More hyperscalers are designing their own AI chips, but a finished chip delivers nothing without the factory-like facility that houses it. In a company blog post, NVIDIA restated the purpose of NVLink Fusion, the technology that connects custom accelerators known as XPUs to NVIDIA rack-scale infrastructure[1]. The idea is to cover the heavy lifting outside of chip design with proven components, so teams can concentrate on the parts they actually want to differentiate.
An AI Factory Is Graded on Output
NVIDIA starts by defining how the economics of an AI factory are measured. The metrics listed are tokens per second, tokens per watt, cost per token, utilization and uptime[1]. Profitability depends on how much output the whole facility delivers, not on the speed of any individual accelerator.
That means a company building a custom XPU cannot stop at chip design. It also has to handle scale-up and scale-out networking, rack-scale architecture, production factory software, and a supplier ecosystem robust enough to keep parts flowing[1]. NVIDIA argues that this burden is what slows XPUs on their way to market.
Seventy-Two XPUs in a Single Domain
At the core of NVLink Fusion is pulling XPUs into the NVLink scale-up domain. Sixth-generation NVLink treats 72 XPUs as one domain and links them with high bandwidth and low latency. According to NVIDIA, end-to-end latency for XPU-to-XPU transfers is 3x lower than alternative solutions based on off-the-shelf Ethernet, and the packet rate is 10x higher[1].
NVIDIA product documentation backs the numbers: NVLink 6 connects 72 accelerators all-to-all at 3.6 TB/s per XPU, for 260 TB/s across an NVL72 domain, with 4x bandwidth efficiency through SHARP FP8 support[2]. Future roadmap configurations extend domain sizes to as many as 1,152 accelerators and include co-packaged optics[1][2].
The other piece is NVLink-C2C, a chip-to-chip link between XPUs and NVIDIA Vera CPUs or other ecosystem CPUs, delivering up to 6x the energy efficiency of a PCIe interface[1]. Agentic systems move constantly between control and compute, so a bottleneck here drags down everything else.
Racks, Cooling and Power Come Off the Shelf
NVLink Fusion adopters can use the NVIDIA MGX rack-scale architecture and the same supply chain behind MGX-based systems such as Vera Rubin NVL72. MGX suppliers provide the building blocks for rack, cooling, power and emerging 800 VDC designs, while manufacturing partners handle design and integration[1].
Jack Luoh, head of product and solution at QCT and Quanta Computer, said that with Vera Rubin NVL72 the company is looking at almost 100 percent automation of system builds on the manufacturing line, and that most of those investments can be leveraged if the XPU uses NVLink Fusion[1]. Tim Wilson, vice president and general manager of data center silicon engineering at Intel, pointed to the ability to choose the CPU architecture, performance level and software capabilities that suit a given workload[1].
The rack interior is built for operations as well. Reference compute trays use 100 percent liquid cooling with no fans, cables or hoses, and trays can be pulled while the rest of the rack keeps running. NVLink Switch trays are liquid cooled too and support continued operation during service[1].
Build the Building Before the Silicon Is Set
Data center construction begins long before the final accelerator mix is decided. Power procurement, facility design, cooling, rack layout and network architecture all come first, and a design locked to a single chip turns any chip delay into a schedule risk for the entire project[1].
NVLink Fusion answers this with a unified architecture. XPU-based systems and GPU-based systems such as Vera Rubin NVL72 can share rack footprints, networking, cooling, power delivery and management systems, so buildout can proceed while the silicon mix is deferred[1][2]. Vince Hu, corporate senior vice president and general manager of the data center and computing business group at MediaTek, said the value of the program is that customers can deploy a rack-level solution with NVIDIA GPUs and then develop their XPU at a different pace[1]. Lie-Szu Juang, chair and chief strategy officer at GUC, highlighted the way NVLink Fusion bridges NVIDIA technology with a third-party process to create a unified rack-scale architecture[1].
CC Lee, senior hardware development manager at Annapurna Labs, an Amazon company, said that using the proven NVL72 rack design helps time to market and opens access to multiple suppliers[1]. AWS is in fact using NVLink Fusion in its Trainium4 deployment[2].
Validate Before Building, Service While Running
Rework on a facility is expensive, so validation before construction matters. NVLink Fusion aligns with the NVIDIA DSX reference architecture for AI factories, codesigning buildings, power, cooling, compute and networking together. The NVIDIA Omniverse DSX AI Factory Blueprint provides a digital twin and open reference design for gigawatt-scale AI factories, letting partners model facilities and technology in one place[1].
The software side is in place as well. NVIDIA NCCL for distributed workloads, NVIDIA Dynamo and NIXL for disaggregation, and NVIDIA Mission Control for cluster management and telemetry let operators run mixed accelerator fleets as a single coordinated system[1].
A Widening Ecosystem
The list of participants keeps growing. On the CPU side are Arm, Intel, Fujitsu and SiFive. Custom silicon partners include Alchip, Astera Labs, GUC, Marvell, MediaTek and SAMSUNG. Design tool vendors Cadence and Synopsys are involved, along with optical interconnect partners Ayar Labs and Lightmatter[2].
The Marvell relationship runs deepest. On March 31, 2026, NVIDIA announced a strategic partnership with Marvell and invested 2 billion USD (about 320 billion yen) in the company. Marvell provides custom XPUs and NVLink Fusion-compatible scale-up networking, while NVIDIA supplies the Vera CPU, ConnectX NICs, BlueField DPUs, NVLink interconnect, Spectrum-X switches and the rack-scale AI compute[3].
※1 USD = 159 JPY
Summary
NVLink Fusion tells companies that want their own silicon that everything outside their point of differentiation can come off the shelf. Sixth-generation NVLink binding 72 XPUs into one domain, MGX racks and supply chain, DSX-based pre-validation, and software through Mission Control arrive as a single package. Being able to break ground on a data center before the chip is finished fits well with an AI buildout where power and land are secured first.
Source: https://blogs.nvidia.com/blog/nvlink-fusion-xpu-ai-factory/
Source: https://www.nvidia.com/en-us/data-center/nvlink-fusion/
