Jan. 18 at 5:40 PM
$SPY $AMD $RIOT AMD signed a data center lease with RIOT. Ten years.
$311 million in base revenue. Expansion options to 200 MW. Right of first refusal on another 100 MW.
The structure is identical to hyperscaler contracts. Same duration. Same economics. Same expansion mechanics. AMD negotiated like Microsoft or Amazon because AMD needs power like Microsoft or Amazon.
This wasn't in the model.
The thesis tracks hyperscaler demand for power-secured infrastructure. The contracts flow from Microsoft, Amazon, Google, CoreWeave. The customer base is understood. The demand is quantifiable. The constraint is clear.
AMD suggests a second demand layer. Chipmakers consuming power directly, not through cloud intermediaries. They're not renting from AWS. They're signing leases with operators. They want the infrastructure, not the service.
The mechanism matters. NVIDIA doesn't just sell chips anymore. DGX Cloud sells compute access. The margin on renting GPUs exceeds the margin on selling them. AMD sees this. They watched NVIDIA build a moat around inference services while everyone focused on chip sales. Going direct to infrastructure is the counter.
If AMD needs 20 MW now with options to 200 MW, the question extends. What does NVIDIA need? They're already running inference at scale through DGX Cloud. What does Intel need if Gaudi gains traction? What does Qualcomm need if edge inference requires centralized validation clusters?
Hyperscaler demand alone justified current valuations. Chipmaker demand stacking on top suggests we've been measuring the wrong ceiling.
The distinction between training and inference matters here. Training gets attention. Massive clusters. Billion-dollar builds. The race to foundation models. But training is episodic. You build the cluster, run the job, iterate. The capacity gets reused or repurposed.
Inference is different. Inference is every API call. Every AI feature in every application. Every autonomous decision made in real time. Training happens once. Inference happens continuously, forever, scaling with adoption.
The power economics diverge. Training demand is bursty and plannable. Inference demand is baseload. Constant draw that grows with usage. If inference is the larger compute need over time, then power constraints are worse than training-focused projections suggest.
AMD going direct to infrastructure might be an early signal of the inference layer materializing at scale. They're not building a training cluster. They're building persistent capacity for workloads that don't stop.
There's a disintermediation question underneath this. Hyperscalers are intermediaries. They aggregate enterprise demand, package it as managed services, build infrastructure to serve it. The assumption is that demand flows through them indefinitely.
AMD bypassing that layer is a tell. They could have contracted with AWS or Azure for capacity. They chose not to. They went to RIOT and signed a lease for the infrastructure itself. They wanted control over the power, not access to someone else's cloud.
Why would a sophisticated buyer choose direct infrastructure over cloud services? Cost is one answer. Cloud margins are substantial. Control is another. Workloads too proprietary to share infrastructure. Availability is a third. Maybe cloud capacity is more constrained than public messaging suggests.
If AMD is going direct, others might follow. Enterprises with sufficient scale. Sovereign AI programs that prefer ownership to dependency on American hyperscalers. AI labs not affiliated with the major clouds. The buyer pool for power-secured operators might be wider than the thesis accounts for.
The power constraint doesn't care who's buying. Megawatts are megawatts. Interconnection queues don't distinguish between hyperscalers and chipmakers. If the competition for constrained power is broader than expected, then the scarcity is deeper than it appears.
One deal is anecdotal. AMD signed one lease at one site. Rockdale was available, which means hyperscalers passed on it. Maybe this is AMD experimenting at the margins.
But the contract structure contradicts that read. Ten years is not an experiment.
$311 million in committed revenue is not a pilot. Expansion options to 200 MW with right of first refusal on another 100 MW is not a hedge. That package went through AMD's legal, finance, and board approval. They modeled forward demand. They negotiated terms. They committed.
Chipmakers are now competing for power-secured infrastructure using the same structures hyperscalers use. The demand pool is wider than hyperscaler-only models suggest. The operators who secured power early have something more buyers want than previously understood.
Whether this is pattern or anomaly depends on what follows. If NVIDIA, Intel, or Qualcomm sign similar deals, the second layer is real. If AMD remains alone, maybe Rockdale was opportunistic.
But AMD moved first. They saw something worth a decade-long commitment. That's signal, not noise.