A company whose software is designed to make graphics processors go further has raised money from the largest maker of graphics processors. That is less contradictory than it sounds.
Simplismart, a Bengaluru startup, said on Tuesday that it raised $14 million in a Series B round led by Dallas Venture Capital, with Nvidia participating alongside Accel, Shastra VC and Micromax. The round values the company at about $100 million after the investment. The company sells a software layer that sits between a trained model and the hardware that serves it, adjusting batching, concurrency and memory use so that a given cluster handles more requests.
Its most quotable customer result is a 56% reduction in GPU serving costs at InVideo, an AI video platform. That figure is reported by the customer rather than audited, and it is the kind of number that depends heavily on what the baseline configuration was. A cluster running at 15% utilization has a great deal of room to improve, and most clusters do.
Utilization is the industry’s quiet problem. Accelerators bought during the buildout of the past two years were sized for peak training demand, and inference workloads are bursty in ways that leave expensive silicon idle for long stretches. The software that fixes this is unglamorous: request scheduling, cache management for repeated prompts, splitting a model across devices, routing simple queries to smaller models and hard ones to larger ones. Each technique returns a percentage point or two. Together they decide whether an AI product has a gross margin.
For Nvidia, the investment fits a pattern that has puzzled outsiders for years. The company’s corporate development arm has put money into dozens of businesses whose software reduces the hardware customers need, on the reasoning that cheaper inference expands the market faster than it cannibalizes hardware sales. A customer who can afford to serve a million users instead of a hundred thousand buys more chips over time, not fewer.
There is a defensive motive as well. The layer Simplismart occupies is where cloud providers have been building their own advantage. Amazon’s inference services, Google’s Vertex platform and Microsoft’s Azure AI stack all bundle optimization with compute, and they can price the software at zero to win the workload. An independent vendor has to be better than free, which means being measurably faster on the same hardware and portable across clouds.
The Indian market adds a specific constraint. Graphics processors are imported, taxed and in short supply, and Indian companies building AI products have complained for two years that allocation goes to larger buyers abroad. Software that extracts more from each card is worth more in a market where the card itself is hard to obtain, and the government’s own compute procurement programme has been built around shared capacity rather than dedicated clusters.
Simplismart is not alone in the category. Together AI and Fireworks in the United States, Baseten and Modal, and a cluster of smaller Indian firms offer variations on managed inference with optimization built in. The differences show up in latency guarantees, in which models are supported on day one, and in whether the customer can run the software inside its own virtual private cloud rather than on the vendor’s.
Analysts said the defensibility question is real. Most of the techniques in this layer are published research, and the large model developers keep folding the best of them into their own serving stacks. What is harder to copy is the operational knowledge of a specific customer’s traffic, which is why these companies tend to grow by holding on to accounts rather than by winning them.
The funding market for this layer has cooled from its peak. Through 2024 and into 2025, investors financed a wave of inference startups on the argument that serving costs would become the dominant expense in AI and that whoever owned the serving layer would take a toll on every request. Several of those companies have since repositioned as model providers or as consultancies, and the ones still raising on infrastructure software have had to show paying enterprise customers rather than benchmarks.
Indian venture funding has been tighter still. Domestic funds have concentrated on consumer and financial technology, and deep technology rounds above $10 million remain unusual outside a handful of names. Simplismart’s raise is small by American standards and significant locally, and the participation of a chip company and a Valley fund gives it a credibility that domestic capital could not.
Nvidia’s check is small by its standards and useful for another reason. A strategic investor in a $14 million round signals to Indian enterprise buyers that the startup is not going to disappear, and it gives the company a channel into the partner network that builds accelerator clusters. Accel’s presence matters similarly: the firm was an early backer of Flipkart and has been the most active American venture investor in Indian software for a decade.
The larger bet, and the reason the round is interesting beyond its size, is that the cost of intelligence will be set by software rather than by the price of the chip. If that is right, the companies that learn to squeeze more out of existing hardware will capture part of the difference, and the chip vendor will sell into a market that grows because the unit economics improved. Nvidia has made that bet repeatedly. It has rarely lost money on it.
