Capital to Compute

When governments and sovereign investors announce large AI investments, I am trying to follow them beyond the press release. When a country commits billions to AI, what will its companies and researchers eventually be able to use, and which parts of that promise can we verify along the way?

My experience at AUTOCRYPT, an automotive cybersecurity company, is part of why I approach the question this way. In investor relations there, I coordinated a KOSDAQ IPO process and, before it, supported a pre-IPO funding round. Between an announcement and a closed transaction came diligence, legal work, approvals and documentation. Until those were resolved, the headline could not tell the whole story.

AI infrastructure adds a physical delivery problem to that financial one. I want to follow a commitment until the equipment is running, then ask who gets to use it. An operating facility reserved for one customer is a different result from computing capacity available to local companies or researchers, even if the construction project succeeds in both cases.

To study those differences through public records, I am developing the Sovereign Compute Delivery Observatory. The first pilot will cover a small set of projects in Korea and the Gulf, with comparisons from the United States or other allied economies. Before attempting a ranking, I want to establish whether their progress can be measured consistently at all.

The delivery milestones

For each project, I want to be able to explain how far the investment has progressed and what still stands between it and a working service. A site agreement is useful evidence, but not proof of a timely grid connection. Likewise, naming a chip supplier leaves open whether anyone has placed a binding order.

The records I am proposing separate those steps, so that evidence of one does not quietly become evidence of another:

I do not expect every project to complete these steps in the same order. Construction may be visible while financing terms remain undisclosed. The record should preserve that uneven picture rather than force the project into a single stage that suggests more certainty than I have.

The later steps matter because delivered accelerators are only part of a computing service. The facility also needs working memory, networking, cooling and software. I am looking for evidence that these have been assembled into a system someone can use.

Access to the completed system

Once I can establish that a facility works, I still want to know what public support made possible for people outside its operator. Did it create additional capacity? Can a local company or research group apply to use that capacity, and on what terms?

I plan to record access separately from delivery. A facility with all its capacity reserved for one customer could meet its commercial objective while offering others little access. Combining those outcomes into one score would make it harder to assess what the government chose to support.

Published allocation terms would help answer the access question. Without them, an operating announcement may establish that the system runs while leaving its wider public benefit unresolved.

Working with incomplete evidence

The hardest part of this project may be deciding how much the public record actually supports. I will prioritize documents close to the event, such as government decisions, permits, utility records and binding procurement disclosures. Company announcements and good reporting can also establish useful facts. Databases and aggregators will be starting points for finding the underlying sources.

I want each important claim to remain traceable to a source and date. Where records disagree about capacity or schedules, I will keep the disagreement. Where I cannot establish a milestone, I will mark it unknown.

That choice matters especially for the country comparisons I am considering. A transparent project may disclose every delay, while another publishes little after its launch announcement. If I treated the second project as either a success or a failure, I would be measuring disclosure habits as much as delivery. The pilot will therefore track disclosure quality alongside the milestones.

My transaction experience can help me recognize which questions to ask, but it cannot supply the missing answers. This research will use public sources and exclude confidential transaction information and security-sensitive facility details. Nor do I expect a small pilot to support causal claims about why one country's projects perform better than another's.

The Korea-Gulf pilot

I am starting with Korea and the Gulf because their proposed AI partnerships bring different resources into the same projects. Korea has semiconductor and manufacturing capability. Gulf states bring capital, energy and centralized coordination. The United States supplies much of the frontier technology and shapes access through export controls and partnerships. Following those relationships may show me where an international agreement runs into a practical delivery constraint.

The first outputs I am planning are:

• A codebook defining the milestones, and a protocol for sources and confidence.
• Project records with dated evidence for the important claims.
• A methodology note and cases following individual projects.
• A bilingual Korea-Gulf brief reporting what the pilot can support.

The evidence may turn out to be too uneven for a comparative score. In that case, paired cases or a smaller set of milestones would be more useful. The result I am working toward is a set of records another person can inspect to understand what became binding, what was delivered and what remains unknown.

Draft research agenda, originally August 15, 2026. The observatory is proposed work; this article does not report a completed dataset or empirical finding. Research method.