d-Matrix licenses Nvidia’s NVLink interconnect for upcoming Raptor chips
AI infrastructure startup d-Matrix has agreed to license Nvidia's NVLink Fusion interconnect technology and adopt its MGX rack reference designs, joining a growing roster of chipmakers—including Qualcomm, Arm, Marvell, Amazon, Fujitsu and MediaTek—that have struck similar deals. The move lets d-Matrix build its future accelerators, including the upcoming Raptor chips, around Nvidia's racks and NVSwitch fabrics rather than developing its own interconnect from scratch, making its inference hardware easier for customers to deploy alongside existing Nvidia infrastructure. It also strengthens Nvidia's position as a broader infrastructure supplier, since it stands to earn revenue from licensing and complementary hardware even when its own GPUs aren't being sold.
By the end of next year, d-Matrix aims to offer systems linking up to 144 Raptor-based accelerators via a single NVLink fabric, delivering roughly 2.3TB of memory capacity and about 7.2 petabytes per second of aggregate memory bandwidth—enough to handle models with more than four trillion parameters at 4-bit precision. The chips use 3D-stacked DRAM bonded directly to compute logic, aiming for SRAM-like bandwidth with much greater capacity than rivals such as Groq, meaning far fewer chips are needed per model. d-Matrix also plans to pair its accelerators with other Nvidia products, including Vera CPUs, NVSwitch appliances, BlueField and ConnectX NICs, and SpectrumX Ethernet gear, deepening its reliance on Nvidia's ecosystem.
- d-Matrix licenses Nvidia's NVLink Fusion and MGX rack designs for its chips.
- Future systems will link up to 144 Raptor accelerators via NVLink.
- Deal expands Nvidia's influence even without direct GPU sales.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Supporters of this approach argue that d-Matrix is making a sensible, pragmatic choice by adopting Nvidia's NVLink and MGX standards rather than reinventing interconnect technology from scratch. By plugging into an already-dominant ecosystem, d-Matrix can get its distinctive memory-dense Raptor chips to market faster and make them immediately deployable alongside the Nvidia infrastructure that most large customers already run, rather than asking buyers to adopt a bespoke, unproven fabric. On this view, competing on genuine technical differentiation, such as 3D-stacked memory bandwidth and capacity, while outsourcing commodity plumbing like interconnects, is a rational allocation of a startup's limited engineering resources.
The case against
Sceptics of this approach would counter that by tying its future chips so closely to Nvidia's proprietary interconnect and rack designs, d-Matrix risks becoming structurally dependent on a company that is also, in effect, a rival supplier of AI compute. Such deep integration could constrain d-Matrix's pricing power, limit its ability to differentiate its platform architecturally in future, and further entrench Nvidia's already dominant position across the AI infrastructure stack, making it harder for genuinely independent alternatives to emerge. From this perspective, the growing roster of chipmakers licensing Nvidia's technology reflects less a series of individually sound business decisions and more a worrying consolidation of control over the industry's underlying standards.
AI Art Business Celebrity Companies Culture Entertainment Technology
Read the full article at the source →
Originally published by The Register as “d-Matrix drinks the Nvidia Kool-Aid with NVLink Fusion and MGX rack designs”.