Simplismart, a Bengaluru startup that builds software to run AI models more efficiently, has raised $14 million in a Series B round with Nvidia among its backers. TechCrunch’s Jagmeet Singh reported the deal on 27 October 2026, describing a bet on software “that helps companies run AI models more efficiently and, in some cases, with fewer GPUs.” Dallas Venture Capital led the round. Nvidia joined Accel, Shastra VC and Micromax, and the deal values the four-year-old company at about $100 million including the new money.
The investment looks odd at first glance. Nvidia sells the GPUs that power almost every large AI deployment, and Simplismart’s pitch is that customers can get more out of the chips they already have. The company’s co-founder and CEO, Amritanshu Jain, argues the two goals are not in conflict. Cheaper, more reliable inference, he says, makes Nvidia hardware more attractive and pushes more companies to put AI into production.
This guide explains what Simplismart actually does, who uses it and what the numbers say. It covers why running a model is now the expensive part of AI, how Nvidia’s own Dynamo software fits in, and how crowded the inference market has become. It finishes with practical lessons for any business paying for GPUs today. For the cost side in more depth, see our guide to FinOps for AI and cutting LLM, agent and GPU costs.
Table of contents
- What Simplismart Raised, and From Whom
- What Simplismart Actually Does
- Why Running AI Models Is the Hard Part
- Simplismart’s Customers and the 56% Claim
- Simplismart’s Growth Numbers
- Why Nvidia Would Back Software That Uses Fewer GPUs
- A Crowded Market for Inference Software
- The Valuation Question
- What Simplismart Means for Businesses Running AI
- Our Verdict on the Simplismart Deal
- Simplismart FAQs
- References
What Simplismart Raised, and From Whom
The round is small by AI standards, but the investor list is the story. Here are the details as TechCrunch reported them.
| Item | Detail |
|---|---|
| Round | Series B, all primary (new shares only) |
| Amount | $14 million |
| Lead investor | Dallas Venture Capital |
| Other investors | Nvidia, Accel, Shastra VC, Micromax |
| Valuation | About $100 million, including the new funding |
| Headquarters | Bengaluru, India |
| Founded | 2022, by Amritanshu Jain and Devansh Ghatak |
| Staff | 58 |
“All primary” means every dollar went to the company as new capital, rather than some of it buying out early shareholders. The valuation includes the round, so on a post-money basis the new investors own about 14% between them ($14 million ÷ $100 million), assuming the reported figure is exact.
The founders
Jain previously worked as an engineer at Oracle, and Ghatak at Google. The two met while studying at BITS Pilani, and TechCrunch says the company grew out of their experience building and deploying machine-learning systems. The company’s legal name, according to SiliconANGLE’s coverage of its earlier round, is Verute Technologies.
The earlier round
In October 2024, Simplismart raised a $7 million Series A led by Accel, with Shastra VC, Titan Capital and Notion co-founder Akshay Kothari also taking part. At the time, SiliconANGLE reported that it had about 30 customers delivering a combined $1 million in annual run-rate revenue. The new round is twice the size of that one ($14 million ÷ $7 million), raised two years later.
What Simplismart Actually Does
Simplismart sells inference software. Inference is the work of running a trained model to answer requests, as opposed to training it in the first place. Every chatbot reply, generated image or transcribed call is an inference job, and each one costs GPU time.
The layer between models and GPUs
TechCrunch describes the software as sitting “between AI models and the computing infrastructure they run on.” It adjusts how a model is deployed and managed so that it meets the customer’s requirements. The company says it can optimise how quickly a model responds, how many requests it handles at once and how much computing power it consumes.
In practice that means choosing settings most teams would otherwise tune by hand. The company’s website lists the levers: the type of GPU, the serving framework, the model backend, the level of quantisation, how the model is split across chips and which caching techniques are used. It says it builds custom CUDA kernels, the low-level code that runs directly on Nvidia GPUs.
Three dials: speed, volume and cost
The core idea is that every workload needs a different balance. Its website gives three examples. Voice agents are tuned for latency, with responses under 100 milliseconds. Document processing is tuned for throughput, handling as many pages as possible. Content generation is tuned for cost.
Jain put it bluntly to TechCrunch: “Nobody actually needs the fastest or the cheapest inference.” What companies need, he said, is to run models at speeds and costs that make sense for their specific applications. A customer-service voice bot that pauses for two seconds feels broken, but an overnight batch job that summarises documents can run slowly on cheaper hardware without anyone noticing.
Your cloud or theirs
There are two ways to deploy. Customers can run models on the startup’s own platform, or inside their existing cloud and on-premises infrastructure. Its website says it supports deployments in private cloud networks or on-site setups, with one control plane across more than 15 clouds, and offers B200, H100, A100, L40S and A10G GPUs.
The platform also includes tools to fine-tune models, scale computing up and down as demand changes, and monitor performance. The startup claims it can scale up in under 500 milliseconds and scale to zero when there is no traffic, so customers are not paying for idle GPUs.
Inference as a service for cloud providers
There is a third business line. The software also lets cloud providers build and run their own AI inference services on GPU infrastructure using its software. That turns it from a tool for AI companies into a white-label engine for data centre operators, which matters in India, where local providers are racing to offer AI services.
Why Running AI Models Is the Hard Part
Training a frontier model makes headlines, but inference is where most companies actually spend. Training happens once, or a few times. Inference runs every time a user asks a question, and it scales with success: more customers means more GPU time.
The trade-offs are real
TechCrunch notes that “different workloads demand distinct trade-offs between speed, cost, and performance,” and that deploying models to meet business requirements “often also requires complex adjustments.” A model that is fast on one GPU may be wasteful on another. Settings that maximise throughput can hurt response time. Getting it wrong either annoys users or burns money.
Why software can mean fewer GPUs
If a model handles twice as many requests per GPU after tuning, a company needs half as many GPUs for the same traffic. That is the simple arithmetic behind the “fewer GPUs” headline. The savings come from techniques such as quantisation (storing the model’s numbers at lower precision), smarter batching of requests, better caching and autoscaling that switches GPUs off when they are not needed. The startup’s role is to choose and combine those techniques automatically, so customers do not need a specialist team to do it.
A large language model sits behind much of what Simplismart serves, but its website says it also supports vision-language, image-generation and speech models. One testimonial on the site, from Tata 1mg, says it used Simplismart to train and deploy a computer vision model that processes medical prescriptions at 95% accuracy.
Simplismart's Customers and the 56% Claim
The startup named several customers to TechCrunch. Most are AI-native companies that generate video or handle large volumes of requests, which is where GPU bills grow fastest.
Who uses the platform
| Customer | What it does | How it uses the platform |
|---|---|---|
| InVideo | AI video generation | Cut GPU serving costs by 56%, the startup says |
| HeyGen | AI video | Named customer |
| Higgsfield | AI video | Named customer |
| Ema | Enterprise AI | Named customer |
| Yotta | Indian cloud provider | Powers the serverless inference service in its Shakti Studio platform |
| Tata 1mg | Online pharmacy | Prescription-reading vision model (website testimonial) |
What a 56% saving looks like
Simplismart says it helped InVideo cut its GPU serving costs by 56%. The figure is the company’s own and InVideo’s actual spending is not public, but the arithmetic is easy to picture. A business spending $100,000 a month on serving would drop to $44,000 ($100,000 × 0.44), saving $56,000 a month or $672,000 a year ($56,000 × 12). For a video company whose costs rise with every clip, that kind of saving changes the economics of the product.
One more testimonial on the website claims image-generation costs fell “from $30,000 to under $1,000” while inference time halved. The site does not say over what period, so treat it as a best case rather than a typical result.
Revenue from the US, customers in India
Jain told TechCrunch that more than half of the company’s revenue comes from the US, even though most of its customers are in India. That pattern is common for Indian infrastructure startups: American AI companies spend more per customer, while the Indian market brings volume. It also explains why the US is the main target for the new money.
Simplismart's Growth Numbers
The headline figure is fast growth from a small base. Jain said Simplismart’s annualised revenue run rate has grown 14-fold over the past six months to an eight-figure sum, meaning at least $10 million. He did not give the exact number.
How the figure is defined
TechCrunch spelled out the caveats. The figure “is based on recent customer usage rather than contracted recurring revenue,” and it excludes revenue passed through to infrastructure providers. In plain terms, the company is taking a recent period of usage and multiplying it up to a yearly rate, and it is not counting the GPU costs it collects from customers and pays straight to cloud providers.
Both choices matter. Usage-based run rates can swing with a single large customer, and including pass-through GPU costs would inflate the number without adding to the company’s own margin. Excluding them is the more conservative choice.
What 14x in six months implies
Growing 14 times over in six months is the same as growing about 55% a month, every month, compounded. The check is 1.55 multiplied by itself six times, which comes to about 13.9. If the current figure sits at the bottom of the eight-figure range, $10 million, the run rate six months ago was about $714,000 ($10 million ÷ 14).
Bars are scaled to $400 million: 300 is 75% of 400, 100 is 25%, 14 is 3.5% and 7 is 1.75%. The two top bars are a reported negotiation, not a completed deal, as the next sections explain.
Two years of change
Put the two rounds side by side and the shape is clear. In October 2024 Simplismart had about 30 customers and around $1 million in run-rate revenue. Two years later it reports an eight-figure run rate, at least ten times the earlier figure ($10 million ÷ $1 million), on a different, usage-based definition. The headcount has reached 58.
Why Nvidia Would Back Software That Uses Fewer GPUs
TechCrunch calls it “an apparent paradox”: the dominant supplier of AI chips is funding software that can help companies run some workloads with fewer of them. There are good reasons it makes sense for Nvidia.
Jain’s argument
Jain argued that making inference more efficient “could ultimately benefit Nvidia by making its hardware more attractive to customers and encouraging wider adoption of AI applications.” If running a model costs less, more products become viable, and more companies move from pilot to production. That means more total GPU demand, even if each job uses fewer chips.
The efficiency paradox
Economists have a name for this pattern. The Jevons paradox describes how making a resource cheaper to use often increases total consumption rather than reducing it. Efficient steam engines led to more coal burning, not less. If Simplismart makes a video model 56% cheaper to serve, the likely result is more video generated, not fewer GPUs sold across the market.
The Dynamo connection
There is also a technical tie. Jain told TechCrunch that Nvidia has been working with Simplismart on integrations, “including around its Dynamo inference software.” Nvidia unveiled Dynamo at its GTC conference on 18 March 2025 as open-source software for “accelerating and scaling AI reasoning models,” and as the successor to its Triton Inference Server.
Dynamo’s headline feature is disaggregated serving. It separates the two phases of running a large language model, processing the prompt and generating the reply, onto different GPUs, so each phase can be tuned on its own. Nvidia said Dynamo raised throughput by up to 30 times when serving DeepSeek-R1 on its Blackwell hardware. A startup that helps customers adopt Dynamo well keeps those customers on Nvidia’s software stack, which is worth more to Nvidia than a single GPU sale.
More than money
Jain said the collaboration involves the startup’s engineers working with Nvidia’s solutions architects, and that Nvidia has introduced the startup to potential customers and partners. For a 58-person company, those introductions may be worth more than the cash. For Nvidia, backing small inference companies spreads its software and builds goodwill with the developers who choose hardware. It fits a wider pattern of Nvidia investing across the AI stack, which we looked at in Nvidia’s $3.5B MediaTek Bet.
A Crowded Market for Inference Software
Simplismart is not alone. TechCrunch notes that it competes with offerings from Nvidia itself and from the biggest cloud providers, including Amazon Web Services, Google Cloud and Microsoft Azure. A growing number of startups are also building tools to make AI models faster and cheaper to run.
Where the startup says it wins
Jain was specific: “Where we tend to win is where people have to be more stringent on their SLAs [service-level agreements] and TCO [total cost of ownership].” In other words, customers with tight promises to keep on response times, and tight budgets, are the ones most likely to pay for careful tuning.
He contrasted this with rivals that offer preset choices. Many inference services give customers a menu of “fast” or “cheap” tiers. The company says it instead aims to find a balance tailored to each customer’s workload, which is the “tailor-made inference, not generic APIs” message on its website.
The risk from the giants
The obvious threat is that the hyperscalers keep improving their own inference tools and bundle them into the cloud bills customers already pay. Nvidia’s own Dynamo is free and open source. A small company can stay ahead on tuning skill and service, but it has to keep moving. Its answer so far has been to work with Nvidia rather than against it, and to sell into the cloud providers as well as compete with them.
The Valuation Question
The funding came with a twist. Before the announcement, India’s Economic Times reported that Simplismart was in talks to raise $40 million to $50 million at a valuation of $300 million to $400 million.
What Jain said
Asked about the report, Jain declined to comment, saying the company was not actively fundraising. He added that the Series B was closed and would not be extended. So the official position is a $14 million round at about $100 million, with nothing more on the table.
Reading the gap
The gap between the two figures is large. A $300 million to $400 million valuation would be three to four times the Series B price ($300 million ÷ $100 million and $400 million ÷ $100 million). That may reflect early conversations that never became a deal, interest from investors that the company chose not to take, or a future round. With revenue growing as fast as Jain describes, it would not be surprising to see a larger raise within a year, but for now it is a report, not a fact.
What the money is for
The startup says it will use the new funding to expand in the US, grow its team and invest in its inference software. Jain told TechCrunch the US remains a key market as the company builds on its customers there.
What Simplismart Means for Businesses Running AI
You don’t need to be a video startup for this story to matter. Any organisation paying for GPUs, directly or through an AI provider, faces the same questions Simplismart is built to answer.
Measure what you actually need
Start with Jain’s point: nobody needs the fastest or cheapest inference, only the right balance. Write down the response time your users will accept, the volume you need to handle at peak, and what you can afford per request. Those three numbers decide most of the technical choices that follow.
Ask about utilisation
Idle GPUs are the most common waste. Ask your provider or your own team how busy your GPUs are on average, and whether capacity scales down at night and at weekends. Scale-to-zero, which the startup advertises, is valuable for spiky or internal workloads that sit idle for hours.
Questions to ask any inference vendor
- What latency and throughput do you guarantee, and is it in the contract?
- Can I run on my own cloud account, or only on yours?
- How do you price: per request, per token, per GPU-hour or a share of savings?
- Are GPU costs passed through at cost, or marked up?
- Which models, GPUs and frameworks do you support, and how fast do you add new ones?
- What happens to my data, and which security certifications do you hold?
Simplismart’s website lists GDPR compliance, ISO 27001 and SOC 2, and claims 99.99% uptime. Any vendor should be able to give equally clear answers. If you are building the business case for AI in the first place, our AI models and tools hub collects guides and reviews to help.
Our Verdict on the Simplismart Deal
The $14 million is modest, but the signal is strong. Nvidia’s participation tells the market that the next stage of AI spending is about running models well, not just buying more chips. Simplismart’s reported growth, 14-fold in six months on a usage basis, shows real demand for that kind of help, especially from AI video companies whose GPU bills grow with every customer.
There are caveats. The growth figure is self-reported and based on recent usage, the 56% saving is a vendor claim, and the market is full of large competitors, including Nvidia itself. But the logic of the deal holds up. Cheaper inference tends to mean more inference, and Nvidia benefits either way. For businesses, the lesson is simpler still: before buying more GPUs, find out how well you are using the ones you have.
Simplismart FAQs
What is Simplismart?
Simplismart is a Bengaluru-based startup founded in 2022 that makes software for deploying and running AI models efficiently. It tunes how models run on GPUs to balance speed, volume and cost for each customer.
How much did Simplismart raise?
It raised $14 million in an all-primary Series B round announced on 27 October 2026, led by Dallas Venture Capital, with Nvidia, Accel, Shastra VC and Micromax taking part.
What is Simplismart’s valuation?
About $100 million, including the new funding, according to TechCrunch. An earlier Economic Times report of talks at $300 million to $400 million was not confirmed, and Jain said the Series B was closed and would not be extended.
Why would Nvidia invest in software that uses fewer GPUs?
Cheaper, more efficient inference tends to increase total AI adoption, which raises overall demand for GPUs. Nvidia is also working with the startup on integrations with its Dynamo inference software, which keeps customers on Nvidia’s platform.
Who are Simplismart’s customers?
Named customers include InVideo, HeyGen, Higgsfield and Ema. Indian cloud provider Yotta uses its technology for serverless inference in its Shakti Studio platform. More than half of its revenue comes from the US.
How fast is Simplismart growing?
Its annualised revenue run rate grew 14-fold in six months to an eight-figure sum, based on recent usage and excluding pass-through infrastructure revenue.
Who founded Simplismart?
Amritanshu Jain, a former Oracle engineer who is now CEO, and Devansh Ghatak, a former Google engineer. They met while studying at BITS Pilani.
