VC & Startups/Nov 21 edition/Reporting & analysis
FundingValuationsM&AEcosystems

FundingRounds, leads & terms

Sarvam launches Token Factory, a rupee-billed pay-per-token inference service on Indian GPUs

Sarvam, valued at $1.5 billion in June, has launched a rupee-billed, pay-per-token inference service on Indian GPUs. Most of its hosted models are Chinese open-weight systems, reviving questions about sovereign AI and whether the business can make money against local rivals.

From the Inc42 article of November 21, 2026 and cited sources: Sarvam launched Token Factory on November 16, 2026. [5] [2] [6] [1]

SOURCES[5] Sarvam: Introducing Sarvam Token Factory sarvam.ai[2] Unite.AI: Sarvam Opens India-Hosted Token Factory unite.ai[6] AlphaSignal: Sarvam AI Launches Token Factory alphasignal.ai

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
Sarvam launched Token Factory on November 16, 2026. [5] [2] [6] [1]
Original source
Inc42 article, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
When
November 21, 2026 article published
Evidence
8 sources across 6 sites

Reach: not measured yet

Illustration from Inc42 Media: Sarvam launches Token Factory, a rupee-billed pay-per-token inference service on Indian GPUs
Image: Inc42 Media — Original article ↗
THE CORE IDEAS4 TAKEAWAYS
01

Sarvam launched Token Factory on November 16, 2026. It serves Sarvam 105B alongside open-weight models such as DeepSeek, GLM, Kimi and Gemma from GPU clusters in India through an OpenAI-compatible API. Introductory prices run to the end of 2026 and are billed in rupees, starting at ₹10 per million input tokens for Sarvam 105B and ₹0.25 per million for cached input. [5] [2] [6] [1]

02

Sarvam's headline usage figure is for a Chinese model, not its own: more than 450 billion DeepSeek V4.1 Flash tokens a day, with about 96% of prompt tokens served from cache. This figure and Sarvam's speed benchmarks are self-reported. Inc42 estimates the volume brings in roughly ₹13.8 Cr a year from prompt tokens alone. Sarvam has not disclosed how many GPUs it uses or how busy they are. [1] [2] [6]

03

The launch comes after a $234M Series B tranche led by HCLTech at a $1.5B post-money valuation. Sarvam's subsidised IndiaAI Mission allocation of H100 GPUs ended in March. Sarvam says Token Factory runs on GPUs it bought privately. Critics point out that six of the eight hosted models are Chinese, which they say sits awkwardly with Sarvam's role in India's sovereign AI push. [3] [4] [1] [7]

04

Competition is crowding in. E2E Networks offers its Blaze service and also runs Jarvislabs, whose assets it bought in August 2025. Global providers such as Together AI and Fireworks already sell token-billed inference. Because they all serve the same open-weight models, customers have little reason to stay with one provider. [1] [8]

WHY IT MATTERS

India now has several token-billed inference providers serving the same open-weight models.

Read the full assessment

Implication: with little lock-in, price, latency and cache pricing will likely decide share, and an analyst says Sarvam must show by end-2027 that DeepSeek users stay for its own models.

Executive brief

Sarvam's most widely cited usage figure is for a Chinese model, not its own. The company says its new Token Factory serves more than 450 billion DeepSeek V4.1 Flash tokens a day, and that 95.7% of prompt tokens on that model come from cache (Inc42; Unite.AI). Sarvam, a beneficiary of the IndiaAI Mission, launched the pay-per-token service on November 16, 2026. It is billed in rupees and runs on GPUs in India. The launch has reopened a debate about what "sovereign AI" means. Competitors E2E Networks (Blaze, Jarvislabs), L&T, Utho and Mirror Security are building similar services.

What changed and event timeline

  1. E2E buys Jarvislabs assets

    E2E Networks acquired Jarvislabs' IP, hardware and customers. Jarvislabs now runs its own token factory and says it served close to 50B tokens in November ().

  2. Subsidised GPU access ends

    Sarvam says its IndiaAI Mission allocation of 4,096 H100s ended in March. The allocation was worth about ₹247 Cr, including a ₹98.68 Cr subsidy ().

  3. Sarvam becomes a unicorn

    Sarvam announced $234M of a $300M Series B at a $1.5B post-money valuation, led by HCLTech with $150M. All figures are company-announced and were reported independently (;).

  4. Token Factory launches

    Sarvam's own models and open-weight models are served from Indian GPU clusters through an OpenAI-compatible API (;).

  5. Sovereignty and economics questioned

    Coverage examines why a Mission-backed firm hosts mostly Chinese models, and whether the service makes money or mainly collects data (;).

Capabilities and access

  • Models: Sarvam 105B, GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash 1017, DeepSeek V4 Pro, Kimi K3.1 Pro and Gemma 4 31B (Unite.AI).
  • Introductory pricing, valid to end-2026, excluding GST:
  • Sarvam 105B: ₹10/M input, ₹25/M output.
Read the full section
  • Models: Sarvam 105B, GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash 1017, DeepSeek V4 Pro, Kimi K3.1 Pro and Gemma 4 31B (Unite.AI). AIM and AlphaSignal count eight models; AlphaSignal also lists Kimi K3.
  • Introductory pricing, valid to end-2026, excluding GST:
  • Sarvam 105B: ₹10/M input, ₹25/M output.
  • DeepSeek V4.1 Flash: ₹14/M input, ₹55/M output.
  • Cached input: ₹0.25/M.
  • Enterprise terms: ISO 27001 and SOC 2 Type II certifications, and a 99.95% uptime SLA (Inc42).
  • Data defaults: zero data retention only on request, and user data is used for training unless the user disables it.

Technical analysis for researchers and developers

  • Stack: vLLM, SGLang, NVIDIA Dynamo and Moonshot's Mooncake KV cache. Prefill and decode run on separate GPU pools.
  • Optimisations: MXFP8 quantisation plus speculative decoding gave Sarvam 105B 2.2× throughput. Speculative-decoding acceptance rose from 38% to 84%.
  • Workload: the average GLM 5.3 agent prompt is about 131K tokens, which explains why caching dominates the economics.
Read the full section
  • Stack: vLLM, SGLang, NVIDIA Dynamo and Moonshot's Mooncake KV cache. Prefill and decode run on separate GPU pools. A Mooncake-based tier keeps prompts cached for hours or days (AlphaSignal; Unite.AI).
  • Optimisations: MXFP8 quantisation plus speculative decoding gave Sarvam 105B 2.2× throughput. Speculative-decoding acceptance rose from 38% to 84%. Sarvam also reports fixing a vLLM decode-kernel clamping bug and an indexer-cache race condition.
  • Workload: the average GLM 5.3 agent prompt is about 131K tokens, which explains why caching dominates the economics.
  • Methodology: benchmarks used B200 GPUs and clients in AWS Mumbai, including a 72-hour median test and runs on November 3–5. AlphaSignal notes that results depend on prompt shape, concurrency, caching and output length.

Claims and evidence

All of these figures are self-reported by Sarvam. None of the reviewed sources independently reproduces them.

Read the full section

All of these figures are self-reported by Sarvam. None of the reviewed sources independently reproduces them.

  • Gemma 4 31B: 0.40s time to first token on a 10K-token prompt (5× faster than one rival) and 308 vs 193 tok/s (AlphaSignal).
  • Sarvam 105B: 0.2s time to first token and 330 tok/s for a single user.
  • Kimi K3.1 Pro: 1.9× more tokens per GPU than InferenceX baselines.
  • Where Sarvam trails: on DeepSeek V4.1 Flash, Together AI's first token came in 0.71s against 0.90s for Sarvam Standard (Unite.AI).
  • Volume: 450B tokens a day. Inc42 estimates about ₹13.8 Cr a year from prompt tokens alone at these rates. GPU count and utilisation are undisclosed (Inc42).

Context and prior work

  • Global peers: Together AI, Fireworks, DeepInfra and Nebius already run token-billed inference (Inc42).
  • E2E Blaze: serves Nemotron, Sarvam, DeepSeek, Qwen, GLM and Gemma through one OpenAI-compatible endpoint. Per-token prices are undisclosed (AIM).
  • Earlier adoption criticism: Sarvam-M (May 2025) drew 334 Hugging Face downloads in two days. Investor Deedy Das called that "embarrassing" (AIM).
Read the full section
  • Global peers: Together AI, Fireworks, DeepInfra and Nebius already run token-billed inference (Inc42).
  • E2E Blaze: serves Nemotron, Sarvam, DeepSeek, Qwen, GLM and Gemma through one OpenAI-compatible endpoint. It runs on a 1,024-GPU Blackwell cluster plus 5.4 MW of additional Blackwell capacity, including HGX B300 systems. Per-token prices are undisclosed (AIM).
  • Earlier adoption criticism: Sarvam-M (May 2025) drew 334 Hugging Face downloads in two days. Investor Deedy Das called that "embarrassing" (AIM).

Limitations, safety and contested findings

  • Sovereignty: AIM counts six of the eight models as Chinese (AIM).
  • Sarvam's response: the service runs on privately procured GPUs, not subsidised ones, and there are no licensing, data-sharing or royalty deals with model developers.
  • Commenters' affiliations: Hemant Mohapatra, who defended the launch, is at Lightspeed, a Sarvam investor.
Read the full section
  • Sovereignty: AIM counts six of the eight models as Chinese (AIM).
  • Sarvam's response: the service runs on privately procured GPUs, not subsidised ones, and there are no licensing, data-sharing or royalty deals with model developers. Inc42 says Sarvam did not answer its questionnaire (Inc42).
  • Commenters' affiliations: Hemant Mohapatra, who defended the launch, is at Lightspeed, a Sarvam investor. Atharva Ingle, who claims comparable or better performance, is at JarvisLabs, a competitor (AIM).
  • Privacy: training on user data by default, with zero data retention only on request, is a risk for sensitive workloads.

Business and practitioner implications

  • Little to lock customers in: the same open-weight models are available from E2E and Jarvislabs, so price, latency and cache pricing decide who wins (Inc42).
  • Utilisation decides profit: Sarvam pays for its B200 cluster whether or not it is busy.
  • The local-hosting edge may shrink as OpenAI moves inference to India.
Read the full section
  • Little to lock customers in: the same open-weight models are available from E2E and Jarvislabs, so price, latency and cache pricing decide who wins (Inc42).
  • Utilisation decides profit: Sarvam pays for its B200 cluster whether or not it is busy. Analyst Sanchit Vir Gogia (Greyhound Research) says Sarvam must show by end-2027 that customers who came for DeepSeek stay for Sarvam's own models.
  • The local-hosting edge may shrink as OpenAI moves inference to India.
  • For developers: benchmark on your own prompts, turn off training on your data, and request zero data retention in writing.
FOLLOW THE EVIDENCE

The source trail.

Sources (8)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.