D-Matrix Adopts NVLink Fusion and Nvidia Slips 2.5%
Nvidia fell 2.5% after inference-chip startup d-Matrix said it will use NVLink Fusion to plug its Raptor processors into Nvidia rack systems, with racks due in 2027.
Nvidia shares fell 2.5% on September 10, 2026 after d-Matrix, a $2 billion inference-chip startup, said it would adopt Nvidia's NVLink Fusion interconnect to link its Raptor inference processors to Nvidia rack-scale systems, with compatible racks expected in 2027.
Nvidia Corp. (NVDA) slipped 2.5% on Thursday after d-Matrix, a chip startup valued at $2 billion, said it would build its Raptor inference processors to connect into Nvidia's rack-scale systems using NVLink Fusion, the company's licensable version of its proprietary high-speed chip-to-chip interconnect. Compatible racks are expected in 2027, according to GuruFocus.
At the last trade recorded at 17:37 GMT on September 10, 2026, Nvidia changed hands at 218.42, down 2.35% from the previous close of 223.67, with an intraday range of 217.20 to 220.99. The move came on a soft tape: the S&P 500 tracker (SPY) was down 0.50% at $758.58, the Nasdaq 100 tracker (QQQ) off 0.84% at $710.32, and the Dow tracker (DIA) down 0.61% at $520.86. Nvidia was falling roughly three times as fast as the broad index and faster than the tech-heavy Nasdaq gauge, which is the part of the day's action that is specific to the company rather than to the market.
What NVLink Fusion actually opens up
NVLink is the fabric that makes a rack of Nvidia GPUs behave, for practical purposes, like one very large processor. Memory and data move between chips at speeds far above what a standard network can carry, and that is much of the reason large training and inference jobs run efficiently on Nvidia hardware rather than on assemblies of cheaper parts. For years the fabric was closed: to use it, you bought Nvidia silicon.
NVLink Fusion changes the terms. It lets third-party accelerators and CPUs attach to Nvidia's rack-scale architecture. That is a deliberate strategic trade. Nvidia keeps the rack, the switching, the software stack and the systems business, and in exchange it gives up part of the exclusivity that made the fabric a wall around its own compute. d-Matrix is the kind of partner that makes the trade visible: it does not sell general-purpose GPUs, it sells inference silicon designed to run already-trained models cheaply, which is precisely the workload where buyers most want an alternative to paying GPU prices.
Why an adoption announcement reads as a negative
On the surface this is a partner win. A startup is standardizing on Nvidia's interconnect, which extends Nvidia's architecture into hardware Nvidia did not build. Ecosystem gravity is worth something, and companies rarely mind when rivals adopt their plumbing.
The market's reflex ran the other way, and the logic is not hard to follow. Nvidia's margins have never rested on the GPU alone; they rest on the fact that the fastest way to get a working AI rack has been to buy an Nvidia rack, whole. Once the fabric admits foreign accelerators, a data-center operator can in principle keep the Nvidia system and swap some of the silicon inside it for cheaper inference parts. Every socket that changes hands is revenue that moves from Nvidia's income statement to somebody else's, even if the rack still carries an Nvidia badge.
Inference is where that risk concentrates. Training runs are enormous, bursty and performance-obsessed; buyers there tolerate premium pricing. Inference is continuous, cost-sensitive and increasingly the larger share of spend as deployed models multiply. A specialist part that is good enough at inference and materially cheaper per query has an obvious commercial argument, and NVLink Fusion removes the integration obstacle that used to blunt it.
The 2027 timeline gives the argument room to age
Nothing in this changes near-term revenue. Compatible racks are expected in 2027, which means design work, qualification and customer testing before a single unit ships. Between now and then, Nvidia has multiple product cycles and the accumulated advantage of a software stack that developers already target.
That distance cuts both ways for shareholders. It means the competitive threat cannot show up in the next few quarterly reports, so anyone selling on Thursday was repricing a structural story rather than reacting to a demand signal. It also means the announcement functions as a marker: the closed-fabric era is being wound down on Nvidia's own initiative, and investors now have a dated event on which to hang the question of what Nvidia's share of an open rack looks like.
The wider context is a market in which inference-specific silicon has been drawing capital and attention. A $2 billion valuation on a startup whose whole proposition is running models more cheaply than a GPU tells you where private money thinks the next cost fight happens. Nvidia's answer — license the fabric, keep the system — is a reasonable one, but it is an answer to a pressure that did not exist a few years ago.
What to watch from here
Three things will determine whether Thursday's 2.5% decline was noise or an early read.
- Who else takes the license. One startup is a data point. If several accelerator vendors and CPU makers plug into NVLink Fusion, the rack becomes a genuinely mixed environment and the pricing conversation changes.
- Whether hyperscalers buy mixed racks. The large cloud operators have their own internal silicon programs. An open fabric makes it easier for them to combine that silicon with Nvidia systems instead of choosing between them.
- How Nvidia prices systems versus chips. If a growing share of rack value sits in switching, networking and software, Nvidia can defend revenue even as some accelerator sockets go elsewhere. The disclosure detail investors will want is how much of the systems business survives that mix shift.
For now, the arithmetic on the day is modest: Nvidia gave back 5.25 points from its previous close, a decline steeper than the benchmark move but well inside the stock's ordinary range of daily travel. The narrative shift is the more durable item. Nvidia has spent this cycle selling an integrated system that was difficult to disassemble. It has now published the instructions for attaching somebody else's chip to it, and the first company to take them up is one that sells the workload Nvidia most needs to keep.
Key facts
- NVDA last trade: 218.42, down 2.35% (as of 17:37 GMT, Sept 10, 2026)
- Previous close: 223.67
- d-Matrix valuation: $2 billion
- Compatible racks expected: 2027
Frequently asked questions
What is NVLink Fusion?
NVLink Fusion is a licensable version of Nvidia's NVLink interconnect, the high-speed fabric that links chips inside a rack so they behave like a single large processor. Fusion allows third-party accelerators and CPUs to attach to Nvidia's rack-scale architecture, rather than restricting the fabric to Nvidia's own silicon as the proprietary version did.
Why did Nvidia stock fall on the news?
Nvidia fell 2.5% because opening its rack interconnect to outside accelerators weakens the exclusivity that made buying a complete Nvidia rack the default choice. If customers can slot cheaper inference chips into an Nvidia system, some silicon revenue shifts to competitors even though the rack and its software remain Nvidia's.
What is d-Matrix and what does Raptor do?
d-Matrix is a chip startup valued at $2 billion that designs processors for AI inference — running models that have already been trained, rather than training them. Its Raptor inference processors are the products it plans to connect to Nvidia rack-scale systems through NVLink Fusion, with compatible racks expected in 2027.
When will these mixed racks be available?
Compatible racks are expected in 2027. That leaves design, qualification and customer testing ahead, so the arrangement has no effect on Nvidia's near-term revenue. Investors selling on the announcement were repricing a longer-term structural question about Nvidia's share of an open rack, not reacting to a change in current demand.
How did Nvidia's move compare with the broader market?
Nvidia was down 2.35% at 218.42 as of 17:37 GMT on September 10, 2026. The S&P 500 tracker SPY fell 0.50% to $758.58, the Nasdaq 100 tracker QQQ fell 0.84% to $710.32, and the Dow tracker DIA fell 0.61% to $520.86, so Nvidia underperformed both the broad and tech-heavy benchmarks.
Why is inference the vulnerable part of Nvidia's business?
Training workloads are large, intermittent and performance-driven, so buyers accept premium pricing. Inference runs continuously at scale and is highly cost-sensitive, which gives specialist chips that are cheaper per query a clear commercial argument. An open interconnect removes the integration barrier that previously made those alternatives harder to deploy.
Sources
Photo: U.S. Department of Energy · CC0 1.0 — source
