Discussion about this post

User's avatar
Adrien's avatar

The framing that "AI leadership is determined by physical infrastructure control as much as algorithmic capability" is the right analytical lens — and the 60/40 split between silicon (TPUs/GPUs) and data center/networking in Alphabet's $175–185B envelope tells you something important about where the constraint actually sits. The bottleneck isn't compute availability anymore; it's permitting, grid interconnection, and liquid cooling deployment, which means the infrastructure supply chain companies — power management, cooling systems, transmission hardware — arguably have more durable pricing power in 2026–2028 than the chip vendors themselves. One nuance worth flagging: Google Cloud's $240B backlog growing 55% sequentially is a more important number than the revenue growth rate, because it locks in utilization well ahead of capacity additions — which is the structural difference between this infrastructure cycle and the telco buildout of the late 1990s, where demand never materialized to fill the pipes. The semiconductor equipment cycle leading indicator is also worth watching here: ASML and Lam Research order books tend to preview silicon availability 12–18 months out, and both are still guiding for elevated demand through 2027.

Dennis Berry's avatar

Great breakdown. So much going on right now.

No posts

Ready for more?