Twilio has been the default choice for real-time communication APIs for over a decade. But default choices deserve scrutiny, especially when the performance gap is measured in orders of magnitude.
We built V100 from scratch in Rust — 20 microservices, each a single binary, deployed on ECS Fargate across three AZs. Then we benchmarked everything. This post presents the numbers side by side. Where Twilio publishes data, we compare directly. Where they do not, we say so.
| Metric | V100 (Benchmarked) | Twilio (Public Data) |
|---|---|---|
| Broadcast pipeline latency | 3.0 µs per tick | Not publicly disclosed |
| Sustained throughput (30s) | 7.4M ops/sec | Not publicly disclosed |
| STUN binding parse | 50 ns | Not publicly disclosed |
| TURN credential validate | 272 ns | Not publicly disclosed |
| TURN credential generate | 525 ns | Not publicly disclosed |
| Live API p50 latency | 54 ms | ~50 ms (typical, per docs) |
| Primary language | Rust | Java / Erlang |
| Signaling binary size | 2 MB | 5–15 MB (SDK) |
| Architecture | 20 Rust microservices, ECS Fargate, 3 AZs | Multi-region AWS (monolithic-ish) |
| Concurrency model | DashMap lock-free | JVM threads / Erlang processes |
| Pipeline headroom over 20Hz director | 16,600x | Not publicly disclosed |
| Video pricing | Contact us | $0.004/min/participant |
V100's broadcast pipeline processes each tick in 3.0 µs — the full decision loop of reading participant state, evaluating layout rules, and emitting directives. At 20Hz director frequency (one decision every 50ms), that is 16,600x headroom before the pipeline becomes the bottleneck.
Twilio does not publish internal pipeline latency. At the API layer, both converge around 50ms, but API latency includes TLS and network hops. The internal pipeline is where architectural choices compound.
V100 parses a STUN binding request in 50 ns, validates a TURN credential in 272 ns, and generates one in 525 ns. These are Criterion.rs measurements on production-identical binaries, not estimates.
Twilio's Network Traversal Service handles STUN/TURN via their global edge network. Their internal credential processing times are not publicly available.
Over a sustained 30-second test, V100's signaling layer held 7.4M ops/sec. This is not a burst metric — the window captures steady-state behavior including cache warming, allocator stabilization, and OS scheduler effects.
Why this matters: Sustained throughput determines how many concurrent rooms and state transitions a single node handles before you scale horizontally. Higher throughput per node means fewer nodes, lower cost, simpler operations.
V100's signaling service compiles to a 2 MB binary with DashMap lock-free concurrency and zero GC. Twilio's SDKs range from 5–15 MB; their servers run on the JVM with hundreds of megabytes baseline memory.
On Fargate, a 2MB Rust binary cold-starts in under a second. JVM services need 5–15 seconds to reach steady-state after JIT warm-up — a real difference when auto-scaling absorbs traffic spikes.
V100 does not match Twilio in every dimension. Twilio has advantages that matter for many teams:
Ecosystem maturity. Twilio has shipped since 2008. SDKs span every major language, documentation is best-in-class, and "drop in and go" integration is hard to beat.
Global edge network. Twilio runs TURN/STUN across dozens of PoPs worldwide. V100 currently targets three AWS AZs. For guaranteed sub-100ms connectivity in Southeast Asia or Africa, Twilio's footprint is broader today.
Phone and SMS. Twilio's core is telephony. PSTN dial-in, SMS verification, SIP trunking — one vendor. V100 is purpose-built for video and real-time media.
Compliance breadth. Twilio holds SOC 2, HIPAA, GDPR, and PCI certifications across their full product line. V100 is architected for compliance, but accumulating third-party certifications takes time.
Microbenchmarks (STUN, TURN, broadcast pipeline): Criterion.rs, 100+ iterations with statistical outlier detection, production binaries compiled with --release and target-cpu=native.
Sustained throughput (7.4M ops/sec): Continuous 30-second window on the signaling service running on ECS Fargate. Operations include room state reads, participant joins/leaves, and layout recalculations through the full pipeline.
Live API p50 (54ms): External clients hitting production, including TLS 1.3 handshake and network transit — an apples-to-apples comparison with Twilio's published ~50ms typical.
Twilio numbers: Sourced from public documentation, blog posts, and pricing pages as of March 2026. Where Twilio does not disclose a metric, we say so rather than estimate.
V100 is not for everyone. Quick prototypes and PSTN integration still favor Twilio's ecosystem. But if video performance is a competitive differentiator — telemedicine latency requirements, large-scale live events, real-time collaboration — these numbers represent a fundamentally different tier: microsecond internal latency, millions of ops/sec sustained, and a deployment footprint in megabytes instead of gigabytes.
Request access to V100's benchmark suite or schedule a technical deep-dive with our engineering team.
Get Started