The Economics
of the Frontier

Cost one: building the model

Frontier Labs sets out to build a model. Two things go in. The first is people — a few hundred researchers and engineers, paid serious salaries, for many months. The second is compute — an enormous cluster of specialised chips, rented or owned, running one long, uninterrupted training run in which the model learns from a vast amount of text and code.

At the end of that run you have exactly one thing: a single trained model — a large file of numbers that can answer questions. Everything spent getting there was a one-time cost. It is much closer to building a factory, or shooting a film, than to a monthly bill. You pay it all up front, before the model has earned a single cent. Real frontier training runs have been described as costing anywhere from roughly $100 million to well over $1 billion, depending on the model’s size and generation.2 For Frontier Labs, say the all-in build cost — team plus the training run — comes to $1 billion.

Cost two: running the model

Here is the part newcomers miss. Building the model was not the end of the spending. Every time a customer asks the finished model a question, the model has to think — and that thinking runs on chips that draw power and cost money. This running cost is called inference, and unlike the build cost it never stops. It is a small charge, paid again and again, on every answer the model ever gives. A lab therefore carries two quite different costs: one huge bill to build, and a stream of little bills to run.

Three ways to sell it

Now Frontier Labs has a model that works and costs a little to run. It opens three doors to revenue.

The metered API. Other companies’ developers connect their software directly to the model and are billed per token, exactly like that water meter — so much per million words in, more per million words out. Pure pay-as-you-go.

Subscriptions. Individual people pay a flat monthly fee — a $20-a-month plan, a heavier $200-a-month plan — for generous but capped access through an app. The lab is making a bet: that the average subscriber will cost less to serve than they pay, even though the heaviest ones cost far more.

Enterprise bundles. Large organisations negotiate packages — blocks of seats for their staff, higher usage limits, security features, committed-use contracts at a discount. These deals are bigger, stickier, and slower to sign than the other two.

The shape of the whole business

Walk Frontier Labs’ one model from cradle to grave and the shape appears. It cost $1 billion to build. Over the roughly two years before a better model replaces it, suppose it earns $2 billion across those three doors — while the compute to serve all those answers costs, say, $0.6 billion. The model, on its own, cleared its build cost and then some.

Illustrative numbers for an imaginary lab. The pattern that matters — lifetime revenue of roughly twice the build cost — is the one Ledger 3 examines in detail.

So the machine has three parts: a large fixed cost to build, a metered margin on every answer sold, and a clock — because a better model, yours or a rival’s, will make this one obsolete. The whole game is to earn back the build, at the metered margin, before the clock runs out.

That is the entire business in one paragraph. Everything that follows is a consequence of it. Ledger 2 shows how labs report these costs and revenues — and how the reporting can mislead. Ledger 3 takes this exact shape and asks what happens when you run it not once but every year, each time building a model ten times larger. The rest of the booklet follows the same machine outward, into the strange economics it produces at scale.

The headline number is a costume

Start with the most quoted figure in the industry: the run rate. When a lab says it has reached a “$30 billion run rate,” it has not earned $30 billion. It has taken its strongest recent month and multiplied by twelve. Run rate is a forecast wearing the costume of a result. It is not an accounting standard, it has no auditor, and in an industry growing this fast it always flatters — because the month it annualises is, by selection, the best one so far.

Set run rate next to the number an accountant would actually sign: recognised revenue, the GAAP figure, the money genuinely booked over a real trailing year. The two diverge sharply. OpenAI exited 2025 describing a $20 billion run rate; the revenue recognised on its books for that year was roughly $13 billion.4 As a rule of thumb for a company compounding this fast, recognised revenue runs near half the run-rate headline. Neither number is a lie. But they answer different questions — “how fast is the last month?” versus “what did a year actually produce?” — and the gap between them is the first thing a careful reader restores.

Gross, net, and the $8 billion argument

The second ruler is gross versus net. Frontier models are sold not only direct but through hyperscaler marketplaces — Amazon Bedrock, Google Vertex, Azure. When a customer buys Claude through Amazon, a cut stays with Amazon. Book the whole sale and your revenue is gross; book only your share and it is net. The choice can swing a top line by 20 to 40 per cent.

This is not abstract. In April 2026 a leaked internal memo from an OpenAI executive argued that Anthropic’s revenue figure was overstated by roughly $8 billion precisely because Anthropic books reseller sales gross while OpenAI’s comparison was net.3 Sort that out and the two companies’ 2026 revenues land far closer together than the duelling press releases suggest. The lesson is not that one side cheated. It is that two true numbers are not comparable until you know each one’s ruler — and that you, the reader, must do that work because the headline will not.

Committed is not spent

The third ruler separates a commitment from a cost. The frontier deals in numbers like “a $300 billion compute deal” or “$700 billion of capital expenditure.” Almost always these are multi-year commitments — a $300 billion cloud contract is spread across five years, so it is nearer $60 billion a year, and even that arrives only if the buyer keeps consuming. A pledge to build capacity is not cash that has left the building. When a headline collapses a five-year commitment into a single staggering figure, your job is to divide it back out and ask what is contractually firm versus contingent on milestones nobody has hit yet.

Whose dollar is it?

The fourth ruler is the subtlest, and Ledger 6 is devoted to it, so here only the flag. A dollar of revenue is supposed to come from a customer — external demand, someone choosing to pay for a service. But at the frontier some revenue dollars began life as investment dollars: a hyperscaler buys equity in a lab, the lab spends that money on the hyperscaler’s cloud, and it returns as the hyperscaler’s revenue. The dollar is real. Whether it represents real external demand is a different question — and one worth holding open every time you see a growth chart.

Why this is not cynicism

None of this means the frontier labs are running a confidence trick. Run rate is a normal way for fast-growing software companies to talk; gross-versus-net is an ordinary accounting judgement; multi-year commitments are how infrastructure has always been financed. The problem is not deception — it is scale and speed. The numbers are large enough to switch off intuition and they move fast enough that the flattering version is always close at hand. So the discipline of this booklet is small and repeatable: before you repeat any frontier figure, ask which of the four rulers it is on, and convert it into something a human can picture. Do that, and the remaining five ledgers become legible. Skip it, and you are just moving very large numbers around.

A vineyard that bleeds cash

Picture a vineyard. Every vintage it bottles sells out at a healthy profit. And every year it loses money — because every year it plants a new field ten times the size of the last. The books show widening losses. The wine is not the problem. The expansion is.

This is, near enough, the argument Anthropic’s chief executive Dario Amodei has made for why his company can be deeply unprofitable and a sound business at the same time.5 Stop looking at the company’s annual profit-and-loss statement, he says, and look instead at a single model as a product with its own lifetime accounts — its own vintage. His toy version runs like this. A model that cost about $100 million to train in 2023 goes on to earn roughly $200 million serving customers. The 2024 model costs around $1 billion and earns about $2 billion. Each vintage, taken alone, returns about twice its training cost — exactly the shape of the single model in Ledger 1.

Yet in any given year the company is also paying to train the next model — and that one costs roughly ten times more. So while the 2024 vintage is happily earning its $2 billion, the company is spending $10 billion training the 2025 model. The annual profit-and-loss statement nets these against each other and shows an accelerating loss: small, then large, then alarming. The conventional reading — “this company loses more money every year” — is arithmetically true and, Amodei argues, completely misleading. The losses are not a failing product. They are the visible cost of a choice to keep planting bigger fields.

The number under the framing: two costs, not one

That story only works if a vintage really does earn back its build, and to see whether it can you have to separate the two costs from Ledger 1 and keep them apart.

The training cost is the one-time bill to build the model — the $1 billion in our example. It is fixed: it does not change whether the model then serves ten customers or ten million.

The serving cost is what it takes to run the model for one customer’s request — a sliver of compute, paid every time. Whether a vintage ever earns back that fixed training bill depends entirely on the gap between what the lab charges for a request and what the request costs it to serve. That gap is the inference gross margin, and it is the number the whole vintage framing quietly rests on.

Working the margin

So work it. Take a million tokens of output from a frontier model — the meter from Ledger 1, clicked over once. The customer is billed on the order of $25. Producing those tokens — the GPU time, the electricity — costs the lab perhaps $6. The lab keeps the difference.

Illustrative, and consistent with the research firm SemiAnalysis, which puts gross margins on frontier models north of 70%.6

A margin near 75 per cent is what makes the vintage arithmetic possible. The training cost is fixed and paid once; the serving margin is collected on every token, forever. Serve enough tokens at 75 cents of profit on the dollar and the cumulative serving profit climbs past the $1 billion training bill — and then past twice it. The famous “2×” is not magic. It is a high-margin meter, running long enough.

Why 2024 broke the framing — and 2026 mended it

The framing is only available in 2026 because that margin recently flipped. In 2024, Anthropic’s gross margin on paying customers was reported at roughly minus 94 per cent — it spent nearly two dollars serving every dollar it billed.7 With a negative margin, no amount of volume rescues you: every additional token served deepens the loss, and the fixed training bill can never be earned back. That was the “selling dollars for cents” the industry’s critics described, and in 2024 they were right.

Then, through 2025 and into 2026, cheaper hardware and far better serving software pushed that margin from deeply negative to comfortably positive — on SemiAnalysis’s reading, Anthropic’s inference margin moved from the high-30s to above 70 per cent inside a year. Only once a vintage clears a healthy profit on every token does “each model earns twice its cost” stop being a hope and start being arithmetic. The vintage framing did not prove itself. A margin flipped, and the framing became possible.

The interrogation: three ways the framing breaks

The vintage story is elegant, internally consistent, and the most attractive idea in this booklet. That is exactly why it deserves pressure rather than applause. It rests on three assumptions, and each one is contestable.

One: the vintage must get its full earning life. “Earns twice its cost” assumes the model serves customers long enough to collect that revenue. But frontier models are made obsolete fast — sometimes by a competitor, sometimes by the lab’s own successor. If a vintage is retired in months rather than years, it never earns out, and the training cost is stranded. The hidden variable is depreciation: how quickly the expensive asset loses its value. The vintage framing tends to assume a generous earning life. The market has not been generous.

Two: each 10×-bigger model must still earn its keep. The cascade only works if a model that costs ten times more can be sold for enough to justify it. That held while each generation delivered a clear capability jump. The critic Gary Marcus, among others, argues the jumps are shrinking — that scaling has entered diminishing returns, and Amodei’s own remark that the field is “near the end of the exponential” concedes the direction of travel. If a $100 billion model is only modestly better than a $10 billion one, customers will not pay ten times more, and the cascade quietly inverts from a growth engine into a trap.

Three: it is a bet that must be won every single round. Amodei is candid that the whole structure depends on forecasting demand correctly — spend the right share of compute on research, keep a serving margin above 50 per cent, predict demand, and you print money; misjudge demand and the result “could swing wildly.” The critic Ed Zitron puts the bear case more bluntly: the losses are not an investment phase, they are the business model, and a business that must raise ever-larger sums to cover ever-larger losses is one bad forecast from a cliff. Sequoia’s David Cahn framed the same worry at the level of the whole industry as the “$600 billion question”9 — the gap between what the sector has spent on infrastructure and the revenue that spending must eventually justify. The vintage framing is the most coherent answer anyone has given to that question. It is not the same thing as a settled one.

A quiet caveat: the framing depends on a choice

One last thing to carry into Ledger 4. “Each vintage earns twice its cost” is not a fact read off a meter; it depends on how you spread — amortise — a training cost across the years a model earns. Stretch the schedule and the vintage looks profitable sooner; compress it and the same model looks like a loss. OpenAI and Anthropic do not make that choice identically, which is one reason their numbers resist clean comparison. The vintage framing is a genuine insight and an accounting lens — and a lens can be adjusted.

Same costume, different company

Two companies are called frontier AI labs. Both are valued in the hundreds of billions. Both train models near the edge of what is possible. And it has become the most common mistake in the room to price them the same way — because they are, increasingly, two different kinds of company wearing one job title.

Anthropic is turning into an enterprise-software business. Around 80 per cent of its revenue comes from companies, not consumers. The number of customers paying it more than a million dollars a year roughly doubled, to over a thousand, in a single eight-week stretch of 2026. The revenue is concentrated, contracted, and sticky — the texture of a company that sells software to other companies.

OpenAI is turning into something else. It has the most famous consumer product in technology — ChatGPT, with hundreds of millions of weekly users — but it has also pushed outward into video generation with Sora, into a web browser, into a custom-chip programme, and into Stargate, a data-centre build-out measured in gigawatts and hundreds of billions of dollars. Those are not the moves of a company optimising a software margin. They are the moves of a company trying to own infrastructure and consumer distribution. For a business like that, the per-product gross margin is almost a rounding error next to the size and timing of the capital bet.

The eight-times gap

The clearest single number separating the two is revenue per user. Anthropic earns on the order of $211 from each of its (far fewer, mostly business) monthly users. OpenAI earns something closer to $25 per weekly user, the great majority of whom pay nothing at all.10 That is an eight-fold difference in how hard each user is monetised, and it is the whole bifurcation in miniature: one company has a smaller number of users it monetises intensely; the other has a vast number of users it monetises lightly and is betting it can convert attention and scale into something valuable later.

The profitable quarter that proves the point

In May 2026 Anthropic told investors it expected its first-ever operating profit — around $559 million on roughly $10.9 billion of quarterly revenue — in the second quarter of the year.11 It is a real milestone. It is also, by the company’s own admission, a blip: scheduled compute costs in the back half of 2026 are expected to push it back into loss, and Anthropic does not expect a full profitable year before 2028. Ledger 3 already explained why a single green quarter is perfectly consistent with structural loss — it is one vintage maturing in the gap before the next big training bill lands. The profitable quarter does not refute the vintage problem. It illustrates it.

OpenAI’s trajectory is not a blip in either direction. Its internal projections, shared with investors, describe losses that widen for years — and that is consistent with the strategy, not a contradiction of it. If you are building infrastructure and consumer distribution at continental scale, sustained loss is the entry fee. The mistake would be to read OpenAI’s losses as the same kind of fact as Anthropic’s.

Reading the two ledgers side by side

These columns are not strictly like-for-like: the two companies recognise revenue and count users differently, and — as the last row notes — the public record is thinner on Anthropic. Read the table as two profiles, not a scoreboard.

Why this booklet does not crown a winner

It is tempting to end a comparison with a verdict. This one does not, and the refusal is deliberate. “Which lab is winning?” assumes they are running the same race. They are not. One is trying to become the next great enterprise-software company; the other is trying to become something closer to a utility with a consumer brand attached. Each could succeed or fail on its own terms. The useful question is not which is ahead — it is which kind of company is each, and which kind do you want as a supplier, a partner, or an investment. That question you can actually answer. A winner you cannot, and anyone who offers you one is selling a ruler they have not shown you.

Tokens as the electricity of the future

Step back from any single lab and ask what the whole industry is actually racing toward. The clearest answer: intelligence is on its way to becoming a metered utility. Ledger 1 already showed the shape — tokens sold through a meter, like water, like power. Carry that forward. Once electricity became a utility, you stopped caring which generator produced the current in your wall; you simply paid per kilowatt-hour and built your life on the assumption it would be there. Tokens are heading the same way: a metered unit of cognitive work, always on, drawn on without a thought.

If that is the destination, then the frontier labs are not, in the end, software companies. They are utilities in the making — and the contest between them is a contest to become the utility company of intelligence, the master vendor of the electricity of the future.

The datacenter is a dark factory

Where is this utility generated? In a building that deserves its industrial name. A “dark factory” is a real term from manufacturing: a plant so fully automated that the lights can be switched off, because no human needs to be on the floor. A modern AI datacenter is exactly that — a vast hall of humming machines, drawing the power of a small city, with almost nobody inside. The difference is only in the product. A dark factory in the old economy stamped out car parts. This one manufactures intelligence itself, by the trillion tokens, and ships it down a wire.

Why nobody can stop spending

This reframes the enormous, relentless capital spending the other ledgers keep brushing against. If the prize is to be the utility — not a utility — then second place is worth a small fraction of first. A land grab for the foundational layer cannot be conducted at half speed: pause, and a rival lays claim to the ground. That is why capex does not respond to losses the way an ordinary company’s would; why the vintage treadmill of Ledger 3 keeps accelerating; why “we could be profitable if we stopped” is technically true and strategically irrelevant. The labs are not behaving like businesses optimising a quarter. They are behaving like contenders in a once-only race to own an entire layer of the future economy.

Chasing verticals to fund the race

And here is the consequence that explains so much of the labs’ restless behaviour. The race to be the utility costs more than the utility business itself yet earns. No lab can fund the journey on raw token sales alone. So whenever a vertical forms — a specific, high-value use of the models that throws off serious cash — the labs go after it, hard and fast.

The clearest example so far is augmented coding. Anthropic’s Claude Code went from public launch to a $1 billion revenue run rate in six months, and past $2.5 billion within a year; OpenAI’s rival, Codex, climbed to several million weekly users in the same window.13 Coding is doing double duty for the labs: it is proof that the technology produces real economic value, and it is an engine — a vertical lucrative enough to help finance the dark factories. Expect the pattern to repeat wherever a vertical’s economics are good enough — customer support, legal work, search, scientific research. When the labs see “ridiculous revenues” forming in a vertical, they will move, because the race needs the fuel.

Utility and product at once

This resolves a contradiction that otherwise makes the labs look incoherent. Are they horizontal infrastructure — utilities selling raw intelligence to everyone — or vertical product companies, building coding tools and browsers and apps? The honest answer is that the hypercompetition forces them to be both, at the same time, by necessity. The utility is the destination. The verticals are the fuel that pays for the trip. A lab that sold only raw tokens could not afford the race; a lab that sold only coding tools would be a feature, not a foundation. So they sprint in two directions at once — and that is why “what is an AI lab, exactly?” is a question that refuses to sit still.

Verticals are one source of fuel for the race. They are not the only one. The next ledger follows the other — the capital that arrives not from customers at all, but from the labs’ own investors, moving in a circle.

Follow one dollar

Amazon wires a dollar of equity investment to Anthropic. Anthropic wires that dollar to Amazon Web Services to pay for computing power. AWS books it as cloud revenue. The dollar has not vanished and nothing improper has happened — but it is now doing double duty: it is counted as Amazon’s investment and as Amazon’s revenue, and it is propping up two growth stories at once. It is the same dollar.

That little circuit, repeated at enormous scale, is the most misunderstood feature of frontier-AI finance. The honest question is not “is this a scandal” — mostly it is not — but “how much of the growth I am looking at is the open market choosing to pay, and how much is an investor’s own money making a lap?”

The deals that fill the ring

Five partnerships fill the loop. In each, an equity investment flows toward a lab and a far larger compute commitment flows back — the round trip the diagram above describes.

Figures as reported, May 2026 — see notes 14–18. Equity flows one way; a much larger compute commitment flows back.

Two of these deserve a closer look, because they are where the loop tightens hardest.

Nvidia and OpenAI. Nvidia announced an investment of up to $100 billion in OpenAI, disbursed in stages as data-centre capacity comes online — capacity OpenAI builds substantially by buying Nvidia’s own chips.17 The chip vendor funds the customer to buy the vendor’s chips. Oracle and OpenAI. Oracle is not an equity investor, but it sits on the same circle: its reported five-year, roughly $300 billion Stargate compute deal with OpenAI18 is a commitment large enough to reshape Oracle’s own financial statements.

How the loop inflates valuations

The circle does more than move cash. It moves valuations. When Nvidia commits capital to OpenAI and OpenAI orders Nvidia chips, Nvidia’s revenue and order backlog swell — and a company with a swelling backlog is worth more, so its market value rises. When Oracle signs the Stargate deal, its contracted backlog leaps; by its own disclosures, Oracle’s remaining performance obligations jumped past half a trillion dollars, and its share price moved with the news. Each deal, in other words, re-rates the worth of more than one company on the circle at the same time.

Stack enough of these and an effect appears that no single deal intends: the whole sector looks larger, busier, and more certain than independently verifiable end-customer demand alone would justify. Valuations lift each other. The capital, the revenue, and the market value all swell together — partly on real demand, and partly because the same money has been counted, in good faith, at several stops around the ring.

The fair rebuttal

Before the caveats, the case for all this, stated properly. A strategic investor seeding a company that will become a major customer is ordinary capitalism, not a trick — telecoms, airlines, and the early cloud were all built this way. The Azure compute OpenAI buys is genuine: real models run for real users on those servers. Microsoft is not paying itself; it is selling a service a willing buyer would want regardless. And there is a sturdier point: the labs are not only spending investor money. A great deal of their revenue is now ordinary customers — enterprises and consumers with no equity stake — paying for a product they chose. The loop is real, but it is not the whole circuit.

Three caveats to carry out of this ledger

So describe the loop soberly and land on three things, no more.

One — endogenous growth. When a portion of a lab’s revenue traces back to its own investors’ capital, the growth curve is partly endogenous: funded from inside the arrangement rather than won in the open market. It is not fake. It is just not the same evidence of demand as a dollar from an unrelated customer, and a revenue chart cannot tell you the mix on its own.

Two — concentration. The loop binds a handful of firms tightly together. If one major lab falters, the damage does not stay contained — it lands on a hyperscaler’s investment line, its revenue line, and its contracted backlog at once. The circle that amplifies growth on the way up amplifies stress on the way down.

Three — opacity. Because of the gross-versus-net problem from Ledger 2, an outside reader cannot reliably size the loop. You can see that it exists and that it is large. You cannot, from the public record, say precisely how much of any headline number it accounts for. Honesty about that limit is part of reading the books well.

If you only watch the two leaders, you mistake the game

Ledgers 3 to 6 examined the American duopoly. But the most instructive economics at the frontier are happening at its edges — where a well-run European lab is losing the race anyway, and where another set of labs has decided that the race the Americans are running was never the point.

Mistral: credible, and structurally behind

Mistral is France’s entrant and Europe’s best one. Its revenue grew impressively in relative terms — from roughly $16 million at the end of 2024 to an estimated $400 million annualised a year later. Set that beside Anthropic’s $30 billion and the relative success disappears: Mistral is on the order of 75 times smaller, and because the leaders are growing faster from a vastly larger base, the gap is widening, not closing.

This is not a talent problem. Mistral ships genuinely good models. It is a capital problem. Its September 2025 Series C raised €1.7 billion, led by the Dutch chip-equipment maker ASML, which became its largest shareholder; in early 2026 it arranged an $830 million debt facility to buy some 13,800 Nvidia chips for a data centre near Paris.19 Those are serious European numbers. They are also roughly two orders of magnitude below the compute commitments Ledger 6 described on the American side. Frontier-tier monetisation is gated by frontier-tier compute, and frontier-tier compute is gated by capital Mistral cannot raise at American scale.

Mistral’s response is to stop competing on the duopoly’s terms: less pure model-API rivalry, more sovereign and on-premise offerings, vertical models for regulated European industries, and the pitch that a European customer may want a European supplier for reasons that are not only technical. That is a sane strategy. It is also, plainly, the strategy of a company that has accepted it will not win the raw-scale race.

China: the model is the bait, the ecosystem is the business

The Chinese labs invite a Western misreading. DeepSeek reported training its V3 model for a marginal cost near $5.6 million — a figure that, taken alone, made the American spend look absurd.20 The honest accounting is less dramatic: that number excludes prior research, failed runs, and the cost of the hardware itself, and the all-in figure is many times higher. But chasing the exact training cost misses the larger point — which is that for the leading Chinese players, the model was never the product.

So what is? The product is the ecosystem, and the ecosystem makes money the way it always has. Alibaba’s Qwen models feed Taobao and its cloud: the money is made on a cut of every commerce transaction and on cloud contracts. ByteDance’s models feed its apps: the money is made on advertising sold against attention. Across the sector the pattern repeats — super-app engagement, payment fees, subscriptions to services that have nothing to do with an API meter. In that design a frontier-grade model given away near-free is not a failed business; it is a customer-acquisition cost for a far larger one. The model is the bait. The ecosystem is the hook, and the hook is where the revenue has always been.

Two of these companies, Zhipu and MiniMax, listed in Hong Kong in January 2026 and saw their valuations multiply within weeks.21 The public market is funding the strategy — and profitability on inference is simply not the case being made to those investors. “Chinese labs can’t make money selling tokens” is true and almost irrelevant: selling tokens was never the plan.

The commodity tier eats the floor

There is a third force, and it presses on everyone. Below the frontier sits a fast-improving commodity tier — cheap open-weight models, many of them Chinese, and discount API offerings. For tasks that do not need frontier capability, these are nearly free. SemiAnalysis’s reading is stark: margins on frontier models sit north of 70 per cent, while margins on trailing models exposed to open-source competition fall below 20.

From the buyer’s side — the subject of the companion booklet, The Token Economics — this is simply good news: cheaper tokens. From the seller’s side, the one this booklet is about, it is a slow squeeze. Every routine workload that drains from a premium API to a near-free open-weight model erodes the pricing power the whole vintage cascade of Ledger 3 depends on. The frontier labs are protected only at the genuine frontier — and only for as long as the frontier stays far enough ahead of the commodity tier to be worth paying for.

Closing the books

Seven ledgers, and a single habit holding them together. A frontier number means nothing until you know which ruler it is on, whose dollar it really is, and what kind of company — or country, or business model — produced it. The figures in this booklet will be stale within months; the way of reading them will not. When the next staggering headline arrives, do what every ledger here has done: find the unit, translate it into something a person can picture, and ask what kind of bet it really represents. That habit, not any number, is what this booklet was for.