AI Accelerator Manufacturing Cost (COGS) — Historical Time Series Data
Estimated manufacturing cost breakdown for NVIDIA A100, H100, H200, B200 and AMD MI300X. Shows HBM shift from 33% to 45%+ of BOM.
AI Accelerator Manufacturing Cost (COGS)
Estimated manufacturing cost breakdown for NVIDIA A100, H100, H200, B200 and AMD MI300X. Shows HBM shift from 33% to 45%+ of BOM.
Data Table
| Period | Series | Value |
|---|---|---|
A100 (2020) | Logic Die | $400 |
A100 (2020) | HBM | $1,350 |
A100 (2020) | Packaging & Test | $750 |
H100 (2023) | Logic Die | $300 |
H100 (2023) | HBM | $1,350 |
H100 (2023) | Packaging & Test | $1,670 |
H200 (2024) | HBM | $1,500 |
H200 (2024) | Logic Die | $300 |
H200 (2024) | Packaging & Test | $2,450 |
MI300X (2024) | Logic Die | $600 |
MI300X (2024) | HBM | $2,900 |
MI300X (2024) | Packaging & Test | $1,800 |
B200 (2025) | Logic Die | $850 |
B200 (2025) | HBM | $2,900 |
B200 (2025) | Packaging & Test | $2,650 |
Methodology & Sources
Sources: Silicon Analysts (H100 Feb 2026), Epoch.ai (B200 Dec 2025), Deep Research (102 sources). NVIDIA/AMD do not disclose COGS; estimates from teardown analysis. Gross margin estimates based on analyst consensus.
Citations:
- Deep Research (102 sources) (Mar 2026)
- Silicon Analysts (Feb 2026)
- Silicon Analysts (Mar 2026)
- Silicon Analysts / SemiAnalysis (Mar 2026)
- Epoch.ai (Dec 2025)
Related Tools
Related Analysis
Market intelligence on this data
All market intelligence →TSMC CoWoS Yield Confirmed at 98-99% in HVM; Nvidia Weighs 8-Stack HBM4E Fallback for Rubin Ultra
CoWoS yield: 98-99% confirmed in HVM (positive); HBM4E 12-stack yield risk flagged — potential stack-count reduction from 12 to 8 layers (est. HBM capacity-per-die reduction of ~33% if downgrade confirmed); TSMC 2026 CapEx raised to $60-64B record (capacity positive).
Nvidia Rubin Ultra HBM Spec Downgrade Under Review as HBM4/4E Supply Crunch Deepens Through 2027
Customer HBM4 fill rates at 60–70% of ordered volumes (Micron-confirmed); HBM4 BOM cost for GPU vendors estimated at ~2x HBM3E per Korean press citing industry sources; CoWoS lead times confirmed at 52+ weeks; ABF substrate orders extending into 2028.
Nvidia–SK Hynix Multi-Year HBM Supply Partnership Confirmed at Up to $500B
Supply commitment: up to $500B multi-year; Capacity: SK Hynix HBM allocation heavily directed toward Nvidia; Lead time risk for non-Nvidia HBM buyers: elevated (magnitude unquantified in sources).
Nvidia-SK Hynix & Samsung-Broadcom Lock In $950B HBM Supply Agreements, Securing Next-Gen Memory Pipeline Through 2030
SK Hynix: $750B long-term HBM supply commitment to U.S. customers (led by Nvidia); Samsung: $200B memory + foundry + packaging commitment to Broadcom through 2030; SK Telecom 2GW AI data center powered by HBM4 targeting 2027 online date; TSMC CoWoS and advanced packaging capacity described as largely booked through 2027.
Frequently Asked Questions
- How much does it cost NVIDIA to manufacture an H100?
- The estimated total manufacturing cost (COGS) for an H100 SXM5 is approximately $3,320: roughly $300 for the 814mm² TSMC 4N logic die, $1,350 for HBM3 memory, $750 for CoWoS-S packaging, and $920 for test/assembly.
- Why is the B200 logic die cheaper than the H100 despite being more advanced?
- The B200 uses a chiplet design with two ~400mm² N4P dies (~800mm² total) instead of a single monolithic 814mm² die like the H100. Each smaller chiplet yields significantly better, driving per-die cost down to roughly $900 total for the logic despite using a more advanced node.
- What percentage of AI chip cost is HBM memory?
- HBM as a share of total BOM has grown from roughly 14% (A100) to 33% (H100) to 43% (H200) to 45%+ (MI300X). The B200 reverses this trend slightly at ~38% because its chiplet logic die is relatively cheap, but the absolute HBM cost continues to rise.