Nvidia investing in fewer GPUs
Nvidia, traditionally seen as the world's largest supplier of AI accelerator chips, has invested in an Indian startup whose goal is precisely the opposite of what you'd expect: helping companies run AI models with fewer GPUs. Simplismart, based in Bengaluru, announced this week a US$ 14 million Series B round with direct Nvidia participation, and was valued at approximately US$ 100 million.
The investment looks like a deliberate paradox. The company most associated with the AI chip boom is betting on software that enables reducing the number of GPUs needed to run artificial intelligence applications in production. But, according to Simplismart's founders and the market's own logic, that's not how it works.
The logic of efficient inference
What Simplismart does, in practice, is mediate the relationship between AI models and the computing infrastructure that hosts them. Its software automatically adjusts how models are deployed — considering response speed, number of requests processed, resource consumption, and cost per request — so that companies get the best possible results with the hardware they already have.
"Nobody actually needs the fastest or the cheapest inference," Amritanshu Jain, the company's co-founder and CEO, told TechCrunch. "Companies need speeds and costs that make sense for their specific applications." A customer service chatbot can tolerate slightly longer latency to prioritize cost savings, while a real-time recommendation engine might prioritize speed above all else.
The four-year-old startup claims it has 14x'd its annualized revenue run rate over the past six months, reaching an eight-figure sum (between US$ 10 million and US$ 99 million). The figure reflects recent customer usage, not firm recurring contracts. More than half of revenue already comes from the United States, despite having more customers in India.
Real-world case studies
The numbers Simplismart presents are concrete. The company says it helped the AI video-generation platform InVideo reduce its GPU serving costs by 56%. Another cited case is Exotel, a voice-communications company: by fractionalizing GPU usage instead of allocating entire units, Simplismart reduced the cost per minute by up to 57% compared to closed-model APIs — even below the cost of self-hosting the same open model — and brought response times from 1.1 seconds to under 300 milliseconds.
The customer portfolio includes heavyweight names in the global AI ecosystem: HeyGen, Higgsfield, InVideo, and Ema. In India, cloud provider Yotta uses Simplismart technology to power its Shakti Studio platform's serverless AI inference service. These cases suggest the software is already in real production operation, not just in testing.
The Nvidia collaboration
Nvidia's investment goes beyond the check. The Santa Clara company is working with Simplismart on technical integrations, including around its own Dynamo inference engine. Simplismart engineers and Nvidia solutions architects collaborate directly, and Nvidia has been introducing the startup to prospective customers and partners.
Nvidia's move is part of a broader consolidation strategy across the inference ecosystem. Over the past years, the company has invested not only in hardware but also in software and ecosystems that make its chips more indispensable — from CUDA to TensorRT and, now, Dynamo. Inference is where the money is: while model training concentrates in dozens of companies worldwide, inference runs continuously at global scale on already-deployed hardware.
The Nvidia-GPU paradox
The most discussed aspect of the investment is the apparent paradox: why would the leading GPU supplier want to help companies use fewer GPUs? The answer, according to Jain, is straightforward. "Where we tend to win is where people have to be more stringent on their SLAs and TCO," he told TechCrunch.
Efficiency doesn't mean reduced demand for hardware — it means expanding the number of economically viable AI applications. If a company can run the same AI application with half the hardware, it can double the number of applications or users without doubling costs. In aggregate, that could increase overall demand for Nvidia chips even as each individual application uses fewer resources.
Where the money goes
With the US$ 14 million raised, Simplismart plans to expand its presence in the United States — already its most profitable market — and grow its engineering team. The company had been in talks to raise between US$ 40 million and US$ 50 million at a valuation of US$ 300 to 400 million in prior rounds; the current Series B, at a US$ 100 million valuation, appears to have been a more conservative round, but with more strategically relevant investors.
In addition to Nvidia, the round was led by Dallas Venture Capital, with participation from Accel, Shastra VC, and Micromax. The funding will be directed to developing the inference engine and orchestration platform, plus expanding partnerships with cloud providers that will turn their GPU capacity into managed AI services under their own brands.
While the largest AI companies race toward ever-larger models, Simplismart shows that the next great challenge of artificial intelligence isn't building more capable models, but making them run reliably, cheaply, and at scale — a market that could grow as much as the inference market itself. Whether Nvidia, by funding efficiency, is preparing the ground for the next wave of AI adoption or simply betting that more efficiency will mean more chip sales in the long run remains to be seen.
Sources: TechCrunch, Dealroom, Simplismart
✓ Independent sources cross-checked and verified before publishing
