The d-Matrix Deal Explains Why Nvidia Dropped — and How It Plans to Own Inference

Nvidia's stock slid about 2.5% on Thursday, and the headline doing the talking — a "$2 billion chip startup" adopting one of Nvidia's own technologies — reads like the threat thesis that has followed this company for three years. d-Matrix, a Microsoft-backed startup valued at $2 billion that makes its own AI chip, said it would use Nvidia's NVLink Fusion to drop that chip straight into Nvidia's data-center racks. On its face, that looks like the moment custom silicon finally starts tearing down the Nvidia moat, coming precisely as the AI market shifts to the one place Nvidia has always been most exposed.

The deal's actual structure points the other way. d-Matrix is not replacing Nvidia's silicon. It is building its chip to plug inside an Nvidia rack, surrounded by Nvidia CPUs, NVLink switches, data-center processors, and Ethernet networking. This one announcement, buried mid-cycle, is a cleaner illustration of how Nvidia plans to monetize the next phase of the AI era than any earnings call so far.

The cycle is flipping from training to inference

The context that makes the news matter is a transition already underway. For most of the AI boom, the money was in training — building the giant models, a months-long process that runs on Nvidia's GPUs and, critically, on CUDA, the software layer that made Nvidia's hardware nearly impossible to swap out. Training is where Nvidia's moat is strongest.

Inference is different. Running a model for every user, every query, every day is a latency- and cost-driven business: the winner is the chip that answers fastest for the lowest cost per token, not necessarily the most flexible one. That is exactly the kind of workload where purpose-built silicon, custom accelerators from hyperscalers and startups, can undercut a general-purpose GPU. Nvidia already acknowledges the shift — the startup's own press materials describe workloads shifting from training AI models to running them, and the chip at the center of Thursday's news, d-Matrix's Raptor, is built purely for inference.

So when a startup with its own inference processor announces it is plugging into Nvidia's infrastructure, the instinctive read is: the dark horse is arriving, and Nvidia is being turned into plumbing. That instinct is wrong about what d-Matrix is doing — but not entirely wrong about what it signals.

Reading the deal as it's actually structured

Here is the detail that carries the whole judgment. A single AI answer is produced in two phases. The "prefill" phase, which reads the question and plans the answer, is compute-heavy and favors Nvidia's GPUs. The "decode" phase, which emits the answer token by token, is latency-sensitive — each new word must return in a fraction of a second. In the d-Matrix system, Nvidia's GPUs handle prefill and d-Matrix's Raptor chip handles decode. The two are wired together in the same rack through NVLink Fusion, a high-bandwidth, low-latency connective technology that lets custom processors sit alongside Nvidia's GPUs, memory, and networking.

That division is complementarity, not competition. d-Matrix gets access to customers that would never have risked a non-Nvidia chip on its own, and Nvidia keeps every piece of the system around the compute — the rack, the interconnect, the CPUs, the networking, the management software. The startup is one partner among several in the same play: Nvidia previously folded Marvell in with a $2 billion investment, and weeks ago it put $3.5 billion into MediaTek's convertible bonds to bring MediaTek's custom design work under the NVLink Fusion umbrella. Nvidia is not fighting custom silicon; it is making itself the connective tissue that custom silicon has to live inside.

This is the training-to-inference transition made physical. Nvidia's response to losing the argument over who makes the inference compute is to own everything around it, and to keep a slice of every rack regardless of which chip inside does the final work. Hardware sets the ceiling; the ecosystem sets the multiple.

What the 2.5% drop actually measures

Set next to that structure, Thursday's selloff is a measure of the narrative, not of the operating result — and the operating result does not even exist yet. Nothing ships for a year. d-Matrix's Raptor chip is slated to tape out by the end of 2026, and the combination racks are not expected until late 2027. The near-term return curve is unchanged by the news.

The reasonable worry underneath the slide is worth naming, because it is real. Inference wants efficiency, and d-Matrix's whole pitch — a 3D DRAM-stacked chip claiming a step-change in low-latency inference, already under evaluation at hyperscalers and frontier labs — is the kind of unit-economics story that has beaten incumbent generations before. If custom inference silicon wins on cost per token, Nvidia's capture per workload shrinks even when its total system revenue grows. NVLink Fusion is a hedge against that, but a hedge that concedes margin: better to make 100% of a smaller slice of the system than fight for a shrinking share of the compute itself.

For a holder, the disciplined reading is simple. Nvidia does not need a monopoly on the accelerators to keep compounding — but this is the clearest confirmation yet that the inference era is contested, and that the company is deliberately trading per-workload margin for system-level position. The number to watch is not next week's price. It is whether those 2027 racks ship on schedule with real hyperscaler commitments behind them, because that is the moment the strategy stops being a claim and starts being revenue. Nvidia is betting a year and a half out that owning the rack beats owning the chip. Thursday's drop says investors haven't fully decided which is worth more.