Ex-Intel, AMD execs’ chip startup Agrani Labs seeks $100m

Semiconductor startup Agrani Labs is in early talks to raise about US$100 million in a series A round.

Founded by four former Intel and AMD executives, the company is building AI inference chips compatible with Nvidia’s CUDA software stack.

Potential backers include Qualcomm, Battery Ventures, and government entities.

Existing investor Peak XV may invest about US$20 million on a pro-rata basis.

The company was launched in 2025 and raised US$15 million from Peak XV last year at a US$36 million valuation.

🧠 Food for thought

Implications, context, and why it matters.

Agrani Labs is building more than a chip. It aims to make a full-stack AI compute platform with its own software, compilers, and libraries 1.
That strategy gives it a shot at challenging Nvidia’s CUDA software ecosystem, a programming platform used to run AI workloads on Nvidia chips that can tie customers to Nvidia hardware and software 2.
The founding team includes former Intel and AMD executives. Vinod Dham, known in the chip industry as the “Father of the Pentium,” is a founding advisor, which adds industry credibility 1.
Government interest could fit with India’s Research, Development & Innovation (RDI) Scheme, a 1 lakh crore rupee fund (US$12 billion) meant to build “technological sovereignty” in sectors such as semiconductors 3.

Cost is pushing demand for new AI hardware. Running inference, using a trained AI model to generate answers or predictions, can cost as much as US$75 per million tokens through APIs 4.
Well-tuned full-stack platforms can raise throughput by 2 to 4 times. If Agrani succeeds, that would support a move toward cheaper inference systems that widen access to AI 4.
An Indian AI chip would add another option to the global tech supply chain. That would ease dependence on a small group of firms and geopolitical flashpoints, while supporting Agrani’s goal of making India a credible player in semiconductor design 2.
Agrani still faces a fast-moving market. Nvidia projects steep gains across generations, with Rubin set for 50 petaflops of FP4 tensor performance versus 10 petaflops for Blackwell B200 5.