Compute access is now gated by silicon contracts, not capital

In 2026, the binding constraint on frontier AI is not capital. It is the time it takes to pour concrete. Samsung Electronics and SK Hynix just committed their high-bandwidth memory, the bottleneck chip on AI accelerators, through 2030 to Nvidia, Broadcom, and a broader US tech coalition, in roughly $950 billion of contracts signed Friday. Headlines will frame it as a procurement win. The structural read is a compute-access stratification.

All three major HBM suppliers were sold out through 2026 before the contracts were signed. New HBM fab capacity does not arrive meaningfully until 2027 or 2028. A new HBM-capable fab takes three or more years to build, which is why offtake certainty, not equity, is the unit these contracts trade in. South Korean presidential adviser Kim Yong-beom, quoted by TechTimes, made the distinction explicit: the deals underwrite demand, they do not invest in the companies.

The mechanism is portable. When the bottleneck component is made in three factories globally, money buys a place in the queue only if it arrives years before the silicon does. The contracts signed Friday set the queue through 2030. Capital without a 2023 signature now buys a 2028 seat at best, and only if a non-Korean fab breaks the timeline.

That's the new shape of the compute divide: not who can pay, but who already locked in the fab time.