Chip Supply Chain
Specimen 00 · the whole chip

Stage 13 of 14High concentrationModerately hard to substitute

Chapter 13: Systems and Networking

Nvidia’s and AMD’s racks wire 72 AI chips together to run as one computer. Many firms can assemble the racks, but lasers for the optical links were scarce in 2026.

2,113 words / 9 minSpecimen: GPU baseboardUpdated 29 October 202644 sources

Link to chapterChapter PDF

Specimen 13 · Systems and Networking

An eight-GPU baseboard, the board that carries eight accelerator modules and the switch chips joining them. It is about 510 by 400 mm and drawn to scale, with the module frames standing about 30 mm above the board. The two nickel-gray squares in the middle are the NVLink switch chips.

In plain terms

One AI is not enough to train a model. So seventy-two of them are wired together in a , a cabinet the size of a wardrobe, and run as one computer. Many such racks are then linked to train a large model. Each chip sits on its own board and draws more than a thousand amps, several times the current a whole house is wired to carry. The chips have to swap results with each other all the time. A copper wire can carry those signals only a few meters before they fade, so the longer links send them as light down glass fiber. Many companies can build the rack, but in 2026 the lasers that send the light were scarce.

In short

In March 2026 Nvidia agreed to invest $2 billion in the laser maker Lumentum, alongside a multibillion-dollar purchase commitment and rights to future capacity for advanced laser components1. is used in high-speed lasers for optical communications, and China produced 70 percent of the world’s indium in 20252.

On 4 February 2025 China added indium phosphide , the the lasers are grown on, to its list3. AXT, a US-listed maker of those wafers, does all its manufacturing in China3. In its filing for the quarter to June 2026, AXT said it could not estimate when it would resume shipping the wafers to the United States3.

Before a package can compute, it has to be built into a machine. Building that machine is a separate industry from making the package.

Chokepoint cardSystems and Networking
ConcentrationFour to six firms could supply this stage today, and the top two take most of the revenue.
SubstitutabilityRated on and its NVSwitch , which Nvidia controls and which join the in a 4. Separately, lasers for the optical links were scarce in 20265. Nvidia secured future capacity rights for them from Lumentum in March 20261.
Price or market size$3.1 millionfor one GB200 NVL72 rack bought by a typical hyperscaler, or about $3.9 million with networking, storage and other items, by SemiAnalysis’s August 2025 estimate6
Who leads
  • USNvidia NVLink, InfiniBand and Spectrum-X Ethernet, which it credited for data center growth in the quarter to April 20267
  • USBroadcom $16.7 billion of AI semiconductor revenue in the quarter to August 2026, up 221 percent8
  • TWFoxconn World’s largest AI server provider, its chairman said in March 20269
  • CNInnoLight High-speed optical modules made in Suzhou and, since 2019, Thailand10
  • USCoherent Lasers and optical networking products, with a $2 billion Nvidia investment agreed in March 202611
Where it is made
  • USUnited States Switch and accelerator silicon, system design, indium phosphide lasers
  • TWTaiwan Contract design, rack assembly and high-layer-count circuit boards
  • JPJapan Low-loss board laminate
  • CNChina Optical module assembly and test, and most of the world’s indium
  • THThailand Chinese module makers’ offshore transceiver plants
Why substitution is possible
UALink is an open specification for the same job as NVLink12. AMD’s Helios rack, launched in July 2026, runs it over 13. Since May 2025 Nvidia’s NVLink Fusion program has let other firms’ custom chips use NVLink, which keeps those firms tied to Nvidia’s interconnect14.
Where China stands

Chinese firms such as InnoLight assemble high-speed , at sites including Suzhou and, since 2019, Thailand10. In February 2025 China put under new export restrictions2. The same month it began requiring export permits for indium phosphide , the that lasers for optical links are grown on3.

Where the US stands

Coherent broke ground in Sherman, Texas, in June 2026 on an expansion of its factory there, which makes optical devices from indium phosphide15. In August 2026 Nvidia said that Wistron’s new plant in Fort Worth, Texas, was producing its GB300 superchips and that Foxconn was building a factory in Houston16.

How AI racks and clusters are wired

Figure 13.1 · Explainer, 20 seconds

How AI chips are wired together

  1. One rack. One AI chip is not enough to train a model. So 72 of them are wired together in one rack, a cabinet the size of a wardrobe, and run as one computer.
  2. Copper inside. Inside the rack, copper wires link each tray of chips to the switch trays, which pass data between any two chips. Copper carries a signal only a few meters.
  3. Light between. Training a large model takes many racks, most of them too far apart for copper. The links between them carry the signal as flashes of light in glass fiber.
  4. Lasers. At each end of a fiber, a laser turns the electrical signal into flashes of light, and a detector turns them back. That lets thousands of chips in many racks work on one model.
Note: Simplified. The trays follow the layout of an Nvidia NVL72 rack. Racks and fibers are not drawn to scale.

An (the AI chip) ships as a module, the chip package on a small board with the circuits that feed it power. Nvidia calls its version , and the open-standards equivalent is the Open Compute Project’s accelerator module. Eight modules bolt onto a baseboard that carries their power and the links between them. On Nvidia’s baseboard, switch chips pass data between any two modules at full speed.

systems drop the baseboard and put the chips on trays stacked in the rack. An Nvidia holds 18 compute trays plus nine switch trays, 72 in all, and draws up to 142 17. The newer Vera Rubin NVL72 uses the same rack footprint18. The research firm SemiAnalysis estimated in August 2025 that a rack of the earlier GB200 NVL72 model cost a typical $3.1 million, or about $3.9 million with networking, storage and other items6.

Inside the rack the trays reach each other over copper17. Between neighboring racks, the signal can travel up to 7 meters over active electrical cables, a type of copper cable sold by firms such as Credo19. Longer links run on glass fiber, with an at each end. A laser in the sending transceiver turns the electrical signal into pulses of light, and a detector in the receiving one turns the pulses back into current.

The accelerators in an AI cluster exchange data over two networks that are not interchangeable. connects anywhere from a few dozen accelerators to thousands closely enough that they can read each other’s memory. It runs at . joins those groups to one another at hundreds of gigabits per second. Scale-up speeds are quoted in bytes and scale-out speeds in bits, and a byte is eight bits. Each GPU in a GB300 NVL72 can send 900 gigabytes per second each way over the scale-up network, or 7,200 gigabits per second20. Over the scale-out network it can send 800 gigabits per second20. So each GPU has nine times as much bandwidth for scale-up as for scale-out, SemiAnalysis calculated in January 202620.

Scale-up bandwidth limits how efficiently a large model trains. A model too big for one accelerator is split across many. The parts that exchange data constantly stay inside the scale-up group, because they need its high bandwidth21. Whole blocks of layers, the stages a model’s calculation passes through in turn, can sit on different groups across the slower scale-out network21.

Figure 13.2

Each GPU has nine times as much bandwidth inside its rack as between racks

One GPU’s links in an Nvidia GB300 NVL72 rack, scale-up inside the rack against scale-out between racks, drawn to scale

A model too big for one accelerator is split across many. The parts that exchange data constantly stay inside the scale-up group, and whole blocks of layers can sit on different groups across the slower scale-out network.

One GPU’s links in an Nvidia GB300 NVL72 rack, scale-up inside the rack against scale-out between racks, drawn to scale
Scale-up, inside the rack (900 GB/s each way)7,200 Gb/s
Scale-out, between racks800 Gb/s
Source: Co-Packaged Optics (CPO) Book – Scaling with Light for the Next Wave of Interconnect 20, Nvidia NVL72 system hardware and components 17 and Nvidia NeMo parallelisms guide 21CSVNote: Simplified. One GPU’s two links are drawn to scale, and the racks are not.

Types of links between AI chips

The designs differ in how many accelerators can share one pool of memory.

  • and NVSwitch. Nvidia’s NVLink bandwidth per GPU, counting both directions, went from 900 GB/s in NVLink 4 to 1,800 GB/s in NVLink 522. Across a whole Blackwell NVL72 rack, NVLink moves 130 TB/s22. Nvidia’s Rubin launch release of January 2026 gives NVLink 6 as 3,600 GB/s per GPU and 260 TB/s for a Vera Rubin NVL72 rack23. As of October 2026, Nvidia’s NVLink product page still lists preliminary figures of 3,000 GB/s and 216 TB/s22.
  • UALink. The open specification, from the UALink Consortium, runs 200 Gb/s per lane, and four lanes, which it calls a station, carry 800 Gb/s12. It can address up to 1,024 endpoints, such as accelerators, in one group12. AMD’s Helios rack, launched in July 2026, joins 72 GPUs and runs UALink over for its scale-up network13.
  • Google’s 3D torus. Google wires 64 of its Ironwood accelerators into each rack, called a cube24. Inside the cube each chip links directly to six others, in a three-dimensional grid that Google calls a 3D torus24. Optical circuit switches, which steer light from one fiber to another, join cubes into superpods of 9,216 chips and can route around a failed cube or link24.

Scale-out: InfiniBand versus Ethernet

Scale-out networks typically use or Ethernet, SemiAnalysis wrote in January 202620. InfiniBand is a specialized network long used in supercomputers. But Ethernet switch capacity has grown quickly. Broadcom’s Tomahawk 4 switched 25.6 Tbps on one chip in December 201925. Tomahawk 5 reached 51.2 Tbps in August 2022, and Broadcom said it had doubled the bandwidth about every two years since 201426. Tomahawk 6 doubled it again to 102.4 Tbps in June 202527. In June 2025 the Ultra Ethernet Consortium published its 1.0 specification, a set of Ethernet-based rules for communication in AI and supercomputing28. It covers how data is carried, how traffic jams are avoided and how one machine reads or writes another’s memory directly28.

Nvidia sells switches and network cards for both InfiniBand and Ethernet. Its data center revenue, which counts its GPUs as well as its networking gear, reached $89.0 billion in the quarter to July 2026, up 117 percent from a year earlier29. In the quarter before, it credited its data center growth to Blackwell 300 products and to demand for its “InfiniBand, Spectrum-X Ethernet, and NVLink” solutions7.

Optics: pluggable, then co-packaged

Broadcom and Nvidia are moving the optics out of pluggable transceivers and into the switch chip’s own package, a design called . A cluster needs several transceivers, the modules that turn electricity into light and back, for each accelerator. Their power cost is one reason for the change. Broadcom says its Tomahawk 6 Davisson cuts the power used by optical links by 70 percent compared with traditional pluggable modules30. It keeps its lasers in field-replaceable modules30. Nvidia’s Spectrum-X Photonics switches, announced in March 2025, have up to 512 ports of 800 gigabits per second each, and Nvidia says they use four times fewer lasers than traditional designs31. Nvidia builds them on its Spectrum-6 switch design23. In August 2026 Nvidia said its Spectrum-6 switch systems, with pluggable or co-packaged optics, were arriving at AI data centers around the world32.

Who makes optics, switch chips and AI servers

Chinese makers of optical modules have built plants outside China. InnoLight has run a plant in Thailand since 2019, alongside its sites in Suzhou and Taiwan10. The Thai subsidiary of Eoptolink, another Chinese module maker, lists sites in Chonburi and Rayong33.

A US rule from July 2026 may reach optical transceivers made by firms on a US security list. The list is the Federal Communications Commission’s Covered List of equipment and services judged to pose an unacceptable risk to US national security or to the security and safety of 34. Under its July 2026 order, the FCC no longer approves new devices for the US market if they contain logic-bearing hardware components made by firms on the list35. The order counts as logic-bearing any part that uses digital logic, from a single chip to a whole module, and leaves out parts with no digital function, such as wiring and housings36. Commenters in the record named optical transceivers as one such component37. As of 16 October 2026, the Covered List did not include InnoLight38.

China produces most of the world’s , the raw material for indium phosphide lasers, and restricts its export. Indium phosphide is used in the high-speed laser diodes and photodetectors for optical communications2. China accounted for 70 percent of world indium production in 20252. In February 2025 placed indium under new export restrictions2.

AI data centers were straining supplies of the lasers inside optical modules and pushing up their prices, Nikkei Asia reported in May 20265. Nvidia had agreed in March 2026 to invest $2 billion in one laser maker, Lumentum, alongside a multibillion-dollar purchase commitment and rights to future capacity for advanced laser components1. Coherent, another laser maker, broke ground in Sherman, Texas, in June 2026 on an expansion of its factory there, which makes optical devices from indium phosphide15.

Two American firms, Broadcom and Marvell, sell both custom AI chips and Ethernet switch chips. In its annual report for fiscal 2025, Broadcom said its AI chips include Ethernet switch chips as well as , which it builds to each customer’s specifications39. Its AI semiconductor revenue reached $16.7 billion in the third quarter of fiscal 2026, up 221 percent from a year earlier8. Marvell also builds custom AI chips to each customer’s specifications and sells Ethernet switch chips, its annual report for fiscal 2026 says40. Marvell’s five-year agreement with Amazon Web Services, announced in December 2024, covers custom AI chips alongside the optical and networking parts that go with them41.

Taiwanese firms do much of the physical assembly. Foxconn booked NT$8.1 trillion of revenue in 2025, up 18 percent, and cloud and networking was its largest product category in the fourth quarter9. In the same results release, its chairman called Foxconn the world’s largest AI server provider9.

Some assembly is moving to the United States. In April 2025 Nvidia said it was building supercomputer manufacturing plants in Texas, with Foxconn in Houston and Wistron in Dallas42. It expected to up in the following 12 to 15 months42. In August 2026 Nvidia said Wistron’s new plant in Fort Worth, Texas, near Dallas, was producing its GB300 superchips, which pair Nvidia’s Grace processor with its Blackwell Ultra GPUs16. It said Foxconn was building a factory in Houston to make Nvidia’s AI systems16.

Lasers and other hard-to-source parts

Many firms can build the racks. In October 2025 Nvidia said more than 50 makers of systems and components were gearing up for the Vera Rubin NVL72, which follows MGX, Nvidia’s standard design for trays and racks18. But in its filing for the quarter to July 2026, Nvidia said the complexity of producing its data center systems had caused production delays29.

Four inputs to the rack are harder to source elsewhere.

  • Indium phosphide laser capacity. Nvidia has secured rights to future capacity from laser makers. Alongside its Lumentum deal, in March 2026 Nvidia agreed to invest $2 billion in Coherent11. The deal also includes a multibillion-dollar purchase commitment and rights to future access and capacity for Coherent’s advanced laser and optical networking products11.
  • Indium phosphide . China added indium phosphide substrates, the the lasers are grown on, to its list on 4 February 20253. AXT, a US-listed maker of these wafers, does all its manufacturing in China3. In its filing for the quarter to June 2026, AXT called the need for Chinese export permits the most significant challenge it faces3. It said it could not estimate when it would resume shipping the wafers to the United States3. The same filing names Sumitomo and JX, both Japanese firms, as AXT’s competitors in indium phosphide3.
  • NVSwitch. Nvidia’s annual report for fiscal 2026 lists NVLink interconnects and switches among its own networking products4. UALink is an open standard designed for accelerator communications within one scale-up group12. Since May 2025 Nvidia’s NVLink Fusion program has let other firms’ custom chips use NVLink, including CPUs that Fujitsu and Qualcomm plan to connect to Nvidia GPUs14.
  • High-layer-count . TTM Technologies, an American circuit board maker, says its circuit board competitors are mostly based in China and Taiwan43. Nitto Boseki, a Japanese glass fiber maker, makes NE-glass, a that absorbs less signal, for the boards in data center switches and routers44. See Substrates and PCBs.

Optical module assembly can move, as InnoLight’s high-volume plant in Thailand, opened in 2019, shows10.

Flashcards

Recall each answer, then turn the card.

Card 1 of 4. Question

What is an AI rack?

Card 1 of 4. Answer

A cabinet of 72 AI chips wired together to run as one computer.

Reread: How it works

Sources (44)

Source tiers A company filings, governments, standards bodies, universities and research institutions. B SemiAnalysis, Reuters, the Financial Times and Nikkei Asia.

  1. ANVIDIA Announces Strategic Partnership With Lumentum to Develop State-of-the-Art Optics TechnologyNvidia · 2 March 2026
  2. AMineral Commodity Summaries 2026: IndiumU.S. Geological Survey · 5 February 2026
  3. AAXT, quarterly report on Form 10-Q for the quarter ended 30 June 2026U.S. Securities and Exchange Commission (filing by AXT) · 13 August 2026
  4. ANvidia, annual report on Form 10-K for the fiscal year ended 25 January 2026U.S. Securities and Exchange Commission (filing by Nvidia) · 25 February 2026
  5. BAI demand strains supplies of lasers, fiber and other optical techNikkei Asia · 27 May 2026
  6. BH100 vs GB200 NVL72 Training Benchmarks - Power, TCO, and Reliability Analysis, Software Improvement Over TimeSemiAnalysis · 20 August 2025
  7. ANvidia, quarterly report on Form 10-Q for the quarter ended 26 April 2026U.S. Securities and Exchange Commission (filing by Nvidia) · 20 May 2026
  8. ABroadcom, current report on Form 8-K dated 2 September 2026U.S. Securities and Exchange Commission (filing by Broadcom) · 2 September 2026
  9. AHon Hai Technology Group (Foxconn) Announces FY2025 & 4Q25 Financial ResultsHon Hai Precision Industry (Foxconn) · 16 March 2026
  10. AInnoLight Technology, company overview pageInnoLight Technology · Accessed 9 September 2026
  11. ANVIDIA and Coherent Announce Strategic Partnership to Develop Optics Technology to Scale Next-Generation Data Center ArchitectureNvidia · 2 March 2026
  12. AUALink™ 200G 1.0 Specification OverviewUALink Consortium · 15 May 2025
  13. AAAI 2026: AMD Launches AMD Helios Rackscale Solution for Frontier AIAMD Newsroom · 23 July 2026
  14. ANVIDIA Unveils NVLink Fusion for Industry to Build Semi-Custom AI Infrastructure With NVIDIA Partner EcosystemNvidia · 18 May 2025
  15. ACoherent Announces a CHIPS Letter of Intent for $50 Million to Expand World-Leading Manufacturing Facility for AI InfrastructureCoherent · 16 June 2026
  16. ANVIDIA and Partners Build in America, for AmericaNvidia · 5 August 2026
  17. ASystem Hardware & ComponentsNvidia · Accessed 9 September 2026
  18. ANVIDIA, Partners Drive Next-Gen Efficient Gigawatt AI Factories in Buildup for Vera RubinNvidia · 13 October 2025
  19. ACredo Technology Group, annual report on Form 10-K for the fiscal year ended 2 May 2026U.S. Securities and Exchange Commission (filing by Credo Technology Group) · 15 June 2026
  20. BCo-Packaged Optics (CPO) Book – Scaling with Light for the Next Wave of InterconnectSemiAnalysis · 1 January 2026
  21. AParallelisms Guide — Megatron BridgeNvidia · Accessed 9 September 2026
  22. ANVLink & NVLink Switch for Advanced Multi-GPU CommunicationNvidia · 20 April 2026
  23. ANVIDIA Kicks Off the Next Generation of AI With RubinNvidia · 5 January 2026
  24. AInside the Ironwood TPU codesigned AI stackGoogle Cloud · 6 November 2025
  25. ABroadcom Ships Tomahawk 4, Industry’s Highest Bandwidth Ethernet Switch Chip at 25.6 Terabits per SecondBroadcom · 9 December 2019
  26. ABroadcom Ships Tomahawk 5, Industry’s Highest Bandwidth Switch Chip to Accelerate AI/ML WorkloadsBroadcom · 16 August 2022
  27. ABroadcom Ships Tomahawk 6: World’s First 102.4 Tbps SwitchBroadcom · 3 June 2025
  28. AUltra Ethernet Consortium (UEC) Launches Specification 1.0 Transforming Ethernet for AI and HPC at ScaleUltra Ethernet Consortium · 11 June 2025
  29. ANvidia, quarterly report on Form 10-Q for the quarter ended 26 July 2026U.S. Securities and Exchange Commission (filing by Nvidia) · 26 August 2026
  30. ABroadcom Announces Tomahawk 6 – Davisson, the Industry’s First 102.4-Tbps Ethernet Switch with Co-Packaged OpticsBroadcom · 8 October 2025
  31. ANVIDIA Announces Spectrum-X Photonics, Co-Packaged Optics Networking Switches to Scale AI Factories to Millions of GPUsNvidia · 18 March 2025
  32. ANVIDIA Announces Financial Results for Second Quarter Fiscal 2027Nvidia · 26 August 2026
  33. AEoptolink Technology, contacts pageEoptolink Technology · Accessed 9 September 2026
  34. AProtecting Against National Security Threats to the Communications Supply Chain through the Equipment Authorization Program: Third Report and Order and Third Further Notice of Proposed Rulemaking (FCC-26-50)U.S. Federal Communications Commission · 23 July 2026 · p. 3
  35. AProtecting Against National Security Threats to the Communications Supply Chain through the Equipment Authorization Program: Third Report and Order and Third Further Notice of Proposed Rulemaking (FCC-26-50)U.S. Federal Communications Commission · 23 July 2026 · p. 6
  36. AProtecting Against National Security Threats to the Communications Supply Chain through the Equipment Authorization Program: Third Report and Order and Third Further Notice of Proposed Rulemaking (FCC-26-50)U.S. Federal Communications Commission · 23 July 2026 · p. 13
  37. AProtecting Against National Security Threats to the Communications Supply Chain through the Equipment Authorization Program: Third Report and Order and Third Further Notice of Proposed Rulemaking (FCC-26-50)U.S. Federal Communications Commission · 23 July 2026 · p. 10
  38. AList of Equipment and Services Covered By Section 2 of The Secure Networks ActU.S. Federal Communications Commission · 16 October 2026
  39. ABroadcom, annual report on Form 10-K for the fiscal year ended 2 November 2025U.S. Securities and Exchange Commission (filing by Broadcom) · 18 December 2025
  40. AMarvell Technology, annual report on Form 10-K for the fiscal year ended 31 January 2026U.S. Securities and Exchange Commission (filing by Marvell Technology) · 11 March 2026
  41. AMarvell Technology, current report on Form 8-K dated 2 December 2024U.S. Securities and Exchange Commission (filing by Marvell Technology) · 2 December 2024
  42. ANVIDIA to Manufacture American-Made AI Supercomputers in US for First TimeNvidia · 14 April 2025
  43. ATTM Technologies, annual report on Form 10-K for the fiscal year ended 29 December 2025U.S. Securities and Exchange Commission (filing by TTM Technologies) · 17 February 2026
  44. ANittobo Group Integrated Report 2025Nitto Boseki (Nittobo) · 16 September 2025 · p. 7