Liquidity doesn't reward speed. It rewards the ability to leave.
Cohere closed a round reported at $2–3 billion at a $20 billion valuation. Run the arithmetic and you land near 83 times price-to-sales. OpenAI, on figures circulating in the same reporting, sits near 34 times. That is not a rounding error. It is a 2.4x premium for a company whose recurring revenue — $240 million as of February 2026 — is roughly one one-hundred-and-fourth the size of OpenAI's operating revenue, with margins the same reporting flags as "below standardized SaaS."
A smaller company, growing more slowly, earning less per dollar of revenue, is priced higher. If that sentence doesn't stop you, nothing below will.
I have spent nine years watching capital move. In 2017 I audited fifty whitepapers from a boutique advisory desk in Vancouver and found that 80% of them had no functioning liquidity model. I stopped asking what the technology could do and started asking where the exit was. That habit is the only reason this number reads as a warning rather than a triumph.
Cohere is not a conventional frontier lab. It is a privatized-deployment business — 85% of revenue comes from on-premise installs, not public APIs. Its customer roster reads like a systems integrator's wish list: Oracle, SAP, Dell, McKinsey, RBC, Fujitsu, LG. It is registered in Canada, runs a second headquarters in Berlin, and absorbed Germany's Aleph Alpha in a 90:10 structure. It has reportedly declined to re-register in the United States.
That last fact is the entire story, and almost nobody in the coverage treated it as one.
Refusing U.S. registration is not a patriotic gesture. It is a deliberate forfeiture of the deepest capital pool on earth — the dollar-denominated fund complex that underwrote every major technology exit of the past two decades. It tells you, before a single term sheet is signed, where Cohere's investor base is heading: Canadian pensions, German industrial capital, Gulf and Southeast Asian sovereign funds. Patient money. Policy-aligned money. Money that does not need a 2027 IPO to return capital.
Crypto has seen this film. I lived a version of it in 2017, when a wave of "national champion" token projects raised on sovereign-adjacent narratives and then discovered that geopolitics generates no order flow. The ICO market had capital without liquidity. Sovereign AI has liquidity commitments without an exit. Both mistakes get made by confusing strategic importance with economic viability.
Start with the business-quality mismatch. 83x for a services business. The comparison set for privatized deployment is not Snowflake. It is Oracle, SAP, IBM — enterprise software-services hybrids that historically trade between 5 and 15 times sales. Cohere is priced at the ceiling of the purest SaaS multiples while wearing the cost structure of an integrator. Its reported 70% gross margin is respectable against traditional IT services (30–50%) but sits below elite SaaS (80%+) and below where public-cloud API businesses converge at scale.
The reporting attributes the gap to a sovereign premium. I don't buy it. A premium requires scarcity the buyer cannot escape. A discount appears when the buyer cannot exit. What Cohere actually has is a reverse hostage situation: it has left the U.S. market, and the U.S. market has left it.
Follow the government money and the leverage stops being subtle. Canada's $240 million CAD commitment — about $175 million USD — sits beneath a $2–3 billion round. That is a leverage ratio near 1:11 to 1:17. Two readings, both true. One: a small public check buys outsized strategic voice over a critical asset. Two: private capital foots the bill for a national strategy it did not design. The second reading decides whether this round prices like a market or like policy.
Schwarz Group's $600 million lead does not rescue the market-pricing story. Schwarz owns STACKIT, the sovereign cloud, and is building an €11 billion data center in Berlin. Its incentive is not a financial return on Cohere equity. Its incentive is an anchor tenant for its own infrastructure. That is vertical integration wearing a venture check. When your lead investor is also your landlord, your landlord is also your price-setter.
Now the part everyone called a merger. Aleph Alpha had raised more than $500 million from Bosch, SAP, Schwarz, and HPE. It now holds 10% of the combined entity. That is not a partnership. That is a mark-to-zero.
I have marked positions to zero before. In 2022 I sat in the TerraUSD pools tracking withdrawal rates in real time as the peg unwound and liquidation cascades turned a "strategic reserve" into a stampede. The lesson was mechanical, not moral: when a structure is described in diplomatic language, the numbers underneath are usually brutal. Aleph Alpha's investors took a 70–80% paper loss, and the transaction was framed as European coordination. Europe's attempt to build a single sovereign champion failed. Cohere is the buyer of that failure, not the proof that the model works.
Run the growth rates side by side and the FOMO becomes arithmetic. Cohere's recurring revenue reached $240 million by February 2026 against a $200 million target — a genuine beat. But that target implies 2024 revenue somewhere between $100 and $150 million, so annual growth lands in the 30–60% band. For a foundation-model company, that is mid-speed at best. The same reporting pegs OpenAI above 300% and Anthropic above 500%. Meanwhile Cohere's valuation moved from $7 billion to $20 billion in twelve months, a near-3x jump on a base growing at most 40%. Valuation compounding at 200% on revenue compounding at 40% is the definition of multiple expansion, and multiple expansion on a decelerating base is the signature of late-cycle capital.
This is where my 2024 work applies. When the spot Bitcoin ETFs launched, I modeled daily creation and redemption flows against traditional equity fund flows. The conclusion that mattered was not that institutions were buying. It was that institutional capital acted as a volatility dampener, not a driver — it entered slowly, stayed longer, and repriced risk rather than momentum. Sovereign capital behaves the same way. It is slow, patient, and politically motivated, which means it will not chase Cohere's multiple higher, and it will not create the liquid float that lets early staff and funds exit. It anchors price. It does not levitate it.
The reporting's sharpest contribution is a three-path split, and it deserves credit. OpenAI as a capital-intensive infrastructure utility, at a valuation and raise I regard with extreme skepticism. Anthropic as a clean-balance-sheet public-market exit. Cohere as a sovereign industrial-policy vehicle. Three labs, three capital structures, three exit paths.
I want to press on that skepticism explicitly. A $122 billion raise at an $852 billion valuation sits an order of magnitude beyond any verifiable data I have seen. When a single figure is that far outside the observable range, treat the entire dataset it belongs to as opinion, not fact. Skepticism isn't cynicism. It is the difference between reading a press release and reading a cap table.
Map the three paths onto crypto, because crypto has run all three experiments. Path one, the capital-intensive utility, is Bitcoin post-ETF: heavy infrastructure, institutional anchor capital, dampened volatility, upside capped by the size of the buyer. Path two, the clean public-market exit, is the ETH-ETF template: institutional acceptance, regulatory clarity, a mature float. Path three, the sovereign industrial-policy vehicle, is the sovereign L1. I have watched that movie. Cosmos's IBC is technically elegant; I have said so for years. But the application layer fragmented across a hundred chains, and ATOM captured almost none of the value that ecosystem produced. Sovereign AI faces the identical trap. Grand narrative, thin value capture. When your customers are governments on fixed fiscal budgets and your moat is a citizenship requirement, you have built a protected market, not a growing one.
This connects to something I modeled this year: a simulation of AI agents transacting through blockchain wallets, micro-payments measured at the protocol layer. The finding that surprised me was that autonomous agents don't care about sovereign boundaries. They route to wherever execution is cheapest and finality is fastest. A model deployed on-premise behind a national firewall serves humans who need compliance. It does not serve an agent economy that is borderless by construction. If the next wave of AI demand is machine-to-machine, sovereign deployment is not a moat. It is a wall around a shrinking room.
Now the counter-intuitive part. Everyone is watching Cohere. The actual winner is not the lab.
Schwarz puts $600 million into Cohere and €11 billion into a Berlin data center. Read those two numbers together and the strategy is transparent. Schwarz is using the AI narrative to finance its own cloud infrastructure. If Cohere succeeds, STACKIT gets its anchor tenant and the data center fills. If Cohere fails, Schwarz still owns the sovereign cloud asset and the physical compute. Downside hedged, upside leveraged — a far better risk-adjusted position than anything Cohere's equity holders carry.
I raised the same flag on liquidity-fragmentation narratives in DeFi. The "problem" that yields a new product is often manufactured by whoever sells the product. Sovereign AI's market-size claim deserves identical scrutiny. If it is genuinely a $600 billion market, that figure needs an independent source. My working hypothesis, built from public-sector IT procurement patterns, is that most of it is existing government budget relabeled. A relabeled budget is a zero-sum reallocation, not a new market. And a zero-sum market does not support 83x multiples.
The other blind spot: the United States is not the loser. Refusing U.S. registration looks like defiance in a pre-2017 frame. In a 2026 frame it looks like opting out of the most liquid capital market on earth while your competitors consolidate it. The real question is not whether sovereign AI can exist. It is whether it can exit.
So here is the trade, and it is not the one being pitched. Sovereign AI is a defensible niche with a low ceiling, priced as if it were a scalable platform. The lab absorbs the political risk; the cloud provider absorbs the upside.
One number decides everything from here: the liquidation preference on this round. If it is 1x participating, the founder and employee pools are already impaired at a $20 billion headline, and the real entry price is a fiction.
Liquidity doesn't reward the sovereign. It rewards whoever holds the exit.