The wiring underneath AI is becoming as contested as the chips themselves.
If you follow AI infrastructure conversations on X (formerly Twitter) right now, you'll notice a shift. The discourse has moved past which GPU is fastest and landed squarely on what connects them. Interconnects — the high-speed fabrics that link GPUs within a node and across a cluster — have become one of the hottest technical debates of 2026. And as that debate heats up, a quiet but powerful force is stepping in to fill the gaps: the compatible hardware market.
The Interconnect Wars: What's Actually Happening
For years, NVIDIA's NVLink was the undisputed king of scale-up interconnects. If you wanted to run a multi-GPU workload inside a single node — training a large language model, running inference at scale — NVLink was your only real option at the performance tier that mattered. It was fast, it was tightly integrated with CUDA, and it was proprietary. That last word is the one infrastructure teams have been wrestling with.
In April 2026, the UALink Consortium dropped its 2.0 specification, and the industry took notice. UALink is designed to be an open-standard alternative to NVLink — a vendor-neutral scale-up interconnect that any chipmaker can implement. Version 2.0 added in-network compute capabilities and chiplet integration support, bringing it meaningfully closer to NVLink parity. For the first time, OEMs designing next-generation AI servers have a credible open fabric option on the table.
At the same time, NVIDIA made a move that surprised many observers: it opened NVLink itself — at least partially — through a program called NVLink Fusion. Rather than fighting the open standard tide entirely, NVIDIA is allowing select third-party silicon vendors to integrate directly into its NVLink fabric. The marquee example is the $2 billion strategic partnership announced with Marvell Technology in March 2026, where Marvell will supply custom XPUs and NVLink Fusion-compatible scale-up networking components that plug directly into NVIDIA's ecosystem.
The message from both developments is the same: the era of a single vendor owning the entire interconnect stack is ending.
Scale-Out Is a Different Fight — And Ethernet Is Winning
While scale-up interconnects (connecting GPUs within a node) get the headline attention, the battle for scale-out fabric (connecting nodes across a cluster) has reached its own inflection point. This is where InfiniBand and Ethernet have been competing for a decade, and the ground is shifting.
InfiniBand — particularly NVIDIA's NDR (Next Data Rate) generation running at 400Gbps per port — has long dominated serious AI training deployments. Its low-latency, lossless transport and tight RDMA integration made it the default choice for hyperscalers building GPU clusters at scale. But the cost and ecosystem complexity of InfiniBand have always been pain points, particularly for mid-market buyers who can't absorb the per-port pricing of OEM-branded Mellanox infrastructure.
RoCE v2 (RDMA over Converged Ethernet) has emerged as the practical answer for many of these deployments. Running RDMA semantics over standard 400G and 800G Ethernet switches, RoCE v2 delivers the low-latency, high-throughput profile that AI workloads need — without locking buyers into a proprietary switch fabric. As networking vendors have hardened their AI-optimized Ethernet NOS offerings in 2025 and 2026, the case for RoCE v2 at scale has become genuinely compelling.
Where the Compatible Market Steps In
Here's where it gets practical for anyone actually building or expanding an AI cluster today: OEMs set the specs, but the compatible market delivers the economics.
Consider cabling and optics. A production InfiniBand NDR cluster running at 400Gbps needs tens or hundreds of QSFP112 optical transceivers and active optical cables (AOCs). NVIDIA's branded Mellanox LinkX optics carry significant price premiums — appropriate for customers who need OEM warranty coverage and seamless support. But for organizations that understand their infrastructure and want to stretch their capital budget, compatible third-party optics provide 100% tested, MSA-compliant alternatives at a fraction of the cost.
The same dynamic plays out in the secondary market for complete InfiniBand switch systems. A Mellanox QM9700 NDR switch fresh from NVIDIA carries a substantial list price. A certified refurbished unit sourced through a reputable secondary market vendor — tested, re-warranted, and ready to rack — can represent 40–60% savings on the same hardware. For teams standing up inference clusters or building out HPC environments on a defined budget, that delta is the difference between a project that gets funded and one that doesn't.
And with the Hopper-to-Blackwell transition pushing large quantities of HDR and EDR InfiniBand infrastructure onto the secondary market — the same wave that's flooding used H100 supply — buyers right now have an exceptional window to source high-quality interconnect hardware at historically attractive prices.
The NVLink Fusion Effect on Third-Party Components
NVLink Fusion is worth examining more carefully, because it has direct implications for the compatible market. By allowing Marvell and other approved partners to build XPUs and networking silicon that interface directly with NVIDIA's NVLink fabric, NVIDIA has essentially created a new category of "compatible" infrastructure at the chip level. This is a significant philosophical shift from the historically closed NVLink ecosystem.
What it signals downstream is that NVLink compatibility — once exclusively an NVIDIA-controlled specification — is becoming an attribute that third-party hardware can carry. As that ecosystem matures, it creates broader opportunities for compatible switch, cable, and optics vendors to develop and test products against a more open specification. The door is cracking open, and the compatible market is already positioning to walk through it.
UALink's Promise — and the Transition Period in Between
UALink 2.0 is genuinely promising, but the honest reality is that v1.0 silicon has barely shipped and v2.0 is a paper specification today. Enterprise buyers building production AI infrastructure in 2026 cannot wait for UALink to mature. They need solutions that work now, at scale, with proven software stacks.
This is precisely the window where the compatible market provides its most valuable service. During transitions — when the next generation is promising but not quite ready, and the current generation is being decommissioned en masse — the secondary and compatible markets bridge the gap. They make proven, battle-tested interconnect technology accessible to organizations that can't afford new OEM pricing, and they do it with hardware that's already running production workloads at some of the largest AI deployments in the world.
What Buyers Should Know Right Now
The interconnect landscape is more dynamic than it's been in a decade. UALink 2.0 is redefining what open scale-up fabric can look like. NVLink Fusion is opening NVIDIA's previously closed ecosystem to third-party silicon. RoCE v2 on 400G/800G Ethernet is making InfiniBand optional for a wider range of workloads. And the mass decommissioning of Hopper-era infrastructure is creating a flood of high-quality used interconnect hardware.
For infrastructure buyers, the actionable takeaway is straightforward: you don't need to wait for the next generation of open interconnect standards to get excellent AI networking at a price that makes sense. The compatible market — through certified refurbished switches, third-party optics, and tested AOC cables — is delivering OEM-grade performance at a cost profile that changes the math on what your team can build today.
At Resilient Tec, we source and supply the interconnect infrastructure that powers serious AI deployments — from Mellanox NDR and HDR switch systems to compatible QSFP and DAC cables that work natively with your existing Blackwell, Hopper, and AMD Instinct environments. If you're building or expanding an AI cluster and want to talk through the options, we're here to help you get more compute connected for less.