Koduri’s Oxmiq Offers IP, Tools for Indian Teams to Design AI Chips
In an exclusive interview, Raja Koduri, founder and CEO of Oxmiq Labs, discussed the company’s India initiative, the role of memory hierarchy in AI computing systems, custom silicon economics, and India’s infrastructure requirements.
Oxmiq Labs has developed tools and intellectual property (IP) to help teams in India design their own AI chips. The company is also working on an India initiative that it expects to detail in December, Koduri told EE Times.
“We have built tools to enable teams in India to make their own chips,” he said. “We are working on an initiative in India, and hopefully, when I return to India in December, I will be able to share some of those initiatives publicly.”
The aim, Koduri said, is to develop state-of-the-art chips designed to produce AI tokens as efficiently as possible.
Separately, Oxmiq is an architecture and engineering partner to AM Intelligence Labs, the AI compute division of AM Group, the parent company of Greenko, India’s renewable energy producer with a portfolio target of 50 GW by 2030. The partnership covers site selection, compute design, and partner selection. The initial build will use Nvidia GPUs, with locally developed designs to follow.
“To begin with, we are bringing the best technology available today, which is Nvidia GPUs,” he said. “But as we move toward 2030, we will also leverage designs developed in India.”
Koduri said India must develop its own designs, chips, and infrastructure to bring down the cost of converting electricity into AI output, even if manufacturing remains overseas.
“Chips designed in India can be manufactured in Taiwan or South Korea,” he said. “That is perfectly fine. But the products should be owned by India.”
From megawatts to intelligence
Koduri said the main constraint in AI computing is how efficiently a data center converts power into AI output. At the technical level, he said, the memory hierarchy matters more than floating-point operations and compute.
“It is all a memory hierarchy problem, from NAND flash storage to double data rate [DDR] memory attached to the central processing unit [CPU], high-bandwidth memory [HBM] attached to the GPU, and static random-access memory [SRAM],” he said.

A startup might focus on one type of memory, such as SRAM, and position it as the answer to a particular computing problem. But Koduri said the system-level requirements are broader.
“We look at the problem as a memory hierarchy and exploit each level as efficiently as possible,” he said.
Asked what changes when energy, compute, cooling, networking, and software are designed together, Koduri said the industry has traditionally operated in separate layers. Chipmakers typically design and sell a chip to a systems company, which builds a system. Another company builds the rack, while a data center operator handles installation, thermal management, and cooling.
“There are multiple layers in the ecosystem before everything comes together,” he said.
However, the scale of AI data center expansion is changing how companies approach the design process. Koduri said engineers must consider the entire data center rather than optimize an individual chip in isolation.
“We have to think about the whole system,” he said. “Not just the system, but the entire data center.”
The objective, he explained, is to convert incoming electricity into tokens as efficiently as possible while managing the heat generated by the chips.
Koduri described this as an optimization problem that will continue to evolve as AI models and computing architectures change. Over the past five years, he said, the industry has moved from basic transformers to mixture-of-experts models, DeepSeek optimizations, and decision-making models. GPU optimization is also continuing.
“The real constraint is that we have megawatts coming in and intelligence coming out,” he said. “How efficiently can we transform those megawatts into intelligence?”
Customers, Koduri said, care about the output of the complete system. “They want to plug in electricity, provide cooling, and get tokens they can sell. That is all.”
Oxmiq’s role at AM Intelligence Labs
Oxmiq’s work with AM Intelligence Labs covers data center planning and the selection of technology partners, Koduri said.
The company’s support begins with site selection and data center design. It then works on compute design, rack placement, networking and storage requirements, thermal management, cooling, and software stack operations.
“We provide all that expertise,” he said.
“We do not necessarily solve every problem ourselves,” Koduri said. “For example, we may select a suitable partner to provide the system, such as an original equipment manufacturer. That is the kind of support we are currently providing to [AM Group], because we come from the technology world, while they come from the energy world.”
Koduri pointed to AM Intelligence Lab’s first AI factory in Hyderabad, which he described as probably the first such deployment in Asia. The company has ordered about 9,000 Nvidia Rubin GPUs for the 30-MW facility, with delivery of Vera Rubin NVL72 rack-scale systems scheduled for the first quarter of 2027. It has since placed an additional order for 20,000 Rubin GPUs for sites in India and Malaysia.
Custom silicon economics and India’s infrastructure needs
Koduri said companies can potentially improve computing efficiency by designing the silicon, systems, and infrastructure for a specific data center rather than relying entirely on components designed for a wide range of customers.
Traditionally, a chipmaker had to sell its product to many data centers to generate revenue. That required designing a flexible product that could accommodate different customer requirements, standards, and features, even if those compromises reduced efficiency.
“If I am designing for one data center and its specific constraints, I can build something that is twice as cost-effective,” he said.
For a 1-GW data center, which Koduri estimated could cost around $55 billion to build, he said custom-designing the system could potentially deliver twice the compute by using power more efficiently.
He said this approach could change both the engineering mindset and the business model. A company developing silicon could now target the data center problem itself, with a single customer potentially generating multibillion-dollar revenue.
Koduri said custom silicon, systems, and software begin to make economic sense for customers planning at least 100 megawatts (MW) of computing capacity.
He put the cost of 100 MW of computing capacity at approximately $5.5 billion using off-the-shelf hardware. By comparison, R&D for a state-of-the-art chip manufactured by TSMC could cost around $100 million to $150 million. Developing a complete custom system could cost up to about $500 million, even with generous spending, he said.
If a custom system roughly halved that cost, a customer could save more than $2 billion even after spending $500 million on development, he said.
However, customers need the IP, tools, and technical know-how to develop custom systems, and Koduri said Oxmiq is building a solution to that problem.
Koduri said Oxmiq’s customers fall into three categories. The first comprises infrastructure companies such as AM Intelligence Labs, which are building computing capacity and need support across the system. The second includes companies designing silicon for applications such as data centers, physical AI, and enterprise AI.
Oxmiq can provide hardware and software IP to startups developing products such as chips for robots, allowing them to focus on their customers’ requirements, Koduri said.
The third category includes companies developing new materials and new physics, such as new forms of memory, silicon, silicon photonics, and photonics-based interconnects. They can license Oxmiq’s IP rather than develop the computing technology alongside their research.
In India, Oxmiq is engaging with customers in the infrastructure and silicon design categories. In the U.S., it works across all three categories, with most customers in infrastructure and new-physics applications, he said.
India’s installed data center power capacity stood at 1.57 GW in August 2026, up from about 375 MW in 2020, and is projected to reach nearly 8 GW by 2030, according to the Indian Press Information Bureau. About $70 billion of investment is underway, with a further $90 billion in announced projects.
Koduri said a 1-GW data center costs approximately $55 billion to build. By comparison, he said capital expenditure per gigawatt in China was moving toward $10 billion. “They can get things done at one-fifth, and perhaps one-tenth, of the cost of the rest of the world,” he said.
He estimated that India would need at least 6 GW of infrastructure to provide intelligence to its entire population, adding that the actual requirement would be much higher. The combined GPU capacity currently available in India, including planned deployments by AM Intelligence Labs, would not reach 100 MW, he said. That suggests India would need to increase that capacity at least 60-fold before the end of the decade.
For Koduri, the opportunity is not limited to building more data centers. It also includes developing the chips, systems, and software that convert electricity into AI output. “The opportunities for design teams, systems teams, and software teams in India are tremendous,” he said. “This is a build-out that India should participate in.”
Also read:
Koduri Unveils GPU Hardware IP and Software Startup
Jensen Resigns Over Second Affair
Nvidia Bets Big on China With H200 Push
Oxmiq Raises $35M for GPU IP, Expands Focus to Data Center Design
RELATED TOPICS: AI CHIP, AI INFRASTRUCTURE, CHIP DESIGN, CUSTOM SILICON, DATA CENTERS, GPUS, INDIA, SEMICONDUCTORS
COMPANIES: NVIDIA, OXMIQ LABS
Yashasvini Razdan is the senior India correspondent for EE Times. With a bachelor’s degree in electronic engineering from Mumbai University and a post-graduate diploma in business journalism, she has served in various senior editorial roles in leading Indian electronics and business publications, where she reported on semiconductor innovations, deep tech startups, and the Indian electronics ecosystem. She is passionate about combining her technical know-how and journalistic acumen to chronicle the evolution of India’s electronics ecosystem for a global audience. Follow Yashasvini on LinkedIn
Source: EE Times




