One of the most stubborn bottlenecks in artificial intelligence hardware is not how fast a chip can compute, but how fast it can talk to its memory. On Thursday, San Francisco-based semiconductor startup Volantis said it has raised $88 million in venture capital to fix exactly that problem by swapping tiny electrical wires for beams of laser light.
The limitation, described by Reuters in its report on the announcement, is physical. Whether the compute silicon comes from Nvidia or Advanced Micro Devices, the AI model lives in the memory chips and must constantly feed data back to the computing chip. Both companies encircle their processors with expensive high-bandwidth memory chips, but even Nvidia's best current GPUs can hold only eight of those memory chips each, because the microscopic wires connecting them simply cannot reach further.
From eight memory chips to 220
Volantis says its optical interconnect removes the reach problem entirely. By sending data between computing chips and memory chips using beams of laser light instead of copper wiring, the company believes it can pack as many as 220 memory chips around a single GPU, a dramatic expansion over today's designs.
The startup's chosen technology is deliberately unglamorous: vertical-cavity surface-emitting lasers, or VCSELs, the same tiny lasers that power facial-recognition features in hundreds of millions of iPhones.
Apple invested heavily in building out that supply chain over the past decade, which means Volantis can tap mature, mass-produced components rather than inventing exotic new ones.
“Advanced packaging is always to be respected - it's never trivial - but it's not necessarily a new thing to do,” said Tapa Ghosh, the company's CEO and co-founder.
“No one is going to win a Nobel Prize if our project works, but the good news is, they won't need to.”
A bet on boring technology
By combining well-understood components, Volantis hopes to sidestep the supply-chain bottlenecks that are choking the rest of the AI chip industry. Memory shortages have already rippled through the consumer market this year, with rising DRAM prices forcing smartphone makers to raise prices on flagship devices. Ghosh told Reuters the company aims to deliver a chip next year that would speed up tasks such as AI coding, where large models must be held close to the compute engine to run efficiently.
Heavyweight backers join the round
The round was led by Stripe veteran Lachy Groom and Abstract Ventures, with participation from John Doerr, VXI Capital, Triatomic and Susa Ventures. Angel investors include AI podcaster Dwarkesh Patel, AI chip veteran Naveen Rao and Anthropic researcher Sholto Douglas, according to Times New.
Investors have poured record sums into semiconductor startups this year as data centers race to build ever-larger AI systems. Volantis is betting that the next big leap in AI performance will come not from faster processors, but from finally solving the memory wall that keeps them waiting for data.
