Simplismart has raised $14 million in a Series B funding round led by Dallas Venture Capital, with NVIDIA, Accel, Shastra VC and Micromax participating. Announced on October 27, the financing backs a business that supplies software for running AI models, rather than building its own fleet of graphics processors.
The company says it will invest in its inference engine and orchestration platform, expand cloud partnerships and pursue growth in North America. Inference is the stage at which an already trained model handles requests. Orchestration manages how those requests reach available computing capacity.
In its funding announcement, Simplismart says its engineers work with NVIDIA Solutions Architects. It also describes a planned scheduling layer to match workloads with capacity across regions while respecting response-time and data-location requirements. That remains a direction for development, not a newly available service announced alongside the round.
Investor commits $5 million
Dallas Venture Capital separately says it is contributing $5 million to the round. Its October 27 investment note identifies Amritanshu Jain and Devansh Ghatak as the founders, who started Simplismart in 2022 after meeting as roommates at BITS Pilani. Jain previously worked on cloud infrastructure at Oracle Cloud, while Ghatak worked on search algorithms at Google.
The investor says the founders secured their first production customers with less than $1 million before a $7 million Series A led by Accel in October 2024. That earlier financing also included Shastra VC, Titan Capital, Dallas Venture Capital and Notion co-founder Akshay Kothari.
DVC describes the platform as working across model serving, infrastructure and the interaction between models and GPU chips. Its note points to deployment across language, speech, vision-language and image-generation workloads. It also describes integration work with NVIDIA Inference Microservices. DVC names HeyGen, Tata 1mg, Mindtickle, InVideo and Dashtoon among customers across several business sectors. The investor describes tools that let enterprises fine-tune, deploy and monitor generative AI models on their own infrastructure.
Routing requests before generating answers
Simplismart also publishes undated case studies about Ema and Invideo. These are company accounts of existing deployment work, not new contracts announced with the funding or independent performance tests.
The Ema account describes an enterprise AI platform whose custom routing models decide which underlying model should process a request. Its two routers have approximately one billion and eight billion parameters. They select where work goes rather than generating the final response that a user receives.
Simplismart says it changed the pipeline to overlap CPU preparation, GPU processing and transfers between them. The work included resequencing the routers, which had operated sequentially, and batching requests with similar sequence lengths. That addresses waiting between operations as well as how requests are grouped for processing.
The account also describes CUDA kernel and compilation changes, including adjustments to how operations are launched and combined on the GPU. Changes to autoscaling triggers and limits addressed how extra capacity becomes ready when demand rises. Simplismart says the work used change isolation and A/B validation, separating alterations and comparing configurations while checking routing quality.
Video workloads present a different problem
The Invideo case study describes a pipeline spanning image enhancement, portrait animation and speech recognition. Traffic ranges from consumer surges to enterprise batch jobs. The work therefore involved different types of requests, rather than tuning one uniform task.
Simplismart says it combined small GPU operations in the Clarity Upscaler and Live Portrait stages. The purpose was to reduce the overhead of launching separate operations and moving data between them. It also cached intermediate results, allowing dependent stages to reuse information instead of repeating computation.
The account describes chunked attention, a memory-management approach intended to reduce peak demand for GPU memory. Its stated purpose was to fit larger batches on the same hardware and reduce retries when a workload exceeded available memory.
Capacity management was another part of the work. Simplismart says load balancing directed short and long, light and heavy requests to suitable replicas, while autoscaling policies addressed sudden traffic increases. The company describes stress-testing the pipeline under steep traffic ramps before checking it again after optimization.
Related startup coverage: Cohere’s Series E financing and Nevis’s Series B round.
