// cosmogenic.org — hardware intelligence — 2026.02

GPU Supply Chain
Bottleneck Analysis

// Semiconductor dependency map · edge hardware alternatives

We have CPUs. We are GPU-constrained. This document maps every choke point from silicon wafer to deployed inference node — and identifies where edge compute can route around the blockage.

Critical concentration risk: TSMC manufactures ~92% of sub-7nm GPUs globally · ASML holds monopoly on EUV lithography · Export controls active · Lead times 12–52 weeks
// supply chain — upstream to downstream — click any node
T1 Raw Materials MODERATE
Silicon Wafers
Shin-Etsu · Sumco · Siltronic
300mm diameter
Japan dominant supply
Concentration60%
Rare Gases
Neon · Argon · Krypton
Ukraine 60%+ neon
supply pre-2022
Concentration75%
refining & purification
T2 Lithography Equipment CRITICAL
ASML
EUV / DUV Lithography
Only EUV manufacturer
€400M per machine
NL export controlled
EUV monopoly100%
wafer exposure & patterning
T3 Semiconductor Foundry CRITICAL
TSMC
Taiwan · 2nm → 5nm
92% advanced node GPUs
Taiwan strait risk
Arizona fab 2025+
Advanced node share92%
Samsung Foundry
Korea · 3nm → 4nm
AMD + Qualcomm client
Yield challenges at 3nm
Korea geopolitical risk
Advanced node share15%
die packaging & testing
T4 GPU Design (Fabless) HIGH
NVIDIA
H100 · H200 · B200 · GB200
~90% AI GPU revenue
CUDA ecosystem lock-in
Export restricted H100
AI GPU market share88%
AMD
MI300X · RX 7900 XTX
ROCm improving fast
MI300X competitive
No CUDA dependency
AI GPU market share10%
Intel Arc / Gaudi
Arc A770 · Gaudi 3
Gaudi 3 = H100 class
Arc A770 = 16GB VRAM
oneAPI not CUDA
AI GPU market share3%
board manufacture · CoWoS packaging
T5 HBM Memory (stacked) CRITICAL
SK Hynix
HBM3E — Primary H100 supplier
H100/H200 exclusive HBM
~50% global HBM share
Korea concentration
HBM market share50%
Samsung
HBM3 — qualification delays
Yield issues 2023–24
HBM4 race ongoing
Same jurisdiction risk
HBM market share35%
Micron
HBM3E — US domestic
US manufactured
Ramping HBM3E 2024
Lower geopolitical risk
HBM market share15%
system integration · thermal
T6 System Integrator / OEM HIGH
Dell / HP / Lenovo
Rack Servers · GPU Nodes
R750xa · DL380 Gen11
12–26 week lead times
GPU allocation queues
Enterprise server share65%
Supermicro
GPU Servers · Direct-liquid
H100 cluster builds
Liquid cooling native
Shorter lead times
GPU server share20%
procurement · deployment
T7 End Deployment HIGH
Hyperscalers
AWS · Azure · GCP · OCI
Absorb ~60% H100 alloc
Reserved capacity lock
Everyone else waits
H100 allocation60%
GPU Cloud
CoreWeave · Lambda · Vast.ai
Spot pricing viable
Vast.ai peer market
No commitment needed
H100 allocation25%
Edge / Fed. Nodes
Dell R740 · RPi · Local
CPU+quant inference
No H100 dependency
Cosmogenic model
Current AI GPU use~10%

Select a node
to view intelligence
briefing

Edge-Viable Alternatives
// route around the GPU bottleneck — no H100 required
RTX 5060 Ti
NVIDIA consumer · in lab
16–24GB GDDR7
~400 TOPS INT8
~£400–600 street
llama.cpp · Ollama ready
Inference perf72%
Availability85%
Cost efficiency90%
AMD RX 7900 XTX
AMD · ROCm
24GB GDDR6
61 TFLOPS FP16
~£800 street
ROCm · llama.cpp
Inference perf68%
Availability80%
Cost efficiency78%
Intel Arc A770
Intel · oneAPI
16GB GDDR6
35 TFLOPS FP16
~£250–300 street
llama.cpp support
Inference perf42%
Availability90%
Cost efficiency95%
CPU Quantised
llama.cpp · Ollama · No GPU
Runs on R740 CPUs
7B INT4 = ~5 tok/s
Zero GPU cost
Today, right now
Inference perf25%
Availability100%
Cost efficiency100%