The Mistral Mystery: How Much is the European AI Champion Worth?

Mistral AI was founded in Paris in April 2023, a few months after OpenAI first released ChatGPT to the public, by Arthur Mensch, formerly of Google DeepMind, and Guillaume Lample and Timothée Lacroix, formerly of Meta's FAIR laboratory in Paris [1]. The founding promise was openness: model weights anyone could download, inspect, and run, a European answer to the closed American laboratories. Three years on, that promise has largely held for the models. It has never extended to the money.

Here is, in essence, everything the company has formally told the public about its finances since it was founded:

Read the table twice and the pattern appears: every public milestone in Mistral's history is a price, not a result.

The revenue figures that circulate are second-hand. Roughly $30 million of revenue for 2024 [6]. Around €300 million of annual recurring revenue as of September 2025 [6]. Something above $400 million by early 2026 [7]. A target, attributed to the company, of more than $1 billion by the end of 2026 [7]. Notice what these figures span: three definitions (revenue, ARR, run-rate), two currencies, and zero audited statements. In three years, Mistral has never published an official revenue figure. Everything above is briefed, leaked, or estimated.

The contrast with its peers is stark. Anthropic publishes its run-rate milestones in its own funding announcements: $1 billion in December 2024, roughly $9 billion at the end of 2025, $47 billion by May 2026 [8]. OpenAI’s numbers reach the press within days of moving. Several Chinese laboratories now print theirs in Hong Kong listing documents [9]. Mistral’s Series C announcement, €1.7 billion led by ASML, contains no revenue figure at all [3].

It is worth being precise about who these peers are. The frontier laboratories build the models: OpenAI, Anthropic and xAI in the United States; DeepSeek, Moonshot, Zhipu and MiniMax in China. The application layer builds on top of them: Cursor, Perplexity. The public giants, Google and Meta, fold their AI economics inside businesses too large to compare. And in Europe, attempts keep coming: Switzerland's Apertus, launched loudly in September 2025 as a fully open public research model and continuing quietly as exactly that [10]; Yann LeCun's Advanced Machine Intelligence in Paris, seeded with $1 billion, though it pursues world models, a bet against the LLM paradigm itself rather than a head-on rival to Mistral [11]; OVHcloud's frontier ambitions announced in June 2026 [12]. Three years in, no European laboratory is close to Mistral in commercial scale. The champion has no domestic rival; its most funded neighbour is running a different race entirely.

What Mistral has instead of disclosed numbers is signal. ASML, Europe’s most valuable company and the world champion of its field, anchored the Series C with €1.3 billion [3]. Nvidia sits in the capital alongside it. And the political support is loud, at least in words; whether words become revenue is a question for later in this piece.

In June 2026, the company's discretion produced its own parable. Days after Mistral renamed its assistant Le Chat to Vibe, a parody benchmark chart went viral on X: a smug white cartoon kitten named Le Chaton Fat, credited with 100 trillion parameters and a thirty-point lead over the strongest American model on a benchmark that does not exist [13]. The illustration at the top of this article is a rendering of that meme; the kitten has, by now, more portraits than the company has published revenue figures. Some readers got the joke. A measurable share did not, concluded the model was real, and blamed European export rules for keeping it from them. Mistral's chief executive replied only that the name was “actually le gros chaton” [13]. It was the most attention the company had drawn all year, and it was about a product that does not exist. An information vacuum does not stay empty. The public filled this one with a fat kitten; prices can be filled the same way. Whether the silence is an omission or a strategy is the right question, and the client list suggests an answer.

Mistral is three businesses wearing one name. The first sells models: La Plateforme, the API through which customers pay per token, plus the open-weight models that anchor the brand. The second sells an assistant: Le Chat, renamed Vibe in May 2026, in consumer tiers at $14.99 a month and an enterprise edition [14]. The third sells computing itself: Mistral Compute, announced with Nvidia in June 2025, a GPU cloud pitched at customers who want their AI infrastructure European-owned [15].

A word on vocabulary, because the industry's main dividing line runs through it. A closed model is rented: the provider hosts the weights, the trained parameters that are the model, and customers reach them only through an API. OpenAI's and Anthropic's frontier models work this way, as does Google's Gemini. An open-weight model is handed over: anyone can download the weights, run them on their own hardware, and fine-tune them on their own data, though the training data and code usually stay private, which is why open-weight is not quite the same as open-source. Meta's Llama family made open weights mainstream; the Chinese laboratories, DeepSeek, Alibaba's Qwen, Moonshot's Kimi, made them frontier-competitive and made openness a geopolitical strategy [37]. The trade is simple to state. Closed buys the best capability with the least effort. Open buys control: on-premise deployment, customisation, no vendor lock-in, and inference costs you own rather than rent.

Mistral's defining choice runs through all three businesses: it sits on the open side of that line, and sells what surrounds the weights. An enterprise can take a Mistral model inside its own walls, fine-tune it on its own data, and run it with nothing leaving the building; the commercial licences, the platform, the support and the compute are what carry the price tag. The American laboratories lead with capability; Mistral leads with control. That is not a slogan, it is the product.

One nuance belongs here, in a single honest sentence: the sovereign champion also distributes through Microsoft, whose Azure platform has carried Mistral’s commercial models since February 2024 [16].

Who buys control? Large European incumbents, to start. CMA CGM, the shipping group, signed a contract reported at around €100 million when it was disclosed in 2025, which made it Mistral's largest known client at the time; whether it still is, nobody outside can say [17]. Airbus followed with a licensing agreement covering its full product suite, announced in June 2026 [18]. Stellantis, BNP Paribas, Capgemini and SAP appear on the company's own customer roster as of mid-2026 [19]. A vertical line, Mistral for Industrial Engineering, launched in May 2026 with BMW, EDF and CMA CGM as first clients [20].

Then the buyers for whom control is not a preference but a requirement. Helsing, the German defence software company, partnered since February 2025 [21]. The French Ministry of the Armed Forces, under a framework agreement signed in January 2026 [22]. The government of Luxembourg and Singapore's Ministry of Defence, both disclosed by mid-2026 [19]. And, announced in June 2026, an assistant powered by Mistral for every French civil servant [23].

Now look back at section 1's question, because this list answers it. Defence ministries do not itemise their AI spend. Intelligence-adjacent buyers do not want their vendor announcing monthly revenue milestones. Multi-year framework agreements produce no ARR press releases. The client mix that makes Mistral valuable is confidential by nature: quiet clients make quiet vendors. The discretion that frustrates the outside investor may be less a communications choice than a structural property of the order book. That is the innocent explanation of the mystery, and it deserves full weight before anyone reaches for a darker one.

What can an outsider still estimate? Third-party work, as of early 2026, puts the revenue mix across API usage, private and on-premise enterprise subscriptions, and consumer tiers, with roughly 60 per cent of revenue coming from Europe [7]. On the consumer side, the honest line is short: Le Chat reached a million downloads in its first two weeks on mobile in February 2025 [24], and no user figure has been published since. The best public signal is regulatory: as of mid-2026, we know Mistral's European user base sits below 45 million monthly actives only because the EU has not designated Vibe a very large online platform under the Digital Services Act [25]. Even the user count is an inference from a threshold not crossed. And a caution that applies to this entire article: every figure is a snapshot at its reporting date, and in a company reportedly growing at this pace, snapshots age in months, not years.

Two recent product moves complete the picture, and both matter for the valuation question ahead. The first is the application layer. The clearest commercial lesson of 2025 and 2026 in the United States is that agentic coding tools, Claude Code, Cursor, Codex, are where model capability turned into revenue fastest. Mistral’s answer arrived late in 2025: Devstral 2, a coding model near the top of the benchmarks, and Vibe CLI, an open-source coding agent claimed to be several times more cost-efficient than its American rivals [26]. It competes the way everything at Mistral competes: on openness, self-hosting and price, not on capability supremacy. Whether a control-led coding tool produces the revenue inflection the capability-led ones did is an open question, and this piece will not pretend to know the answer.

The second is compute. In June 2026, SpaceX began renting its xAI data-centre capacity to Google for $920 million a month, and to Anthropic for $1.25 billion a month through 2029: roughly $26 billion a year, for computing power, from two customers [27]. The lesson cuts both ways. Selling compute can out-earn selling models; and renting compute is now the largest cost line an AI laboratory carries. Anthropic's xAI contract alone equals roughly a third of its reported run-rate revenue, paid to a direct competitor [27]. Mistral has read the same lesson: its $830 million of debt buys 13,800 Nvidia GPUs outside Paris, with a target of 200 megawatts across Europe by the end of 2027 [4] [15]. The scale is one to two orders of magnitude below where the money currently is, and it is financed with debt rather than cash flow. But it points at the one market where the small player holds a card the giants cannot match: a European buyer who needs European-owned compute cannot rent from Memphis.

Whatever multiple the next section reaches for, one caution should already be visible: part of this company is becoming capital-intensive infrastructure, and that part should not carry a software multiple at all.

Which numbers should an investor even look at? Three definitions first, because the jargon hides the problem. Trailing revenue, also written LTM for last twelve months, is what the company earned over the past year: a fact, but an old one. ARR, annualised recurring revenue, takes the latest month or quarter and multiplies it out to a full year: a snapshot dressed up as a year, and a loose one, because companies compute it differently; Anthropic's ARR arithmetic is reportedly more aggressive than OpenAI's, which alone can move a multiple materially [51]. NTM is the estimate of the next twelve months: not a fact at all, but the nearest of the guesses. For companies at this growth rate, the backward-looking pair is nearly useless: it measures a company that no longer exists by the time the figure reaches you. Forward multiples are speculative by construction, though NTM is the least speculative estimate available, and, as we will see, the market visibly coordinates on it. And the metrics that will actually decide the outcome sit further out still and are barely estimable at all: the size of the addressable market in three, five, ten years and beyond, the share of it each laboratory captures, and, above all, how profitable any of this proves once price wars, training costs and compute economics settle. On that last question, the profitability path, there is very little visibility for any frontier laboratory today, Mistral included. Hold that hierarchy in mind, because what follows works with prices and revenue multiples not because they are the right metric, but because they are the only ones the public record permits.

Read the chart the way a buyer reads it. Paying 20 times next-twelve-month revenue sounds indefensible next to the S&P 500 at three and a half to four times, or even the Magnificent 7 at roughly nine [50]. But the comparison only means something once you attach growth rates. The Magnificent 7's revenues are growing at roughly 15 to 20 per cent a year on current analyst estimates, and a large slice of that average is Nvidia alone [50]. The frontier laboratories live on a different curve entirely: Anthropic multiplied its run-rate nine-fold during 2025 and five-fold again in the first five months of 2026 [8]; OpenAI roughly tripled in 2025 [28]; and Mistral, if its own guidance holds, would take estimated revenue of $400 to 500 million to $1 billion within about a year, better than doubling [7]. That is what the middle line of the chart assumes: 100 per cent growth, held for three years. Hold the price still and the multiple falls to 2.5 times, far below where the Magnificent 7 trades today. And if, three years out, the company still has strong growth in store and a clearer path to profitability, the market may well still pay 10 to 20 times: applied to a revenue base eight times larger, that is an exit worth four to eight times the entry price. That is the whole thought process behind paying 20 times forward, and every link in it is an if: the growth, its persistence, the exit multiple, the profitability that is supposed to have arrived by then. A later chart in this article will put numbers on those exit scenarios. For now, notice only that the chain is coherent. Whether it is true is a different question.

Begin, then, with what the market is paying for the companies Mistral is measured against. Anthropic's Series H closed on 28 May 2026 at $965 billion post-money, against a run-rate the company itself put at $47 billion: roughly 21 times revenue [8]. OpenAI's last primary round valued it at $852 billion against a reported run-rate near $25 billion in early 2026: roughly 34 times [28]. And Moonshot, the maker of Kimi, raised in May 2026 at $20 billion against roughly $200 million of annual recurring revenue reported in April 2026: about 100 times [29]. Note the last one, because it is the arresting comparison: Moonshot carries the same price as Mistral's rumoured round, on roughly half the revenue. A hundred times its last reported revenue looks, on its face, like a number that makes no sense. It is more rational than it appears: investors are pricing the future, and if Moonshot's revenue multiplies the way its American peers' revenue just did, the multiple collapses toward the ordinary within a year or two. Hold that pattern for the rest of this section: in this industry, multiples fall not because prices fall, but because revenue catches up.

The Multiple Spectrumas of early July 2026

Company	Last mark	Latest revenue report	Multiple on latest report	Implied NTM at 20x
Anthropic Frontier lab · US	$965bn28 May 2026 · closed round	$47bn run-rateMay 2026 · company-stated	~21x	~$48bn
OpenAI Frontier lab · US	$852bnearly 2026 · round	~$25bn run-rateearly 2026 · reported	~34x	~$43bn
Mistral Frontier lab · France	~€20bn ≈ $23bnJun 2026 · talks reported, not closed	~$400-500mearly 2026 · third-party estimates	~45-55x	~$1.15bn
Perplexity Application layer · US	$23bn9 Jan 2026 · round	>$450m ARRMar 2026 · reported	~51x	~$1.15bn
Anysphere (Cursor) Application layer · US · acquired by SpaceX	$60bn16 Jun 2026 · acquisition, option exercised	~$4bn ARRJun 2026 · reported	~15x	~$3bn
Moonshot (Kimi) Frontier lab · China	$20bnMay 2026 · closed round	~$200m ARRApr 2026 · reported	~100x	~$1bn
DeepSeek Frontier lab · China	~$45bnmid-2026 · talks reported	never disclosed	unknowable	~$2.25bn
Zhipu / Knowledge Atlas Frontier lab · China · public, HK	~HK$586bn ≈ $75bnmid-Jun 2026 · market cap	CNY 724m ≈ $100mFY2025 · listing disclosures	~750x	~$3.75bn

All figures as reported at the dates shown. The multiple on the latest report is largely a product of when that report happened; it is context, not a basis for comparison. Private marks are not market prices; the 20x EV/NTM anchor is a market convention, not a law. Sources 8, 28, 29, 40, 41, 42, 43.

Four things in this table deserve a sentence each. Perplexity, an application company rather than a laboratory, carries almost exactly Mistral's price on almost exactly Mistral's estimated revenue; the market is charging the same for a lab and an app, and at least one of those prices is asking the wrong question. Anysphere, the maker of Cursor, is the only row where real money bought a whole company rather than marking a sliver of one: SpaceX's $60 billion acquisition, agreed as an option and exercised on 16 June 2026, closed at roughly 15 times the latest reported ARR, below the private-round convention, though the option's price was fixed earlier, when revenue was smaller [43]. Set that against Anysphere's trajectory, ARR up roughly forty-fold in the eighteen months to June 2026, and SpaceX appears to have bought its target at a very attractive multiple: on any plausible next twelve months, the price is cheaper still. DeepSeek's row is the purest expression of this article's theme: a $45 billion negotiation over revenue nobody outside has ever seen. And Zhipu's row shows what happens when a public market, rather than a lead investor's convention, sets the number: the Hong Kong crowd pays a multiple that makes every private mark in the table look sober, though note the timing mismatch, a mid-2026 market price against full-year 2025 revenue in a company reportedly growing at triple digits [41][42].

Now invert the question. At €20 billion, roughly $23 billion at the June 2026 exchange rate, what revenue would make Mistral merely fairly priced against its peers? At Anthropic's 21 times, the answer is about $1.1 billion. Read that number twice, because it is doing quiet work: $1.1 billion is almost exactly the revenue target Mistral set itself for the end of 2026 [7]. In other words, the rumoured price pays today for the year-end target, at the multiple of Anthropic's May 2026 round. Handle that benchmark with care, for two reasons. Anthropic's multiple looks low mainly because of timing: its round closed just after its revenue inflection, when the denominator had freshly quintupled [8]. And its ARR arithmetic is reportedly more aggressive than OpenAI's, which flatters the same ratio further [51]. One more caution, this time about the target's vintage: it was given around the turn of the year, before the industry's monetisation inflection [8]. A target set before the inflection may simply be stale. If Mistral is riding the same wave as its peers, its true next twelve months could sit well above the old guidance; if it is not riding it, that would be information too.

That is also why the multiples in the table above should be read gently. A multiple computed on the latest reported revenue depends as much on when the report happened as on what the company is worth: at these growth rates, a figure from six months ago describes a much smaller company than today's, so the same price looks far more expensive against it. Comparing companies on such multiples compares calendars, not businesses. The instrument the market actually appears to use cuts through this: enterprise value against the next twelve months of revenue, with 20 times as the number that keeps reappearing. Watch it work in reverse. OpenAI's jump from $300 billion in March 2025 to $500 billion by October 2025 shocked commentators at the time [38]; at 20 times, the two prices implied $15 billion and $25 billion of forward revenue, and by early 2026 the company's reported run-rate had reached roughly $25 billion [28]. The anticipation was right. Anthropic's $380 billion round of February 2026 implied $19 billion forward at 20 times, at a moment when its last disclosed run-rate was $9 billion: the price assumed a doubling, which sounded bold then and proved conservative within months [39]. In both cases the valuation grew because expected revenue grew, not because the multiple expanded. Two honesty clauses belong here: private rounds report equity value rather than enterprise value, and nobody outside observes a private company's next twelve months, so the 20 times is a convention the market coordinates on, not a law. But conventions are information.

Apply the convention to Mistral and something odd appears. €20 billion at 20 times implies a little over $1 billion of next-twelve-month revenue: the same number the trailing arithmetic already found, and, almost to the digit, the target the company set itself for the end of 2026 [7]. Two different roads arrive at the same place, which strengthens the reading. Priced this way, the rumoured round assumes Mistral merely meets guidance it has already given, with zero anticipation beyond it, in an industry where its American peers have just passed through revenue inflection points with growth accelerating. And if you believed Mistral's next twelve months held $2 billion, the same price would sit between 10 and 12 times: roughly half the convention. Either the market quietly doubts the guidance, or the discount prices something else, the disclosure vacuum itself perhaps, or the tightly controlled capitalisation table. This article does not know which. Neither, from outside, can you.

The Chinese laboratories test the convention from the other side. Moonshot's $20 billion of May 2026, against $200 million of April 2026 ARR, implies a five-fold revenue anticipation at 20 times forward [29]. DeepSeek is reportedly raising at about $45 billion, as of mid-2026, without ever having disclosed revenue at all: at the convention, that price implies $2.25 billion of forward revenue that nobody outside has ever seen, a mystery deeper than Mistral's [40]. And Zhipu and MiniMax, listed in Hong Kong since January 2026, are priced daily by a public market rather than negotiated in a round [9]. Whether the 20 times magnet holds there is settled by trading, not convention: a useful reminder that the multiple is a social fact, not a physical one.

4. TAM and the Path to Profitability

TAM, the total addressable market, is the metric section 3 set aside as barely estimable. How barely? Take the forecasts the industry itself cites. The direct market for AI is projected to roughly quadruple from about $260 billion in 2025 to $1.2 trillion by 2030 [44]. McKinsey has put AI's annual productivity impact at $2.6 to 4.4 trillion and the total addressable market as high as $15 trillion [45]; PwC's much-quoted figure has AI adding $15.7 trillion to global GDP by 2030 [45]; and Goldman Sachs expects token consumption, the raw unit of AI usage, to grow 24-fold through 2030 on the rise of enterprise agents [46]. Notice that these numbers differ by an order of magnitude depending on what is being counted: software sold, infrastructure built, or economic value diffused. The spread is not a flaw in the forecasts; it is the measure of how unformed the market still is. For Mistral, the relevant slice is European and sovereign, and there the estimates are more concrete: Europe's AI market at roughly $86 billion in 2025, projected toward $548 billion by 2032 [47]; European tech spend crossing €1.5 trillion in 2026, with AI, cloud and sovereignty named by Forrester as the drivers [48]; and Gartner projecting European sovereign-cloud spending to more than triple between 2025 and 2027, to $23 billion, with 61 per cent of Western European CIOs telling its November 2025 survey they plan to increase their use of local providers [49]. Set €20 billion against those numbers and the constraint is visibly not the size of the pond. The question the price actually rides on is the one no report can settle: what share, and at what margin.

And profitability, the second half of this section's title? It is the least visible metric in the industry. No frontier laboratory publishes audited accounts; what circulates instead are reported losses that would sink any ordinary company, funded by ever-larger rounds, and profitability milestones that are projections rather than closed periods. The best-documented case makes the point: Anthropic told investors in May 2026 to expect its first profitable quarter, roughly $559 million of operating profit on about $10.9 billion of second-quarter revenue [51]. Even that milestone comes with an asterisk the size of a data centre: critics noted the figure is flattered by a temporary compute discount during the ramp-up of its SpaceX contract, which will cost $1.25 billion a month through 2029, and the company itself has signalled it may not stay profitable through the year as those costs land [51]. Two readings of that number coexist, and both are correct. Five per cent is a thin margin, and it may vanish under the compute bill. Yet breakeven crossings in software are rarely plateaus; they are inflection points that move fast, because once revenue outruns a largely fixed cost base, each additional dollar falls through to profit. Which reading wins depends mostly on the compute contracts, which is precisely the line item outsiders see least. More broadly, the optimist's case for the industry rests on inference costs, which have fallen fast and keep falling; the pessimist's case is the price war, the open-weight wave, DeepSeek above all, repricing the product toward its marginal cost. For Mistral, nothing is disclosed at all, but the revenue mix described in section 2 implies a split structure: licences and on-premise subscriptions carry software margins, while debt-financed Compute carries data-centre margins, thinner and slower to arrive. The honest summary is uncomfortable for every company in the comparison table: on the metric that will ultimately decide each of these valuations, the public record offers projection, briefing, and hope, and the single best-documented profit in the industry arrives with a footnote saying it may already be gone.

The honest caveats take one paragraph. The €20 billion is a reported negotiating position, not a closed round. Run-rate, ARR and revenue are three different objects, and the figures above mix all three because that is all the public record offers. Private marks are not market prices; nobody has to clear a trade at these levels. The currencies wobble. And part of Mistral, the debt-financed Compute business, is capital-intensive infrastructure that should arguably not carry a software multiple at all.

5. The Sovereignty Premium

The arithmetic of section 3 says the price assumes two things at once: the target is met, and the peer multiple holds. Is there any reason Mistral should command more than the peer multiple? There is one candidate argument, and June 2026 made it loudly: sovereignty. The buyers described in section 2 are pricing two distinct risks. Exposure: what a foreign provider can see. And dependence: what a foreign government can take away. Start with exposure, and start with how mundane it is. Everything typed into an AI application, every question, every draft, every uploaded document, passes through the provider's infrastructure in readable form, and the provider's own policies state that conversations can be reviewed and, where its rules require, escalated. No hacking is involved; the access is the product working as designed. Enterprise tiers promise stricter handling, but the promise is contractual, not physical. June 2026 supplied the demonstration: OpenAI flagged a user's conversations to the FBI, which passed them to Brazilian authorities, who arrested a man planning to have his eight-year-old son killed [30]. A life was probably saved, and many will judge that an ethically good way to breach the privacy of one conversation. Now run the same mechanism against different content: a government's classified briefings, a defence contractor's designs, a company's negotiation strategy, or simply the private information that matters most to any of us. You see how easily it can be extracted, and how little would need to change for it to be used against you. That is why many governments, companies and institutions have concluded they cannot take the risk that an American, Chinese or any foreign AI application sits between them and their own secrets.

Exposure is only half of it, and the other layers are just as documented. The statute: under the CLOUD Act and FISA Section 702, a US provider cannot lawfully refuse its own government, whatever its preferences [31]. The directive: on 12 June, Washington ordered access to the most capable American models suspended for foreign nationals, and they went dark worldwide within days, for nineteen days [32]. And Europe has seen where concentrated access leads before: one HSBC insider in Geneva stripped confidentiality from a hundred thousand clients and helped end Swiss banking secrecy, and European states used, and in the German cases paid for, the stolen data [33]. The appetite is documented; AI multiplies the access. That is the premium's case, and the procurement shift of June 2026, from Palantir to ChapsVision, from Azure to Scaleway, a Mistral assistant for every French civil servant, is its early evidence [23].

The premium's limits deserve equal time, and three sentences carry them. Mistral is a French champion more than a European one, and the same month that produced the sovereignty consensus watched France and Germany terminate their joint fighter programme after €3.3 billion of sunk costs, a reminder that Europe co-funds champions more readily than it co-builds them; the ASML anchor investment, a capital alliance rather than a joint venture, is the counter-model that works [34]. Sovereignty itself is layered: what Mistral sells is jurisdictional and operational sovereignty, weights you host on chips you own, and that genuinely protects against the June scenario; supply-chain sovereignty, against the fact that those chips are American designs fabbed in Taiwan with Korean memory, is available to no one at any price [35]. And political consensus has failed to become revenue before: Quaero and Gaia-X were both announced as Europe's answers to something, and neither's revenue ever mattered [36].

What does €20 billion encode, then? The end-2026 target met, at a multiple borrowed from Anthropic's post-inflection round, plus the belief that sovereign demand makes the target durable. Each leg is arguable; none is knowable from outside. Which raises the question the last section owes an answer to: if the decisive facts cannot be known from outside, and some cannot be known from inside either, how much information does an investment decision actually need?

6. How Much Information Does an Investment Decision Need?