Buying the power, not the chip
An accelerator vendor takes a fifth of a power developer. A memory maker commits $38bn to two fabs. Capital is moving into the parts of the stack you cannot conjure.
Nvidia is investing about $2bn for roughly 20% of a Stargate power developer at a $10bn enterprise value, extendable to $3bn and around 30%. Chip companies do not naturally buy electricity developers.
They do it when the constraint moves
Every large accelerator deployment now meets interconnection queues, substation lead times and grid capacity planned for a different decade. When the limit on your sales is your customers' ability to plug the product in, upstream is where the money goes.
An offtake agreement secures supply. Equity secures influence over what gets built, and where.
The same week, at the other end
SK Hynix approved ₩54tn — about $38bn — for two new fabs. Memory capex at foundry scale, for capacity that will not produce a wafer for years.
Fabs do not respond to quarters. Approving two encodes a belief that current AI infrastructure buying is a floor, backed with balance sheet that is painful to reverse. Memory is also the most cyclical business in semiconductors, and its history is a sequence of capacity additions that landed exactly as demand turned.
Construction, not deployment
A general engineering contractor — ports, refineries, metros — just won up to ₹15,000 crore to build an AI data centre. That is what capacity expansion looks like once it stops being specialist work.
The binding constraint was never racking servers. It was land, power connection, cooling and the project management to run a multi-year build, which is precisely what heavy-engineering firms already do.
What this asks of you
If you are planning compute two years out, price power and delivery schedules, not accelerators. The chip is the part of this supply chain that responds fastest to money. Everything around it responds to permits.
And note who ends up with leverage: a supplier holding stakes in the electricity its customers need occupies an unusual position, and no regulator has yet decided to find that interesting.
Sophic Capital — August 15, 2026: Flattish Week, Decent AI Datapoints → · Silicon Analysts — AI Data Center Value Chain (2026) →