The AI bottleneck map

Who supplies whom in the AI build, from indium phosphide wafers to hyperscalers. Every link comes from a dated filing in our notes.

By A (Akash Chaudhari)

Who supplies whom in the AI build. Layers: Materials & foundry → Chips & parts → Systems & power → Sites & compute → Who pays.

• AXT → Coherent · Lumentum: InP substrates for lasers
• AXT → Nokia: InP wafers for Nokia's fabs
• Coherent · Lumentum → 800G / 1.6T modules: Lasers into 800G and 1.6T modules
• MaxLinear → 800G / 1.6T modules: PAM4 DSP inside each module
• 800G / 1.6T modules → OpenAI · Meta · Microsoft: Optics for AI clusters
• 800G / 1.6T modules → Google: Optics for AI clusters
• Nokia → Google: 800G optics (Google in Finland)
• TSMC → AMD: All chips at 7nm and below
• TSMC → Astera Labs: Sole-source wafers
• TSMC → Amkor: Wafers to package; CoWoS short
• Amkor → NVIDIA: Packaging; NVIDIA prepaid $1.5B
• Rambus → OpenAI · Meta · Microsoft: RCD, PMIC chips in server memory
• Astera Labs → OpenAI · Meta · Microsoft: Rack connectivity; one buyer = 53%
• AMD → OpenAI · Meta · Microsoft: MI450 and Helios racks
• AMD → Oracle: Helios racks
• NVIDIA → SharonAI: GPUs ($303M prepaid)
• SharonAI → NVIDIA: Rents compute back ($0.82B/yr min)
• Engine parts → INNIO: Castings, generators
• INNIO → AI campuses: Gas engines via VoltaGrid and others
• Bloom Energy → Oracle: On-site power for AI data centers
• Bloom Energy → Google: 700 MW via AEP
• AI campuses → OpenAI · Meta · Microsoft: Powered capacity
• Cipher → AWS: Black Pearl, Stingray leases
• Cipher → Google: Barber Lake (Google backstop)
• NVIDIA → OpenAI · Meta · Microsoft: GPUs