Last month, a company with roughly $240 million in annual recurring revenue was valued at $20 billion. That is a multiple of about 83 times sales. OpenAI, on the same ledger, was reportedly priced at roughly 34 times its operating revenue. Read those two sentences again, because the smaller, slower, lower-margin business is being priced at two and a half times the rate of the one everyone agrees is winning.
The difference, we are told, is sovereignty.
I have spent nine years explaining cryptography to people who will never read a whitepaper, and my instinct when a single word is asked to carry an entire valuation is to reach for a calculator rather than a flag. So let me do what I always do with a token narrative: follow the money, follow the keys, and ask who actually holds them when the press cycle ends.
Because "sovereignty" in AI right now means precisely what "decentralization" meant in crypto from 2020 to 2022 — a genuine technical property being used to disguise an entirely conventional capital structure. And if we in the blockchain community cannot see our own history repeating itself in this deal, we will repeat the painful part of it too. — Root: DeFi Summer.
Here is the shape of the thing. Cohere, a Canadian-founded AI lab, raised a round valuing it near $20 billion, up roughly threefold from about $7 billion twelve months earlier. The lead investor was the Schwarz Group, the German retail giant behind the STACKIT sovereign cloud and a planned 11 billion euro data center complex in Berlin. The Canadian government contributed roughly 240 million Canadian dollars, a figure that reportedly helped lever between two and three billion dollars of total capital — a ratio somewhere between one to eleven and one to seventeen. In parallel, Cohere absorbed Aleph Alpha, the once-celebrated German champion, in a deal that left Aleph Alpha's existing investors holding 10% of the combined entity.
The reporting around this is chronologically strange and I will say so plainly: the timeline does not fully cohere, several sources are missing, and at least one figure attached to OpenAI sits an order of magnitude beyond anything in my own records. I am not endorsing the numbers. I am interested in the logic, because the logic is what will be copied by France, Japan, Korea, and every finance ministry that wants a domestic AI champion on its soil.
The pitch is simple and emotionally powerful. Governments and regulated industries do not want their data flowing into American clouds, subject to American law and American subpoenas. They want a model they host themselves, a vendor they can audit, a jurisdiction they can vote in. That is a real need. I have argued the same thing about self-custody for a decade — and I still believe it.
But the mechanism by which that need is being financed is not decentralized at all. And the market is being asked to believe that a change of address is the same thing as a transfer of power.
Let me start with the multiple, because the multiple is the whole story.
Sovereign AI customers — governments, banks, defense-adjacent enterprises — buy private deployments. That means customized integration, on-site infrastructure, and a support team per contract. This is the business model of Oracle, SAP, and IBM. Historically, the market pays those companies somewhere between five and fifteen times sales. Cohere is being priced at 83 times. That is the multiple of a hypergrowth public SaaS company at the peak of a bull market, applied to the revenue profile of an enterprise systems integrator.
The valuation is not a forecast about growth. It is a price paid for scarcity and political optionality. There is exactly one credible, Western, non-US sovereign AI vendor at scale. When you are the only asset a government can buy to satisfy a strategic requirement, price discovery stops working. This is not a market clearing. It is a procurement decision wearing a term sheet.
And the comparison itself deserves scrutiny, based on my audit experience reviewing how revenue is classified across deployment models. The OpenAI figure appears to use operating revenue. The Cohere figure uses annual recurring revenue. ARR is a prospective measure; confirmed revenue is retrospective. If you normalize both to the same definition, Cohere's real multiple is likely north of 100 times. The gap between the two companies is even wider than the headline suggests — not narrower.
Now look at the leverage. When a state contributes 240 million dollars and that unlocks two to three billion, the state is buying enormous strategic influence for a small fraction of the capital. The private investors are putting up real money to fund a national industrial policy. That is their choice, and for a pension fund with a thirty-year horizon it may even be rational. But let us be honest about what it is: not venture capital making an independent judgment, but patient institutional capital buying a policy outcome.
And then there is the detail that the coverage buried, which I think is the most telling sentence in the entire affair: Cohere reportedly declined to re-register in the United States. Refusing American incorporation is not a patriotic gesture — it is a decision to seal off the primary exit path for technology investors. It means the future cap table will be dominated by Canadian pensions, German industrial capital, and sovereign wealth funds. It means the eventual liquidity event is a strategic sale or a regional listing, not a Nasdaq debut. You do not get to call yourself the neutral infrastructure of the free world and simultaneously deny your investors the deepest capital market on earth. Something has to give.
The Aleph Alpha merger is the part the ecosystem should study most closely, because the comfortable word "merger" is doing a great deal of work. Aleph Alpha raised well over half a billion dollars from some of the most serious industrial names in Europe. It ends with a 10% stake. If that ratio is anywhere near accurate, its investors absorbed a seventy to eighty percent write-down, delivered as a fait accompli and dressed as diplomacy. This is not a coming-together of equals. It is the orderly liquidation of a failed national champion, and the receiving entity inherits both the German government contracts and the losses attached to them.
I have seen this exact choreography before. During the DeFi Summer of 2020, I led a volunteer research team of fifteen developers to audit early governance mechanisms, and we published a fifty-page study on how liquidity was really distributed. Time and again we watched small protocols "merge" into larger DAOs at ratios that quietly annihilated the smaller community's holdings, while the messaging emphasized partnership and shared vision. — Root: DeFi Summer. The token holders found out months later, usually in a forum post written in the passive voice.
Which brings me to the governance parallel that should make every DAO delegate uncomfortable.
Delegation makes governance more centralized, and sovereign AI financing makes capital more centralized, and both are sold to us as the opposite. In a DAO, the theory is that token holders delegate to experts who vote on their behalf. In practice, most holders do not research, so they delegate to the loudest KOL, and voting power pools into a handful of addresses. The mechanism designed to distribute voice ends up concentrating it. Now translate that into this deal. The mechanism designed to distribute AI capability away from a single American center is concentrating the ownership of that capability into a small circle of state funds and industrial conglomerates who will, quite reasonably, act in their own interest.
The cloud provider is the entity that actually holds the keys. The Schwarz Group does not merely invest six hundred million dollars; it also supplies the infrastructure on which Cohere runs. That is a vertical integration, not a financial bet. And in a vertically integrated structure, the investor controls the cost of the very input the company depends on. When your landlord also sits on your board, your independence is a matter of goodwill, not of architecture.
Here is what I think almost everyone is getting wrong about this story.
The consensus reading is that Cohere has proven a new path — that labs can escape American dominance by pivoting to sovereign customers. I want to push back. The pattern I see is the opposite: the lab is not the protagonist. The lab is the demand anchor for a cloud. The 11 billion euro Berlin data center, the STACKIT sovereign cloud, the six hundred million dollar check — these are infrastructure plays. Cohere's model is what fills the seats. If Cohere thrives, STACKIT thrives. If Cohere dies, the Schwarz Group still owns the concrete and the GPUs and the grid connection. That asymmetry tells you who bears the risk and who collects the rent.
We made this mistake in crypto for years. We celebrated "decentralized" applications that ran on three centralized node providers, or bridged through one dominant multisig. The narrative measured decentralization as a metric; the reality was a small set of critical dependencies that, if they failed, took everything with them. Decentralization is a mindset, not a metric — and sovereignty is the same. A government can mandate that data stays inside its borders and still hand total control to a single vendor. Geography is not autonomy.
There is a second blind spot, and it is generational. Sovereign AI customers are buying defensibility, not capability, and defensibility erodes the moment open models catch up. Llama descendants, Mistral, Qwen, DeepSeek — the open ecosystem keeps closing the gap in local deployment. A finance ministry that can run an open checkpoint on its own metal at a fifth of the price does not need an eighty-three-times-sales vendor forever. The moat here is political trust, and political trust is exactly the kind of moat that a change of government can drain overnight. The client list — Oracle, SAP, Dell, McKinsey, RBC, Fujitsu, LG — is impressive, but those logos are distribution channels as much as they are customers. When your model is white-labeled inside someone else's private cloud, your brand is being diluted with every seat you fill.
I facilitated a working group of thirty ethicists and developers in 2026 to draft an accountability charter for autonomous agents transacting on public ledgers. The single hardest question we faced was: when a machine acts and harms someone, who answers? We kept arriving at the same uncomfortable conclusion — accountability requires a named human in a jurisdiction that can be held to account. The sovereign AI model actually has an advantage here, because there is a state standing behind the deployment. But it also introduces a new failure mode: if that state chooses to reach into the deployment, the neutrality it sold you evaporates, and you have no cryptographic recourse. Cloud is not custody.
So what do we do with this?
For builders: watch who owns the infrastructure, not who owns the narrative. The interesting question is never "is this decentralized?" — it is "if the loudest funder changes their mind, what breaks?" If the answer is "everything," you are not sovereign. You are a tenant.
For investors: a valuation multiple is a story with a number attached. When the story is "this is the only asset of its kind," be careful, because that sentence has preceded almost every peak I have lived through. — Root: The 2022 Bear Market.
And for all of us who believe in this technology: the real test of the next decade is not whether we can route value around the banks. It is whether we can route decision-making around the entities that own the rails. Code is law, but people are the protocol. The question I keep returning to, and the one I will leave with you, is not whether sovereign AI can escape American gravity. It is whether any of us — nations, DAOs, or individuals — can actually hold our own keys when someone else owns the driveway.