It is a strong case, but not one that settles the argument. The GPU is still the part that does the computing. What competitors find hardest to match today is the system around it: NVLink and NVSwitch joining 72 GPUs into one rack-scale domain, NVIDIA's InfiniBand and Spectrum-X networking between racks, and the CUDA and NCCL software that makes it all behave as one machine. Industry analysts such as Dell'Oro say NVLink leads scale-up deployments today. Competitors, open standards bodies and the same analysts also expect the gap to narrow, with Dell'Oro forecasting that Ethernet will become the leading scale-up technology by the end of the decade. Here is the evidence on both sides.
How does NVLink compare with competing scale-up interconnects?
| Interconnect | Bandwidth per accelerator (as published) | Accelerators per domain | Openness | Status (Nov 2026) |
|---|---|---|---|---|
| NVIDIA NVLink 5 (GB300 NVL72) | 1.8 TB/s bidirectional (900 GB/s each way) | 72 per rack | Proprietary; NVLink Fusion licenses it to partner CPUs and accelerators | Shipping |
| NVIDIA NVLink 6 (Vera Rubin NVL72) | 3 TB/s (revised down from 3.6 TB/s) | 72 per rack | As above | "Available Now", ramping |
| AMD Helios (UALink over Ethernet) | 260 TB/s per 72-GPU rack (AMD; direction not stated); 3.6 TB/s bidirectional per GPU per StorageReview | 72 per rack | Open standards, Broadcom Tomahawk 6 switches | Volume deployments from 2H 2026 |
| Google TPU7x Ironwood ICI | 1,200 GB/s bidirectional | 9,216 chips per pod (3D torus with optical switching) | Google Cloud only | Available in Google Cloud |
| AWS Trainium3 NeuronSwitch | Not published per chip | 144 chips per UltraServer | AWS only; Trainium4 to use NVLink Fusion | Available in AWS |
| UALink 1.0 specification | 800 Gb/s per direction per x4 port | Up to 1,024 | Open consortium standard | Chips in labs 2H 2026, products 2027 |
Source: NVIDIA NVLink and Vera Rubin NVL72 pages (15 Sep and 21 Oct 2026; revision reported by TechRadar, 31 Oct 2026), AMD Helios page (1 Sep 2026), StorageReview (23 Jul 2026), Google Cloud TPU7x documentation (29 Oct 2026), AWS Trainium page (26 Oct 2026), UALink 1.0 white paper (Apr 2025) and The Register (7 Apr 2026). Vendors count bandwidth differently; figures are not directly comparable without converting to per accelerator, per direction.
What is the case that NVLink is the real advantage?
Three arguments come up most often.
It turns a rack into one large GPU. In a GB300 NVL72, each GPU has 18 NVLink links, one to each of 18 NVSwitch chips, so any GPU reaches any other in a single switch hop at full speed. NVIDIA's ConnectX-8 network adapter, by comparison, carries 800 Gb/s, or 100 GB/s per direction, about one ninth of a GPU's 900 GB/s NVLink bandwidth per direction. Communication-heavy techniques such as tensor parallelism and wide expert parallelism for mixture-of-experts models need that bandwidth. DeepSeek, working on H800 systems, described a computation-to-communication ratio of about 1:1 for cross-node expert parallelism and asked hardware vendors to unify scale-up and scale-out networks. See how GB300 NVL72 connects 72 GPUs.
It is a system, not a part. ABI Research argued in August 2024 that UALink is "years behind NVLink" and that NVIDIA's moat is system-wide: NVLink plus networking plus software. The software layer includes CUDA, the NCCL collective communication library that frameworks use to move data between GPUs, SHARP in-network reductions in NVSwitch and InfiniBand switches, and the management tools for each. A competitor needs all of these to work together, not just a fast link.
Networking is now a large business for NVIDIA. NVIDIA reported US$15 billion of data-centre networking revenue for the quarter to 26 April 2026, "nearly tripled" year over year, against US$60 billion of compute. Networking grew another 18% sequentially the following quarter. The dollar figure for that quarter was not in NVIDIA's results release, which moved to a new reporting split.
NVLink's strength is less the link speed itself than the fact that the switch, the network, the libraries and the GPU are designed together and ship together.
What is the case against?
Competitors are building equivalent domains. AMD's Helios rack puts 72 MI455X GPUs in one scale-up domain over UALink running on Ethernet, using 12 Broadcom Tomahawk 6 switches. Google's Ironwood links 9,216 TPUs per pod over its own inter-chip interconnect and optical circuit switches. AWS links up to 144 Trainium3 chips through NeuronSwitch. These are proprietary or vendor-specific too, but they show that large scale-up domains are not unique to NVIDIA.
Open standards are arriving. UALink 1.0 was published on 8 April 2025, targeting up to 1,024 accelerators, and version 2.0 on 7 April 2026. Its backers include AMD, AWS, Google, Intel, Meta and Microsoft. The Ultra Ethernet Consortium's 1.1 specification of 12 October 2026 adds a profile for short-range scale-up links. Dell'Oro expects Ethernet to lead scale-up by the end of the decade. First UALink chips are only reaching labs in the second half of 2026, though, so this is a future challenge rather than a current one.
NVIDIA is opening NVLink itself. NVLink Fusion, announced in May 2025, lets partners' CPUs and custom accelerators join an NVLink domain. AWS says it is designing Trainium4 to use NVLink 6 and NVIDIA's MGX rack architecture. Jensen Huang presented it as opening "NVIDIA's AI platform and rich ecosystem for partners". It can also be read as a response to large customers that want to use their own silicon alongside NVIDIA's networking and racks.
The GPU still matters. No interconnect helps if the accelerators are slower or harder to program, and NVIDIA's lead in software support is an argument about the GPU platform as much as the link. Roadmaps also change: NVIDIA's current pages list NVLink 6 at 3 TB/s per GPU and 216 TB/s per Vera Rubin rack, down from the 3.6 TB/s and 260 TB/s shown at CES in January 2026, without a published explanation.
How much does the interconnect matter in practice?
It depends on the workload. Microsoft's DeepSpeed team measured tensor-parallel communication at 22–47% of training iteration time across GPT-3 and Llama 2 sizes on one to four DGX H100 nodes. Inference decode, by contrast, is mostly limited by each GPU's memory bandwidth, so a fast scale-up domain matters most for very large mixture-of-experts models served across many GPUs. For a team running a single GPU, NVLink to other GPUs matters not at all; what matters is the CPU-to-GPU link, which on Grace Blackwell systems is the 900 GB/s NVLink-C2C. See NVLink vs InfiniBand vs PCIe and what 800 Gb/s and 130 TB/s really mean.
What should buyers take from the debate?
- For rack-scale training and large model serving, NVLink-based systems are the most widely deployed option today, and the comparison to make is total system cost and delivered throughput, not link speed alone.
- For cloud users, TPU and Trainium offer alternative scale-up domains without owning hardware; portability of your code between them is the practical question.
- For multi-vendor strategies, watch UALink and UEC products through 2027 and ask suppliers for interoperability evidence rather than specification compliance alone.
- For deskside development, DGX Station's GB300 has the same Blackwell Ultra GPU and NVLink-C2C as an NVL72 node, but no NVLink to other GPUs; two units connect over 400 Gb/s Ethernet. See DGX Station's networking.
For the infrastructure beyond the interconnect, see why an AI data centre needs more than GPUs.
Buying in Singapore
SourceIT lists six DGX Station GB300 systems, all price on request: the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01, Supermicro ARS-511GD-NB-LCC and Dell Pro Max with GB300. For PCIe-based workstations and servers, the NVIDIA RTX PRO 6000 Blackwell Workstation Edition is S$29,100 and the Server Edition S$32,999 (listed 6 Nov 2026); these cards connect over PCIe. Multi-GPU servers such as the Lenovo ThinkSystem SR780a V3 are quoted on request. We do not list NVL72 racks or InfiniBand equipment; ask us to source them. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
What is NVLink? NVIDIA's high-speed GPU-to-GPU interconnect. On Blackwell and Blackwell Ultra it provides 1.8 TB/s of bidirectional bandwidth per GPU, and NVSwitch chips join up to 72 GPUs in one rack.
Is NVLink an open standard? No. It is proprietary, though NVLink Fusion licenses it to partners' CPUs and custom accelerators. UALink is the main open alternative for scale-up links.
Does AMD have an NVLink equivalent? AMD uses Infinity Fabric within its accelerators and, in the 72-GPU Helios rack, UALink running over Ethernet with Broadcom Tomahawk 6 switches. AMD quotes 260 TB/s per rack without stating the direction.
When will UALink products be available? According to The Register's report on the 2.0 specifications in April 2026, version 1.0 chips reach labs in the second half of 2026, with products during 2027.
Do I need NVLink for AI inference? Not for models that fit on one GPU. It matters most for very large models split across many GPUs, especially mixture-of-experts models that exchange data between GPUs at every layer.
