Volantis raises $88 million to link one GPU to 220 memory chips with light

At A-1's target of 240 terabytes per second, a user receiving 10,000 tokens per second gets 24 gigabytes of memory traffic per token ^[18][19]. That is about a quarter of one percent of the roughly 10 terabytes the system is meant to hold ^[20]. A model that had to read all 10 terabytes for every token would top out near 24 tokens per second on the same bandwidth ^[21]. Neither report says whether the 20-trillion-parameter figure and the speed figure are meant to hold on the same model at the same time.

The chip count shows a similar gap. Going from about 8 memory chips per GPU to as many as 220 is a 27.5-fold increase ^[8][15]. Volantis says A-1 raises capacity and bandwidth by nearly two orders of magnitude at once ^[22]. If the company means something close to 100 times, roughly another factor of 3.6 has to come from denser memory chips or faster links per chip ^[24].

The energy claim can be checked against the bandwidth. At 240 terabytes per second, A-1 moves 1,920 terabits per second, and at the company's figure of under one picojoule per bit for an end-to-end link ^[11], carrying all of it would take less than about 1.9 kilowatts ^[16]. That covers the optical links only.

Where the money goes is narrow by design. According to Crypto Briefing, Volantis is not necessarily trying to out-muscle Nvidia and AMD on raw compute, and its bet is that memory access is where the next inference gains come from ^[12]. It uses custom micro-VCSELs in place of external lasers and draws on the existing gallium arsenide VCSEL supply chain ^[7]. The same laser type already ships in the iPhone's Face ID ^[6]. The company says its approach aggregates memory without radical changes to packaging or manufacturing ^[23]. Proceeds are earmarked for A-1, a larger engineering team and preparation for customer deployments ^[13]. According to pulse2.com, the founders' earlier work includes the first CoWoS product, high-volume tunable VCSELs and early co-packaged silicon-photonics systems ^[9].

The step from a $9 million seed in 2025 to this round is close to tenfold ^[10][17]. Angels Dwarkesh Patel, Naveen Rao and Sholto Douglas also came in ^[4].

Tapa Ghosh, the chief executive, built the pitch on a single trade-off. "Today's hardware forces a tradeoff between running the largest, most sophisticated models and running them fast," he said. "We started Volantis to eliminate that tradeoff." ^[14]

I see three ways this goes. A-1 reaches customers in 2027 ^[5] at close to 240 terabytes per second, and the memory-first bet holds. It arrives late or well under spec, and the $97 million ^[2] bought a prototype. Or it arrives on spec and buyers find their models read too much memory per token for the 10,000 figure to apply. I think bandwidth is the fair test, because it does not depend on which model a customer runs. A delivered A-1 measuring far below 240 terabytes per second would prove the thesis wrong. One round at one company is also too little evidence to show that investors in general are moving money from compute to memory.