D-Matrix has spent more than seven years building a memory-centric inference chip, and its chief executive says no other company will be within two years of matching it.
Rather than compete with Nvidia's footprint, D-Matrix is plugging directly into it.
"This is probably, most certainly, the largest deployed infrastructure base for AI in the world."
Sid Sheth runs D-Matrix, which just closed a six-month-plus partnership putting its inference chips inside Nvidia's own ecosystem.
I listened to the full segment so you can skip it.
Here are the 3 takeaways that matter.
👤 Guest: Sid Sheth, chief executive of D-Matrix, a memory-centric AI inference chip company he has spent more than seven years building
🎙️ Host: Ed Ludlow, who anchors Bloomberg Tech from San Francisco
📰 Published: 10 September 2026 on the Bloomberg Tech YouTube channel
🔴 YouTube | ⏱️ 8 min
Key Takeaways
D-Matrix is plugging into Nvidia's install base rather than competing with it
Sheth calls it "the largest deployed infrastructure base for AI in the world" and says the six-plus-month partnership gets D-Matrix to market faster than building its own footprint
Its Raptor XPU skips HBM entirely and 3D-stacks DRAM directly onto the chip instead
Sheth says this addresses what he calls the "memory wall," the industry's biggest constraint on inference latency, energy efficiency and cost
He says no rival will have equivalent memory technology within a two-year window
Raptor is due to reach the market in about 12 months, with customer pull already coming from hyperscalers, frontier labs, sovereigns, inference clouds and high-frequency traders
1. Riding Nvidia's Footprint
Ludlow opened by asking why D-Matrix wants to plug into what is largely Nvidia's own architecture rather than build a rival ecosystem.
The scale of Nvidia's footprint is the whole argument. "This is a landmark announcement for Dematrix, obviously, and the biggest reason we embarked on this partnership with Nvidia, and we've been working on this for six-plus months with them, is exactly what Jensen said in his comments earlier. This is probably, most certainly, the largest deployed infrastructure base for AI in the world."
Riding that base beats building a new one. "Nvidia hardware is in every cloud, every country, pretty much every data center in the world. At D-Matrix, we have a very unique memory-centric inference computing technology. We have been working on this for seven plus years, really targeted for ultra-low latency inference."
He ties the demand directly to agentic coding tools. "Ultra low latency inference computing just took off in a massive way in the last 12 months, post GPT Codecs and Cloud Code and with the arrival of agentic coding, everyone is looking for how they can quickly get access to low latency compute so that they can really interact with these agentic coding tools."
Asked whether Nvidia is making an equity investment in D-Matrix, the way it did with MediaTek through convertible notes, Sheth declined to say: "At this point, Ted, I couldn't comment on the investment piece specifically, but look, this is a partnership that is very strategic to Nvidia. And they are putting a lot of resources, engineering resources."
2. Punching the Memory Wall
Ludlow asked what the partnership tangibly buys D-Matrix that it would not otherwise have.
The headline claim is a first. "Look, the Raptor XPU, which is the specific XPU that's in partnership with Nvidia and will be launched as part of this partnership first, is going to be the world's first 3D-stacked DRAM XPU, right? And just let me touch upon why that is so important. It is important because memory is really the biggest bottleneck for AI right now."
The technical choice is skipping HBM altogether. "And what we have done is instead of using HBM, we don't use any HBM at D-Matrix. What we have done is we have taken DRAM and found a way to package it directly."
He describes the result as solving three problems at once. "And it is 3D stacked exactly like you just saw in the picture there. And that allows for the architecture to punch through the memory wall, which is really the biggest bottleneck in AI today in terms of addressing latency, addressing energy efficiency, addressing cost."
He claims a lead he expects to hold. "We have found a recipe. We have cracked the code on how to do it. And we'll be the first to bring that solution to market. So there is tremendous customer pull for that product right now because no other company is going to be within a two-year window of getting access to that technology."
3. Raptor Ships in 12 Months
The near-term goal is getting the product into customers' hands. "Our first goal is to really make sure that the Raptor XPU that is going to be integrated into the Nvidia ecosystem is going to be ready to come to market in about 12 months. So the biggest opportunity here is really about getting that product in customers' hands because we've got a lot of customer pull for that specific product in the Nvidia ecosystem."
He would not name customers, only categories. "It's a combination, Ed, and we'll be announcing these customers. Stay tuned. It is a combination of hyperscalers, frontier labs, sovereigns, inference clouds, high-frequency traders." A formal announcement is coming "in pretty short order"
Bonus Insights
This is Sheth's own segment of a longer Bloomberg Tech episode built around Apple's iPhone Duo launch, which also carried separate interviews with Mark Gurman on the phone itself and with Positron's chief executive on a separate inference-chip funding round
Sheth's bottom line is that D-Matrix is betting its differentiation is in the memory architecture rather than the compute itself, and that riding Nvidia's existing footprint gets that architecture to market faster than fighting it would.
Products, Companies & Tools Mentioned
D-Matrix (Sheth's company, whose Raptor XPU is the first chip in this Nvidia partnership and is due to reach the market in about 12 months)
Nvidia (Its install base is the reason for the partnership; chief executive Jensen Huang is quoted separately welcoming outside chipmakers onto Nvidia's platform)
MediaTek (Cited by Ludlow as the earlier Nvidia XPU partner that took a direct equity investment through convertible notes, unlike D-Matrix's arrangement)
If this was worth your time, send it to someone who follows the name.
Get the latest market chatter as it happens: