Altman-backed Volantis bets $88 million that lasers, not copper, will break AI's memory wall

Volantis Semiconductor wants to solve the memory wall by ripping the wires out of the problem entirely. The San Francisco startup, which counts Sam Altman among its early backers, announced an $88 million Series A on October 1, co-led by Lachy Groom and Abstract Ventures, bringing its total raised to $97 million. The pitch, detailed in coverage from Unite.ai, is a chip package called the A-1 that uses micro-VCSEL lasers and optical waveguides instead of copper traces to move data between processors and memory.

The problem Volantis is chasing is the one that increasingly caps how fast an AI accelerator can actually run a model: not raw compute, but how quickly data can shuttle between the processor and the memory stacked around it. Copper interconnects hit physical limits on bandwidth and power as chips pack more high bandwidth memory closer to the compute die. Volantis claims its photonic approach can connect as many as 220 memory chips around a single processor, packing more than 10TB of memory onto one package at up to 240TB/s of bandwidth, enough, the company says, to serve trillion parameter mixture of experts models at up to 10,000 tokens per second per user.

Those are target specifications from a company that has not shipped a product. Volantis is aiming its first customer deployments at 2027, which means the headline comparison it is making now, claiming 15 times better tokens per dollar and 6 times better tokens per watt than Nvidia's Rubin platform on large low latency MoE models, pits an unreleased chip against another chip that is itself still ramping. Rubin is slated for a second half 2026 launch with 288GB of HBM4 per GPU and 22TB/s of memory bandwidth, specifications detailed in Nvidia's own developer blog. Both sides of Volantis's marketing comparison are essentially paper specs.

The harder question is whether photonic interconnects can be manufactured at the yields and reliability that a data center actually needs. Laser attachment, waveguide signal loss, thermal drift, and repairability at scale are the kinds of problems that sink promising interconnect technologies before they reach volume production, and Unite.ai's reporting on the raise flags all four as unresolved. Swapping a failed HBM stack on a conventional accelerator is already a delicate operation; doing the same on a package wired with hundreds of optical waveguides and micro-lasers introduces failure modes the industry has little production experience fixing at scale.

Volantis's executive team reads like a bet that this problem is solvable by people who have already built pieces of it. Chief executive Tapa Ghosh, a Thiel Fellow and Y Combinator alum with a background in 3D chip technology, founded the company in 2022. Chief technology officer Roy Meade joined as the first employee after leading HBM development at Micron and then working at Ayar Labs, a rival silicon photonics interconnect company. That lineage matters more than it might look: Meade is applying memory manufacturing experience from inside one of the industry's largest HBM suppliers to a photonics approach built, in part, on what he learned at a direct competitor.

The timing of the raise is what gives it weight beyond one startup's funding round. In December 2025, Marvell announced it would acquire Celestial AI, the best funded independent company building photonic memory interconnects, for as much as $5.5 billion including earnouts, a deal that closed February 2, 2026, according to CNBC's reporting. Celestial AI had raised $589 million and built its Photonic Fabric technology around 14.4Tbps of bandwidth before Marvell folded it into its own roadmap. A large chipmaker paying up to $5.5 billion to own this exact technology category, eight months before Volantis closed its much smaller round, functions as a market signal that the memory wall thesis itself is no longer speculative, even if any individual company's chip is.

That consolidation also narrows the field Volantis has to compete in. Lightmatter is still building its Passage photonic interconnect independently, but Celestial AI, the company closest in scale and ambition to what Volantis is proposing, is now a Marvell asset rather than a free agent. Volantis, at $97 million raised against Celestial AI's $2.5 billion valuation before acquisition, is roughly two funding stages behind where its most direct comparable company stood when a strategic buyer decided the technology was worth owning outright.

What Volantis actually has to prove before any of its bandwidth numbers matter is manufacturability at the volumes AI data centers require, not a demo board. The company's own roadmap puts that test off until 2027. Until then, the $88 million buys engineering time to attach lasers to silicon and run waveguides around packages that, if they work, promise to make the copper wire on an accelerator board look like the bottleneck it was all along.