Volantis Secures $88M to Break AI's Memory Wall Using Light-Based Chip Architecture

Volantis Tackles AI's Memory Bottleneck with Photonic Interconnects

San Francisco-based AI chip startup Volantis has raised $88 million in Series A funding to address one of the most pressing challenges in AI hardware: the memory wall2

2

4

4

. The funding round was led by Stripe veteran Lachy Groom and Abstract Ventures, with participation from prominent investors including John Doerr, who backed Google and Amazon in their early days, VXI Capital, Triatomic, and Susa Ventures2

2

. Angel investors include AI podcaster Dwarkesh Patel, AI chip veteran Naveen Rao, and Anthropic researcher Sholto Douglas5

5

. Founded in 2022 by CEO Tapa Ghosh, a Thiel fellow and former Y Combinator founder, and CTO Roy Meade, who previously ran Micron's high-bandwidth memory programme and served as vice president at Ayar Labs, Volantis is developing photonic inference systems that could reshape AI accelerator performance3

3

.

Breaking Through the Shoreline Constraint with Laser-Based Communication

The fundamental limit for AI chips from Nvidia and Advanced Micro Devices centers on memory bandwidth—how efficiently computing parts communicate with memory chips where AI models reside2

2

5

5

. Current solutions encircle computing chips with expensive high-bandwidth memory, but Nvidia's best offerings can hold only eight memory chips per GPU due to the limited reach of electrical wires that connect them2

2

. These tiny electrical interconnects extend just 5 millimeters, creating what's known as the "shoreline" bottleneck4

4

. AMD's MI455X currently caps at 432 GB because of this physical perimeter constraint1

1

. Volantis is developing technology to send data between computing chips and memory chips using beams of laser light, eliminating the reach problem and enabling the system to pack 220 memory chiplets around a GPU2

2

4

4

. The photonic interconnects can extend memory placement from a few millimeters to over 200 millimeters, effectively eliminating the shoreline as a bottleneck1

1

4

4

.

VCSELs Power the Photonic Memory Layer Strategy

Volantis is tapping vertical-cavity surface-emitting lasers, or VCSELs, to power its photonic memory layer2

2

4

4

. While these tiny lasers may sound exotic, they already power facial recognition features in hundreds of millions of iPhones, with Apple having invested heavily in bolstering the supply chain over the past decade2

2

5

5

. A VCSEL comprises three main components: a quantum well that converts electricity into light, and two mirrors that amplify it4

4

. Rather than relying on external lasers, Volantis integrates micro-VCSELs directly into the photonic interposer, with the light source baked into the chip architecture1

1

. VCSELs are easier to manufacture and often cost less than lasers used in optical networking devices4

4

. The design deliberately uses gallium arsenide lasers to sidestep constraints in indium phosphide, of which China holds 70% of supply3

3

. Indium phosphide wafer prices rose approximately 250% by June, with IQE's chief executive calling access a key industry risk3

3

.

The A-1 Chip Targets Real-Time Frontier AI Inference

Volantis plans to ship its chips with a data center inference appliance called the A-1, about a third the size of a standard server rack4

4

. The company believes it can pack upwards of 10 terabytes of memory at speeds up to 240 terabits per second onto a single massive package1

1

4

4

. This should be enough to serve 20 trillion parameter models at 10,000 tokens per second per user1

1

4

4

. According to CEO Tapa Ghosh, the technology provides more than 30 times the memory bandwidth of current accelerators4

4

. "This will enable real-time frontier inference, restart scaling laws & enable entire code bases in context windows," Ghosh wrote, adding, "As a starting point, imagine a coding agent that completes a task in 30 seconds rather than 30 minutes"4

4

. The company targets 1 picojoule per bit per 24 Gbps lane and total system power around 20 kilowatts1

1

. Volantis plans first customer deliveries in 20273

3

4

4

.

Licensing Strategy Differentiates Volantis from Photonics Competitors

Volantis is taking an unusual approach by focusing differentiation primarily on the optical interposer while licensing everything else from IP providers, even if it's not the highest performance available1

1

. "The differentiation is primarily coming from the optical interposer. So why don't we license absolutely everything we possibly can, even if it's not the highest performance in the world, because we're already biting off a lot of risk with the interposer and the integration," CEO Ghosh explained1

1

. Unlike competitors Celestial AI and Lightmatter, which license their technology to other chipmakers, Volantis aims to develop a fully fledged inference chip rather than pursuing an IP play1

1

. "It is, at least on day one, a way easier problem to demonstrate a solution that ships," Ghosh said1

1

. The founding team includes engineers who previously worked at Nvidia and Broadcom, with technical achievements including the first commercial implementation of CoWoS, an interconnect widely used in GPUs4

4

.

Wide-and-Efficient Philosophy Challenges Fat Pipe Mentality

CTO Roy Meade describes Volantis's approach as "wide and efficient," arguing that interconnects can achieve higher bandwidth through more links rather than just increasing link speed1

1

. "People are still constrained to this, what I would call a fat pipe point-to-point mentality, and naturally, if that's the case, you're going to go towards SerDes," Meade said, pointing to HBM as an example where higher bandwidth is achieved using an extremely-wide bus but slower data rate per pin relative to other memory technologies1

1

. By bringing together relatively well-known components, Volantis hopes to avoid supply chain bottlenecks in developing AI chips and deliver a chip next year that would speed up tasks like AI coding2

2

5

5

. "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"2

2

5

5

. The technology's reliability has improved significantly in recent years, with higher reliability achievable by sacrificing some speed1

1

. Volantis still needs to raise substantially more venture capital beyond the $88 million to tape out its first production-ready chip1

1

. The company's success could accelerate the shift toward photonic solutions in AI hardware, particularly as demand for processing massive parameter models continues to grow and traditional electrical interconnects reach their physical limits.