A capacity plan's "clusters live in Q3" line usually traces back to a vendor lead time. Outside a small group of trusted countries, the lead time does not cover the step that decides everything: advanced Nvidia processors are export-controlled, and Washington's approval of the sale is discretionary [7].
Buyers file that approval as paperwork: a queue, owned by legal, tracked in a spreadsheet. In practice, per The Next Web's reading of the mechanism, it behaves differently: it can be granted, withheld, expanded or slowed without a law being passed or a dollar being spent, which makes it closer to a visa than to a subsidy [8].
The arithmetic shows how much rides on the stamp. Follow-up coverage of the Journal's paywalled report puts roughly 70,000 Nvidia processors behind a proposed $5bn Firebird project in Armenia [9]. Divide the second figure by the first and every approved processor carries about $71,400 of project capital [14]. What the approval gates is not only the chip invoice but the land, the substation, the cooling and the financing that were all committed on the assumption the hardware would clear. Treat both numbers as second-hand: The Next Web says they reached it through other outlets' coverage rather than the Journal article itself [9].
The same currency is being spent in the other direction. G42, the Abu Dhabi AI company, is exploring the sale of a majority stake to American companies to secure chip access past April 2027, according to The Next Web's Thursday reporting [11]. Armenia was offered approvals to sign something; G42 is weighing control of itself to keep approvals it already has. Firebird, meanwhile, signed a $10bn agreement with Kazakhstan in June to build what it called Data Center Valley [10]. For scale on the other side of the Armenian table, Azerbaijan's startups raised $2.6m last year [13].
What this reporting does not contain is the downside case. There is no approval timeline in it, no published conditions attached to the Armenian licence, and no instance of an approval being narrowed or withdrawn after a political outcome changed [16]. The upside is documented, including by Nvidia's own executive [12]. The revocation risk that operators would actually have to underwrite is inference rather than evidence.
So the question for a compute plan is not who the vendor is. Draw two axes. One: does the capacity need a US licence, or does it sit inside the trusted group. Two: does the project give Washington something it wants, or is it incidental to every open file. Licence-needed and useful is the Firebird quadrant, where approvals can move faster than any procurement process would explain, and where the schedule moves with someone else's negotiation. Licence-needed and incidental is the harder place to sit, because you are in the queue rather than in the deal, and nothing in the supplier contract reaches the office that sets the order. The honest version of that Q3 date names the approval it depends on, and names what the approver gets from the building being finished.
