The SpecialistThe man who keeps building the same thing, only bigger.
Inside every chip is a city. Cores are neighborhoods. Memories are warehouses. The interconnect is the road system. Sailesh Kumar designs roads.
He is on his fourth iteration of the same essential problem. At Cisco and Huawei, he laid down the routing logic for a 400-gigabit network processor. At NetSpeed Systems, the company he co-founded in 2011 with Sundari Mitra, he turned that thinking inward and pioneered what the industry now calls software-defined fabric - the idea that a chip's internal network should be configured the way you'd compose software, not the way you'd lay bricks. Intel bought the company in 2018 and promptly made him a Fellow, a title that the company hands out sparingly to people whose ideas it wants to keep in the building.
He stayed five years. He worked on the Xeon platform, helped shape Intel's data-center compute protocols, and watched the industry careen into a problem none of the old playbooks were built for: AI accelerators with appetites so large that the chips couldn't be made big enough to feed them. The answer, increasingly, was to stop making one big chip and start gluing many smaller chips together. The chiplet era. Which meant that the road system - the fabric, the interconnect - had to do something it had never had to do before. It had to leave the chip.
In March 2023 he left Intel and started Baya Systems.
The pitch, compressed
Baya's flagship is called WeaveIP. The metaphor is intentional. A modern AI system-on-chip might contain a half dozen distinct on-chip networks - one for the CPU complex, one for the GPU cluster, one for memory, one for I/O - and each of those networks was historically designed by hand, by a different team, in a different style, with bespoke protocols. WeaveIP threads those separate fabrics into a single intelligent global fabric, generated from software the way an FPGA bitstream is generated from Verilog. You describe the constraints. The tool emits the road system. Then it does the same trick across multiple chiplets, so two dies on a package can talk to each other as if they were one.
Kumar's preferred line about this - delivered at industry events with the dry patience of someone repeating an obvious thing - is that the current approach is brute force. “Designing increasingly complex combinations of CPUs, GPUs, neural network accelerators and other processors is a brute-force solution that the industry cannot rely on forever,” he told the press when Baya emerged from stealth. “It simply comes with too many risks: high re-engineering costs, difficulty scaling and potentially hitting the market with sub-par metrics.”
That a former Intel Fellow is, in effect, telling the industry it builds the wrong way is not lost on his investors. In January 2025, Baya closed a $36 million Series B led by Maverick Silicon and Synopsys, with participation from Matrix Partners and Intel Capital. Synopsys is the dominant EDA vendor on the planet. Intel Capital is, of course, Intel's. The cap table is a kind of unspoken endorsement: even the incumbents think the fabric problem needs to be solved differently.
Two hundred patents and counting
Kumar is what the academic-industrial complex used to politely call a prolific researcher. He has more than 150 patents to his name and roughly two dozen highly cited papers on switch fabrics, deadlock-free routing, and network-on-chip microarchitecture. He holds a B.Tech. from IIT Kanpur and a Ph.D. from Washington University in St. Louis, where his thesis work focused on the algorithmic side of high-performance packet processing - the kind of math problems where a single clever data structure can save you a 20% area penalty on a chip the size of a thumbnail.
His career is unusual in that it has the shape of a single curiosity, pursued through escalating venues. The 400G processor at Huawei was the same kind of routing problem at line-rate scale. The software-defined fabric at NetSpeed was the same problem moved inside a system-on-chip. The Xeon platform work at Intel was the same problem at server-class scale. The chiplet fabric at Baya is the same problem, again, this time spanning multiple pieces of silicon. He has been, in effect, writing the same book in larger and larger fonts for two decades.
Customers who matter
The earliest visible win is Tenstorrent, the RISC-V AI startup Jim Keller runs. Tenstorrent is building accelerators it hopes will be cheaper and more open than what Nvidia ships, and a critical part of that bet is the fabric between cores. Tenstorrent uses Baya. There are other customers - the company has talked about “breakout customer growth” through 2025 - but most are still under NDA, which is what you would expect from a company whose product is buried so deep inside other peoples’ chips that no one outside the engineering team ever sees it.
This is the part of the semiconductor industry that operates in near-total silence. The interconnect inside an AI accelerator does not have a brand. It does not get a launch event. It does not show up on a spec sheet. It is the dark matter of silicon, invisible until it isn’t. Kumar has been working in this exact place for his entire career.
A founder who is not in a hurry
If you watch Kumar speak - the SBT C-Suite Spotlight interview is the longest piece of him on record - he comes across as the engineer he is. He talks in clean sentences, slightly slower than the host. He is comfortable with technical specifics and faintly impatient with hand-waving. He is, in the way of certain second-time founders, completely uninterested in performing founderhood. There is no founder mythology, no origin story about a moment of clarity in a coffee shop. He saw the problem. He left. He built the company.
What he wants, he has said in different ways across different interviews, is for chiplet integration to feel less like surgery and more like assembly. Designers should be able to pick a CPU complex from one vendor, a memory subsystem from another, an accelerator from a third, and have them work together at full bandwidth on day one - without a six-month interconnect respin. That is the world Baya is trying to manufacture into existence. Whether the industry gets there depends on a great many things; whether Kumar keeps shipping the pieces that make it possible is, on the evidence of the last twenty years, the closest thing to a sure bet in the EDA stack.
In the chip industry, the people who matter most are the ones whose names you never have to learn. Sailesh Kumar is one of them.