Published: October 4, 2026 | Category: Startup | By Mahesh
Volantis, a San Francisco chip startup, announced an $88 million Series A on October 1, 2026 to build optical links between AI processors and memory [4]. The round was co-led by Lachy Groom, co-founder of Physical Intelligence, and Abstract Ventures, with John Doerr also taking part [1][2]. Angel investors include OpenAI's Sam Altman, Google's Jeff Dean and SemiAnalysis's Dylan Patel, and total funding now stands at $97 million [2]. The company's own announcement was made on X, so Depth Grid relied on press coverage of it for the details below.
The problem Volantis targets is the memory wall. Large models need memory capacity to hold parameters and context, and they need bandwidth to move that data to the processor fast enough to keep it busy [1]. Today's accelerators top out at around eight memory chips each, because those chips must sit physically close to the processor [3][6]. Volantis proposes replacing part of the electrical link with optics so one GPU can reach as many as 220 memory chips [3]. Reuters, as relayed by Hoodline, put that at nearly 28 times what Nvidia's current chips hold [5].
The first product is an inference system called A-1, with targets of models above 20 trillion parameters, up to 10,000 tokens per second per user, about 10 terabytes of memory capacity and 240 terabytes per second of bandwidth [3][4]. Customer deliveries are planned for 2027 [3]. Every figure is a company target, not a measurement. They are also not consistent across coverage: Hoodline cites Embedded.com for a design that pools more than 100 terabytes of memory per chip at above 30 terabytes per second [5]. Until Volantis publishes a datasheet, treat the specifications as a direction rather than a spec sheet.
"We are solving AI's memory bottleneck by using optics."Tapa Ghosh, chief executive, Volantis, on X, via Dealroom [1]
The technical bet is specific. Instead of keeping lasers outside the chip, Volantis built custom micro-VCSELs, a type of small laser, directly into its silicon and dropped conventional optical fiber for a structure that packs in more bandwidth [2]. Because its design pools memory across chips, it can use cheaper memory while still raising throughput, which the company says also lowers cost per token [2]. One report adds that Volantis chose gallium arsenide over indium phosphide, the material most photonic startups prefer [6]. For readers following what inference actually costs, see our explainer on OpenAI's two newest models.
The field is crowded and well funded. Ayar Labs has lifted its 2026 funding total to $650 million, and Nvidia put $6.5 billion into photonics in just three months [2]. Volantis tries to stand apart by focusing on the link between a processor and its memory rather than links between separate chips or systems [4]. The commercial test is whether it can turn an optical interconnect into a manufacturable system with the promised capacity, bandwidth, cost and reliability [4]. In the coverage I reviewed, I found no independent benchmark of the A-1, and none should be expected before a working system exists.
What to Watch Next
Four signals matter from here. First, a published datasheet that reconciles the conflicting capacity and bandwidth numbers [3][5]. Second, a working A-1 prototype with independent benchmarks, since the whole case depends on measured tokens per second at a real cost per token. Third, named customers for the 2027 deliveries, which would show whether any cloud or model builder is willing to adopt an unproven memory architecture [3]. Fourth, how Nvidia's heavy photonics spending changes the competitive picture [2]. For Indian startups that buy inference through APIs, the useful number to track is whether any of this ever reaches per-token prices, which on the company's own 2027 timeline is still some way off.
Sources
- Dealroom. "Volantis raises $88M to break AI's memory bottleneck with photonics," October 2026. dealroom.co
- WOWTALE. "AI Inference Memory-Wall Breaker Volantis Raises $88M Series A," October 2, 2026. en.wowtale.net
- Quest Review Center. "Volantis secures $88m Series A to revolutionise AI memory access with photonics," October 2026. questreviewcenter.com
- Runtime Wire. "Volantis raises $88M to put optical memory links inside AI systems," October 2026. runtimewire.com
- Hoodline. "San Francisco Startup Volantis Lands $88M to Zap AI's Memory Bottleneck With Light," October 2026, citing Reuters and Embedded.com. hoodline.com
- Byteiota. "Volantis Raises $88M to Give GPUs 220 Memory Chips Instead of 8," October 2026. byteiota.com
Read More on Depth Grid
- What an LLM Actually Is, Explained Through the Two Models OpenAI Just Shipped
- OpenAI Shelved Its Next Model Over Lying. Then Shipped Agents on the Old One Anyway.
- OpenAI Just Refused to Ship Its Own Flagship Model
Article by Mahesh | Depth Grid
Discussion (0)
Comments are disabled for this post.