Silicon Analysts
Loading...

How Much Does It Cost to Make a Semiconductor Chip?

A complete guide to semiconductor chip manufacturing costs in 2026. Covers wafer pricing by process node, HBM memory economics, advanced packaging costs, and full cost breakdowns for leading AI accelerators.

Last updated: · Data sources: Industry estimates, TSMC/Samsung public filings, teardown analysis · Model your own chip costs →

Quick answer: Semiconductor chip manufacturing costs range from $1–5 for a simple IoT chip (28nm) to $3,000–13,000+ for a cutting-edge AI accelerator (3nm with HBM and CoWoS packaging). The three biggest cost drivers are the wafer ($3,000–$20,000 per wafer depending on node), HBM memory ($200–$300 per stack × 6–8 stacks), and advanced packaging ($70–$120 per unit).

Wafer Costs by Process Node

Wafer cost is the foundation of chip economics. Prices vary by foundry, node, volume, and contract terms. Leading-edge nodes (3nm, 5nm) use EUV lithography, which drives costs significantly higher than mature nodes.

NodeFoundryAvg Wafer CostRange
3nmTSMC N3/N3E$19,500$17,000–$22,000
3nmSamsung 3GAE$15,000$13,000–$17,000
5nmTSMC N5/N4$18,500$16,000–$21,300
5nmSamsung 5nm$13,000$11,000–$15,000
7nmTSMC N7$9,500$8,100–$10,900
16nmIntel$4,500$3,800–$5,200
28nmTSMC$3,000$2,550–$3,450

Source: TrendForce, Morgan Stanley, CSET, industry consensus estimates. Model specific wafer costs →

Packaging Costs

Advanced packaging has become a major cost component and supply bottleneck. AI accelerators require 2.5D or 3D packaging to integrate HBM memory stacks alongside logic dies.

Package TypeCost per UnitTypical Use
QFN$0.80Cost-optimized IoT/MCU
FC-BGA (flip-chip)$8Standard CPUs, GPUs
FOWLP/InFO$15Mobile SoCs (Apple A-series)
Intel EMIB$352.5D bridge (Ponte Vecchio)
TSMC CoWoS$702.5D interposer (H100, MI300X)
TSMC SoIC (3D)$1203D stacking (AMD chiplets)

CoWoS demand exceeds supply by 40–50% with lead times of 40–52 weeks. Model packaging costs →

Memory Costs (HBM, DRAM)

High Bandwidth Memory (HBM) is the fastest-growing cost component for AI chips. Each stack uses through-silicon vias (TSVs) to vertically connect multiple DRAM dies, delivering 5–10× the bandwidth of standard GDDR.

Memory TypeCost / StackCapacity / StackBandwidth / Stack
HBM2e$12016 GB460 GB/s
HBM3$20024 GB819 GB/s
HBM3E$30036 GB1.18 TB/s
HBM4 (est.)$50048 GB>2 TB/s

SK Hynix leads with ~50–55% market share, followed by Samsung (~35–40%) and Micron (~5–10%). Full HBM market analysis →

AI Chip Cost Breakdowns

Estimated manufacturing cost breakdowns for major AI accelerators. All costs are bill-of-materials (BOM) estimates excluding design (NRE) costs.

NVIDIA H100 SXM5

ComponentEst. CostDetails
Logic Die$300814mm², TSMC 4N
HBM3 Memory$1,3506 stacks × $200, 80GB total
CoWoS-S Packaging$7502.5D silicon interposer
Test & Assembly~$920Substrate, test, assembly
Total Manufacturing~$3,320Sell price: ~$28,000 (88% margin)

NVIDIA B200

ComponentEst. CostDetails
Logic Dies (2×)~$6002× ~800mm², TSMC 4NP (Blackwell)
HBM3E Memory~$2,4008 stacks × $300, 192GB total
CoWoS-L Packaging~$1,000Organic interposer + LSI bridges
Test & Assembly~$2,400Multi-die module complexity
Total Manufacturing~$6,400Sell price: ~$40,000 (84% margin)

AMD MI300X

ComponentEst. CostDetails
Logic Chiplets~$1,500Multi-die, TSMC N5/N6, 1,725mm² total
HBM3 Memory~$1,6008 stacks × $200, 192GB total
SoIC/3D Packaging~$1,2003D stacking + CoWoS integration
Test & Assembly~$1,000Complex multi-die module
Total Manufacturing~$5,300Sell price: ~$15,000 (65% margin)

Source: Epoch AI, Raymond James, TrendForce, SemiAnalysis teardown data. Compare any two chips side by side →

Why Chip Manufacturing Costs Are Rising

Moore's Law historically delivered both more transistors and lower cost per transistor. That economic equation broke at the 5nm node. Three structural shifts are driving costs higher:

  1. EUV lithography costs: Advanced nodes require multi-pass EUV patterning. TSMC 3nm wafers cost ~6.5× more than 28nm. Each EUV scanner costs $350M+.
  2. Advanced packaging overhead: CoWoS, SoIC, and HBM integration add $500–$1,500+ per chip — cost layers that didn't exist at older nodes.
  3. Memory bandwidth demands: AI workloads require 6–8 HBM stacks per chip at $200–$300 each, adding $1,200–$2,400 to BOM. This is the fastest-growing cost component.

The result: while a 28nm chip's BOM might be $3–10, a leading-edge AI chip's BOM is $3,000–$13,000 — a 1,000× increase driven by packaging and memory, not just the silicon.

Interactive Cost Modeling

Use our free tools to model specific chip cost scenarios:

Methodology & Data Sources

Cost estimates on this page are derived from multiple sources:

  • Epoch AI Monte Carlo cost models
  • Raymond James semiconductor equity research
  • TrendForce quarterly foundry reports
  • SemiAnalysis teardown and cost analysis
  • Official vendor spec sheets (NVIDIA, AMD, Intel)
  • CSET (Georgetown) semiconductor cost research

Data last updated: . All costs are estimates and may vary by volume, contract terms, and market conditions. This data is available programmatically via our free JSON API.

Frequently Asked Questions

How much does it cost to make a semiconductor chip?
Semiconductor chip manufacturing cost varies enormously. A simple IoT chip on 28nm costs $1–5 per die. A mainstream mobile SoC on 5nm costs $20–50. An advanced AI accelerator like the NVIDIA H100 costs approximately $3,320 to manufacture (including die, HBM memory, packaging, and test). The most complex chips like the NVIDIA GB200 superchip cost over $13,000 to manufacture.
How much does a TSMC wafer cost in 2026?
TSMC wafer costs in 2026: 3nm (N3/N3E) averages ~$19,500 per wafer (range $17,000–$22,000). 5nm (N5) averages ~$18,500 ($16,000–$21,300). 7nm (N7) averages ~$9,500 ($8,100–$10,900). 28nm averages ~$3,000 ($2,550–$3,450). Samsung 3nm is approximately 20–25% cheaper at ~$15,000 per wafer.
How much does HBM memory cost?
HBM3 costs approximately $200 per 24GB stack. HBM3E costs approximately $300 per 36GB stack. HBM2e costs approximately $120 per 16GB stack. HBM4 (next generation) is expected to cost approximately $500 per 48GB stack. HBM commands a 5–6× premium over equivalent DDR5 capacity due to 3D stacking complexity.
How much does CoWoS packaging cost?
TSMC CoWoS-S 2.5D packaging costs approximately $70 per unit for standard interposer configurations. CoWoS-L (for larger multi-die designs) costs approximately 20–40% more. Intel EMIB costs approximately $35 per unit. 3D stacking (SoIC) costs approximately $120 per unit. CoWoS lead times are 40–52 weeks as of early 2026.
What is the cost breakdown of an NVIDIA H100?
The NVIDIA H100 SXM5 has an estimated manufacturing cost of ~$3,320: logic die cost ~$300 (814mm² on TSMC 4N), HBM3 memory ~$1,350 (six stacks, 80GB), CoWoS-S packaging ~$750, plus test and assembly. NVIDIA sells the H100 at approximately $28,000, yielding a gross margin of approximately 88%.
Why are chip manufacturing costs increasing?
Three main factors: (1) EUV lithography — advanced nodes require multi-pass EUV patterning, dramatically increasing wafer cost. TSMC 3nm wafers cost roughly 6.5× more than 28nm. (2) Advanced packaging — CoWoS, HBM integration, and multi-die architectures add $500–$1,500+ per chip. (3) Memory bandwidth — AI workloads require 6–8 HBM stacks costing $1,200–$2,400 per chip. These three costs didn't exist or were negligible at older nodes.
How much does it cost to design a chip?
Chip design costs (NRE) have escalated dramatically. A 28nm chip design costs approximately $30–50M. A 7nm design costs $100–200M. A 5nm design costs $200–400M. A 3nm design costs $400–600M+. These costs include EDA tools, IP licensing, mask sets ($10–20M at 3nm), engineering headcount, and verification/validation.
What determines chip manufacturing yield?
Yield depends on die size, defect density (D0), and process maturity. TSMC 3nm has a D0 of ~0.09–0.14 defects/cm² (early vs mature production). Larger dies have exponentially lower yields — an 814mm² H100 die yields roughly 40–50%, while a 100mm² chip yields 85–95%. Yield directly determines cost per die since lower yield means fewer good chips per wafer.