> ## Content Index > Fetch the complete content index at: https://www.sotkn.com/llms.txt > Use this file to discover other available public pages before exploring further. # What Is NVLink? The Switch Inside the AI Rack, and Who Gets Paid for It - URL: https://www.sotkn.com/explained/nvlink-explained/ - Published: 2026-11-05T10:28:08.000Z - Updated: 2026-11-05T10:28:08.000Z - Description: NVLink is NVIDIA’s link that joins GPUs inside a server or rack so they share memory and act like one large GPU. A Vera Rubin rack links 72 GPUs at 3.6 TB/s each. Hyperscalers building their own chips must rent NVLink, build their own, or wait for open UALink. The switch is the stuck step. - Author: Sean - Tags: Mechanism, #desk-explained EXPLAINED · Mechanism · Last reviewed: 5 Nov 2026 · Next review: after NVIDIA’s third-quarter fiscal 2027 results (November 2026) and the first Vera Rubin rack deliveries NVLink is NVIDIA’s high-speed link that joins GPUs inside one server or rack so they can read each other’s memory and work like a single large GPU. Engineers call this scale-up, as opposed to scale-out, the network that links racks together. **In short.** NVLink is NVIDIA’s link that joins GPUs inside a server or rack so they share memory and act like one large GPU. A Vera Rubin rack joins 72 GPUs at 3.6 terabytes per second each, about 18 times the speed of the link between racks by our arithmetic. Hyperscalers building their own chips have three choices: rent NVLink through Fusion, build their own link as Google has, or wait for open UALink parts. Our read: the switch is the stuck step, and whoever supplies it is paid on every rack. Why an investor cares: the rack, not the chip, is now the unit NVIDIA sells, and NVLink is what makes 72 GPUs a rack rather than 72 boxes. Our [guide to rack-scale AI](https://www.sotkn.com/explained/rack-scale-ai-explained/) covers the rack; this page explains the link that holds it together, and why hyperscalers want a say in who supplies it. ## What is NVLink? NVLink is a direct, very fast link between GPUs. A GPU can use it to read from or write to another GPU’s memory as if that memory were its own, without going through the server’s main processor or the network card. NVIDIA first used it to join GPUs on a single board. On an eight-GPU server board, such as the one in the photograph above, the GPUs connect through NVLink switch chips so that each can reach every other at full speed. NVIDIA still sells that design: its HGX Rubin NVL8 board links eight Rubin GPUs through NVLink. The bigger change came when NVIDIA stretched NVLink across a whole rack. In its NVL72 systems, 72 GPUs connect through NVLink switch trays in the middle of the rack. NVIDIA said on its August 2026 call that it is now on its third generation of these NVLink-72 rack-scale systems. ## How is scale-up different from scale-out? A large AI cluster has two networks. The scale-up network joins chips inside one server or rack so they share memory. The scale-out network, built on InfiniBand or Ethernet, links those racks to each other so tens of thousands of chips can work on one job. | | Scale-up (NVLink, ICI, UALink) | Scale-out (InfiniBand, Ethernet) | | ------------------------ | ----------------------------------------------------------------------------------------------------- | ------------------------------------------- | | **What it joins** | Chips inside one server, rack or pod | Racks and pods across a data hall | | **What the chips share** | Memory: one chip can read and write another’s directly | Messages: data is packed up and sent | | **Speed per chip** | 3.6 TB/s on NVLink 6, both directions | Up to 800 Gb/s per network card | | **Typical domain** | 8 GPUs in a server; 72 in an NVIDIA rack; 9,216 in a Google TPU superpod | Thousands of chips or more | | **Main medium** | Mostly copper inside the rack | Mostly optical cables between racks | | **Who sells it** | NVIDIA for NVLink; Google builds ICI for itself; UALink is an open standard with parts in development | NVIDIA, Arista, Cisco, Celestica and others | *Speeds and domain sizes from NVIDIA (5 January 2026 and 18 May 2025) and Google (9 April 2025). Copper and optics split is the general design pattern, not a rule for every system.* The two differ mainly in speed and reach. NVIDIA rates NVLink 6 at 3.6 terabytes per second per GPU, counting both directions. That is 1.8 terabytes, or 14.4 terabits, each way: about 18 times an 800 gigabit scale-out link, by our arithmetic. Scale-out reaches much further, across a whole data hall. Our [comparison of InfiniBand and Ethernet](https://www.sotkn.com/explained/infiniband-vs-ethernet/) covers that second network. ## Why does a rack need NVLink? Because the largest models no longer fit on one GPU. A model with hundreds of billions of settings has to be split across many chips, and those chips swap results many times for every answer they produce. The faster they swap, the less time each GPU spends waiting. Inside an NVLink domain that swap runs at memory speed. NVIDIA says a Vera Rubin NVL72 rack, 72 GPUs and 36 CPUs, delivers 260 terabytes per second of NVLink bandwidth in total. Our check: 72 GPUs at 3.6 terabytes each is 259.2, which NVIDIA rounds to 260. This matters most for inference on mixture-of-experts models, where each request is routed between parts of the model held on different chips. The bigger the domain, the more of the model sits inside it. Our [guide to AI inference](https://www.sotkn.com/explained/ai-inference-explained/) covers why serving models is now the larger workload. ## How do hyperscalers link their own chips? NVLink only comes with NVIDIA. A hyperscaler that designs its own AI chip has to solve scale-up some other way, and there are three routes. - **Build your own link.** Google’s TPUs use its Inter-Chip Interconnect, ICI. Its Ironwood TPU, announced in April 2025, scales to 9,216 chips in one superpod, about 128 times an NVL72 rack by our arithmetic, at 1.2 terabytes per second per chip, a third of NVLink 6’s per-GPU figure. That compares a 2025 chip with a 2026 one, so treat it as a shape, not a contest. - **Rent NVLink.** NVIDIA launched NVLink Fusion in May 2025 so other companies’ chips can join an NVLink domain. AWS said in December 2025 that it is designing Trainium4 to use NVLink 6 and NVIDIA’s rack design, linking up to 72 custom chips at 3.6 terabytes per second each. - **Wait for the open standard.** The UALink Consortium, more than 85 companies, published its first specification in April 2025: 200 gigabits per lane for up to 1,024 accelerators in a pod, about 14 times an NVL72 rack. It said multiple vendors are developing UALink chips and switches. Our read: the AWS choice is the telling one. A hyperscaler with its own chip team still chose to rent NVIDIA’s scale-up link rather than wait. Our [guide to who makes custom AI chips](https://www.sotkn.com/explained/who-makes-custom-ai-chips/) covers the chips themselves, and our [guide to TPUs](https://www.sotkn.com/explained/tpu-explained/) covers Google’s. ![NVLink explained: largest published scale-up domains, 8 GPUs for an HGX Rubin NVL8 board, 72 for a Vera Rubin NVL72 rack, up to 1,024 for UALink 1.0 and 9,216 for a Google Ironwood superpod](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/11/nvlink-explained-domain-sizes.png) Largest published scale-up domain by design. NVIDIA, UALink Consortium, Google; Second Order. ## Who gets paid for NVLink? Our read: NVIDIA first, the switch-chip makers second, and the rack builders on every unit. - **NVIDIA.** It sells the GPUs, the NVLink switch trays and, through Fusion, the link inside other companies’ chips. NVIDIA named NVLink among the drivers of its record $75.2 billion of data centre revenue in the quarter to 26 April 2026\. Data centre networking was a record $14.8 billion, about 19.7% of data centre revenue by our arithmetic; NVIDIA does not report NVLink on its own. - **Switch and link chip makers.** Astera Labs, a Fusion partner, said its Scorpio fabric switches became its largest product line in the third quarter of 2026, led by volume production of its Scorpio X scale-up switch. Its revenue rose 154% from a year earlier to $584.9 million, and it guided to $730 to $750 million for the fourth quarter. - **Custom-chip design partners.** Marvell, MediaTek and Alchip were among the first NVLink Fusion adopters, building custom AI chips for others that can plug into NVLink. - **Rack builders.** Every NVL72 rack needs its switch trays, copper links, cooling and power assembled and tested as one machine before it ships. Why the bill grows: NVIDIA said on its August 2026 call that its revenue opportunity is about $40 billion per gigawatt with Vera Rubin, against $25 billion with Blackwell and about $18 billion with Hopper. That is 60% more than Blackwell, by our arithmetic, and NVIDIA listed NVLink as part of what Vera Rubin spans. Treat that as NVIDIA’s own framing, not a printed revenue line. ![NVLink explained: how a scale-up domain gets paid, from chip choice and link choice to the switch as the stuck step, then rack build, then acceptance and revenue](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/11/nvlink-explained-how-scale-up-gets-paid.png) How a scale-up domain gets paid. Second Order. ## Where is the stuck step? Our read: the switch. In a scale-up domain every chip must reach every other at full speed, and that runs through switch silicon. For NVIDIA GPUs, NVIDIA supplies it. For everyone else, the switch decides how big the domain can get and who is paid for it. That is why the open route is slow. UALink’s specification is published, but the consortium said vendors are still developing the chips and switches. Until those ship in volume, a custom chip that wants a large domain either rents NVLink or relies on a link its owner built, as Google did. Astera Labs shows the build-up of supply behind it. Its inventory was $200.9 million at 30 September 2026, about 3.4 times the level at the end of 2025, by our arithmetic. Our read: switch makers are stocking ahead of rack orders they expect to come. The clock is the Vera Rubin ramp. NVIDIA said Rubin-based products would be available from partners in the second half of 2026, with AWS, Google Cloud, Microsoft and Oracle among the first clouds to deploy them. Every Vera Rubin NVL72 rack needs its NVLink switch trays before it can earn. Our [guide to GPU cluster acceptance](https://www.sotkn.com/explained/gpu-cluster-acceptance-explained/) covers what has to pass before a rack is paid for. ## What would change the view? - **UALink ships early and in volume.** If open switches reach hyperscalers soon, custom chips gain large domains without NVIDIA, and the scale-up cheque splits. - **Ethernet takes scale-up.** If Ethernet tuned for the job proves good enough inside the rack, the scale-up switch market starts to look like scale-out, with many suppliers. - **Fusion stalls.** If Trainium4 or other Fusion chips slip, the rented-link route looks weaker and in-house links like Google’s look stronger. ## Common misconceptions - **People assume NVLink is a network between racks.** Actually it joins GPUs inside a server or rack; InfiniBand or Ethernet link the racks. - **People assume NVLink only matters for NVIDIA GPUs.** Actually NVLink Fusion lets custom chips join, and AWS plans to use it for Trainium4. - **People assume the GPU is the whole story.** Actually a rack of 72 GPUs only works as one machine because of the switch trays between them. - **People assume bigger domains always win.** Actually domain size, speed per chip and cost trade off: Google’s superpod is far larger than an NVL72 rack but slower per chip. ## FAQ **What is NVLink in simple terms?** NVLink is the wiring NVIDIA uses to join GPUs inside a server or rack so closely that they can read each other’s memory. Software can then treat eight, or 72, GPUs almost like one very large GPU. **What is the difference between NVLink and InfiniBand?** NVLink joins GPUs inside one server or rack, the scale-up network. InfiniBand and Ethernet join racks to each other, the scale-out network. A large AI cluster uses both: NVLink within each rack and InfiniBand or Ethernet between them. **How fast is NVLink?** NVIDIA’s sixth generation, NVLink 6, gives each Rubin GPU 3.6 terabytes per second of bandwidth, counting both directions. A Vera Rubin NVL72 rack of 72 GPUs totals 260 terabytes per second. **What is an NVLink switch?** A chip that lets every GPU in the domain talk to every other at full speed. In a 72-GPU rack, the GPUs are not wired to each other directly; they are wired to NVLink switch trays in the middle of the rack. **What is NVLink Fusion?** A way for other companies’ chips to join an NVLink domain. Announced in May 2025, it lets a hyperscaler put an NVLink chiplet inside its own custom AI chip. AWS said in December 2025 that Trainium4 will use it. **What is UALink?** Ultra Accelerator Link, an open scale-up standard from a consortium of more than 85 companies. Its first specification, in April 2025, allows up to 1,024 accelerators in one pod. It is the open alternative to NVLink. **Is this a list of stocks to buy or sell?** No. This page explains NVLink and where the money goes. It does not rank companies or give buy or sell calls. Not investment advice. Do your own research. Related: [GPU clusters](https://www.sotkn.com/explained/gpu-cluster-explained/); [who makes AI cluster networking](https://www.sotkn.com/explained/who-makes-ai-cluster-networking/); [optics and networking in AI clusters](https://www.sotkn.com/explained/optics-networking-ai-cluster-explained/); [TPU vs GPU](https://www.sotkn.com/explained/tpu-vs-gpu/). ## Stack doors Layer: [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/). Also: [Semiconductors](https://www.sotkn.com/stack/semiconductors/); [who the hyperscalers are](https://www.sotkn.com/explained/who-are-the-hyperscalers/). Markets: [United States](https://www.sotkn.com/markets/us/). [Get the Wire](https://www.sotkn.com/#/portal/signup) — short daily notes when AI chips, racks and the links between them move. Not investment advice. Do your own research. **How we check this** NVLink 6 speeds, rack sizes and Rubin timing come from NVIDIA’s Rubin launch release of 5 January 2026\. NVLink Fusion comes from NVIDIA’s release of 18 May 2025, and the AWS Trainium4 plan from NVIDIA’s technical blog of 2 December 2025\. NVIDIA’s data centre and networking revenue come from its first-quarter fiscal 2027 CFO commentary, and the revenue opportunity per gigawatt from its second-quarter call of 26 August 2026\. UALink’s specification comes from the consortium’s release of 8 April 2025, Google’s Ironwood figures from its post of 9 April 2025, and Astera Labs’ results from its release of 3 November 2026\. The 72 × 3.6 rack check, the 18-times comparison with an 800 Gb/s link, the domain-size ratios, NVIDIA’s networking share, the change in revenue per gigawatt and Astera’s growth and inventory figures are our arithmetic, checked in code and labelled as ours on the page. Who gets paid and the stuck step are our read. The featured photograph is credited in its caption. This page explains one technology. It does not rank companies or give buy or sell calls. Not investment advice. Do your own research. **Last reviewed:** 5 Nov 2026 · **Next review:** after NVIDIA’s third-quarter fiscal 2027 results (November 2026) and the first Vera Rubin rack deliveries Spot an error? [Tell us](mailto:sean@thekopinotes.com) and we will correct it and note the change here. [Our editorial standards →](https://www.sotkn.com/editorial-standards/) ## References 1. [NVIDIA — CFO commentary, first quarter fiscal 2027 (quarter ended 26 Apr 2026)](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000051/q1fy27cfocommentary.htm?ref=sotkn.com) — data center revenue a record $75.2 billion, driven by Blackwell 300 and demand for InfiniBand, Spectrum-X Ethernet and NVLink; networking revenue a record $14.8 billion; hyperscale about 50% of data center revenue 2. [NVIDIA — second-quarter fiscal 2027 earnings call transcript (26 Aug 2026)](https://s201.q4cdn.com/141608511/files/content%5Ffiles/TRANSCRIPT%5F-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5%5F00-PM-ET.pdf?ref=sotkn.com) — revenue opportunity per gigawatt about $18 billion with Hopper, $25 billion with Blackwell and $40 billion with Vera Rubin, which spans Vera CPU, Rubin GPU, NVLink and InfiniBand or Ethernet; third generation of NVLink-72 rack-scale systems 3. [NVIDIA — Rubin platform launch at CES (5 Jan 2026)](https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer?ref=sotkn.com) — NVLink 6: 3.6 TB/s per GPU and 260 TB/s per Vera Rubin NVL72 rack of 72 GPUs and 36 CPUs; HGX Rubin NVL8 links eight GPUs; Rubin products from partners in the second half of 2026; AWS, Google Cloud, Microsoft and OCI among first to deploy 4. [NVIDIA — NVLink Fusion for semi-custom AI infrastructure (18 May 2025)](https://nvidianews.nvidia.com/news/nvidia-nvlink-fusion-semi-custom-ai-infrastructure-partner-ecosystem?ref=sotkn.com) — NVLink Fusion lets hyperscalers and others connect custom chips to NVLink scale-up; MediaTek, Marvell, Alchip, Astera Labs, Synopsys and Cadence first adopters; Fujitsu and Qualcomm CPUs; scale-out networking up to 800 Gb/s 5. [NVIDIA Technical Blog — AWS integrates NVLink Fusion for Trainium4 (2 Dec 2025)](https://developer.nvidia.com/blog/aws-integrates-ai-infrastructure-with-nvidia-nvlink-fusion-for-trainium4-deployment/?ref=sotkn.com) — AWS designing Trainium4 to integrate with NVLink 6 and NVIDIA’s MGX rack; NVLink Fusion chiplet dropped into custom chips; up to 72 custom chips all-to-all at 3.6 TB/s each, 260 TB/s total 6. [UALink Consortium — UALink 200G 1.0 Specification released (8 Apr 2025)](https://ualinkconsortium.org/wp-content/uploads/2025/04/UALink-1.0-Specification-PR%5FFINAL.pdf?ref=sotkn.com) — open scale-up interconnect from a consortium of more than 85 member companies; 200G per lane for up to 1,024 accelerators in a pod; multiple vendors developing UALink accelerators and switches 7. [Google — Ironwood, the first Google TPU for the age of inference (9 Apr 2025)](https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference/?ref=sotkn.com) — seventh-generation TPU scales to 9,216 liquid-cooled chips linked by Google’s Inter-Chip Interconnect (ICI), nearly 10 MW; ICI bandwidth 1.2 TB/s bidirectional per chip 8. [Astera Labs — third-quarter 2026 results (3 Nov 2026)](https://ir.asteralabs.com/news-releases/news-release-details/astera-labs-reports-third-quarter-2026-financial-results?ref=sotkn.com) — revenue $584.9 million, up 154% from a year earlier; Scorpio fabric switches became the largest product family, led by volume production of the Scorpio X-Series 320-lane switch; fourth-quarter guide $730 to $750 million; inventory $200.9 million against $59.0 million at end-2025; portfolio spans NVLink Fusion, UALink, PCIe and Ethernet S **[Sean](https://www.sotkn.com/author/sean/)** Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck. Covers AI infrastructure, semiconductors, and second-order stack reads. [Author page →](https://www.sotkn.com/author/sean/) · [Editorial standards →](https://www.sotkn.com/editorial-standards/)