Nvidia, the company that sells the picks and shovels of the AI boom, has joined a $14 million bet on needing fewer of them. The chipmaker participated in a Series B round for Simplismart, a Bengaluru startup founded in 2022 whose software helps companies run AI models more efficiently — in some cases, on fewer GPUs. It looks, at first glance, like a company funding its own disruption. Simplismart’s chief executive says it is the opposite.
The round, led by Dallas Venture Capital with Accel, Shastra VC and Micromax also participating, values Simplismart at about $100 million including the new money. The funding will go toward expanding the startup’s presence in the United States, growing its 58-person team, and investing further in its inference software. More than half of Simplismart’s revenue already comes from the U.S., even though it has more customers in India — a geography the new money is meant to close.
The paradox of fewer GPUs, stated plainly
TechCrunch, which reported the round, called the investment an apparent paradox: the dominant supplier of AI chips backing software that can help companies run some workloads with fewer GPUs. It is a fair description. Simplismart’s product exists to squeeze more work out of the GPUs a customer already owns. In some cases, that means running a workload on fewer GPUs — but whether efficiency ultimately means fewer GPUs sold, or more AI built and more chips bought, is an open question.
Amritanshu Jain, Simplismart’s co-founder and chief executive, has an answer for that. “Nobody actually needs the fastest or the cheapest inference,” he told TechCrunch — companies need speeds and costs that fit their applications. And making inference more efficient, he argued, could ultimately benefit Nvidia by making its hardware more attractive to customers and encouraging wider adoption of AI applications. That is Jain’s argument: if cheaper inference means more AI gets built, then more AI getting built could mean more chips — eventually — for everyone, including Nvidia. It is a possibility, not a proven mechanism — and whether Nvidia’s own rationale matches Jain’s reading is unknown, since no Nvidia spokesperson is quoted in the coverage.
The software layer between the model and the GPU
Simplismart’s pitch is about the running of AI models: doing it reliably, economically, at scale, under the service-level agreements and cost constraints real businesses live with. “Where we tend to win is where people have to be more stringent on their SLAs and TCO,” Jain said. The company’s software sits between the model and the GPU, tuning how each workload is deployed: how fast it responds, how many requests it handles, how much compute it burns. It also offers fine-tuning tools, automatic scaling as demand changes, and performance monitoring.
Crucially, Simplismart owns no compute of its own. Unlike inference providers that buy or reserve GPU capacity and resell it, the company sells only the software layer — the code that lets other people’s GPU fleets run production AI workloads. Cloud providers can use it to offer managed AI services under their own brands without building the stack from scratch. “We have spent four years building the software layer between the model and the GPU,” Jain told YourStory. Indian cloud provider Yotta already uses the technology to power the serverless inference service inside its Shakti Studio platform.
The evidence it works — with caveats
The company says its software helped AI video-generation platform InVideo cut GPU serving costs by 56%. Its customer list runs through the current AI video boom — HeyGen, Higgsfield, InVideo — plus enterprise AI startup Ema and Indian firms Shiprocket and Apna. And the commercial momentum is real enough to quote: Simplismart says its annualised revenue run rate has grown 14-fold over the past six months, to an eight-figure sum.
The caveats matter as much as the numbers. The 56% figure is Simplismart’s own claim. The 14-fold growth is based on recent customer usage rather than contracted recurring revenue, excludes money passed through to infrastructure providers, and the exact figure was not disclosed. These are company-reported numbers, and they should be read as such. But the direction of travel — a software layer for inference, selling into a market where every GPU hour has a price — is not in doubt. The AI inference market is projected to grow from $106.15 billion in 2025 to $254.98 billion by 2030, according to MarketsandMarkets.
Partner, investor, competitor
The relationship is closer than a cheque. Alongside the investment, Nvidia has been working with Simplismart on technical integrations, including around Dynamo, its open-source inference software — with Simplismart’s engineers and Nvidia’s solutions architects involved — and has introduced the startup to potential customers and partners. Dynamo 1.0 is the company’s bid to be the “operating system” of the AI factory: open-source software that orchestrates inference across GPU fleets and, in Nvidia’s benchmarks, boosts Blackwell inference performance by up to seven times. It is already integrated by Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure, and adopted by companies from Cursor and Perplexity to ByteDance and PayPal.
Which makes the tension sharper, not softer: Simplismart competes with offerings from Nvidia itself, as well as the big cloud providers and a growing field of inference startups. Nvidia is simultaneously Simplismart’s investor, its integration partner, and one of its competitors. Whether that amounts to a hedge across the inference stack, a venture-style option, or — as Jain argues — a bet that efficiency expands the market for its hardware, is an open question. If inference is the business, as Jensen Huang put it — “Inference is the engine of intelligence, powering every query, every agent and every application” — then the efficiency layer is where that business is increasingly decided.
The case against
There are honest reasons to doubt the story. Fourteen million dollars is a rounding error for Nvidia, and a minority stake in a $100 million startup is not a strategic declaration — it may be a venture-style option on an inference vendor, plus a few customer introductions, and nothing more. Efficiency gains could just as easily let existing customers do more with the GPUs they already own, delaying new purchases instead of creating them. And the round itself was smaller than the figures in an earlier report: The Economic Times had reported Simplismart was in talks to raise $40 million to $50 million at a $300 million to $400 million valuation. Jain declined to comment on that report, said the startup was not actively fundraising, and said the Series B was closed and would not be extended. The gap between the reported talks and the closed round is, at minimum, a question the company has not answered.
But strip away the caveats and the shape of Jain’s argument is clear enough: the next wave of AI demand may come not from the companies that can afford the fastest inference, but from the ones that can finally afford inference at all — and making the chips less necessary per workload could be what makes them necessary for everyone else. Whether Nvidia’s own rationale matches Jain’s reading, and whether efficiency creates demand or simply delays the next GPU purchase, remains unknown. That is the open question at the heart of the bet.
