Sandlogic Pursues Nvidia-Scale Stack With In-House Chips

Sandlogic, a Bengaluru AI startup founded in 2018 by Kamalakar Devaki, is shifting from B2B enterprise AI services to building a vertically integrated edge AI stack that includes custom silicon, models, and a runtime platform. The company reports total external funding around $4 million plus government grants, and last reported revenue near Rs 1 crore. Its stated goal is to approach the performance profile of Nvidia's A100 using a bespoke Krsna AI co-processor while delivering developer portability through a Cuda-like runtime, EdgeMatrix.

Sandlogic is combining three technical pillars to deliver a full-stack edge offering.

• •A custom co-processor, Krsna AI, positioned for inference and edge workloads.
• •Small and vision-language models, marketed as Shakti SLMs and VLMs, optimized for on-device deployment.
• •A portability/runtime layer, EdgeMatrix, intended to let developers "build once, run anywhere" across heterogeneous edge hardware.

The stack targets low-latency, resource-constrained inference scenarios typical in manufacturing, logistics, fintech, and healthcare. Existing customer projects include document parsing, lead scoring for fertility clinics, and voice agents for call centers, which demonstrate Sandlogic's data and domain expertise. Matching A100 class throughput or versatility is a high bar; the A100 is a datacenter GPU optimized for large-batch training and large models. Sandlogic's realistic near-term path is to optimize for edge-specific performance metrics: power efficiency, quantized inference, model compression, and low-latency I/O.

The move reflects two converging trends: enterprises want tighter hardware-software integration for edge AI, and regional startups are seeking sovereignty over AI infrastructure. Nvidia's advantage is not only silicon but also developer lock-in through Cuda and a vast ecosystem. Sandlogic acknowledges this and is explicitly attempting to replicate the software portability and toolchain benefits that make Nvidia ubiquitous. However, competing with incumbent chipmakers requires deep IP in RTL, fabrication partnerships, compilers, runtime optimization, and long sales cycles for edge devices.

Financially, Sandlogic is making a classic deep-tech tradeoff: short-term profitability for long-term platform value. The company has operated profitably in the past but has recorded losses in the last two years while investing in chips and models. With $4 million in external funding, the runway for a chip-first strategy is limited; chip development typically demands tens to hundreds of millions to reach production at scale.

Execution hinges on three variables: fundraising or strategic partnerships for silicon tapeout and manufacturing, demonstrable performance gains on real edge benchmarks versus existing accelerators, and adoption of EdgeMatrix by third-party developers or OEMs. If Sandlogic secures fabrication partners, a clear cost-performance lead on targeted workloads, and early OEM integrations, it could become a credible regional alternative for edge AI. If not, the pivot risks draining resources and eroding the enterprise business that funded earlier growth.