Two numbers, pulled from the same page.
An 83x price-to-sales multiple for Cohere. A 34x price-to-sales multiple for OpenAI.
The first company is smaller. It grows slower. It reports a lower gross margin. And it is priced at 2.4x the multiple of the second.
That does not reconcile.
When two figures refuse to reconcile, I do not assume the market is irrational. Markets are lazy, not stupid. I assume I am missing an entry. Something else is doing the work. Here, that something is not revenue, not growth, not margin. It is a government, a cloud provider, and a category label: sovereign AI.
This is the instinct that carried me through six weeks decompiling MakerDAO's CDP system as an undergraduate. The whitepaper said one thing. The assembly said another. I trusted the assembly. The same rule applies to a funding round. Read the bytecode, not the press release.
The Cohere round is an assembly I cannot fully disassemble. But the parts I can see are worth tracing.
Ghost in the audit: finding what wasn't declared.
Context: what was handed to me
Cohere, a Canadian AI lab, reportedly raised $2-3 billion at a $20 billion valuation. Twelve months earlier, the same company was valued near $7 billion. A near-tripling. Over roughly the same window, ARR reportedly moved from a $200 million target toward $240 million as of February 2026. Growth of 20 to 40 percent. Valuation growth of 200 percent.
A German retail conglomerate, Schwarz Group, led roughly $600 million of the round. Schwarz owns STACKIT, a sovereign cloud business, and is building an eleven-billion-euro data center campus in Berlin. The Canadian government put in 240 million Canadian dollars, about $175 million, as strategic backing.
Around April 2026, Cohere and Aleph Alpha merged. The German lab, which had raised more than $500 million from Bosch, SAP, Schwarz, and HPE, was folded in at a 90:10 split. Ninety for Cohere, ten for Aleph Alpha.
And at some point in the arc, Cohere declined to reincorporate in the United States.
Three labs, three exit paths, one page. OpenAI is being financed as capital-intensive infrastructure — a utility with a valuation in the hundreds of billions. Anthropic is being positioned for a clean-balance-sheet public listing. Cohere is being financed as sovereign industrial policy, with state capital and a dual-headquarters structure. Three different customer bases, three different liquidity paths, three different risk profiles. That divergence is itself the story. A single AI sector does not behave this way. Three sectors wearing one label do.
Before I go further, a disclosure that matters more than any of the above.
My verifiable data ends around early 2025. Every 2026 event here sits outside that range. The source material is a 35-point dossier in which 19 points carry no attribution at all — including the merger terms, the Anthropic IPO ambition, and Cohere's refusal to register in the US. That is a high-conviction opinion wrapped around low-confidence facts. It is the same structure most token sale decks used in 2017.
I am not endorsing the numbers. I am auditing the logic the numbers are asked to carry. If the numbers are wrong, everything below collapses. If they are right, the structure is more interesting than the headline.
What sovereign AI actually describes
Strip the branding. Sovereign AI means a government wants its model, its data, and its inference to sit inside its own jurisdiction, under its own law, on infrastructure it can physically point to. No data leaving for a US hyperscaler. No dependency on a vendor that could be sanctioned, deprioritized, or subpoenaed.
That is not a new idea. It is the same argument every sovereign layer-1 made in crypto. Every government-backed chain, every national stablecoin pilot, every domestic chain proposal since 2019 has used the identical sentence: we cannot depend on foreign infrastructure for critical systems. The sentence is true. It has always been true. It is also extraordinarily convenient for whoever sells the alternative.
I have watched that argument fund a lot of mediocre code. I do not want to let it excuse an 83x multiple by default.
But I have to be fair. The sovereign AI buyer is not a retail speculator. It is a procurement office with a legal mandate. That changes the demand curve. A government does not buy on price. It buys on admissibility. Cohere's Canadian registration plus its dual Berlin headquarters makes it admissible in a way OpenAI and Anthropic structurally are not.
Admissibility is a real moat. It is just not a growth moat. The market priced it like growth.
The valuation inversion is the entry that does not balance
Let me do the arithmetic the way I would do it on a cap table.
83x on $240 million ARR implies $20 billion. Fine. Now look at what the revenue is made of. Eighty-five percent of Cohere's revenue comes from private deployments. Not API calls. Not self-serve. Deployments. Someone installs the model inside a customer's environment, integrates it, supports it, maintains it.
That is not a SaaS business. That is enterprise software services. The closest comparables are not Snowflake or Datadog. They are Oracle, SAP, IBM. Historically those trade between 5x and 15x sales. At the generous end, 15x on $240 million is $3.6 billion. The round asks for 5.5x that.
The 70 percent gross margin sharpens the point instead of softening it. Seventy percent beats traditional IT services. It loses to top-tier SaaS, which clears 80. And it is achieved partly because inference runs on the customer's hardware. The cost moves off Cohere's books and onto the buyer's. That is a legitimate structural choice. It is not margin expansion. It is margin relocation.
So: smaller than OpenAI by roughly two orders of magnitude on revenue, growing slower, with a services-shaped margin profile, at 2.4x the multiple.
Reconcile that.
You cannot reconcile it with growth. So the market is paying for something that does not appear on the income statement. Call it the sovereign premium. I would break it into three stacked components: non-market capital, scarcity, and an option on a category that may not exist.
None of those three are revenue. All three are narrative.
Trust is math, not magic: stripping the sovereign premium apart
Take the capital first.
The Canadian government put in $175 million. The round is $2-3 billion. That is leverage of roughly 1:11 to 1:17. Government money does two jobs at once. It signals national priority, and it de-risks private capital that would otherwise demand a lower entry price. This is the same structural role a state grant played in early semiconductor fabs, and the same role a foundation treasury played in early layer-1 ecosystems. The state absorbs first-loss narrative risk. Private capital buys the upside the narrative creates.
Is that irrational? No. It is deliberate. The question is who pays for it. If the round is led by pension funds and industrial strategics spending voluntarily, fine. If they are being steered, the market price is a fiction and 83x is not a market signal at all. It is a memo with a number on it.
Take scarcity second.
There are very few admissible sovereign AI vendors. The buyer pool is small. The seller pool is smaller. Any auction with three bidders and one qualified seller produces a price reflecting the auction, not the asset. This is what happened to every compliant token in 2018 — the ones that survived a regulatory filter traded at a premium that had nothing to do with usage. Compliance scarcity is real scarcity. It is also thin. It evaporates the moment a second qualified seller appears.
Take the option third.
Schwarz's $600 million is not a financial bet. Schwarz owns STACKIT. It is building billions in data center capacity. Cohere is the anchor tenant that makes that infrastructure legible to everyone else. This is vertical integration dressed as venture capital. The $600 million buys a captive workload for a cloud business and a story for a fundraising deck. When one counterparty is investor, landlord, and vendor at the same time, you are not looking at a round. You are looking at an org chart that has not been drawn yet.
When the vault opens itself, the value flows away from the AI lab
Follow the incentives and the picture inverts.
Schwarz benefits if Cohere succeeds, because STACKIT gets workload. Schwarz also survives if Cohere fails, because it still owns the cloud assets. Cohere carries all the model risk. Schwarz carries none. That asymmetry is the entire story, and it is invisible on the 83x headline.
I have seen this shape in market structure before. During DeFi summer in 2020, I isolated Compound's cToken implementation on a testnet and manipulated the interest rate model until a rounding error surfaced. It was small — a few basis points. I spent two weeks scripting an exploit proof-of-concept and concluded early users could lose about $45,000 in aggregate. I sent it in anonymously. The fix landed within 48 hours. The lesson was not that rounding errors are dangerous. The lesson was that theoretical security models routinely break against practical edge cases, and the people closest to the model are usually the last to notice.
The sovereign premium is a rounding error the size of a round. It lives in the gap between what the multiple implies and what the business earns. Nobody in the room has an incentive to notice.
Reconstructing the cap table as a ledger
Here is where I apply the method I used after FTX.
I did not write opinion pieces in November 2022. I downloaded the public blockchain data from the exchange's hot wallets and traced fund movements across three months. I mapped 1,200 transactions until the commingling between customer funds and Alameda was visible as a shape rather than a claim. The discovery was not the fraud. The discovery was that the fraud was legible on-chain months before it was legible in the news, and nobody was reading the ledger.
The same rule applies here, with one caveat: this ledger is private. There is no explorer for a sovereign AI cap table. My instruments are the disclosure and the silence around it. So I read the silence, and I read the flow.
Trace the flow. Canadian government in at $175 million. Schwarz in at $600 million. Pensions and strategics fill the rest. Out: compute, headcount, integration costs across a growing enterprise customer base, plus whatever liabilities came attached to the Aleph Alpha assets. The single largest line on the out side is almost certainly the build-out the round is financing. A company scaling private deployments does not scale linearly. Every new government customer adds integration labor that refuses to amortize the way an API customer does. You can hide that on a slide. You cannot hide it on a P&L.
The compute layer decides the model layer
There is a dimension of this deal that the headline ignores, and it is the one that will set the ceiling.
Cohere's compute strategy runs through STACKIT, the sovereign cloud. That is coherent on geopolitical logic. It is weaker on the three axes that decide model trajectories: efficiency, cost, and chip supply.
European GPU density trails US hyperscalers by a year or more. European clouds lack the scale effects that make AWS, Azure, and Google profitable at the margin. And Europe does not make AI accelerators. ASML builds the lithography machines, not the chips. Cohere's training and inference will depend on NVIDIA or AMD supply regardless of how sovereign the narrative is. If export controls tighten even against allies, that supply chain has a single point of failure located outside the jurisdiction it claims to be independent of.
The absence that should make you pause is NVIDIA itself. Anthropic's round carried a rumored ten-billion-dollar anchor from NVIDIA. Schwarz led Cohere's. NVIDIA is not at the table. That is not proof of a disadvantaged position in the compute queue, but it is consistent with one.
This is where my ZK work becomes relevant, and it is the part most analysts skip. In 2024 I spent three months profiling the constraint generation phase of a Plonk-based proof system, hunting bottlenecks in the arithmetization process. I rewrote the field arithmetic in Rust and cut proof generation time by 15 percent on a standard 10,000-transaction suite. The headline number was small. The lesson was large: theoretical complexity does not translate into practical performance. A protocol can be elegant on paper and cache-hostile in production. The gap between the two is where every real engineering cost lives.
Private AI deployment has the same gap. On paper, a model running inside the customer's environment looks like infinite margin. In practice, every deployment is a small bespoke engineering project. The customer's infrastructure is different. Their compliance constraints are different. Their integration surface is different. The cost does not disappear. It moves to a line item that is harder to scale. That is the structural reason an 85-percent-private-deployment revenue mix behaves like a services business, not like software, no matter how it is labeled.
So when the dossier reports an ARR of $240 million, I want to know one thing above all others. Is that contracted recurring value, or annualized first-year value on multi-year deployment contracts? In private-deployment businesses those two numbers can differ by a factor of two. The 83x is computed on the friendlier one. That is not an accusation. It is a measurement standard.
The 90:10 split is not a merger. It is a write-down
Aleph Alpha raised more than $500 million from Bosch, SAP, Schwarz, and HPE. It ends up with 10 percent of a combined entity. Its investors accepted something on the order of 70 to 80 percent of book value gone. There is no other reading of the ratio. A merger between peers splits near 50:50. A 90:10 split means one side had no leverage left.
I have seen this exact shape at a smaller scale. In 2021, during the NFT run, I analyzed the Ethereum sidechain behind Axie Infinity and found a discrepancy between the advertised minting logic and the actual bytecode. I wrote a node script to trace minting transactions and showed the contract permitted mints under specific block conditions the documentation did not describe. I published the breakdown. The team hard-forked shortly after. Digital beasts, fragile code: the story said one thing, the constructor said another.
The pattern here is identical at a higher altitude. The public story says European consolidation. The mechanics say one asset was marked to zero and the surviving entity absorbed its obligations. What obligations? German federal contracts, Bosch and SAP relationships, and whatever burn and liabilities traveled with them. That is a plausible explanation for why a multi-billion round is needed at all. You do not raise $2-3 billion because things are going well. You raise it because things are expensive.
Aleph Alpha is not a footnote. Aleph Alpha is the control group. It is what happens when a sovereign AI lab takes industrial capital, fails to reach scale, and gets absorbed. The market is pricing Cohere at 83x. The market also watched Aleph Alpha's investors take a 70-80 percent haircut on the same thesis one quarter earlier. Those two facts should be read together, and they usually are not.
Silence speaks louder than the proof
There are things the round does not say, and I trust the absences more than the disclosures.
Cohere reportedly declined to reincorporate in the United States. Everyone read that as a sovereignty statement. It is also a capital-markets statement. Refusing US registration removes the company from the standard path to a US listing and from the pool of US growth funds that would otherwise anchor the cap table. Read the cap table forward. Canadian pension funds. German industrial capital. Possibly Gulf sovereign wealth. This is a shareholder base that has to hold for a decade, not one that wants to. The liquidity path is not sealed by accident. It is sealed by design.
Second absence: the terms. A $2-3 billion round in this environment almost always carries milestone conditions — ARR targets, government contract counts, a listing timeline. None are disclosed. When a term sheet is silent on structure, you have two options. Either the terms are unfavorable and under NDA, or the structure is unfavorable by being simple. Either way, understand what you are not being told: pre-money or post-money, whether there is a 1x participating liquidation preference, whether there is an IPO ratchet. Those three items decide whether founders and employees keep their equity or watch it dilute toward zero.
I have a specific reason to care about liquidation preference language. This is the tokenomics lesson crypto learned the hard way. A token sale with a 1x participating preference and a vesting cliff is functionally a senior secured claim with a marketing wrapper. The retail buyer sees upside. The structure sees a stack of priority. Sovereign AI financing is running the same playbook with pension capital instead of retail, and the seniority is buried in a term sheet nobody will publish.
There is a related pattern I keep returning to, and it is why I distrust lock-in dressed as permanence. Soulbound tokens have been a three-year concept precisely because nobody wants an immutable record they cannot exit. Sovereign AI sells a version of the same promise: a model, a jurisdiction, a record that does not move. That is attractive to a state and corrosive to a market. When your data and inference live somewhere permanent by design, you have not bought a service. You have bought a relationship with no exit clause.
And the deepest absence: there is no net revenue retention. No churn. No customer acquisition cost. No average contract value. Every metric that would tell me whether sovereign AI is a durable business or a procurement cycle is missing. You can build a compelling narrative without any of them. You cannot value a company without them.
This is the same substitution I watch in stablecoins. Tether has sat near 70 percent of the stablecoin market for years with no fully independent audit of its reserves, and the industry has agreed to act as if that is fine. The market accepts it because the demand for a dollar in an account is stronger than the demand for proof. The same trade is happening here. The demand for a sovereign AI story is stronger than the demand for a reconciled round.
I am not saying the Cohere numbers are false. I am saying that on the current evidence they are unaudited. A majority-unsourced fact pattern priced at 83x is not a valuation. It is a bet on the absence of a correction.
The category is manufactured, and that is the point
I have a standing suspicion about categories that arrive fully formed with a market size attached. It is the same suspicion I hold about liquidity fragmentation in DeFi, which is not a problem so much as a phrase that sells the next product. Sovereign AI is a real demand signal wrapped in a manufactured category. The demand is real: governments do want local control. The category is manufactured: the $600 billion market forecast circulating alongside this round has no independent verification in the dossier, and market-size forecasts in a category with three buyers and one qualified seller tend to be aspirational.
Recognize what the category does. It converts a procurement preference into an investable theme. Once it is a theme, capital allocates to it whether or not unit economics justify it. The theme then defends itself, because questioning it reads as questioning national sovereignty, which nobody wants to do in public. That is the failure mode of this category. Not fraud. Compliance. The kind of environment where nobody asks whether the model is competitive because asking sounds disloyal.
The industry effects are real but lopsided. Sovereign AI absorbs government IT budgets that might otherwise have gone to general cloud and consulting spend. It keeps high-skill integration jobs close to the customer — dozens per deployment rather than thousands of engineering roles in a central lab. It shifts capital toward infrastructure owners and away from model owners. If sovereign AI becomes a permanent category, the durable winners are the clouds and the systems integrators, not the labs. The labs are the branding layer. The compute is the asset.
Contrarian: the risk is not in the model race
The consensus read is that this round validates sovereign AI as a category and marks Cohere a winner. My read runs the opposite sequence. The round validates that sovereign AI can be financed. It does not validate that it can be profitable. Those are different claims, and only the second one reaches a clean exit.
Here is the blind spot. Everyone is watching the model race — benchmarks, funding rounds, parameter counts. The risk is not in the model race. The risk is in the capital structure, and there are three exposures under the 83x that have nothing to do with AI capability.
First, the anchor-tenant problem. Schwarz is simultaneously investor and cloud provider. Cohere's most important relationship is also its pricing counterparty. A cloud provider that owns part of its biggest customer can move margin between the two balance sheets. Over three to five years, that is not a partnership. That is an internalization path. Cohere does not get acquired. It gets absorbed, the way a key supplier becomes a division.
Second, the shareholder-base problem. Refusing US registration removed the standard exit. The cap table that remains is patient capital, which sounds good until you remember why it is patient: it has no fast exit available. Patient capital eventually wants its return through one of two doors. Either a listing in a secondary venue that will price this like enterprise software, not like a frontier lab, or a strategic sale to one of the very corporates already on the cap table. Both doors point below $20 billion.
Third, the category problem. Sovereign AI is a political construct as much as a market one. Government demand for it is a function of the political cycle. A change of government in Ottawa or Berlin can reprioritize the budget line overnight. When I mapped where the value actually accrues in this deal — to the infrastructure owner, not the model owner — I appreciated the structure more. It also confirmed that the model owner carries the least defensible risk.
Now add the clock nobody is watching: open-weight progress. Cohere's Command R line trails the frontier on public benchmarks by a year or more. In a market where government buyers want the best model for the money, a trailing model has to be compensated by something, and that something is admissibility. Admissibility is a defensible moat precisely until open-weight models strong enough for local deployment arrive. Llama onward, Mistral, Qwen, DeepSeek — every one of them compresses the moat. A government with data-sovereignty needs and a competent IT department can, in time, assemble an open-weight stack without paying an 83x multiple to anyone.
I watched this exact dynamic in NFT infrastructure. The moat was never the token. It was the ecosystem, and the ecosystem leaked. The market caught up eventually.
Takeaway: three things to watch
Forward-looking, then. Three signals over the next 18 months.
Watch whether any independent source appears for the 19 unsourced claims. Until one does, treat the round as unaudited. An unsourced majority in a fact pattern this load-bearing is not a documentation problem. It is the whole product.
Watch whether a second qualified sovereign vendor emerges in the EU. Mistral is the obvious candidate. The moment a real competitor exists, the scarcity premium propping up 83x compresses toward the enterprise software range, which is 5x to 15x, and the math does the rest.
And watch open-weight progress, because that is the clock on the entire category. The day a government can assemble sovereign inference from open weights plus local infrastructure, the premium sovereign AI charges for admissibility stops being a moat and becomes a line item a procurement office can cut.
The symmetry here is uncomfortable and worth naming. The entire thesis of sovereign AI is that a nation should not depend on foreign infrastructure it cannot audit. That is a good argument. It is the same argument that would tell a pension fund not to underwrite an 83x multiple built on 19 unsourced facts. Sovereign capital should hold sovereign investments to sovereign-grade verification.
The ledger is open on one side and closed on the other. That asymmetry has a price. We do not know yet who pays it.