OpenAI Realtime API WebRTC vs FlashInfer 2.0

“The primitive here is clean: WebRTC peer connection directly to OpenAI's edge, which means browser-native ICE negotiation handles NAT traversal and the audio path skips your server entirely. The DX bet they made — offload transport complexity to the browser's WebRTC stack instead of making developers manage WebSocket keepalives and audio buffering — is exactly the right call. First 10 minutes is legitimately just grabbing a session token from your backend and calling the peer connection API; the VAD controls mean you're not building your own endpointing logic either. The specific technical decision that earns the ship: billing at the token level on audio streams instead of per-minute flat rates means you're not getting charged for silence, which is the kind of thing that matters the moment you build anything with real pauses in conversation.”

“The primitive here is a CUDA kernel library for attention computation and KV-cache management — not a platform, not a wrapper, an actual low-level building block you can drop into vLLM or SGLang. The DX bet is correctness and composability over abstraction: they expose the knobs (speculative decoding thresholds, LoRA batching configs) without hiding them behind a config YAML that pretends the complexity doesn't exist. The moment of truth is swapping in the FlashInfer attention backend in an existing serving stack, and from what the repo shows, that's genuinely a few lines. The 40% latency claim needs a methodology cite — they show specific token generation benchmarks on H100s with prefill/decode separation, which is at least a real number attached to a real setup, not a vibe. This is infrastructure that a competent team could not replicate in a weekend; the CUDA work is deep and the speculative decoding integration is non-trivial. Ships because the craft is demonstrably in the kernels, not the landing page.”

“Direct competitors here are Deepgram + ElevenLabs in a pipeline, Hume AI's empathic voice interface, and Groq's low-latency audio stack — all of which require more integration work. The scenario where this breaks is multi-tenant applications where you need per-user audio isolation and compliance logging: WebRTC direct-to-OpenAI means your audio never touches your server, which is a privacy feature until your enterprise customer asks for a SOC2 audit trail of every utterance and you realize you've built yourself into a corner. What kills this in 12 months isn't competition — it's OpenAI's own pricing volatility; audio token costs have moved twice in 18 months and any product with margins built around current rates is one pricing page update away from a rebuild. That said, for the majority of voice-in-browser use cases, nothing ships faster right now, so ship.”

“Category is LLM inference optimization, direct competitors are FlashAttention-3, vLLM's built-in attention kernels, and NVIDIA's TensorRT-LLM — none of which are sleeping. The 40% latency claim is real in a narrow regime: it applies to specific decode-heavy workloads on Hopper-generation GPUs with prefill-decode disaggregation; swap in an A100 cluster doing long-context prefill and the number shrinks. What kills this in 12 months is not a competitor — it's NVIDIA shipping optimized kernels directly into cuDNN or the next-generation attention primitives landing in TensorRT-LLM, at which point the delta collapses. What earns the ship anyway: multi-LoRA batching at scale is a genuinely underserved problem that the big players haven't prioritized, and Together AI has production traffic to validate these claims against real workloads, not synthetic benchmarks. The open-source release is credible signal that they're playing for ecosystem, not just headlines.”

“The thesis this bets on: within 2 years, voice becomes the default interface for a class of ambient computing applications — in-browser, in-app, on device — and the architectural bottleneck isn't model quality but transport latency and server cost. Removing the relay tier collapses infrastructure costs by ~30-40% for high-volume voice apps and enables deployment in contexts where standing up a relay server is a blocker (edge deployments, client-side-only apps, WebAssembly contexts). The second-order effect that matters: this shifts power from infrastructure middleware vendors who built businesses on being the relay layer — companies like Daily.co and LiveKit as voice-AI relay brokers — to application developers who can now go direct. The trend line is WebRTC adoption in AI interfaces, and OpenAI is on-time, not early; Twilio and others have been here for calls, but nobody owned the AI voice path specifically. The future state where this is infrastructure: every SaaS product has a voice command surface that costs pennies per session to run.”

“The thesis here is specific and falsifiable: inference compute will remain the dominant cost in LLM deployment for at least the next three years, and kernel-level optimization will continue to yield meaningful gains even as hardware scales. What has to go right is that the prefill-decode disaggregation architecture becomes the dominant serving pattern — if monolithic batching stays standard, FlashInfer's architectural assumptions become a liability rather than an asset. The second-order effect that matters most isn't latency reduction for Together AI's own platform — it's that cheap, reliable multi-LoRA serving changes the economics of fine-tuning. If you can serve 50 LoRA adapters off one base model at acceptable latency, the cost of domain-specific fine-tuning drops by an order of magnitude, which shifts power toward the fine-tuning layer and away from base model providers. FlashInfer is riding the prefill-decode disaggregation trend, and it's on-time rather than early — vLLM and SGLang have already moved this direction, which means the ecosystem is ready to absorb this rather than resist it.”

“The buyer is any developer building voice-first applications, but the budget question is complicated: audio token costs at scale are brutal, and there's no pricing tier that rewards high-volume committed usage the way AWS Reserved Instances do. The moat analysis is the core problem — this is OpenAI's own API, which means the 'product' for any startup building on top of it has exactly zero defensibility against OpenAI shipping a higher-level voice product that obsoletes your integration entirely; the relay-less architecture actually makes that MORE likely because OpenAI now owns the full audio session and can see every interaction. What happens when a platform player ships 80% of this for free? It already happened — this IS the platform player, and anything you build on it is a feature, not a business. I'd ship if you're using this as infrastructure inside a product with a different moat, but as a standalone voice-AI product, you're building on a foundation that can be pulled at any time.”

“The buyer here is infrastructure engineers at companies running self-hosted LLM inference at scale — a real buyer with a real budget (GPU compute costs), not a vague enterprise persona. The open-source release is a distribution play, not a charity: Together AI captures value through their managed inference platform, where FlashInfer improvements directly reduce their per-token compute cost and become a credible differentiator in a market where Fireworks, Groq, and Anyscale compete on latency benchmarks. The moat question is the hard one — open-sourcing the kernel library means competitors can adopt it too, so the defensibility is execution velocity and production integration depth, not the code itself. What happens when NVIDIA ships this natively is the real stress test, and the honest answer is that Together AI's moat shifts entirely to their managed platform and the workflow integrations built on top of it. Still a ship because the business logic is coherent: they're using open source to build pipeline credibility while monetizing on the managed layer, which is a proven playbook.”