San Francisco semiconductor startup Volantis said Thursday it has raised $88 million in venture capital to attack one of the central bottlenecks in AI chip design: the physical distance over which computing chips can talk to memory.

The problem is well known in the industry. The fundamental limit for AI chips from Nvidia and rivals such as Advanced Micro Devices is how efficiently the computing portions of a chip communicate with the memory chips where an AI model resides and feeds data to the compute logic. Nvidia and AMD address the issue by encircling their computing chips with expensive high-bandwidth memory.

Even Nvidia's best current offerings, however, can accommodate only eight high-bandwidth memory chips per graphics processing unit. The constraint comes from the limited reach of the tiny electrical wires that physically connect the memory chips to the GPU package.

Volantis claims a way around that ceiling. The company is developing technology that sends data between computing chips and memory chips using beams of laser light. Optical links eliminate the reach problem of electrical interconnects, and Volantis says the approach would let it pack 220 memory chips around a single GPU — nearly 28 times the count achievable in Nvidia's current top-end configurations.

Familiar components, aggressive packaging

To build the optical links, Volantis is tapping vertical-cavity surface-emitting lasers, or VCSELs. The devices sound exotic, but they already ship at massive scale: they power the facial recognition features in many Apple devices. Apple invested heavily in bolstering the VCSEL supply chain over the past decade, which means the components are relatively well understood and available.

By combining established parts rather than inventing new ones, Volantis hopes to sidestep some of the supply chain bottlenecks that have slowed AI chip development across the industry. The company plans to deliver a chip next year that would speed up tasks such as AI coding, said Tapa Ghosh, CEO and co-founder of Volantis, in an interview.

"Advanced packaging is always to be respected - it's never trivial - but it's not necessarily a new thing to do," Ghosh said. "No one is going to win a Nobel Prize if our project works, but the good news is, they won't need to."

The framing signals the company's strategy: the engineering risk lies in integration and packaging, not in inventing fundamentally new physics. VCSELs, advanced packaging and high-volume memory are all mature disciplines. The bet is that combining them correctly yields a step change in memory bandwidth per GPU.

Investor lineup

The $88 million round was led by Lachy Groom, a Stripe veteran, together with Abstract Ventures.

John Doerr — the venture capitalist who backed Google and Amazon.com in their early days — also participated. Additional institutional investors include VXI Capital, Triatomic and Susa Ventures.

The round also drew angel investments from figures close to the AI engineering community: AI podcaster Dwarkesh Patel, AI chip veteran Naveen Rao, and Anthropic researcher Sholto Douglas.

Why memory reach matters

The funding lands on a specific pressure point in AI hardware economics. AI models live in memory and must stream data continuously to compute logic. When bandwidth between the two falls short, expensive GPUs sit idle waiting for data. That inefficiency drives the industry's heavy spending on high-bandwidth memory, which has become one of the tightest components in the AI supply chain.

Nvidia's current packaging limit of eight memory chips per GPU reflects the physics of electrical signaling over short copper interconnects. Signal integrity degrades with distance, which forces designers to place memory as close as possible to the compute die. Optical interconnects sidestep that constraint, since light can carry data over longer distances within a package with far less loss.

If Volantis can validate its approach — packing 220 memory chips optically around a GPU — the architecture would represent a significant departure from the packaging strategies of the incumbent vendors.

Timeline and open questions

Volantis has committed to delivering a chip next year. The company has not yet publicly detailed the chip's bandwidth specifications, memory type, power budget, or target workloads beyond AI coding.

Key engineering hurdles remain. Integrating hundreds of optical links into a production-grade package raises questions of yield, thermal management and cost that the company will need to answer before customers can evaluate the technology against established high-bandwidth memory solutions from the major vendors.

Still, the round closes with backing from investors with a track record of identifying infrastructure shifts early, and the startup enters a market where every percentage point of memory bandwidth improvement carries direct commercial value for AI operators.

The announcement was made Oct. 1 from San Francisco.