Volantis, a San Francisco semiconductor startup, raised $88 million in Series A funding to build a photonic interconnect it says will break the memory wall that limits AI inference. Lachy Groom and Abstract Ventures co-led the round, with participation from John Doerr, VXI Capital, Triatomic and Susa Ventures. Angel investors include Dwarkesh Patel, Naveen Rao and Sholto Douglas.
The round brings total funding to $97 million. Earlier backers include Sam Altman, Jeff Dean and Dylan Patel.
The company plans to deliver its first integrated inference engines to customers in 2027. Its first system, called A-1, is designed to run models exceeding 20 trillion parameters at up to 10,000 tokens per second per user while cutting the cost of inference per token. Those targets sit well past what today’s GPU clusters offer, and independent verification of them is still some way off.
The memory wall, explained
AI inference is increasingly bottlenecked by memory rather than compute. Model weights are large, and GPUs hold only so much memory on chip and in attached stacks. When that pool fills, operators chain accelerators across networks and pay in latency and bandwidth for every hop.
Memory capacity and bandwidth have historically been a trade-off. Add more memory through off-chip DRAM and bandwidth per byte drops. Add high-bandwidth memory and capacity stays limited and expensive. Every major accelerator vendor lives inside this constraint, and data center builders pay for it in floor space and power.
Volantis claims to sidestep it with light. The company builds a photonic interconnect designed specifically to connect compute chips to memory. Its optical fabric ties large numbers of memory chips into a unified pool, aggregating bandwidth as memory is added, so capacity and bandwidth grow together. SiliconAngle reports the design delivers more than 30 times the memory bandwidth of current accelerators.
How the hardware works
The interconnect wires stretch over 200 millimeters, a range long enough to attach more than 220 memory chiplets to a single accelerator. Because the links transmit data as light, distance stops being the tax it is on copper runs.
The light comes from custom micro-VCSELs, tiny vertical-cavity surface-emitting lasers. A VCSEL uses a quantum well to turn some of the electricity running through the host chip into light, with two mirrors amplifying the signal. Volantis draws on the existing gallium arsenide VCSEL supply chain rather than indium phosphide, avoiding a known supply bottleneck in photonics. The company says its micro-VCSELs are small, temperature-stable and low power, enabling end-to-end links below one picojoule per bit.
The A-1 appliance is about a third the size of a standard server rack. It carries 10 terabytes of memory and 250 terabits per second of memory bandwidth. That footprint matters for data center operators, since memory capacity is a big driver of server real estate in inference fleets. A rack that serves more users per square meter is worth more than one that serves fewer.
| A-1 specification | Value |
|---|---|
| Target model size | 20 trillion+ parameters |
| Generation speed | Up to 10,000 tokens per second per user |
| Memory | 10 TB |
| Memory bandwidth | 250 Tb/s |
| Size | About a third of a standard server rack |
| Customer delivery | 2027 |
The team behind it
The founding team includes semiconductor and photonics veterans from Nvidia, AMD, Broadcom and Ayar Labs. Their prior work spans the first CoWoS chip packaging product, the first high-volume tunable VCSELs and early silicon photonics co-packaged optics systems.
CEO and co-founder Tapa Ghosh leads the company. The founder lineup explains the round’s roster. Patel, Rao and Douglas are among the most watched names in AI infrastructure investing and research, and each has argued publicly that inference economics, not training, will decide which AI companies survive the next funding cycle. Douglas came from DeepMind’s training infrastructure work, and Rao was the founding CEO of Nervana Systems, the deep learning chip startup acquired by Intel in 2016. Their presence cuts both ways: it signals serious conviction, and it raises the bar for what counts as proof.
Why investors are betting on photonics now
The raise tracks a shift in where AI spending is going. Training builds dominated capital flows for years. As frontier models move into production, the bill has shifted to inference, and heavy request queues against oversized models have made memory bandwidth the binding constraint at major providers. Serve a frontier model cheaply and fast, or lose the user to whichever provider does.
Several companies are attacking the same wall from different directions. Nvidia’s Rubin generation pushes HBM4 stacks and NVLink 6 bandwidth. Google’s TurboQuant work attacks memory use through compression, a software answer to a hardware problem. Memory makers, whose quarterly results beat expectations this week, are ramping capacity through 2028 on the strength of AI demand. Each approach shrinks the problem from a different side.
Volantis takes the architectural route: change how chips reach memory, rather than cramming more memory onto the accelerator. If it works, it reshapes what a data center buys. If it does not, the 2027 ship date and the targets above will be the reason named in the postmortem. Longer supply chains have unmade many a hardware startup, and photonic interconnects have several years of overpromising behind them. Co-packaged optics has been two years away for most of the last decade.
What happens next
The company says the new capital funds development and commercialization of A-1 and the photonic memory architecture, including expanding the engineering team and moving toward customer deployments. No pricing, pilot customers or foundry partner have been named yet. Data center operators have seen ambitious memory specs before, and the test will be whether A-1 runs real frontier workloads rather than marketing benchmarks.
The broader question is whether inference really becomes a hardware market distinct from training. If model sizes keep doubling, the answer is probably yes, and companies like Volantis will have a market to sell into. If compression and smaller efficient models keep pace with demand instead, the memory wall becomes negotiable and the premium for new hardware thins out. Both outcomes remain live. The $88 million bet says one of them pays.