Cohere is selling insurance against export controls at a $20bn mark

Take the ratio before the story: 20 over 7 is 2.9 times in a year ^[11], and the two marks are not measuring the same company, because the April combination with Germany's Aleph Alpha folded another builder's assets into the second number ^[3]. Semafor does not split how much of the added $13bn ^[12] is Aleph Alpha and how much is investors paying up for a thesis, and no revenue figure appears anywhere in the interview, so anyone sizing the mark has to supply the missing denominator themselves: at thirty times forward revenue, my assumption and not a disclosure, $20bn wants roughly $667m of it ^[18].

The load-bearing number in the pitch is the estimate that sovereignty requirements could influence up to 40% of AI spending by 2030 ^[16], and the verb is doing the work. Influence is a test an incumbent can pass with a region, a contract clause and a key it does not hold, whereas Aidan Gomez is selling the stronger version, in which a deployed customer system is one his company "can't see in", "can't switch it off" and cannot intervene in at all ^[15]. Those are two different products sold into the same procurement line.

What lifts any of this above a risk memo is a dated event: Semafor reports that Washington imposed export controls on Anthropic's Mythos model in June ^[5], which hands a risk committee something specific to cite when it asks for a second supplier. Gomez says outreach has risen steeply this year, and that standing outside the US-China duopoly makes big global companies much easier to reach ^[17].

This is probably wrong, but the interesting reading of the $20bn concerns procurement rather than model quality: it prices the chance that the signature on enterprise AI purchases moves from the technologist to the risk function, or rather, that the same budget gets spent against different criteria. The counter-thesis sits inside Cohere's own evidence. The IDC study it released found that most executives across four countries call sovereignty important while few can define it or hold a firm plan ^[7], and a requirement nobody can define does not convert into a purchase order on schedule; the skeptics quoted alongside Gomez argue that duplicating US and Chinese technology will be prohibitively expensive ^[6], and he half concedes the point himself, calling a world where everyone builds a copy of everything extraordinarily inefficient ^[9].

The mark also fixes what Cohere does not get to spend on. Gomez has given up the outspending race outright, saying the contest is not about building the biggest model or spending the most money ^[14], so the capital goes into telecoms, financial services, healthcare, the grid and water treatment ^[8], and into the jurisdictional plumbing those buyers audit, rather than into the cluster that would let a Canadian firm Semafor still calls a minnow next to DeepSeek and Moonshot ^[4] compete on raw capability. Interest is cheap, and the steep outreach becomes evidence only when it arrives attached to a contract value.