NVIDIA divides AI GPU cluster interconnects into two layers: scale-up within a compute domain and scale-out across data-center serversCITE:E1. In NVIDIA's framing, scale-out networks connect servers across the data center, while scale-up networks let the GPUs inside a single domain behave as one engine of computeCITE:E1. NVIDIA positions NVLink as "the purpose-built scale-up networking fabric for AI factories"CITE:E1.
NVIDIA states that this layer is not a peripheral detail — all-to-all bandwidth and latency are critical to AI factory performanceCITE:E6. The company specifically warns that if experts in a model sit behind a low-bandwidth or high-latency fabric, the gains from expert parallelism can be erased by communication overheadCITE:E6. That framing is why scale-up bandwidth, not just raw compute, has become a competitive front.
NVIDIA's sixth-generation NVLink provides 3.6 TB/s of bidirectional GPU-to-GPU bandwidth per GPU, scaling to 260 TB/s of rack-level bandwidth across a 72-GPU domainCITE:E2.
At the system level, the NVIDIA GB200 NVL72 connects 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale, liquid-cooled designCITE:E3. NVIDIA states its NVLink Switch System provides 130 TB/s of low-latency GPU communication for AI and high-performance computing workloads within that rackCITE:E3.
| Building block | Entity | Scope | Key figure |
|---|---|---|---|
| NVLink (6th gen) | NVIDIA | Per-GPU, bidirectional | 3.6 TB/s |
| NVLink Switch System | NVIDIA | 72-GPU domain | 260 TB/s |
| GB200 NVL72 rack | NVIDIA | 36 Grace CPUs + 72 Blackwell GPUs | 130 TB/s |
| UALink 1.0 (200G) | UALink Consortium | Up to 1,024 accelerators/pod | 200G per lane |
UALink Consortium's 200G 1.0 specification supports up to 1,024 accelerators within a single AI computing pod, while Ultra Ethernet Consortium addresses a separate layer of the networkCITE:E4CITE:E5.
UALink Consortium describes UALink as "the open scale-up interconnect for next generation AI workloads"CITE:E4. Its UALink 1.0 Specification, published April 8, 2025, "enables 200G per lane scale-up connection for up to 1,024 accelerators within an AI computing pod"CITE:E4. That positions UALink as a direct open alternative to NVLink's scale-up role described above.
Ultra Ethernet Consortium, by contrast, states its goal is to "deliver an Ethernet based open, interoperable, high performance, full-communications stack architecture to meet the growing network demands of AI & HPC at scale"CITE:E5. The consortium explicitly defines "AI at scale" as the "GPU scale-out network"CITE:E5 — the layer NVIDIA itself separates from scale-upCITE:E1. In other words, UALink competes with NVLink at the scale-up layer, while Ultra Ethernet occupies the scale-out layer alongside it rather than against it.
UALink Consortium counts more than 85 member companies backing its open scale-up standardCITE:E7. The consortium states its board includes Alibaba, AMD, Apple, Astera Labs, AWS, Cisco, Google, HPE, Intel, Meta, Microsoft, and SynopsysCITE:E7.
That roster spans cloud operators, chipmakers, and networking vendors, positioned around a single open industry standard defined by the UALink ConsortiumCITE:E7.
What this means: NVIDIA has already shipped a scale-up fabric — sixth-generation NVLink at 3.6 TB/s per GPU and 260 TB/s per domain, running inside the GB200 NVL72 at 130 TB/sCITE:E2CITE:E3 — for the exact layer it identifies as critical to AI factory performanceCITE:E6. UALink Consortium's published 200G-per-lane, 1,024-accelerator specification targets that same scale-up layerCITE:E4, and it carries the backing of 85-plus companies including several of NVIDIA's largest customers and rivalsCITE:E7. Ultra Ethernet Consortium's open stack, meanwhile, is defined against the adjacent scale-out layerCITE:E5, so it does not directly contest NVLink's scale-up bandwidth figures — it sits beside that fight rather than inside it.
NVIDIA divides AI GPU cluster interconnects into two layers: scale-up within a compute domain and scale-out across data-center serversCITE:E1.
NVIDIA's sixth-generation NVLink provides 3.
UALink Consortium's 200G 1.0 specification supports up to 1,024 accelerators within a single AI computing pod, while Ultra Ethernet Consortium addresses a separ…
UALink Consortium counts more than 85 member companies backing its open scale-up standardCITE:E7.
The numbers show why this is a two-front contest, not a single race. NVLink's 3.6 TB/s per-GPU bandwidth and 260 TB/s domain-level figure are already running in production inside the GB200 NVL72, which delivers 130 TB/s of low-latency GPU communication at rack scale. UALink's 200G-per-lane, 1,024-accelerator specification, backed by a board of 85-plus companies including AMD, AWS, Google, Intel, and Microsoft, is by comparison still a published standard rather than a shipping system. Ultra Ethernet is not a direct NVLink rival at all — its own consortium frames its job as the scale-out layer, not scale-up, so it competes alongside UALink and NVLink rather than against them. The metric worth watching is whether UALink's member companies ship silicon that actually hits the 200G-per-lane, 1,024-accelerator spec in real clusters, since NVIDIA's own framing of all-to-all bandwidth and latency as critical to AI factory performance is the bar any open alternative has to clear.
Utility-scale solar has become one of the cheapest new power sources, with costs down roughly 90% since 2010, while 2024 deployment volume outpaced every other generation technology. Paired with record-low battery prices, solar-plus-storage is already powering an AI data center in Arizona — though it still falls short of full 24-hour dispatchable baseload.
2026 marks the European Union AI Act's real enforcement start: the AI Office gains fining power over general-purpose AI on August 2, while high-risk system deadlines are pushed to 2027 and 2028. California activates two new laws on January 1 covering frontier-developer safety disclosure and training-data transparency. The federal government moves the opposite direction, ordering a Justice Department task force to challenge state AI laws, naming California's SB 53 as a target.
China's EHang (億航) is the only eVTOL maker actually flying paying passengers today, holding a full Chinese type, production, and operating certificate set and delivering 221 aircraft in 2025. US rivals Joby and Archer remain in certification or pre-launch stages, while Germany's Lilium and Volocopter both went insolvent in 2024–2025, with Volocopter absorbed by a Chinese buyer.