Essay No. 072  ·  AI Infrastructure / Supply Chain / Platform Risk
Nvidia AI Infrastructure Semiconductors Supply Chain Power Delivery Blackwell Hopper Vera Rubin Data Centers Platform Risk

Nvidia's Hopper Supplier Cut Was a Warning: AI Hardware Design Wins Are Not Moats Hopper H100 Blackwell GB200 Vera Rubin Power delivery Liquid cooling NVLink Rack-scale AI Supplier concentration Platform risk

A supplier socket can look like a moat until the platform owner changes the architecture. Hopper showed the risk. Blackwell and Vera Rubin made it much bigger.

PM
PUGALENTHI MAGENDRAN
May 27, 2026  ·  Research memo  ·  Updating a 2022 supplier-risk thesis
16 MIN
Thesis
The 2022 Hopper supplier cut was not just a stock-specific short report. It was an early warning about AI hardware platform risk. As Nvidia moved from H100 GPUs to GB200 rack-scale systems and then Vera Rubin AI supercomputers, the supplier question changed. It is no longer enough to win one component socket. Suppliers now have to survive Nvidia's full-stack control over power, cooling, networking, cabling, packaging, software, and rack-scale architecture. In AI infrastructure, a design win is valuable, but the platform owner owns the roadmap.
Editorial note
This essay is a strategic analysis of AI hardware platform risk. It is not investment advice. It does not recommend buying or shorting any stock. The 2022 short report is used as a historical case study about supplier concentration, not as a current trade idea. The unnamed long-time Nvidia supplier referenced in the 2022 piece is not identified here because the cited source does not name it. A separate, public Vicor 2024 Form 10-K example is used only to illustrate that customer concentration disclosure remains a real issue across hardware supply chains, and is explicitly not used to identify the 2022 supplier.
Executive summary
  • In 2022, a SemiAnalysis short report said a long-time Nvidia supplier had seemingly been cut out of Hopper, while Nvidia represented up to 55% of that supplier's revenue.
  • Hopper was already a major platform transition: H100 used TSMC 4N, 80B transistors, HBM3, fourth-generation NVLink, and a high-power SXM form factor.
  • The real lesson was supplier concentration and platform risk, not the specific short idea.
  • By Blackwell, Nvidia shifted the product boundary from GPU board to rack-scale AI system, with GB200 NVL72 combining 36 Grace CPUs and 72 Blackwell GPUs.
  • By Vera Rubin, Nvidia's platform control extends deeper into CPUs, GPUs, NVLink, NICs, DPUs, networking, power, cooling, and rack-scale AI factory architecture.

Section 1  ·  Historical frameWhat the 2022 report got right

The 2022 SemiAnalysis short report carried a deliberately blunt headline. A long-time Nvidia supplier had seemingly been cut out of the next-generation Hopper platform, and Nvidia represented up to 55% of that supplier's revenue.[1] The report framed Hopper as a major architecture transition: an 80B transistor GPU, an up to 6x Transformer performance claim, fourth-generation NVLink, DPX instructions, and a top-spec SXM-style package at 700W. It also noted that the stock fell more than 20% on publication day.[1]

Page 2 of the public excerpt showed the Hopper board and package framing alongside the short-report thesis. The actual identity of the supplier is not disclosed in the public excerpt. This essay deliberately does not name it, because the source does not name it. The strategic lesson is large regardless: one lost Nvidia design win can matter enormously when customer concentration is high, and architecture transitions are exactly the points where the supplier map can change without warning.[1]

Section 2  ·  HopperHopper was the first warning

Hopper was not just a chip refresh. It was a platform transition for AI and HPC. Nvidia's Hopper architecture deep dive describes the GH100 as fabricated on TSMC 4N, with 80B transistors on an 814 mm² die, HBM3 memory controllers, fourth-generation Tensor Cores, fourth-generation NVLink, and PCIe Gen 5 support.[2] The 700W top-spec SXM point flagged in the 2022 short report mattered for a different reason. It showed that power delivery was becoming strategic. Bigger silicon at higher TDP raised the cost of being wrong on the power architecture.

Process
TSMC 4N, customized for Nvidia.
Transistors
Approximately 80 billion.
Die area
Approximately 814 mm².
Memory
HBM3 memory controllers integrated on package.
Tensor Cores
Fourth-generation, with Transformer Engine support.
Interconnect
Fourth-generation NVLink and PCIe Gen 5.
Form factor
SXM5 module at up to 700W for the top-spec part, per public reporting.

Hopper made the supplier-risk problem visible because the architecture changed. When the architecture changes, the supplier map can change with it.

Section 3  ·  The moat questionA design win is not a moat

A supplier design win can look like a moat for two or three years and then disappear in a single architecture transition. The mechanism is simple. The platform owner controls the system. The supplier controls a component inside the system. If the platform owner changes the board, the power architecture, the cooling approach, the module design, the packaging strategy, or the supplier mix, the old design win can vanish even when the part itself still works.

In AI infrastructure, the design win is valuable, but the platform owner owns the roadmap.

What a design win gives a supplier
  • Revenue at scale, often quickly.
  • Credibility with other potential customers.
  • Validation across reliability, yield, and field data.
  • Scale advantages on cost and learning curve.
  • Visibility into roadmap and reference designs.
What platform risk takes away
  • Customer concentration that compresses bargaining power.
  • Redesign risk at each architecture transition.
  • Dual-source pressure from the platform owner.
  • Margin pressure as the platform commoditizes adjacent layers.
  • Roadmap dependency on a single customer's strategy.
  • Architecture displacement when the platform changes shape.

Section 4  ·  Power deliveryPower delivery became AI infrastructure

The 700W Hopper SXM point was a signal of the future. AI accelerators need massive current delivered efficiently, reliably, and with minimal heat. The simple chain is: higher GPU power means higher current; every conversion loss becomes heat; heat affects rack density and total cost of ownership; rack power limits determine how many GPUs customers can deploy. Power modules, bus bars, voltage regulators, cooling loops, and cables are not generic components anymore. They are part of AI system architecture.

Old view
Power delivery = board component.
New view
Power delivery = rack-level architecture affecting density, cooling, uptime, and deployment economics.

Power delivery is no longer a boring component category. It is AI infrastructure.

Section 5  ·  BlackwellThe product boundary moved to the rack

With Blackwell, Nvidia shifted the product boundary from a GPU board to a rack-scale AI system. Nvidia's GB200 NVL72 product page describes a rack-scale, liquid-cooled design that connects 36 Grace CPUs and 72 Blackwell GPUs through a 72-GPU NVLink domain that acts as one massive GPU, with the claim of 30x faster real-time trillion-parameter LLM inference compared with prior generations.[3] Nvidia's OCP technical blog describing the same rack design adds the manufacturing and infrastructure detail: more than 5,000 copper cables, a new bus bar architecture supporting 1,400 amps, 130 TB/s all-to-all bandwidth, 260 TB/s AllReduce bandwidth, and direct liquid cooling supporting up to 120 kW of rack cooling capacity.[4]

GB200 NVL72 GPUs
72
Blackwell GPUs in a single NVLink domain acting as one massive GPU.
Grace CPUs
36
Arm-based Grace CPUs paired with Blackwell GPUs in the same rack.
Copper cables
~ 5,000+
Internal copper cabling per rack, per Nvidia's OCP technical disclosure.
Bus bar current
~ 1,400 A
New bus bar architecture supporting 1,400 amps of rack-level current.
All-to-all bandwidth
~ 130 TB/s
All-to-all bandwidth across the 72-GPU NVLink domain.
Rack cooling capacity
~ 120 kW
Direct liquid cooling capacity per Nvidia's OCP reference design.

The right way to read those numbers is as a redefinition of the supplier question. It is no longer only who supplies a component on a GPU board?. It is also who fits inside Nvidia's rack-scale power, cooling, cabling, and manufacturing architecture?. Suppliers that win at the GPU board level may still lose if they cannot fit at the rack level. Suppliers that win at the rack level may not survive the next architecture revision either, because the platform owner can change the rack again.

The AI hardware battleground moved from the GPU board to the full rack.

Section 6  ·  Vera RubinPlatform control goes deeper

Vera Rubin pushes platform control further. Nvidia's Vera Rubin newsroom announcement describes Vera Rubin NVL72 as a rack-scale AI supercomputer platform combining 72 Rubin GPUs, 36 Vera CPUs, Nvidia NVLink 6, ConnectX-9 SuperNICs, and BlueField-4 DPUs, positioned for agentic AI workloads.[5] Nvidia's Vera Rubin NVL72 product page reinforces the rack-scale framing and aligns the platform with Quantum-X800 InfiniBand and Spectrum-X Ethernet networking.[6]

Rubin GPUs
72
Rubin GPUs per Vera Rubin NVL72 rack, per Nvidia.
Vera CPUs
36
Arm-based Vera CPUs paired with Rubin GPUs in the same rack.
NVLink
Gen 6
Next-generation NVLink fabric inside the Vera Rubin NVL72 platform.
NICs and DPUs
CX-9 + BF-4
ConnectX-9 SuperNICs and BlueField-4 DPUs as part of the platform.

Nvidia is not just scaling chips. It is absorbing more of the system-design problem.

Section 7  ·  EvolutionFrom Hopper to Blackwell to Rubin

Era Main object Key architecture shift Supplier implication
Hopper GPU accelerator H100, HBM3, NVLink, high-power SXM Board-level suppliers can be displaced during an architecture transition
Blackwell Rack-scale system GB200 NVL72, liquid cooling, NVLink domain, bus bars, dense copper cabling Suppliers must fit rack-scale power, cooling, and cabling architecture
Vera Rubin AI supercomputer platform Rubin GPUs, Vera CPUs, NICs, DPUs, NVLink 6, networking, system-level integration Nvidia controls more of the full AI factory stack

Section 8  ·  Control surfaceNvidia's platform control is the supplier risk

The reason supplier risk is bigger in 2026 than it was in 2022 is that Nvidia controls or strongly influences more of the system than ever before. A supplier may have signed a Hopper-era contract for a specific module. The Blackwell- and Rubin-era contract is for a position inside a much wider platform that Nvidia is designing top to bottom.

What Nvidia controls or strongly influences
GPU architecture
HBM qualification and allocation
Board design
NVLink and NVSwitch topology
Power delivery architecture
Liquid cooling design
Rack mechanics
Cabling and bus bars
Networking (InfiniBand, Ethernet, NICs, DPUs)
Software stack (CUDA, runtimes, orchestration)
MGX and NVL reference systems
Hyperscaler deployment patterns

The closer Nvidia gets to shipping the whole AI factory, the less protection any single component supplier has.

Section 9  ·  ConcentrationCustomer concentration is still dangerous

The 2022 short report said Nvidia represented up to 55% of the unnamed supplier's revenue.[1] That is an extreme concentration-risk example. The general pattern shows up across the industry, including in public 10-K filings of unrelated suppliers. As a separate, careful illustration, Vicor's 2024 Form 10-K discloses that one customer accounted for approximately 12.1% of net revenue in 2024, compared with approximately 10.7% in 2023 and approximately 12.4% in 2022.[8]

Important framing
The Vicor 10-K reference is used here only to show that customer concentration is a real, ongoing disclosure issue across hardware supply chains. It is not used to identify the 2022 unnamed supplier. The 2022 SemiAnalysis report does not name the supplier, and this essay does not infer the identity. The strategic point stands independently of any specific company: if one customer controls the roadmap and a large share of revenue, the supplier is not fully independent.

If one customer controls the roadmap and a large share of revenue, the supplier is not fully independent. It is attached to the customer's architecture decisions.

Section 10  ·  ScaleNvidia's scale makes every supplier decision larger

Nvidia's Q1 FY2027 results report quarterly revenue of approximately US$81.6 billion, with data center revenue of approximately US$75.2 billion. Under the prior reporting structure, data center compute revenue was approximately US$60.4 billion and data center networking revenue was approximately US$14.8 billion. Jensen Huang described the AI infrastructure buildout as the largest infrastructure expansion in human history.[7]

Q1 FY2027 revenue
~ US$81.6B
Quarterly Nvidia revenue, per the Q1 FY2027 release.
Data center revenue
~ US$75.2B
Q1 FY2027 data center revenue, per Nvidia.
Data center compute
~ US$60.4B
Under the prior reporting structure.
Data center networking
~ US$14.8B
Under the prior reporting structure.

The interpretation is straightforward. A small component inside Nvidia's architecture can become a large business. Losing that component can destroy the revenue story for a single-customer supplier. The asymmetry runs in both directions: Nvidia's scale turns supplier wins into growth engines and supplier losses into existential shocks. That is exactly why the 2022 short report drew a 20%+ stock reaction on publication day.[1]

Nvidia's scale turns supplier wins into growth engines and supplier losses into existential shocks.

Section 11  ·  Proof pointsWhat suppliers must prove now

The honest version of the supplier-strategy question is not how do I lock in Nvidia?. Suppliers cannot lock in a customer that designs more of the system every year. The version that survives the platform is closer to how do I remain useful even when Nvidia redraws the system?.

What credible AI hardware suppliers must prove
  1. Multiple customers across hyperscalers, AI accelerator vendors, and ODMs.
  2. Multiple sockets per customer, not a single one-shot win.
  3. Roadmap alignment with Nvidia and other platform owners.
  4. Ability to support liquid cooling and rack-scale architecture.
  5. Strong cost and efficiency advantages, not just feature parity.
  6. Dual-source resilience inside customer programs.
  7. Manufacturability at Nvidia-scale volumes and reliability.
  8. Survival across reference architecture changes (MGX, NVL, future revisions).
  9. Value that Nvidia cannot easily internalize or commoditize.

The safest supplier is not the one with one big Nvidia win. It is the one that remains useful even when Nvidia redraws the system.

Section 12  ·  AI infrastructureWhat this means beyond power delivery

The lesson applies far beyond power delivery. The same platform-control dynamic shapes how value gets distributed across AI infrastructure as Nvidia, hyperscalers, and OCP-aligned partners absorb more of the stack.

HBM suppliers
Advanced packaging suppliers (CoWoS-class)
Power module companies
Connector and cable companies
Thermal management companies
Server OEMs and ODMs
Networking vendors
Optical interconnect vendors
Board suppliers
Mechanical and rack suppliers

AI infrastructure is becoming a platform war, and platform wars are dangerous for suppliers that only own one piece of the stack.

Section 13  ·  EvidenceEvidence ledger

Claim
Evidence
Interpretation
Hopper created a supplier-risk event
The 2022 SemiAnalysis short report said a long-time Nvidia supplier was cut out and Nvidia represented up to 55% of revenue, with a 20%+ stock reaction on the day.
AI supplier concentration can be dangerous.
Hopper was a major architecture transition
Nvidia describes GH100 as TSMC 4N, approximately 80B transistors, HBM3, fourth-gen Tensor Cores, fourth-gen NVLink, and PCIe Gen 5.
Supplier changes happened during a major platform shift.
H100 was already power-intensive
The 2022 short report highlighted a 700W top-spec SXM package.
Power delivery was becoming strategic.
Blackwell moved the product boundary to the rack
GB200 NVL72 connects 36 Grace CPUs and 72 Blackwell GPUs in a liquid-cooled NVLink domain, with Nvidia claiming 30x real-time trillion-parameter LLM inference.
The platform is now rack-scale.
Rack power and cooling became core architecture
Nvidia's OCP blog cites a 1,400A bus bar, more than 5,000 copper cables, 130 TB/s all-to-all bandwidth, 260 TB/s AllReduce bandwidth, and 120 kW direct-liquid cooling.
Power and cooling are part of AI system design.
Rubin expands platform control
Vera Rubin NVL72 combines 72 Rubin GPUs, 36 Vera CPUs, NVLink 6, ConnectX-9 SuperNICs, and BlueField-4 DPUs.
Nvidia controls more of the AI factory stack.
Nvidia's scale makes design wins huge
Q1 FY2027 revenue was approximately US$81.6B with data center at approximately US$75.2B.
Supplier wins and losses inside Nvidia have massive impact.
Concentration risk remains real
Vicor's 2024 Form 10-K discloses one customer at approximately 12.1% of net revenue in 2024 (10.7% in 2023, 12.4% in 2022).
Supplier dependence can persist even when specific customers change.

Section 14  ·  Risk registerRisks and limitations

This essay is an analysis of public disclosures and historical context. It is not investment advice. It does not recommend buying or shorting any stock. The honest risks against the read above run in several directions, and they are listed here so the argument can be stress-tested.

The 2022 short report is a single source describing a specific situation. The unnamed supplier and the precise scope of the architecture change are not fully disclosed publicly.
Platform control does not equal full vertical integration. Nvidia still relies on TSMC, HBM suppliers, packaging partners, networking ecosystems, and OEM/ODM manufacturing.
Supplier strategies vary widely. Some are deeply embedded across multiple programs and platforms. Others depend more narrowly on specific sockets.
AI accelerator competition (AMD, custom hyperscaler silicon, ASIC platforms) reduces Nvidia's relative platform leverage in some segments.
Rack-scale architecture revisions can also create new supplier opportunities, not only displace existing ones.
Hyperscalers may push standardisation (OCP, MGX) that re-opens the supplier ecosystem even as Nvidia integrates more of the stack.
AI demand normalisation or model-efficiency improvements could reduce the absolute scale of supplier opportunities tied to Nvidia.
Power, cooling, and cabling claims in vendor materials are vendor claims. Field performance at hyperscale may differ.
The Vicor 10-K reference is used solely to illustrate customer concentration as a general disclosure issue and should not be read as a comment on any specific supplier's relationship with Nvidia.
Nvidia roadmap claims (Vera Rubin timelines, architecture details) are Nvidia's claims, and product schedules can change.

Section 15  ·  Bottom lineBottom line

Bottom line

The 2022 Hopper supplier cut was not just a stock-specific short report. It was an early warning about AI hardware platform risk. A design win inside Nvidia can be enormously valuable, but it is not a permanent moat.

As Nvidia moved from H100 GPUs to GB200 rack-scale systems and then Vera Rubin AI supercomputers, the supplier question changed. Suppliers now have to survive Nvidia's full-stack control over power, cooling, networking, cabling, packaging, software, and rack-scale architecture.

In AI infrastructure, the component supplier can win the socket. But the platform owner owns the system.

Section 16  ·  DefinitionsGlossary

Design win
A successful selection of a supplier's part into a customer's product. Typically generates revenue once the customer ships, but can be displaced at the next architecture transition.
Platform risk
The risk that a platform owner (here Nvidia) changes the architecture in ways that remove, reduce, or commoditize a supplier's position inside the platform.
Customer concentration
The extent to which a supplier's revenue depends on a small number of customers. High concentration creates leverage for the customer and risk for the supplier.
GPU
Graphics processing unit. Highly parallel accelerator used for AI training and inference, scientific compute, and graphics.
SXM
Nvidia's high-power GPU module form factor used in data center systems. SXM modules typically support higher TDPs and integrated NVLink than PCIe-style cards.
HBM
High Bandwidth Memory. Stacked DRAM integrated with logic dies through advanced packaging. Central to AI accelerator memory bandwidth.
NVLink
Nvidia's high-bandwidth GPU-to-GPU and GPU-to-CPU interconnect. Multiple generations (4, 5, 6) extend bandwidth and domain size.
NVSwitch
Nvidia's switch silicon for NVLink fabrics. Enables many GPUs to communicate as a single logical domain across a rack or larger system.
Power delivery network
The set of converters, regulators, capacitors, and traces that supply current to chips. At AI scale, this extends from the rack down to the package.
Voltage regulation
Conversion of higher distribution voltages down to the precise low voltages chips need. Critical for AI accelerator efficiency and stability.
Bus bar
A thick conductor used to distribute large currents efficiently inside a rack. Rack-scale AI systems push bus bar architecture into new territory.
Liquid cooling
Cooling that uses circulating liquid rather than only air to remove heat. Standard for high-density AI racks like GB200 NVL72.
Rack-scale system
A computing platform whose primary product is an entire rack, not individual servers or boards. GB200 NVL72 is a rack-scale AI system.
AI factory
An informal term used by Nvidia and others to describe large-scale data centers dedicated to AI training and inference, treated as a single integrated production system.
MGX
Nvidia's modular server reference architecture program that allows OEMs and ODMs to build a range of AI servers from common building blocks.
OCP
Open Compute Project. An industry initiative for open-source data center hardware designs. Nvidia has contributed GB200 NVL72 designs to OCP.
Supplier concentration
The extent to which a customer depends on a small number of suppliers. The mirror of customer concentration. Both can be sources of risk.
Dual sourcing
Qualifying two or more suppliers for the same critical component. Used by platform owners to reduce dependency and increase pricing leverage.

Section 17  ·  MethodSources and method notes

How this essay reads sources

The 2022 SemiAnalysis short report is treated as a historical case study about supplier concentration and platform risk. The public excerpt does not name the supplier, and this essay does not infer or attribute the identity to any specific company. The Vicor 2024 Form 10-K reference is used solely to illustrate that customer concentration disclosure is a real, ongoing issue across hardware supply chains; it is not used to identify the 2022 supplier.

The 2026 read is built from primary Nvidia sources: the Hopper architecture deep dive, the GB200 NVL72 product page, Nvidia's OCP technical blog on GB200 NVL72 designs, the Vera Rubin platform announcement, the Vera Rubin NVL72 product page, and Nvidia's Q1 FY2027 results. Company claims about performance, bandwidth, cooling capacity, and roadmap are treated as company claims, not as endorsed forecasts. The structural arguments that design wins are not moats, that the product boundary moved from board to rack to platform, and that supplier risk increases as Nvidia integrates more of the stack are independent analysis.

Footnotes  ·  primary sources

  1. SemiAnalysis, “Short Report: Nvidia Supplier Cut Out Of Next Generation Hopper GPUs — Nvidia Represents Up To 55% of Revenue,” 2022 (PDF supplied by author). Historical anchor used in this essay for the claim that a long-time Nvidia supplier had seemingly been cut out of the next-generation Hopper platform, the up to 55% Nvidia share of supplier revenue, the Hopper launch context, the approximately 80B transistor claim, the up to 6x Transformer performance claim, the 700W top-spec SXM package framing, the more than 20% stock reaction on publication day, and the page 2 Hopper board and package visual framing. The public excerpt does not name the supplier, so this essay does not identify it either.
  2. Nvidia Developer Blog, “Nvidia Hopper Architecture In-Depth,” developer.nvidia.com/…/hopper-architecture-in-depth. Source for the Hopper architecture context used in this essay, including TSMC 4N, approximately 80B transistors, approximately 814 mm² die, HBM3 memory controllers, fourth-generation Tensor Cores, fourth-generation NVLink, and PCIe Gen 5 support.
  3. Nvidia, “GB200 NVL72,” nvidia.com/…/gb200-nvl72. Source for the GB200 NVL72 framing as a rack-scale, liquid-cooled system connecting 36 Grace CPUs and 72 Blackwell GPUs through a 72-GPU NVLink domain that acts as one massive GPU, with Nvidia's claim of 30x faster real-time trillion-parameter LLM inference.
  4. Nvidia Developer Blog, “Nvidia Contributes GB200 NVL72 Designs to Open Compute Project,” developer.nvidia.com/…/gb200-nvl72-ocp. Source for the more than 5,000 internal copper cables, the new approximately 1,400 A bus bar architecture, 130 TB/s all-to-all bandwidth, 260 TB/s AllReduce bandwidth, 120 kW direct liquid cooling capacity, and the rack-scale power, cooling, and cabling architecture used in this essay.
  5. Nvidia, “Rubin Platform AI Supercomputer” nvidianews.nvidia.com/…/rubin-platform-ai-supercomputer. Source for Vera Rubin NVL72 combining 72 Rubin GPUs, 36 Vera CPUs, Nvidia NVLink 6, ConnectX-9 SuperNICs, and BlueField-4 DPUs, with the rack-scale AI supercomputer and agentic-AI framing used in this essay.
  6. Nvidia, “Vera Rubin NVL72,” nvidia.com/…/vera-rubin-nvl72. Source for the Vera Rubin NVL72 product framing including Rubin GPUs, Vera CPUs, ConnectX-9 SuperNICs, BlueField-4 DPUs, NVLink 6, Quantum-X800 InfiniBand, and Spectrum-X Ethernet networking, used in this essay's rack-scale AI infrastructure framing.
  7. Nvidia, “Nvidia Announces Financial Results for First Quarter, Fiscal 2027,” nvidianews.nvidia.com/…/q1-fy2027. Source for Q1 FY2027 quarterly revenue of approximately US$81.6B, data center revenue of approximately US$75.2B, data center compute of approximately US$60.4B and data center networking of approximately US$14.8B under the prior reporting structure, and Jensen Huang's framing of the AI infrastructure buildout used in this essay.
  8. Vicor Corporation, 2024 Form 10-K sec.gov/…/vicr-20241231. Used in this essay only as a general illustration of customer-concentration disclosure in hardware supply chains, citing one customer at approximately 12.1% of net revenue in 2024 (compared with approximately 10.7% in 2023 and approximately 12.4% in 2022). Explicitly not used to identify the unnamed 2022 supplier in the SemiAnalysis short report.
← All essays