The demand for local AI inference is changing the requirements of AI-enabled SoCs. Indian startup HrdWyr is building its SoCs around the interaction of AI, real-time control, signal processing, and data movement.
The Bengaluru-based fabless semiconductor company, founded by Ramamurthy Sivakumar and Guruswamy Ganesh in 2023, is developing AI-native SoCs initially aimed at power and battery management applications. EE Times spoke with CEO Sivakumar, who believes AI workloads require architectures designed around AI from the outset rather than adding machine-learning capability to a conventional processor.
“A conventional SoC with an ML accelerator cannot deliver the level of performance required for advanced AI workloads,” Sivakumar said. “More importantly, it cannot provide the power efficiency needed to run those workloads effectively.”
HrdWyr closed a $13 million Series A funding round, led by Ideaspring Capital, with participation from Singularity AMC, Avatar Growth Capital, and Persistent Systems in May 2026. In the press statement, the startup said the funding will accelerate development of its AI-native SoCs and expand customer engagements across key global markets.
View All HrdWyr’s architecture is designed to keep deterministic functions on predictable hardware paths while using AI for adaptation, optimization, and diagnostics. “Most conventional SoCs treat AI like an afterthought, simply bolting an AI accelerator onto a standard CPU setup,” Sivakumar said. “We’ve built AI, real-time control, signal processing, and data movement into one seamlessly integrated system.”
The architecture is tailored to specific applications rather than designed as a general-purpose AI platform. Sivakumar said the company’s approach can be tailored to applications such as power conversion, motor control, and security, allowing the AI models to be optimized for the target workload.
He described the company’s approach through an internal framework called True Edge Autonomy (TEA). This framework is anchored by an agentic AI stack built on reinforcement learning to extract the full performance of its AI SoCs.
“The real power of AI will be unlocked as we enter the era of physical AI, where advanced intelligence integrates with real-world systems,” Sivakumar said. “This inflection point demands a fundamental rethinking of how computing systems are conceived, architected, and deployed. It calls for a structural reset of the semiconductor stack, with IP and products designed for AI processing from first principles, not as an afterthought.”
How HrdWyr builds AI-native SoCs
Sivakumar said HrdWyr’s differentiation comes from combining multiple functions, including system integration and its AI engine, within a single device. “Our approach may not necessarily be different in every respect, and that is perfectly acceptable,” he said. “As this industry develops, there is room for many successful companies.”
Sivakumar identified system integration as HrdWyr’s first differentiator. Instead of relying on separate battery charger ICs, battery fuel gauges, and microcontrollers, the HrdWyr chip combines multiple functions into a single mixed-signal SoC.
“We have integrated functions that would otherwise require multiple chips,” he said. “We have also added analog capabilities that are not available in existing products.”
The company said this tighter integration is intended to improve performance at the system level. For system designers, the expected benefits include lower control-loop latency, reduced central processing unit (CPU) utilization, less data movement, and lower power consumption. The architecture is also intended to adapt to component aging, changing operating conditions, and product-to-product variations. According to Sivakumar, these benefits could translate into lower thermal requirements, smoother control, and less overall system complexity.
The startup sees its second differentiator in HrdWyr’s AI engine, which is designed around its initial target applications in battery and power management. These systems can change continuously with temperature, battery age, load, and user behavior. Sivakumar said fixed rules and static lookup tables can struggle to account for these changes over a product’s lifetime.
HrdWyr proposes using reinforcement learning to manage long-term tradeoffs, such as charging speed versus battery life or peak performance versus thermal stress. For battery management, the startup sees reinforcement learning as a way to make charging and power behavior adapt to actual operating conditions and usage patterns rather than relying solely on fixed rules.
However, the approach is not intended to replace conventional safety mechanisms. Sivakumar said HrdWyr’s reinforcement learning operates within a verified safety envelope, while standard protection circuits and deterministic control continue to handle safety-critical functions.
Battery management is HrdWyr’s first application, but Sivakumar said the same reinforcement learning approach could be used in other systems that need to continually adapt to changes caused by internal and external conditions.
“Our chip is a highly mixed-signal design,” he said. “It contains extensive digital circuitry along with substantial analog circuitry, all integrated onto a single piece of silicon using the same manufacturing process.”
Sivakumar said the real engineering challenge lies less in designing AI algorithms than in integrating multiple IP blocks into a manufacturable chip.
“Selecting AI models and partitioning workloads between edge devices and the cloud is comparatively straightforward,” he said. “The harder task is building multiple complex IP blocks and integrating them into a complete SoC. Even after the design is complete, converting it into a GDSII database is a highly demanding process.”
The startup is also developing its own software stack to make its different processing elements easier for system designers to use. The compiler is intended to distribute workloads across the CPU, control, math, and AI engines. The SDK provides optimized libraries and reference applications, while the runtime handles scheduling, synchronization, and data movement. HrdWyr is also developing tools for profiling, debugging, simulation, and model deployment.
Sivakumar believes this will allow engineers to focus on their algorithms rather than manually manage each hardware block.
He declined to identify the external foundry that manufactures these chips, along with the process node, citing commercial sensitivity. HrdWyr selected its foundry nearly a year before tape-out because preparing a GDSII database for manufacturing requires close collaboration with the foundry. The startup has completed tape-out of its first chip and is now in the board development stage, awaiting packaged silicon from its outsourced semiconductor assembly and test (OSAT) partner, Tata Electronics, for mounting on reference boards.
Sivakumar declined to disclose unit volumes, revenue, or the size of the initial production batch. HrdWyr has about 15 employees, including its engineering team. He described HrdWyr as a fabless semiconductor product company developing end-to-end products for global markets rather than licensing IP or operating as a services business.
From battery management to physical AI
The startup’s first product family, known as Indus, starts with a power and battery management IC designed for true wireless stereo (TWS) headsets. More broadly, the device targets products powered by single-cell lithium-ion batteries.
A second member of the Indus family, targeting brushless direct current (BLDC) motor control applications, is already under development.
HrdWyr’s first customer is boAt, a popular Indian consumer electronics brand that designs, markets, and sells personal audio equipment and smart wearables. Sivakumar said about half of HrdWyr’s customer engagements are in India and half are overseas, although he declined to identify the remaining customers.
For OEMs, HrdWyr is targeting applications where AI needs to work alongside physical systems, including motors, batteries, power electronics, and sensors. It is positioning its products against the more general-purpose platforms offered by established suppliers such as Texas Instruments, NXP, Renesas, Qualcomm, and Nordic Semiconductor. Rather than replacing the deterministic control methods already used in such systems, HrdWyr said it intends to add adaptive intelligence to them.
“We aren’t here to replace the deterministic control the industry relies on—we are here to supercharge it with adaptive intelligence, all under one roof,” Sivakumar said.
Sivakumar said the company is deliberately keeping its initial customer list focused rather than spreading its resources across a large number of engagements. The aim is to make those early customers successful and then build a broader merchant semiconductor business.
“The longer-term opportunity lies in the move from screen-based, data center-driven AI to physical AI,” he said. “The latter increasingly need to sense, make decisions, and respond locally.”
He pointed to applications ranging from drones and humanoid robots to manufacturing systems, security cameras, and other devices that interact continuously with their surroundings.
For HrdWyr, this makes market selection as important as the underlying technology. Sivakumar said one of the biggest strategic risks is building a product for a market that has already moved on by the time the product is ready. “If you make the wrong product decision, this industry is unforgiving,” he said.
The startup also faces the more immediate challenge of recruiting experienced engineers and building an international presence as it grows. India does not have a long track record of producing semiconductor end products at scale, and HrdWyr is among a small group of startups attempting to build products and brands from India for global markets.
Beyond HrdWyr itself, Sivakumar said he would “much rather see 40 companies like HrdWyr emerge in India” than have HrdWyr remain the only one. For him, the development of a broader semiconductor product ecosystem is part of the opportunity, rather than simply a backdrop to HrdWyr’s own growth.
For now, HrdWyr’s business is concentrated on power and battery management, with motor control next. Its larger ambition is to turn that application-specific approach into a global semiconductor product business. Whether it can do so will depend on execution, customer adoption, and the pace at which [AI moves into physical systems](https://www.embedded.com/the-moment-your-edge-ai-becomes-physical-ai/).
For Sivakumar, however, the objective remains unchanged. “If we are not ultimately competing globally with our own products, then we are simply creating value for somebody else.”
Read also:
[SUSE Positions Hardware Choice as a Core Part of Sovereign AI](https://www.eetimes.com/suse-positions-hardware-choice-as-a-core-part-of-sovereign-ai/)
[India’s OSAT-ATMP Build-Out: From Legacy Packages to 2.5D](https://www.eetimes.com/indias-osat-atmp-build-out-from-legacy-packages-to-2-5d/)
[Meta Cuts Server Count 25% by Reusing Old Memory: Can Anyone Else Do It?](https://www.eetimes.com/meta-cuts-server-count-25-by-reusing-old-memory-can-anyone-else-do-it/)
[STMicroelectronics Bets on Hardware-Based Post-Quantum Cryptography](https://www.eetimes.com/stmicroelectronics-bets-on-hardware-based-post-quantum-cryptography-with-st54m/)