare metal allows them to do it. "Running your own AI models is technically not so complex," Granig says. "You just need to have the right hardware."

Sipfront must also test their customers’ AI-driven systems for compliance with new regulations. Provisions of the EU AI Act that took effect in August, 2026 require Sipfront’s customers to disclose when an EU-based caller is talking to an AI. The Act also includes provisions for capturing consent and accommodating callers’ rights to human handoff and dialect fairness. These are all things Sipfront tests for today. As those workloads grow, the GPU footprint grows with them, on the same bare metal partnership that carries the call traffic.

"We go to Servers.com and say, look, we need these kind of servers with these kind of GPUs, can you provide them in these regions?" Granig says. "The support is always super fantastic. We get hardware we need at a price that is a no-brainer."

For the business story behind this architecture, read the Sipfront case study. To talk through a workload that other providers flag instead of fit, contact the team.

Thanks to Andreas Granig for sharing the Sipfront x servers.com story.