Chip Supply Chain
Specimen 00 · the whole chip

Chapter 15The economics

Chapter 15: The Economics of AI Chips

Epoch AI estimated in December 2025 that Nvidia’s B200 AI chip cost about $6,400 to build, mostly for memory and packaging. Its processor silicon cost about $900.

2,120 words / 9 minSpecimen: wafer costUpdated 29 October 202633 sources

Link to chapterChapter PDF

Specimen 15 · The Economics of AI Chips

Epoch AI’s December 2025 estimate of what it cost to build one B200, drawn as a column with each layer as tall as its share of the bill. From the bottom: $2,900 of memory in garnet, $1,100 of packaging, $1,000 of packaging yield loss drawn hollow, $900 of logic dies, $480 of power delivery, board and test. The total is about $6,400, against a selling price of $30,000 to $40,000 per chip reported in March 2024.

In plain terms

Most of what it costs to build Nvidia’s B200, an AI from its Blackwell line, goes to memory and packaging that a few firms supply. A is the list of parts inside a product and what each one costs. Epoch AI estimated in December 2025 that the B200 cost about $6,400 to build1. The two , the pieces of silicon that do the calculating, cost about $900 of that1. The beside them accounted for about half, and the packaging that wires the and memory together cost more than the dies1. Three firms, SK hynix, Samsung and Micron, dominate the world market for that memory2. One Taiwanese firm, TSMC, does that packaging1.

In short

Epoch AI says Nvidia’s margin on Blackwell may be lower than the gap between the B200’s build cost and its selling price suggests1. It cites reports from March 2024 that put the B200’s selling price at $30,000 to $40,000 per 1. Epoch AI notes that most Blackwell revenue comes from servers and systems, which may carry lower margins1.

In 2026 Micron’s , the share of sales left after production costs, passed Nvidia’s. Micron’s rose from 39.8 percent for the year to August 2025 to 80.7 percent for the year to September 20263. It reached 86.8 percent in the last quarter of that year3. Nvidia’s was 75.0 percent in the quarter to July 20264. Epoch AI’s estimate priced memory at $14 to $17 per gigabyte, before the 2026 rise in memory makers’ sales and margins5.

A B200, one of Nvidia’s Blackwell AI , sells for far more than it costs to build. In December 2025 Epoch AI estimated that building one cost about $6,4001. Epoch AI cites reports from March 2024 that put its selling price at $30,000 to $40,000 per chip1. Most B200s are sold inside Nvidia’s GB200 systems6. Epoch AI’s chip sales model, with data through October 2025, puts what Nvidia earns per for the computing hardware in those racks at $33,000 to $42,0006. That figure leaves out the networking gear in the racks6.

What an Nvidia B200 costs to build

Three numbers set the cost of a , the finished rectangle of circuitry cut out of a .

  • Wafer price. TSMC does not publish it. TSMC makes Nvidia’s Blackwell chips on a custom-built process called 4NP7. Epoch AI estimates $17,000 for a 12-inch () wafer on that process1.
  • Dies per wafer. A wafer is a circle and dies are rectangles, so the edge of the wafer goes to waste. The bigger the die, the fewer fit. Each of the two in a B200 is about 800 1. Nvidia calls them , meaning they are near the largest size a can print in one shot7.
  • . Yield is the share of dies that work. Defects land at random, so a bigger die is more likely to catch one. Epoch AI models Blackwell die yield at 40 to 70 percent, centered on 601.

Together, those three numbers put the two dies inside a B200 at about $9001. Epoch AI’s estimate for the whole B200 counts only variable costs, which rise with each chip made1. It leaves out fixed costs such as research and development1.

The B200’s parts and rack costs

Memory is the largest line

The largest line on the bill is the 192 GB of beside those dies, a generation called HBM3E1. At $14 to $17 per gigabyte it comes to about $2,900, which Epoch AI counts as half of the build cost1. Epoch AI says HBM prices are negotiated deal by deal and can vary significantly with volume and contract length1. The $14 to $17 range dates from December 2025, before the 2026 rise in memory makers’ sales and margins described below5. SK hynix said and prices rose significantly in the second quarter of 2026 from the quarter before, led by high-performance products for AI servers5.

Few firms make this memory, and one of them says demand has outrun its supply. SK hynix, Samsung and Micron dominate the world market for it2. SK hynix said in July 2026 that demand for its products had exceeded its available supply in recent quarters8.

Packaging, substrate and test

-L, the TSMC process that mounts dies and memory on one carrier and wires them together, adds about $1,100, more than the logic dies it carries1. Epoch AI models packaging yield at 65 to 95 percent1. A package that fails testing loses its logic dies and memory stacks1. Spreading that loss over the good units adds about $1,000 to each B2001. Power delivery hardware, the , and final assembly and testing add $4801. See Advanced packaging for how tight packaging capacity has been.

The B200 bill of materials

LineEstimateRange
HBM3E, 192 GB$2,900$2,800–3,100
CoWoS-L packaging$1,100$1,000–1,200
Packaging yield loss$1,000$430–1,700
Logic die, two on TSMC 4NP$900$720–1,200
Power delivery, board, assembly and test$480$370–600
Manufacturing cost~$6,400$5,700–7,300
Reported selling price per chip, March 2024$30,000–40,000
Nvidia compute revenue per GPU in GB200 racks, data to October 2025$33,000–42,000

Epoch AI models each line as a range, deriving the packaging line from TSMC’s revenue and Nvidia’s share of CoWoS capacity1.

Epoch AI says Nvidia’s actual margin on Blackwell may be lower than the gap between the B200’s build cost and its reported chip price suggests1. The estimate covers the module only and leaves out servers and rack-scale systems1. Epoch AI notes that most Blackwell revenue comes from those systems, which may carry lower margins1.

Epoch AI’s line-by-line estimate covers the B200, which is no longer Nvidia’s newest chip. In its results for the quarter to July 2026, Nvidia said its Vera Rubin platform was into full production4. Each Rubin GPU carries up to 288 GB of HBM4, a newer generation of the stacked memory9.

Figure 15.1

Memory, $2,900, is the largest line in Epoch AI’s 2025 B200 estimate

Where the ~$6,400 of B200 manufacturing cost goes, Epoch AI estimate, December 2025

Where the ~$6,400 of B200 manufacturing cost goes, Epoch AI estimate, December 2025
HBM3E, 192 GB$2,900
CoWoS-L packaging$1,100
Packaging yield loss$1,000
Logic die, two on 4NP$900
Power delivery, board, test$480
Source: Epoch AI B200 cost model, from a $17,000 4NP wafer, 40-70% die yield, $14-17 per GB of memory, and a packaging line derived from TSMC’s advanced packaging revenue 1CSV

From package to rack to cluster

Each new generation of Nvidia rack costs its owner more per hour than the last10. At rack scale, cost is measured per GPU-hour, one GPU running for one hour. For an operator that owns the machine, SemiAnalysis put a at $3.57 per GPU-hour in July 2026, against $1.84 for a GB200 and $2.36 for a GB30010.

But SemiAnalysis found Rubin cheaper than GB200 and GB300 per token, a short chunk of text10. Rubin stayed cheaper at every level of interactivity SemiAnalysis compared, where interactivity is the number of tokens per second each user receives10. That finding rests on early results from the cloud provider CoreWeave running the DeepSeek R1 model, and SemiAnalysis said it had not verified them10.

One level up, a cluster needs power and a network as well as servers. SemiAnalysis estimated in June 2024 that a cluster of 100,000 of Nvidia’s earlier H100 GPUs cost over $4 billion in servers11. The cluster needs more than 150 of data center capacity and uses 1.59 a year, which costs $123.9 million at $0.078 per kWh11. In the network design Nvidia recommended, 98,304 link the cluster’s GPUs to the first layer of switches, called leaf switches11.

Profit margins along the supply chain

FirmStageQuarter ended
MicronHBM and DRAMSeptember 202686.8%3
NvidiaAccelerator designJuly 202675.0%4
TSMCWafers and CoWoSSeptember 202666.4%12
KLA (non‑GAAP)InspectionJune 202662.4%13
ASMLSeptember 202657.5%14
Lam ResearchJune 202651.7%15
Applied MaterialsJuly 202650.3%16
ASE TechnologyPackaging and test segmentJune 202627.3%17
AmkorJune 202616.8%18

Gross margin is highest at the memory maker Micron and lowest in assembly and test. The firms in design, equipment and wafers earn gross margins of 50.3 to 75.0 percent, and those in assembly and test 16.8 to 27.3 percent. Micron’s gross margin rose from 39.8 percent for the year to August 2025 to 80.7 percent for the year to September 20263. It reached 86.8 percent in the last quarter of that year3.

Each figure except KLA’s comes from the firm’s own earnings release. KLA’s shareholder letter gives only a non-GAAP gross margin for the quarter, 62.4 percent13. KLA’s non-GAAP figures adjust its results under US accounting rules, known as GAAP, to leave out items such as acquisition-related charges and restructuring, severance and other charges19. For ASE the figure is its packaging and test segment17. The company as a whole had a gross margin of 21.0 percent, because about a third of its revenue came from electronics manufacturing services, which earned 8.9 percent17.

Two other firms’ results give no gross margin. SK hynix’s was 76 percent in the quarter to June 2026, and its revenue was up 257 percent from a year earlier5. The maker Ibiden’s operating profit was 21.8 percent of net sales in the quarter to June 202620.

Figure 15.2

Micron’s gross margin, 86.8 percent, was highest of nine firms in 2026

Gross margin, quarters ended June to September 2026

Gross margin, quarters ended June to September 2026
Micron86.8%
Nvidia75%
TSMC66.4%
KLA (non‑GAAP)62.4%
ASML57.5%
Lam Research51.7%
Applied Materials50.3%
ASE packaging and test27.3%
Amkor16.8%
Source: Each firm’s own quarterly earnings release, for quarters ending between June and September 2026 (TSMC’s linked). ASE’s bar is its packaging and test segment. KLA’s bar is the non-GAAP figure from its shareholder letter 12CSV
Figure 15.3

Memory, $2,900, is the largest cost of building a B200, which sells for $30,000 to $40,000

One B200, drawn to scale. Epoch AI’s December 2025 estimate of what it cost to build, inside the selling price reported in March 2024

One B200, drawn to scale. Epoch AI’s December 2025 estimate of what it cost to build, inside the selling price reported in March 2024
Memory, HBM3E 192 GB$2,900
CoWoS-L packaging$1,100
Packaging yield loss$1,000
Logic dies, two on 4NP$900
Power delivery, board, test$480
Build costabout $6,400
Selling price per chip$30,000–40,000
Source: Epoch AI B200 cost model, December 2025, and the March 2024 reports of its selling price that Epoch AI cites 1CSV

World chip sales are forecast to jump in 2026, with memory growing far faster than logic. The Semiconductor Industry Association, using monthly data compiled by World Semiconductor Trade Statistics, counted $791.7 billion of global chip sales in 202521. In its June 2026 forecast, World Semiconductor Trade Statistics put 2025 slightly higher, at $795.6 billion22. It forecast $1.51 trillion for 2026, with memory growing about 250 percent and logic 37 percent22.

In 2025 at five large tech firms far exceeded the world’s sales of . The industry group SEMI counted $135.1 billion of semiconductor equipment billings in 202523. Capital spending at Alphabet, Amazon, Meta, Microsoft and Oracle came to $448.2 billion in 2025, the sum of Epoch AI’s quarterly figures24. Epoch AI says that spending would reach $770 billion in 2026 if the trend continues24. These totals overlap, so they cannot be added together.

Which stages stay expensive and which get cheaper

Each stage falls into one of three classes, by whether its cost falls as volume grows. A structurally expensive stage stays costly because the work is hard, capacity is short or one firm sets the price. A stage compressible with scale gets cheaper as volume grows. An already commoditized stage has several competing suppliers. The classes are a judgment drawn from the sources in this chapter and the stage chapters.

StageClassWhy
lithographyStructurally expensiveOne supplier, ASML. See Lithography.
and Structurally expensiveEvery design needs its own mask set, and every EUV layer a flawless mask. See Photomasks and pellicles.
Structurally expensiveKLA held 56.8 percent of the market for these tools in 2025, by CSET’s analysis of TechInsights data25. See Metrology and inspection.
HBMStructurally expensiveSK hynix says HBM needs more wafer input than ordinary DRAM8. See Memory and HBM.
Advanced packaging, CoWoS-LStructurally expensiveModeled packaging yield is 65 to 95 percent, and a failed package loses its dies and memory.
Structurally expensiveInvestment can add substrate lines, but Ajinomoto said in June 2026 that the insulating film it makes for these substrates has held over 95 percent of its market since launch26. See Substrates and PCBs.
Power and cooling plantStructurally expensiveIn the United States, grid connections face long queues, so at least three of OpenAI’s seven Stargate sites will use on-site natural gas plants27. GE Vernova’s backlog of gas power equipment, together with agreements reserving future production slots, grew from 100 GW to 116 GW in the second quarter of 202628. See Data centers and power.
Optical transceiversStructurally expensiveIn May 2026 Nikkei reported that building AI data centers had caused shortages and price rises for lasers, optical fibers and other optical parts29. Nvidia agreed in March 2026 to invest $2 billion in the laser maker Lumentum, with rights to future capacity30. It agreed a similar $2 billion deal with Coherent31. See Systems and networking.
Leading-edge logic wafersCompressible with scaleA very large die yields well only once a process matures. See Foundries and fabs.
Design and Compressible with scaleA design is paid for once, and its cost is spread across every chip sold. See Chip design, EDA and IP.
Deposition and etch toolsCompressible with scaleMost deposition and etch steps have three or four established tool makers. See Deposition and etch.
Test and assembly (OSAT)Already commoditizedGross margins of 16.8 to 27.3 percent at Amkor and ASE’s packaging and test segment in the quarter to June 2026.
Wafers and Already commoditizedSeveral firms make approved 300 mm wafers. By Siltronic’s estimate, the five largest wafer makers served around 75 percent of the market in the first three quarters of 202532. See Silicon and wafers.
Rack integration ()Already commoditizedIn October 2025 Nvidia said more than 50 makers of systems and components were gearing up for the Vera Rubin NVL7233. See Systems and networking.

Flashcards

Recall each answer, then turn the card.

Card 1 of 4. Question

Where does most of the cost of building an Nvidia B200 go?

Card 1 of 4. Answer

To its stacked memory and its packaging, which a few firms supply.

Reread: How it works

Sources (33)

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

  1. ANVIDIA’s B200 costs around $6,400 to produceEpoch AI · 10 December 2025
  2. AUnderstanding the Biden Administration’s Updated Export ControlsCenter for Strategic and International Studies · 11 December 2024
  3. AMicron Technology, Inc. Reports Record Fiscal Fourth-Quarter and Full-Year 2026 ResultsMicron Technology · 30 September 2026
  4. ANVIDIA Announces Financial Results for Second Quarter Fiscal 2027Nvidia · 26 August 2026
  5. ASK hynix Announces 2Q26 Financial ResultsSK hynix · 29 July 2026
  6. AAI chip sales documentationEpoch AI · 8 January 2026
  7. ANVIDIA Blackwell ArchitectureNvidia · 18 December 2025
  8. ASK hynix, prospectus on Form 424B4 dated 10 July 2026U.S. Securities and Exchange Commission (filing by SK hynix) · 10 July 2026
  9. AInside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AINvidia · 21 July 2026
  10. BVera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture AnalysisSemiAnalysis · 23 July 2026
  11. B100,000 H100 Clusters: Power, Network Topology, Ethernet vs InfiniBand, Reliability, Failures, CheckpointingSemiAnalysis · 17 June 2024
  12. ATSMC Reports Third Quarter EPS of NT$29.43U.S. Securities and Exchange Commission (filing by TSMC) · 15 October 2026
  13. ALetter to Shareholders Q4 Fiscal 2026KLA Corporation · 28 July 2026 · p. 3
  14. AASML reports €11.7 billion total net sales and €5.0 billion net income in Q3 2026U.S. Securities and Exchange Commission (filing by ASML Holding) · 14 October 2026
  15. ALam Research Corporation Reports Financial Results for the Quarter Ended June 28, 2026Lam Research · 29 July 2026
  16. AApplied Materials Announces Third Quarter 2026 ResultsApplied Materials · 13 August 2026
  17. AASE Technology Holding Co., Ltd. Reports Its Unaudited Consolidated Financial Results for the Second Quarter of 2026PR Newswire · 30 July 2026
  18. AAmkor Technology, current report on Form 8-K dated 27 July 2026U.S. Securities and Exchange Commission (filing by Amkor Technology) · 27 July 2026
  19. AKLA Corporation Reports Fiscal 2026 Fourth Quarter and Full Year ResultsKLA Corporation · 28 July 2026
  20. AOutline of Consolidated Financial Results For the First Quarter Ended June 30, 2026Ibiden · 4 August 2026
  21. AGlobal Annual Semiconductor Sales Increase 25.6% to $791.7 Billion in 2025Semiconductor Industry Association · 6 February 2026
  22. AGlobal Semiconductor Market Surges Beyond USD 1.5 Trillion in 2026 Driven by Extraordinary Memory ExpansionWorld Semiconductor Trade Statistics · 2 June 2026
  23. ASEMI Reports Global Semiconductor Equipment Billings Reached $135 Billion in 2025, Up 15% Year-on-YearSEMI · 7 April 2026
  24. AHyperscaler capex has quadrupled since GPT-4’s releaseEpoch AI · 26 February 2026
  25. ASupply Chain Explorer: Advanced ChipsEmerging Technology Observatory (ETO) · Accessed 9 September 2026
  26. AAjinomoto, Growth Strategy for the Electronic Materials Business, 30 June 2026Ajinomoto · 30 June 2026 · p. 3
  27. AOpenAI Stargate: where the US sites standEpoch AI · 17 April 2026
  28. AGE Vernova, current report on Form 8-K dated 22 July 2026U.S. Securities and Exchange Commission (filing by GE Vernova) · 22 July 2026
  29. BAI demand strains supplies of lasers, fiber and other optical techNikkei Asia · 27 May 2026
  30. ANVIDIA Announces Strategic Partnership With Lumentum to Develop State-of-the-Art Optics TechnologyNvidia · 2 March 2026
  31. ANVIDIA and Coherent Announce Strategic Partnership to Develop Optics Technology to Scale Next-Generation Data Center ArchitectureNvidia · 2 March 2026
  32. AInvestor Presentation: Foundation of Digital LifeSiltronic · April 2026 · p. 6
  33. ANVIDIA, Partners Drive Next-Gen Efficient Gigawatt AI Factories in Buildup for Vera RubinNvidia · 13 October 2025