Elena Voss
$100 billion. That's the size of Alat, the Saudi Arabia Public Investment Fund vehicle launched to build the country's electronics manufacturing industry. On April 9, 2026, Alat fired its CEO and dropped its plans to invest in chip production. The richest sovereign wealth fund on earth tried to build sovereign chip manufacturing. After two years and a $100 billion mandate, it quit on chips.
The Merger
On April 10, Reuters reported that Canada's Cohere and Germany's Aleph Alpha were in advanced talks to merge, with the German government willing to become a key customer of the combined company. These are the two most prominent frontier AI labs outside the US and China. Cohere is Canada's. Aleph Alpha is Germany's. Each is the thing its national tech press has spent three years calling a "sovereign AI champion."
Cohere raised $270M in June 2023 at a $2.2B valuation — the round that made it Canada's official AI bet. By 2025 it had raised roughly $1.5B total and was generating about $100M in annualized revenue. Aleph Alpha raised €500M in 2023 from a German industrial consortium led by Schwarz Group and Bosch, at what was then one of the largest European AI rounds ever assembled. Both companies existed to demonstrate that the frontier AI stack could be built and sustained outside the US. On April 10, they acknowledged it can't be — not as two separate companies, and possibly not as one combined company without a national government offering itself as the anchor customer.
That is not how you announce a merger when you expect the combined entity to dominate. That is how you announce a merger when you're trying to keep it alive.
The Subsidy
The same week, Japan approved an additional $4B in state subsidies for Rapidus, the state-backed chipmaker building a 2nm fab in Hokkaido. Total Japanese state commitment to Rapidus, including prior rounds and fees: $16.3 billion. The company has not yet shipped a chip.
$16.3 billion sounds like a lot of money. In the context of leading-edge foundries, it isn't. A single TSMC Arizona fab runs roughly $40B. Intel's Ohio project is north of $28B. Samsung's Texas site is projected at $44B. Rapidus needs to match the process node of companies that have been doing it for a decade, with a budget that's a fraction of theirs. The April top-up is the fourth since launch. Each one is described, in the same language, as the commitment that finally gets Rapidus over the line.
The Migration
On April 10, France announced that government computers running Windows would be moved to Linux, explicitly to reduce reliance on US technology. Scale: 2.5 million devices. Timeline: not published.
This is the cheapest sovereign tech story of the week, which is exactly the problem. Linux migration reduces the Microsoft dependency. It does not reduce any of the other ones. The hardware French government computers run on is still US-designed. The Python package index they pull from is still hosted in US data centers. The certificate authorities they trust are still US corporate products. The frontier AI models they increasingly depend on are still Anthropic or OpenAI, neither of which runs on a European server by default. Switching the OS is the smallest surface area of dependency in the stack. France chose it because it is one of the few surfaces still available to move.
The Ratio
Five sovereign tech stories from five countries in one week. None of them are successes. The same day the FT reported a reverse flow of top AI researchers from the US back to China — driven by pay, quality of life, and a hostile US immigration regime — the Western-democratic side of the ledger was producing mergers, subsidies, and Linux migrations. The capital going in has never been higher. The independence coming out has never been lower.
Count the capital going in. Saudi Alat: $100B committed, now withdrawn from chips. EU Chips Act: €43B committed over the 2023-2030 window. Japan Rapidus: $16.3B committed so far. Cohere and Aleph Alpha: roughly $2B raised between them. Mistral (not in this week's news but running the same playbook): another $1B+. Call the total Western-democratic sovereign tech commitment of the past four years, generously counted, around $200 billion.
Now count the capital going in on the other side. In February, Alphabet, Amazon, Meta, and Microsoft forecast a combined ~$650 billion in 2026 capex, most of it AI infrastructure. Meta signed an additional $21B AI cloud deal with CoreWeave on April 10. That is not a slipped decimal. Four companies. One year. $650 billion.
The entire sovereign-tech commitment of the Western democratic world over the past four years is a fraction of one year of US hyperscaler capex. The ratio is roughly 1:13. And it is the wrong 1:13, because the sovereign side is spread across a dozen countries and four categories — chips, AI labs, operating systems, research talent — while the private side is four companies chasing one category. Concentration beats dispersion, every quarter.
The Denominator
Sovereign tech used to be a political project. Countries built their own railways, their own steel mills, their own radio networks, their own radar, their own satellite launchers — because the capital required was affordable at national scale, and the strategic value was obvious. Sovereignty meant having your own industry. Having your own industry was a matter of political will.
That era is over. The reason isn't that political will has weakened — it has, if anything, strengthened. Alat was $100B of political will. Rapidus is $16.3B and counting. France booting Microsoft off government desktops is one of the more politically motivated tech decisions of the past two decades.
The will is there. What's gone is the math.
Sovereign tech now has to operate at commercial scale to be relevant, because anything less is obsolete on arrival. Commercial scale, in 2026, is defined by four companies that spend $650 billion on capex in a single year. No national treasury outside the US and China can sustain that. The old denominator — your own industry's capital cost divided by your GDP — used to produce numbers small enough to be affordable. The new denominator — your national strategy's capital cost divided by Alphabet's 2026 capex forecast — produces numbers that are, in polite language, tragic.
This is the quiet thing about the Cohere–Aleph Alpha merger talks. It is not a story about two struggling companies. It is a story about the price of a non-US frontier AI effort being too high for Canada alone, too high for Germany alone, and possibly still too high for both of them combined with a government contract backstop. That price wasn't set in Ottawa or Berlin. It was set in Mountain View and Menlo Park, by companies that can commit $30B in a single quarterly earnings call and ask for the next tranche in the next one.
The Number, Recontextualized
$100B was supposed to be enough for Alat to build a Saudi electronics industry. It wasn't. $16.3B is supposed to be enough for Rapidus to catch TSMC. It probably isn't. $1.5B was supposed to be enough for Cohere to be Canada's OpenAI. It wasn't, and the April 10 merger report confirms it.
The numbers are not small. They are historically unprecedented for sovereign tech commitment. They just aren't big enough relative to the commercial denominator that every sovereign tech strategy now has to clear to matter. The price hasn't gone up because the technology got harder. It has gone up because the private-sector alternative has outpaced the public one in ways that compound quarterly. The compounding is not reversible from anywhere on a national balance sheet outside Beijing and Washington.
In 2015, a country could commit $5 billion to a national semiconductor program and be taken seriously. In 2026, $100 billion is what you spend before admitting you can't. The ratio moved. The pile didn't.
More on Cohere, Rapidus, and sovereign tech. Explore entity coverage via the Pulse API.