NetworkNews

Anthropic-Lambda: decoding the reported $35bn deal

On this page
  1. Separate the reporting from the announcement
  2. Why the megawatts cannot simply be added
  3. What a customer can ask a provider
  4. Sources

Press reports link a $35 billion Anthropic-Lambda cloud agreement to Beacon Point in Texas. Hut 8 confirms the campus leases and planned power capacity, but its announcement does not name the tenant.

Two announced 352 MW IT tranches sit within Beacon Point’s 1,000 MW utility envelope. The remaining envelope is a subtraction, not measured overhead or available IT capacity.
Two announced 352 MW IT tranches sit within Beacon Point’s 1,000 MW utility envelope. The remaining envelope is a subtraction, not measured overhead or available IT capacity. Chart : PeopleAreGeek. Data source.
View full-size image

Separate the reporting from the announcement

Data Center Dynamics, citing the Wall Street Journal, reported on September 1 that the agreement concerns roughly 350 MW and that Nvidia would hold the site lease. Those are reported deal details. The accessible Hut 8 release does not independently confirm the full Anthropic-Lambda-Nvidia chain. We have not inspected the contracts.

Hut 8’s July 20 announcement does confirm two 352 MW IT leases with the same unnamed tenant: 704 MW IT in total, against 1,000 MW of utility capacity. It targets initial energization in Q1 2027. A target is not evidence that customers can already run workloads there.

Why the megawatts cannot simply be added

The cover diagram shows the two confirmed IT tranches within the campus’s utility power envelope. IT power concerns the computing equipment; utility power also has to accommodate the facility. They are different accounting boundaries. Adding 704 to 1,000 would count overlapping capacity twice.

The reported 350 MW cloud agreement should not be added either: the reporting places it at this same campus. Without the actual allocation contract, it is also inappropriate to assign it to a particular phase or treat 350 and 352 as an exact match.

The same caution applies to dollars. A cloud-service commitment and a landlord’s lease value pay for different layers over potentially different periods. Dividing one headline amount by another does not reveal the operator’s margin, and dividing by site power does not reveal a price per GPU-hour.

What a customer can ask a provider

For a workload that needs dependable capacity, the useful questions concern delivery: which location is committed, when the service becomes available, what happens if it is late, and what alternate capacity is reserved. Ask who operates the compute and who owes the service-level commitments. A recognizable chip supplier does not replace those answers.

For resilience, draw the dependency chain for each supposedly independent provider. Two API endpoints can share a facility, power connection or hardware supplier. That is a reason to examine correlated failures, not proof that this particular agreement is unsound. These announcements alone cannot forecast spot GPU pricing or the availability of Claude this month.

Sources

Separated reported deal terms from Hut 8’s confirmed campus figures; removed unsupported GPU price forecasts.