Self-service · Rack Teardowns
What's inside an AI rack — and who gets paid.
Analyst-grade cost anatomy for 34 rack architectures (lifecycles 2023–2031): every BOM line priced at the current component book, cost-pool rollups, ASP-vs-BOM margin delta, and modeled top suppliers per pool. $499 per rack.
- Rack architectures
- 34
- Vendors covered
- 20
- Chips per rack
- 1–144
- Power density (kW)
- 0.5–600
What every teardown breaks out
Each rack is decomposed to the BOM line, then rolled up into cost pools — with the ASP-vs-BOM margin delta and the modeled top suppliers behind each pool:
- Compute / accelerators
- Memory (HBM)
- Networking & interconnect
- Power delivery
- Cooling & thermal
- Chassis, mechanical & integration
Rack architectures covered
| Architecture | Vendor | Chips/rack | Power density | Lifecycle |
|---|---|---|---|---|
| AMD MI500 / Helios v2 rack | AMD | 72 | 250 kW | 2028–2031 |
| AMD Instinct MI455X / Helios NVL72-class rack | AMD | 72 | 150 kW | 2027–2030 |
| AMD Instinct MI325X 8-GPU (OEM) | AMD | 8 | 9 kW | 2025–2030 |
| MI300X 8-GPU (OEM) | AMD/Supermicro | 32 | 35 kW | 2024–2027 |
| AWS Trainium3 UltraServer | AWS (captive) | 144 | 60 kW | 2026–2029 |
| AWS Trainium2 rack | AWS (captive) | 64 | 40 kW | 2025–2028 |
| Cerebras CS-3 | Cerebras | 1 | 23 kW | 2025–2030 |
| Dell PowerEdge XE9680 | Dell | 8 | 10.2 kW | 2025–2030 |
| Dell PowerEdge XE9685L | Dell | 8 | 14 kW | 2025–2030 |
| TPU v6p (pod fragment) | Google (captive) | 30 | 25 kW | 2026–2029 |
| Google TPU v7 (pod fragment) | Google (captive) | 32 | 30 kW | 2026–2030 |
| Groq LPU rack (GroqRack) | Groq | 72 | 20 kW | 2025–2030 |
| HPE ProLiant DL384 Gen12 | HPE | 2 | 2 kW | 2025–2030 |
| HPE Cray XD670 | HPE | 8 | 10.2 kW | 2025–2030 |
| Huawei Ascend 950 rack | Huawei (captive) | 32 | 45 kW | 2027–2031 |
| Huawei Ascend 920 rack | Huawei (captive) | 32 | 30 kW | 2025–2030 |
| Intel Gaudi 3 8-accelerator (OEM) | Intel | 8 | 7.6 kW | 2025–2030 |
| Lenovo ThinkSystem SR685a V3 | Lenovo | 8 | 9 kW | 2025–2030 |
| Meta MTIA-3 rack | Meta (captive) | 32 | 45 kW | 2027–2030 |
| Meta MTIA-2 rack | Meta (captive) | 32 | 35 kW | 2025–2028 |
| MSFT Maia-300 rack | Microsoft (captive) | 32 | 60 kW | 2027–2030 |
| MSFT Maia-200 rack | Microsoft (captive) | 32 | 50 kW | 2026–2029 |
| NVIDIA IGX Orin | NVIDIA | 1 | 0.5 kW | 2025–2030 |
| NVIDIA DGX B200 | NVIDIA | 8 | 14.3 kW | 2025–2030 |
| Vera Rubin Ultra VR300 NVL576 (Kyber) | Nvidia | 144 | 600 kW | 2027–2031 |
| Vera Rubin VR200 NVL72 | Nvidia | 72 | 180 kW | 2027–2031 |
| GB300 NVL72 | Nvidia | 72 | 140 kW | 2026–2029 |
| GB200 NVL72 | Nvidia | 72 | 120 kW | 2025–2028 |
| DGX H100 (5-server) | Nvidia | 40 | 50 kW | 2023–2026 |
| Ruggedized MIL-SPEC tactical AI rack | OEM (Crystal Group / Klas) | 4 | 3 kW | 2025–2030 |
| Compact sovereign AI pod (self-contained) | OEM (self-contained) | 8 | 12 kW | 2025–2030 |
| SambaNova DataScale SN40L | SambaNova | 8 | 7 kW | 2025–2030 |
| Supermicro SYS-821GE-TNHR | Supermicro | 8 | 10 kW | 2025–2030 |
| HGX H200 (OEM) | Supermicro/Dell | 40 | 60 kW | 2024–2027 |
Sign in to run a teardown
Get the full priced BOM, cost pools, margin delta and modeled suppliers for any rack above — $499 per rack.
Create account →Modeled estimates from public data and the Mintok component price book — methodology and calibration published. Not investment advice. Methodology.