Volantis raises $88 million Series A to attack the AI memory wall with light

Artificial intelligence models have been growing far faster than the hardware that feeds them data. For several years the binding constraint on AI inference has not been raw arithmetic but memory: processors can execute calculations faster than memory can hand them the numbers those calculations require. Volantis, a San Francisco semiconductor company, is betting that the way out is to replace part of the copper between compute chips and memory with beams of light.

Volantis announced on October 1, 2026 that it raised an $88 million Series A round co-led by investor Lachy Groom and Abstract Ventures. The company said the financing brings its total raised to $97 million. The money is earmarked for development and commercialization of its first system, the A-1, a photonic inference machine that connects compute chips to memory optically rather than electrically.

The company was founded in 2022 and is led by CEO and co-founder Tapa Ghosh and CTO Roy Meade. Its founding team includes semiconductor and photonics veterans from NVIDIA, AMD, Broadcom and Ayar Labs. Their previous work, according to Volantis, includes the first CoWoS product, the first high-volume tunable VCSELs and early silicon photonics co-packaged optics systems.

What that team is attacking is often called the AI memory wall. Volantis argues that on-chip SRAM provides high bandwidth but limited capacity, that HBM-based GPUs provide capacity but limited bandwidth, and that even 3D DRAM stays on the same tradeoff curve. The A-1 is designed to increase memory capacity and bandwidth simultaneously, by nearly two orders of magnitude, by pooling many memory chips into a single optical domain.

Key Facts

PR Newswire reported on October 1, 2026 that the round was co-led by Lachy Groom and Abstract Ventures, with participation from John Doerr, VXI Capital, Triatomic and Susa Ventures. The company added that angel investors Dwarkesh Patel, Naveen Rao and Sholto Douglas also took part. Unite.AI reported on October 1, 2026 that the $88 million Series A brings Volantis to $97 million raised to date, a total that also reflects backing from Sam Altman, Jeff Dean, Dylan Patel and John Doerr.

The A-1 is specified as a rack-scale inference appliance. Unite.AI reported on October 1, 2026 that the product page lists 10 terabytes of memory capacity, 240 terabytes per second of memory bandwidth, a 20 kilowatt power envelope, 10 terabytes per second of off-wafer input output bandwidth and a 15U form factor intended to fit inside existing data center racks. Volantis says the system is being designed to run models exceeding 20 trillion parameters at up to 10,000 tokens per second per user while reducing the cost per token.

SiliconANGLE reported on October 1, 2026 that the A-1 appliance is about a third the size of a standard server rack and is described as capable of processing up to 10,000 tokens per second on a 20 trillion parameter model. The same report noted that Volantis claims more than 30 times the memory bandwidth of current accelerators, because its optical interconnect wires reach beyond 200 millimeters, against the electrical wiring that today runs up to 5 millimeters and limits how many memory modules can sit on a GPU.

Crypto Briefing reported on October 1, 2026 that the architecture is designed to connect a GPU to as many as 220 memory chips, versus around 8 chips in a conventional setup, and that the underlying laser technology is already proven in consumer devices such as the iPhone Face ID sensor. Volantis builds custom micro-VCSELs, microscopic vertical cavity surface emitting lasers made from gallium arsenide that fire light straight up off a chip surface and enable end-to-end links consuming less than one picojoule per bit. Using gallium arsenide also lets the company lean on an established VCSEL supply chain and avoid indium phosphide constraints.

The performance claims are comparative. Volantis states the A-1 delivers 15 times better tokens per dollar than NVIDIA Rubin, six times better tokens per watt for low-latency mixture-of-experts models above one trillion parameters, and more than 30 times lower latency for large-model inference. CEO Tapa Ghosh framed the latency point in human terms in his announcement, writing that the improvement is like a coding agent finishing a task in 30 seconds rather than 30 minutes. The $88 million round follows a $9 million seed round in 2025 backed by names including Alex Wang and Trevor Blackwell, and Volantis plans to deliver its first integrated inference engines to customers in 2027.

Analysis

What this really means is that the AI industry next competitive front may be the wire, not the transistor. Compute density has improved steadily, but the distance between a processor and the memory it reads has barely changed, and that distance now sets the ceiling on how fast large models can answer. Volantis is not claiming a faster arithmetic unit. It is claiming a longer, thinner, faster pipe, and selling that pipe as a whole system whose memory pool is measured in terabytes rather than in the handful of stacks that fit around a package.

The assembly of backers is as interesting as the technology. A round co-led by Lachy Groom and Abstract Ventures, with John Doerr, VXI Capital, Triatomic, Susa Ventures, Naveen Rao, Dwarkesh Patel and Sholto Douglas involved, and a cumulative $97 million that traces back to Sam Altman and Jeff Dean, reads like a bet placed by people who think about inference economics rather than chip design. Rao previously ran Intel artificial intelligence products group, so part of that cap table has operated a large accelerator business before.

The bigger picture here is that optical interconnect has spent two decades arriving and never quite landing outside specialized networking. What changed is the workload. Mixture-of-experts models above a trillion parameters are memory bound to an almost absurd degree, and the payoff for moving data optically grows with every parameter added. Volantis is positioning the A-1 as a complement to, or a replacement for, conventional accelerators from NVIDIA and AMD, and it is promising shipments next year. That is a short runway between an $88 million check and customer hardware, and the company has not named customers for the 2027 engines.

There is also a supply chain angle worth watching. By choosing gallium arsenide micro-VCSELs instead of external lasers, Volantis says it draws on an established consumer photonics supply chain and sidesteps indium phosphide constraints. That choice ties an AI infrastructure roadmap to a component base manufactured today at consumer electronics volumes, which is either a clever hedge or a scaling risk depending on how fast demand arrives.

Why It Matters

The AI memory wall is not an abstract engineering concern. It sets the cost of every token an agent generates, and therefore the price of every product built on top of one. If a system can hold a 20 trillion parameter model and serve it at up to 10,000 tokens per second for a single user, the economics of long running agents, code generation and real time reasoning change shape. The claimed 15 times better tokens per dollar than NVIDIA Rubin, and more than 30 times lower latency for large model inference, are the numbers that would decide whether photonics graduates from a research curiosity to a purchase order.

It also matters because the market is not short of capital for anyone with a plausible answer to the bottleneck. The same week saw $88 million flow to a company with no shipping product, on the strength of a founding team, a specification sheet and a supply chain argument. That is a signal about how urgently inference capacity is being sought.

Finally, the outcome will be measured in rack space and kilowatts as much as in bandwidth. A 15U form factor and a 20 kilowatt power envelope mean the A-1 has to fit facilities that already exist, competing for power and cooling with everything else in the building. Optical links that save energy per bit only help if the whole box, including its lasers and its optics, stays inside an operator budget.

Next Up

The near term test is delivery. Volantis says it plans to deliver its first integrated inference engines to customers in 2027, and it has said the A-1 will ship as a system rather than as a standalone chip. Between now and then the company must turn micro-VCSEL production, optical packaging and system software into something a data center operator can install and run.

Watch for named customers, independent measurement of the tokens per dollar and tokens per watt claims, and any detail on how the optical memory pool scales beyond the first configuration. The $88 million buys Volantis time to answer those questions. Whether the memory wall actually cracks will be decided by what shows up in racks.