The missing conversion
A building is not a running workload. A robot is not a dependable service. A functioning prototype is not a business. These distinctions are familiar; the harder work is keeping them intact when a striking announcement invites a single large conclusion.
The research unit should be a conversion: what moved from a stated capability to an accepted result, over what time, under which constraints and at whose cost?
A megawatt needs a location and a date
A data-center investigation should distinguish announced, permitted, financed, built, energized, equipped and accepted capacity. Define whether a number describes a facility, an IT load, a phase or an entire campus.
Epoch AI’s methodology describes source-based research followed by models that estimate capacity and costs. That is useful evidence with an estimation boundary; it is not a direct customer-acceptance measurement. Track each provider’s definition rather than combining incompatible totals. [1]
A fleet is not yet a theory of intelligence
NOAA’s 2026 ocean-observation mission offers a concrete setting for inspecting coordinated uncrewed systems and the data they provide. It does not, by itself, establish decentralized collective reasoning or commercial margins. [2]
A fleet can create value through spatial coverage, endurance and reliable measurements, even if each vehicle uses modest intelligence. The economic question is what useful observation or service the arrangement makes possible compared with alternatives.
What an orbital demonstration proves
DARPA describes an automated in-space assembly proof of concept. Caltech’s account includes the limited initial section achieved and the ground commands and checks involved. These details make the result more interpretable, not less interesting. [3] [4]
The claim to investigate is which physical constraint this process can remove. Repeatability, mass, energy, communications, intervention and customer requirements belong in the next evidence. A space demonstration does not automatically establish economical large-scale construction or a new AI revenue pool.
Options before cash flow
The counterargument deserves weight: a successful prototype can expand what is possible long before a service becomes profitable. Requiring current revenue for every frontier project would miss the value of creating a viable option.
That option should still have a mechanism. What uncertainty was reduced? What next milestone is affordable? Who can exercise the capability? What would prevent it becoming useful? Distinguish the option from the later profit being imagined.
Follow the conversion
A fixed cohort and a stable definition make progress visible. Record the last evidenced milestone, its date, the next constraint and the evidence needed to cross it. Keep delayed and failed cases in the denominator.
The most interesting bottleneck may move. When a constraint clears, examine accepted output and who retains the benefit after pass-through, competition and cost. This is a research framework, not a national capacity inventory or a measured cross-sector productivity result.
Sources & scope
Sources checked: 13 November 2026.
Source-based research and estimation; estimates are not metered customer acceptance.
Mission objectives and observations; not proof of decentralized reasoning or commercial margins.
An engineering proof of concept; not a commercial deployment or general-intelligence claim.
Describes the limited assembly demonstration, ground commands and operational constraints.