Topics In Demand
Notification
New

No notification found.

Middleware’s Moment: The Quiet Layer Making Loud Noise in 2025 VC Circles
Middleware’s Moment: The Quiet Layer Making Loud Noise in 2025 VC Circles

44

0

AI Startups captured a record 46.4% of the total $209 Bn raised last year, compared to less than 10% a decade earlier. Also, more than 75-77% of investments by value have gone to LLM makers in H2FY2025. So, we are witnessing a record number of total investments that is happening in LLM Makers recently. These foundational model makers are working hard to come up with a better model at decreasing prices. They are also gearing up to keep their inference cost down.

In addition to model making, however, more model-agnostic capabilities in core data infrastructure, middleware integration, and model optimization activity have picked up. There are three AI infrastructure segments that are gaining grounds looking at the funding scenarios in 2025 till now.

  1. Hosting and optimization infrastructure startups picking up:

These AI Infrastructure layer startups are making it easier and more cost-effective for businesses and developers to deploy and use various AI models for language processing, image generation, video creation, and other AI-powered applications. They help developers and corporations integrate AI capabilities into their applications without having to build the underlying infrastructure themselves.

For example, Silicon Flow is a China based startup that recently raised a series A round in June, after raising a $13.8 Mn seed round in Feb 25. Silicon Flow is essentially a "middleware" company - they sit between the AI models (like ChatGPT, Claude, diffusion models, etc.) and the end users who want to build products with those models. They make AI more accessible by handling the integration of AI application with the technical infrastructure layer. Silicon Flow handles the complexities like model hosting, scaling, optimization, and API access.

  1. Data Infrastructure is complementing model makers:

Several recent investments are going to data infrastructure startups that are complementing model makers. We have seen several Big Tech acquisition investments in 2025 that highlight that the need for clean and real time data is more important than ever. There is a similar trend reciprocated in VC investments. These startups serve as a critical building block for any machine learning model. Data infrastructure solutions serve the purpose of data pipeline management to get the right data to the right model at the right time. For example, Chalk AI that got a series A funding of $50 Mn. It is an AI based model agnostic platform for managing data pipelines. Another example is Deccan AI which provides high-quality data labelling and model evaluation with expert human-in-the-loop support. Its market are the model makers that need data for training their models. And of course, the high-profile $14.3 Bn acquisition of Scale AI by Meta to boost data annotation efforts for next generation models.

  1. Machine unlearning platform for model makers:

Another set of model-agnostic solutions that are getting investor attention are the ones that solve the problem of AI hallucinations, AI bias and data privacy. For example, Hirundo, founded in 2023, got a seed round funding of $8Mn in June 25. It provides platform for machine unlearning solutions. Its solutions are unlike the traditional masking method; it goes to the root cause of hallucinations or logical gaps and removes deviant data from the model in production. The platform removes any possibility of retraining or surface level fixes like filtering.

Conclusions: So, what does it mean for GenAI advancements and adoption?  The implementation of current large language models is posing several challenges in the enterprise-wide deployment and scaling. The models should be trustworthy, pose minimum AI bias, hallucinations, and be safer to implement. The continuous flow of real time clean data could improve the efficiency of these models. Since, enterprises need to work on multiple language models for various tasks, switching between models should not only be automated, but also cost-efficient and in a way that optimizes load while giving the desired output. It is these integration aspects to make model-based AI usage smoother that these startups draw attention to, and will likely increasingly be in demand as more integration is needed to build AI agents.

Opportunity for India: India also needs to step up the game and come up with these middleware solutions that could complement our Indic foundational models. AI Kosha will not be enough to provide the real time data required for some of these models; we need more solutions in data labelling and model agnostic real time data providers across text, audio, image and 3D segments. In 2020, with no GenAI in picture, we had estimated that data annotation could be a $1 Bn opportunity for India by 2025, but with GenAI and AI agents devouring data at exponential rates, this opportunity is significantly bigger and impactful.


That the contents of third-party articles/blogs published here on the website, and the interpretation of all information in the article/blogs such as data, maps, numbers, opinions etc. displayed in the article/blogs and views or the opinions expressed within the content are solely of the author's; and do not reflect the opinions and beliefs of NASSCOM or its affiliates in any manner. NASSCOM does not take any liability w.r.t. content in any manner and will not be liable in any manner whatsoever for any kind of liability arising out of any act, error or omission. The contents of third-party article/blogs published, are provided solely as convenience; and the presence of these articles/blogs should not, under any circumstances, be considered as an endorsement of the contents by NASSCOM in any manner; and if you chose to access these articles/blogs , you do so at your own risk.


images
Madhumay
Deputy Manager - Research

© Copyright nasscom. All Rights Reserved.