Sarvam's token factory vs OpenRouter: vertical integration vs aggregation explained
itsOmSarraf_ · x · 2026-11-18
A detailed analysis breaks down why Sarvam's "token factory" is fundamentally different from OpenRouter:
- Vertical integration vs aggregation: OpenRouter owns no models, GPUs, or training pipeline — it's a routing/distribution layer whose assets are unified API schema, fallback handling, and billing normalization. Sarvam owns the whole production chain: pretraining, tokenizer, inference runtime, and silicon.
- Tokenizer economics (the underrated part): vanilla tokenizers are inefficient for Indic scripts — a Devanagari sentence can cost 3–5x more tokens than English. Sarvam trained Indic-specific tokenizers, structurally changing inference cost in a way a pass-through aggregator cannot.
- Co-design across the stack: owning model + tokenizer + serving enables custom FP8/int8 quantization, speculative decoding with own draft models, and continuous batching tuned to Indic code-switched workloads. OpenRouter optimizes which endpoint serves a request; Sarvam optimizes how the token is produced.
- What OpenRouter has: multi-model fallback, per-request cost/latency routing, provider price competition, neutral billing. Sarvam is single-provider — control at the cost of redundancy.
- Margin structure: OpenRouter takes a thin pass-through margin; Sarvam captures the full token margin, which only works if tokenizer + serving efficiency undercuts renting foreign APIs.
TL;DR: OpenRouter is a logistics company for tokens produced elsewhere; Sarvam is trying to be the plant itself. History suggests distribution layers are durable — but they can't create token efficiency, only arbitrage it.
More from Infra
- DeepSeek-V4.1-Flash hits 1,129 output tokens/sec on four AMD MI355X GPUs — Scobleizer · 2026-11-18
- x402 opens four public working groups: domain discovery, identity, tax, card acceptance — kleffew94 · 2026-11-18
- Chutes partners with Alibaba Cloud to offer GLM-5.3 and DeepSeek V4.1 at 35% off — markjeffrey · 2026-11-18
- Bloomberg: New US Chip Rule Freezes $480M Nvidia GPU Loan for ByteDance in Malaysia — rohanpaul_ai · 2026-11-18
- One RTX 4090 runs Qwen3.8-Flash-Next: a hand-built inference engine, fully gated — Distinct-Pie2389 · 2026-11-18
- GPT-6.2 Astra as planner only: same 3D scene for $1.62 instead of $7.40 — GapNew4766 · 2026-11-18