With US restrictions on graphics processing units exports, my stockpile has become invaluable: Sunil Gupta of Yotta

Once the tenders are awarded, and GPUs are procured and supplied to startups and other stakeholders, a comprehensive AI dataset platform will be created. The platform will house the largest collection of anonymised data, driving innovation and enhancing the capabilities of AI applications.

In a conversation with Forbes India, Sunil Gupta, co-founder and CEO of Yotta, talks about moving into B2C with myShakti, the upcoming mobile app version, security concerns surrounding DeepSeek, plans for listing, and more. Edited excerpts:

On owning the highest number of GPUs in the country

Having a large capacity of GPUs in India is like having a gold mine. Currently, I possess around 90 percent of the available GPU capacity in India. Of the 16,000 GPUs I ordered, 4,000 are already installed, and another 8,000 are partly in my storage or in my partner’s warehouses. I haven't deployed them yet to avoid unnecessary power consumption. Other than E2E Networks, which has some spare GPUs (around 200 to 300), and Tatas, which ordered 250 GPUs, no other company in India has a significant GPU capacity. The other bidders for the government contract don't have GPUs and are yet to place orders with Nvidia, Intel or AMD.

Implementing GPUs with InfiniBand, CUDA and other software layers is a complex task. Most bidders don't have their own data centres, so they'll need to outsource from someone like CtrlS datacenters or Yotta.

Given the time required to procure GPUs, set up data centres, and implement the infrastructure, it will likely take seven to eight months for the other bidders to get empanelled. In contrast, Yotta, Tata and E2E can provide GPUs immediately.

I took a significant exposure on GPUs in the last five to six months, but due to global demand towards India being more towards short-term bulk GPU requirements for training of LLMs and not being a sustained long-term demand, and India demand growing slowly due to it being limited to development of SLMs, fine-tuning, and POCs and not for building LLMs, I could not monetise them to the level we foresaw. However, with the India AI mission and the US restrictions on GPU exports, my GPU stockpile has suddenly become invaluable.

On progress under the IndiaAI mission

I have been appreciating the IndiaAI mission since its inception. By providing funding to users, they are addressing the primary hindrance: Cost. Despite aggressive pricing, users found AI solutions cost-prohibitive. Even with reasonable pricing, deploying AI models for production-grade use required substantial investment, often in crores.

Many of end enterprise CIOs for whom my customers, including Sarvam AI, develop and deploy AI models for, struggled to justify AI investments to their CEOs and CFOs.

The government's approach, instead of subsidising service providers, puts money directly in the hands of end-users. This creates a strong business case for us, as we receive payment and benefit users and the government by growing the ecosystem.

However, since the IndiaAI mission was announced nine to 10 months ago, potential consumers have been holding back their requirements. With the mission now taking shape, I've signed large contracts with institutes that include clauses allowing them to exit if they receive IndiaAI funding.

In the next 10 to 15 days, the government will finalise empanelment, and the real demand will begin. Users will start requesting IndiaAI to fund their GPU purchases, and that's when the excitement will start.

On launching a chatbot utilising DeepSeek’s open-source model

MyShakti is based on DeepSeek's 70B model. To be frank, we've simply put an app envelope on top of this base model. I haven't trained or fine-tuned the model, nor have I added any data or attachments. It's the original model, trained on DeepSeek's dataset, which they've open-sourced.

I've containerised, secured and deployed the model on my GPUs. Our current model delivers mixed results. For some queries, it provides thought processes and then answers. For others, it gives inconsistent responses. However, for many queries, it provides answers comparable to OpenAI. We are still working on improving it.

Following the success of MyShakti, we're now developing a solution based on DeepSeek's 670 billion parameter model. Within the next seven or eight days, you'll see significant progress.

We're not simply using the model in its raw form. Instead, we're working with Nvidia to containerise it using their Nvidia Cloud Functions (NVCF) platform. NVCF is a serverless API that allows users to deploy and manage AI workloads on GPUs. In fact, Nvidia has informed us that we're the first company globally to utilise this platform.

The 670 billion parameter model is massive and requires multiple H100 servers to run it. This means it can't be executed on a single server or desktop, unlike our existing model. We're collaborating closely with Nvidia to bring this model live in some days.

On privacy concerns with DeepSeek

DeepSeek has open-sourced their model, providing detailed documentation on their training methods. In fact, this transparency issue is more applicable to closed-source models like OpenAI, which are facing lawsuits due to their unclear training data.

A comparative analysis of DeepSeek and ChatGPT models shows that while ChatGPT has guardrails to ensure generative AI, DeepSeek's 670B model has no such restrictions. This model excels in reasoning, coding and analysing large documents.

DeepSeek's open-source nature means it's up to the users to decide how to utilise it. With the model running on Indian servers, we have complete control over root access, network access and admin access.

To address concerns about security threats related to China, we have restricted all admin and user level access to Chinese IPs completely, and ensured that not only we block any inferencing requests from there but also block any telemetry data to be routed outside, using firewalls to block such requests.

Some people worry that using AI models like DeepSeek means their data will be used by the company. However, that's not the case when you run the model on your own servers.

Think of it like MS Office. When you use the cloud-based version, your data is stored on servers in the US or elsewhere. But when you use the desktop version, your files are stored locally on your laptop, unless you choose to share them.

Similarly, with our AI model, we've deployed it on our servers in Mumbai, using our own GPUs and storage. When users send inferencing queries, the data is processed and stored on our servers, not sent to any external servers, including those in China.

It's essential to distinguish between using a public SaaS application like DeepSeek's app, which may store data on Chinese servers, and running the DeepSeek model on your own servers in India. We're proud to be the first in India to deploy this AI model locally, ensuring that all data processing and storage happen within our servers.

On the current usage of GPUs

You previously asked me about the usage of our GPUs, and I mentioned that most of our capacity was being utilised internationally, with domestic usage being relatively low. However, the scenario has changed.

Internationally, our GPUs are still being used for training models, but for inferencing, companies prefer to use their own GPUs in their home countries due to data localisation and security concerns. As a result, our GPUs are now being increasingly utilised domestically.

Our focus has always been on serving Indian users, and we've been building our business to cater to their needs. With the growing demand for AI and data localisation, I'm confident that our GPUs will be fully utilised, primarily driven by steady-state demand from Indian companies.

The timing is perfect—with Indian companies scaling up, DeepSeek entering the market and the US restrictions creating new opportunities. Our GPUs, which were previously underutilised, will now be fully utilised, and I'm excited about the prospects.

On Yotta’s upcoming plans

We have several exciting developments in the pipeline. First, we're shifting our focus towards AI, which will be a huge area of growth for us. We're also expanding our offerings beyond B2B to include B2C services.

Our AI platform MyShakti was initially intended for B2B clients, but we've decided to take it to the B2C space as well. We've acquired a software company, which gives us the capabilities to develop and deploy AI models. We're also running our own sovereign cloud, which enables us to offer more services.

In the next few days, you can expect to see a mobile app for MyShakti, and we'll be integrating more open-source models, including LLaMA, into our platform. We'll also be working with our clients, such as Sarvam and Hanooman, to containerise their models and make them available on our platform.

Another significant development is our planned listing, which is expected to happen soon.