Skip to content

Sarvam’s Token Factory Bets Big on AI Inference—But Can It Monetize Idle GPUs?

·StartupReader editorial desk· 4 min readReviewed by our editors
LAUNCH. Token Factory from Sarvam AI, pay-per-token inference platform. taking on: Jarvislabs, Mirror Security. Bengaluru.

Sarvam AI has introduced Token Factory, a pay-per-token inference platform that lets developers and enterprises use AI models without handling GPUs or deployment. This move places the Bengaluru-based startup at the center of India’s growing competition in AI infrastructure, where other players like L&T, E2E Networks’ Jarvislabs, and Mirror Security are also developing similar token-based services. However, Sarvam’s approach—running on privately owned GPUs and pricing in Indian rupees—raises concerns about long-term profitability, strategic alignment with India’s AI goals, and whether the company can actually turn its hardware into a sustainable business.

The platform, launched November 16, offers access to Sarvam’s proprietary Sarvam 105B model alongside open-weight alternatives like DeepSeek V4.1 Flash, Kimi K3.1 Pro, and Google’s Gemma 4 31B. Pricing starts at ₹10 per million input tokens for Sarvam’s model, with output tokens billed at ₹25 per million—a structure that encourages cached requests (₹0.25 per million cached tokens). According to Inc42, Token Factory is already processing 450 billion tokens daily through DeepSeek alone. Still, Sarvam hasn’t shared details on GPU usage or how much of its revenue comes from full-price billing, leaving key questions about financial health unanswered.

Why this matters: Token factories represent a shift from raw GPU rental to managed inference services, similar to global players like Together AI or DeepInfra. For Sarvam, the platform could help fill gaps between training cycles for its trillion-parameter model while generating income. But the business model depends on one key factor: how efficiently the GPUs are used. Sanchit Vir Gogia, chief analyst at Greyhound Research, told Inc42 that Sarvam’s leadership will need to demonstrate by 2027 whether customers who come for DeepSeek stay for Sarvam’s own models. As he put it, “It’s either a revenue play or a data play”—meaning the company must prove it’s making money, not just collecting usage data while covering GPU costs.

The launch also sparks discussions about India’s AI strategy. Sarvam was part of the IndiaAI Mission, which provided ₹247 crore in GPU subsidies (including ₹98.68 crore in government funding). By hosting open-weight models, including Chinese-developed DeepSeek, Sarvam avoids licensing issues but risks alienating stakeholders pushing for more self-reliant AI solutions.

Context from StartupReader’s coverage: Sarvam’s focus on inference-as-a-service follows its 2026 dominance in Indic AI funding (per Tracxn) and partnerships with India’s Ministry of Defence, where its Prism AI platform is being tested for decision-making automation. Yet Token Factory’s reliance on global models—rather than homegrown alternatives—may clash with India’s push for indigenous AI, especially in sectors like defense where sovereignty is a priority. As Anthropic’s potential IPO (covered by StartupReader on November 17) shows, the market is scrutinizing how AI companies balance innovation with commercial viability. Sarvam’s experiment with inference monetization could serve as a test case for whether Indian infrastructure players can profit from AI’s growth—or if they’re just optimizing for scale at the expense of sustainability.

What to watch next: Financial transparency: Sarvam hasn’t disclosed how many GPUs it operates or how much revenue comes from full pricing. The shift from introductory rates (set until late 2026) to market prices will show whether Token Factory is a money-maker or a loss leader. Customer retention: DeepSeek’s 450B daily tokens are mostly cached (95.7%), meaning low margins. Sarvam must convince users to switch to its proprietary models to justify the cost of keeping GPUs running. Strategic trade-offs: By avoiding licensing deals with model developers, Sarvam minimizes risk—but its use of open-weight models (including non-Indian ones) could limit its appeal in government or defense contracts, where local AI development is favored. Competitive pressure: L&T, Jarvislabs, and Mirror Security are also building token-based inference platforms. Sarvam’s advantage lies in its B200 GPU cluster and early entry, but long-term success will depend on whether it can stand out beyond pricing.

The launch of Token Factory highlights a broader shift in AI infrastructure: the focus is no longer just on hardware, but on who controls the delivery of AI services—and whether they can charge for it. For Sarvam, the coming months will reveal whether its factory runs on profit or just keeps the lights on.

Sources: inc42.com

“Sarvam’s Token Factory launch shows how AI inference services are evolving—but the real question is whether providers can turn GPU capacity into lasting revenue or if they’re just building pipelines for data collection.”
— StartupReader
ShareLinkedInXWhatsApp

Read the original reporting

The outlets below did the original reporting.

Related briefs

This brief was drafted automatically from the sources above and published under our editorial policy. Spotted an error? Tell us.