This Hardware Fix Will Make AI Reasoning Instant | Tapa Ghosh
MTS
October 18, 2026
Volantis co-founder and CEO Tapa Ghosh discusses how using microscopic optical interposers instead of electrical wires breaks the memory bandwidth and capacity bottleneck in AI inference, dramatically speeding up complex reasoning models and coding agents. Turn ideas into software people love.
Tapa Ghosh
If you think about a coding agent run that might take, let's say, four to six hours today, that gets compressed to something like on the order of like five to 15 minutes, which is a pretty colossal difference in like the user experience, at least in my view. And then the other thing you would see for people using chat bots is you would have instantaneous answers instead of it taking a long time to kind of think through things.
SPEAKER_2
All right, we're back. I'm here with Tapa Ghosh, who is the co-founder and CEO of Volantis, a San Francisco chip startup. He's using photonics to break AI's memory bottleneck for inference. He's also a Teal fellow and went through Y Combinator previously, so welcome Tapa. Great to have you on.
Great to be here. Yeah, so tell us about, you just raised $88 million, series A, for Volantis. So tell us more in like plain terms, sort of how photonics can break the memory bottleneck that we see in AI.
Tapa Ghosh
Yeah, so the main reason why memory is a problem at all is because we want to run these giant models, and we also want to run them really fast, which means we basically have to stream the weights and the KB cache really fast from memory. And so the industry solution to this is basically to build this technology called HBM, where we essentially take it and we basically stack up as much memory as we can and we put it right next to the chip. And that actually works pretty well. It's also enormously profitable for the HBM and DRAM industry.
But the problem is kind of reaching limits. And so the reason why it's reaching limit is because fundamentally, if you just think about, like really there's only two things that matter in computers, there's compute memory. And the way that we today connect that memory to compute is we use these itsy bitsy electrical wires that are very small, so you can pack them very tightly. You can get very high density. It's like something like approaching like the entire Internet traffic of like an entire region in such a small area.
But that's also where the drawback is. These electrical wires only travel over a very short distance, and so it really limits the amount of memory you can put on a single chip. And because each memory chip only runs so fast, it also limits how fast you can access its memory. And so like the simplest way to think about what we've done is we've basically built a photonic or optical version of these little electrical wires that are just as tiny. They're the same size, roughly, about half a micron wide, but they really don't have a reach limit. They go 100 times further. And so you can take a lot more memory and attach it to a single chip, and now you can parallelize the speed and bandwidth across all of these different memory chips. And so you basically get to have your cake and eat it too. You can have more memory. You can also have it faster. And also interesting is you get to use slower memories for these memory chips that can be much cheaper potentially in the long run and alleviate some of the memory bottleneck that the broader industry faces.
SPEAKER_2
Right. So when did you first recognize how big of an issue this was and how much of a bottleneck it would end up being? And how did you sort of bring this team together, walk us through that?
Tapa Ghosh
Yeah. So ironically, I think this was in some primitive way, like visible to me as early as 2016 when it was like the first advent of the early TPUs and where the industry is going and making compute really fast. But memory really wasn't improving as much, right? I think that was kind of okay because in those days we were building what would essentially be the infant version of the models we have today.
And then in 2022, I started Volantis when this was still pre-Chat GPT, but it was obvious to anyone who kind of talked to the AI world, these models were going to get really, really big. And so that was really kind of when it was like, okay, we should really do something about this and we should do something about it in a scalable way like instead of a 2X in a one-time improvement or a 4X one-time improvement, it needs to be something that actually fundamentally solves the physics bottleneck and scales for cloud applications. And so, yeah, so it turns out that when it comes to like these little optical wires or optical interposers, my co-founder Roy had previously led the HVM team at the program at Micro-IoN and was a VP at a company called IR Labs, which kind of pioneered early type of integrated optics with the processor. And so he had also been thinking kind of independently along these same lines of how do we solve these memory bottlenecks by using new technologies like optics. And so, yeah, we got to talking, it turned out like this was the first time something that he thought would work and was actually scalable, which was pretty cool.
Read the other 9 minutes of this episode here.
$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.
Not what you expected? Full refund within 14 days.
Thousands of transcripts fetched by people building searchable podcast archives
Fetch transcripts with code
The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000796976546