Every large language model training run lives or dies on one unglamorous piece of plumbing: the wires connecting thousands of GPUs together. Nvidia's NVLink has quietly become the toll booth on that highway, and a wave of well-funded startups now believes they can tear it down. For any founder or CTO planning a multi-year AI infrastructure budget, that fight is not a technical curiosity — it is a direct line item.
What is the Concept
NVLink is Nvidia's proprietary interconnect that lets GPUs inside a server, and increasingly across servers, share memory and data at speeds far above standard networking gear. It is fast, but it is closed: buying into NVLink means buying into Nvidia's roadmap, pricing, and supply timelines. Startups such as Enfabrica, Celestial AI, and Ayar Labs, alongside the broader UALink industry consortium backed by AMD, Broadcom, and hyperscalers, are building open alternatives using optical interconnects, silicon photonics, and Ethernet-based scale-up fabrics. The pitch is simple: match NVLink-class bandwidth without the single-vendor tax.
For a decision-maker, this is less about chip architecture and more about a classic build-versus-buy dependency problem, playing out at the physical layer of the data center.
Why It Matters Now (2025–2026 Context)
GPU scarcity defined 2023 and 2024. In 2026, the constraint is shifting to interconnect bandwidth and power budget, because clusters are now large enough that the network between chips, not the chips themselves, is the bottleneck. Nvidia's pricing power over NVLink-connected systems has become a board-level concern for any company running large training or inference clusters, since it compounds every hardware refresh cycle.
The UALink 1.0 specification, ratified with backing from AMD, Google, Meta, Microsoft, and Broadcom, is the clearest signal yet that the industry wants a second option. When hyperscalers fund a standard, it usually ships.
How AI Is Changing This
Ironically, the AI workloads driving demand for these interconnects are also what makes switching away from NVLink harder in the short term. Distributed training frameworks are tuned against Nvidia's software stack, and moving to an open interconnect requires re-validating performance at scale, not just swapping cables. That said, inference workloads — which now represent a growing share of AI compute spend — are more portable, and that is exactly where these startups are landing their first deals.
Contrarian take: the real winner of this race will not be whichever startup builds the fastest interconnect. It will be whichever one builds the most boring, drop-in-compatible one, because enterprise infrastructure teams do not reward speed alone — they reward the lowest-risk migration path.
Real-World Examples
Microsoft and Meta have both publicly signaled interest in scale-up fabrics that do not depend on a single GPU vendor, partly to preserve negotiating leverage with Nvidia on future chip pricing. Broadcom's Ethernet-based Tomahawk switches are being positioned explicitly as an NVLink-scale alternative for AI clusters. Smaller players like Astera Labs have built a multi-billion-dollar valuation almost entirely on connectivity chips that sit adjacent to, not inside, Nvidia's stack — proof that the interconnect layer alone is now investable as its own category.
Practical Insights / Actions
The founder mistake we see most often: teams negotiate hard on GPU unit price but never model the interconnect and networking spend that rides along with it, then get surprised when the total cost of a cluster refresh comes in 20 to 30 percent over budget.
Future Outlook
By 2027, expect most hyperscale AI clusters to run a hybrid interconnect strategy — Nvidia's proprietary fabric for the highest-performance training pods, and open, UALink-style fabrics for inference and mid-tier training. The hidden opportunity for SMEs and mid-market AI companies is that this competition will push interconnect pricing down across the board, even for buyers who never touch a UALink product directly, simply because Nvidia will have to respond.
Conclusion
The race to replace NVLink is really a race to decide who controls the economics of AI infrastructure for the next decade. Companies that start modeling interconnect risk and cost today, rather than treating it as an afterthought to GPU procurement, will be the ones negotiating from strength when the next cluster refresh comes due. If your team is planning a 2026 infrastructure buildout, an interconnect-aware cost audit is a far cheaper insurance policy than a mid-contract vendor renegotiation.