Table of Contents
Around 550 BC, Croesus of Lydia, the king whose name still means rich, faced the ancient version of a capital allocation question, whether to invade Persia. Before committing, he did what boards still do and commissioned an outside opinion from the most prestigious advisory house available. His envoys hauled a fortune to Delphi. Herodotus itemizes it like an invoice – 117 ingots, four of refined gold and the rest a gold-silver alloy, a golden lion of ten talents, great bowls of gold and silver – and then records the question. The answer came back: if Croesus marched against Persia, he would destroy a great empire.
He marched, and he destroyed a great empire. His own. Delphi kept the fee, kept the reputation, and kept a defense that has aged better than the kingdom did.
I recently spent an afternoon taking apart a McKinsey slide that projects $400 billion to $600 billion of quantum computing value in financial services by 2035, and the deeper I got, the less the story was about McKinsey. The demand for a number that makes the unknowable feel decidable is ancient, human, and close to universal. That part is anthropology, and there’s no shame in it. The newer part should bother us more: we exempt consulting forecasts, the modern suppliers of that comfort, from every audit we impose on any other number in the building.
What an afternoon with a calculator found
The teardown is published separately, so one paragraph will do here. Multiply the 2035 baselines on McKinsey’s own slide by its own impact percentages and you reproduce the chart exactly; the affected-share assumptions printed beside them never enter the arithmetic, and applying them cuts the total to $175–270 billion at most. The estimate merges quantum computing’s contribution with AI’s and never allocates between them.
And the range has an edition history. It stood at $394–700 billion in the 2023 Monitor and at $400–600 billion since June 2025, a $100 billion trim at the top that no edition explains. For precedent, look to the metaverse, which McKinsey valued at up to $5 trillion by 2030 in June 2022, an estimate that was never reconciled or revised in public – simply left to age in place once sentiment turned.
So this isn’t the sequel where I keep swinging at one firm. Bankers, journalists, and conference speakers carried the number through decks and articles before anyone, so far as I can find, multiplied it, and the interesting question is why. Why did we need it so badly that nobody checked? My answer, and the rest of this piece defends it: boards buy these numbers as licenses to act, and the information content is close to incidental.
What the number is actually for
In 1921, Frank Knight split not-knowing in two. Risk is the measurable half, the kind an actuary can price, and uncertainty is the half no measurement reaches. Insurance exists for the first category, and no actuarial instrument exists for the second, which is why the second frightens us in a way the first doesn’t. In 1961, Daniel Ellsberg posed the choices that made the fear measurable. Offered a gamble with known odds and an identical gamble with unknown odds, people prefer the known odds so consistently that no single assignment of probabilities can explain their choices.
Whether a cryptographically relevant quantum computer arrives in 2030 or 2040, and what it earns or destroys when it does, is uncertainty in Knight’s strict sense, with no stable distribution anyone can price. And a board can’t act on dread; dread has no cell format. A range like $400–600 billion by 2035 is the conversion device. With it, an executive can turn an unmeasurable question into something a spreadsheet accepts, something a committee can vote on, something a strategy document can cite.
A board meeting is a machine that accepts particular shapes of input, and a judgment isn’t one of them. An agenda line has room for a figure, a range, an owner, and a date, and for nothing else. The relief in the room when the figure finally exists is real, and I’ve watched it arrive.
None of this is fraud. It’s what our species does when it must move before it can know, and the quantum snake oil peddlers I catalog elsewhere exploit the reflex rather than create it. Ban every exaggerator from the market tomorrow and the demand remains, because the demand comes from us.
The oldest service industry
Croesus wasn’t buying an innovation, because paid divination is among the oldest documented professions we have. Babylonian kings kept the bārû at court, liver-readers who examined the entrails of sacrificed sheep before major campaigns and answered in writing. Chinese courts employed astronomer-astrologers for three thousand years.
Delphi ran a paid consultation practice with a priority tier, charging a standing fee called the pelanos and granting promanteia, first place in the queue, to favored cities and individuals. After Croesus’s gold arrived, the Delphians voted him and all Lydians promanteia and exemption from the fee, which is as clean a loyalty program as antiquity records.
In each case the product was more than prediction. Kings asked about the future, and what the ritual reliably delivered was permission, the entitlement to act despite not knowing. The ritual let the king move the unbearable part of the decision, the commitment under uncertainty, onto an institution built to hold it. The god could carry what the king couldn’t.
And the famous ambiguity, whether engineered or selected for by the retelling, insulated the oracle from failure. The prophecy named no empire, so a great one would fall either way and the oracle would be right either way. Twenty-five centuries later, the standard caveat, that a value estimate is approximate rather than definitive, is the same engineering with a slide template. A headline built this way commits its author to nothing that the fine print can’t take back.
Why it has to be a number
The oracle spoke in verse, the modern version speaks in ranges. Theodore Porter answered this in Trust in Numbers, still the best account of why. Quantification, he argues, is a ‘technology of distance’: in institutions too large for personal trust, numbers become the only currency of credibility that moves between strangers. A senior banker can trust a colleague’s judgment because she has watched it perform for a decade.
The board can’t, the regulator can’t, and the investor on another continent has never met anyone involved. What all of them can accept is a figure with a brand attached.
People can’t forward an argument intact. Whoever heard it firsthand passes on a summary, the next person a bullet, and by the third meeting the reasoning is gone and only the conclusion remains, unsupported. People forward a number without loss, and the tenth deck shows the same $400–600 billion, per McKinsey, as the first.
There’s a rank mechanic too. The manager who says “I believe” speaks with one person’s authority. The manager who says “McKinsey estimates” borrows an institution’s, and the borrowed rank outranks everyone physically in the room. Contesting it costs an afternoon of verification that nobody present has done. I know the price because I paid it, and an afternoon turned out to be exactly what it cost.
Wrong together
Keynes named the career logic in 1936: it is “better for reputation to fail conventionally than to succeed unconventionally.” A decision that fails alongside everyone else’s reads as bad luck. One that fails alone reads as folly, and folly is what ends careers. Citing the consensus number therefore buys insurance. If quantum disappoints, the executive who quoted McKinsey wasn’t wrong. Ten thousand decks were wrong together, and no one is fired in a crowd.
Thomas Schelling supplies the other half. Industries need focal points, shared reference values that let thousands of independent actors plan as if coordinated, and a focal point’s job is to be commonly known, not correct. An industry that must plan together against an unknowable date will pay for a shared fiction before it pays for a private truth.
Once every bank knows every other bank has seen the same slide, banks organize budgets, vendor pitches, and analyst questions around the figure whether or not anyone believes it.
I’ve seen the same coordination effect with other numbers, including the time a threat-timeline tool’s cautious default output, a Q-Day estimate of 2051, showed up in a Fortune Global 500 board-level session as if it were a finding. And the mirror image arrived in January 2025, when one dismissive sentence from a chip-industry CEO erased more than $8 billion of quantum market value in a single session. No model accompanied the sentence, and the market didn’t ask for one. It needed a verdict and took the loudest on offer.
The critic is anchored too
Amos Tversky and Daniel Kahneman showed in 1974 that anchors bias judgment even when everyone involved knows the anchor is arbitrary. In the famous version, a rigged wheel of fortune shifted people’s estimates of the African share of United Nations membership by twenty points. The anchor’s obvious arbitrariness didn’t blunt the effect, which brings me to an admission about my own work. My recomputation of McKinsey’s slide, the $175–270 billion, is denominated entirely in their currency, and it exists because their number exists. I can win the argument on every technical point, and the anchor still sets the field on which the argument happens. The debate becomes which hundreds of billions, when the honest answer might be that no defensible figure exists at any size.
(Thirty years in and around advisory work shows both sides of that counter. I’ve been asked for the one number more times than I can count, and I’ve watched relief cross a boardroom when somebody supplies it, and I’ve been the supplier. The pull is real, and it doesn’t care which side of the table anyone sits on.)
Croesus tested his oracle first
Herodotus preserves an interesting detail. Before trusting Delphi with the invasion, Croesus ran a validation exercise. He sent envoys to the seven leading oracles of the age with one instruction. On the hundredth day after departure, each oracle was to be asked what the king of Lydia was doing at that hour.
Then he did something no envoy could report by guesswork, chopping up a tortoise and a lamb and boiling them together in a covered bronze cauldron. Delphi answered in verse, and the verse described the tortoise, the lamb, and the bronze. The other six failed or scraped by, and Croesus took his gold, and his real question, to the two he judged to have passed.
Croesus did what a modern procurement team would call vendor validation, and he did it well, but his error was one level up. He validated the oracle on a checkable question, the kind an answer can be scored on within a season, and then spent the earned trust on an uncheckable one, the kind no score would ever arrive for. The trust was earned and real, but it wasn’t transferable, and nothing in the ritual told him so.
We run the same leak at industrial scale. McKinsey earned its reputation in feedback-rich work like operations, procurement, and post-merger integration, where mistakes show up in the numbers within quarters and clients renew or don’t.
Philosophers call our reliance on specialists epistemic dependence, and it isn’t a defect. Nobody re-derives medicine or metallurgy from first principles, and John Hardwig made the case decades ago that deferring to experts is how knowledge functions at all. The defect is in our calibration machinery, which assumes that somewhere, somebody audits the expert. For the tortoise-and-cauldron questions, somebody does, and for the 2035 questions, nobody ever has.
The last unregulated forecast
Almost every other profession that sells the future to serious money picked up a scorekeeper. Sell-side analysts certify under SEC rules that their published views are genuinely their own, and commercial scorers rank their accuracy publicly, year after year. Public-company auditors answer to a board created by Sarbanes-Oxley in 2002 and overseen by the SEC.
Credit rating agencies, whose ratings the official crisis inquiry called key enablers of the financial meltdown, now answer to the SEC’s Office of Credit Ratings, which Dodd-Frank ordered built for them.
Weather forecasters have been scored for calibration with Brier’s method, which dates to 1950, and the habit of scoring is one reason the forecast on your phone became reliable enough to take for granted. And Philip Tetlock’s forecasting tournaments showed that predictive skill can be measured, and that some forecasters beat the rest consistently once someone keeps score.
The regulated three came to their scorekeepers through scandal. Analysts got theirs after the dot-com research settlements, auditors after Enron, raters after 2008. Weather verification and forecasting tournaments took the other route, through research and professional pride, and either route ends in a trail somebody can check. Value-at-stake forecasting has never had its scandal and has never built the trail, because a wrong decade-out estimate ruins no identifiable victim on any identifiable day. The cost arrives as misallocated budgets, deferred migrations, and strategy premised on a figure nobody can point to in court.
Value-at-stake estimation is the exception, and I mean the genre rather than one firm. Consulting forecasts from Gartner, IDC, BCG, Bain, and McKinsey headline markets, and no accuracy ranking, regulator, or public reconciliation follows. Consulting is hardly lawless, since contracts, securities rules, and professional liability all still apply, but nothing anywhere scores the forecasts. The genre pairs research-grade prestige with marketing-grade accountability, and its unfalsifiability is built into the format rather than stumbled into.
Forecast horizons of a decade mean no estimate gets scored inside anyone’s tenure. Annual editions replace their predecessors instead of reconciling with them. The caveats double as the retreat, available when challenged and invisible in the headline. A $100 billion revision between editions, the kind McKinsey’s own finance estimate went through between 2023 and 2025, would come with a documented reconciliation in any controlled forecasting process. Here it came with nothing, because no office receives the correction and no client asks for one.
What an honest number looks like
Honest numbers about the future exist, and I stake my own work on the difference. Five properties separate estimation from divination. The model is disclosed, so a reader can rerun it. The assumptions are auditable, each carrying a source a skeptic can check. The editions reconcile, so a changed number arrives with the reason it changed. The resolution is defined in advance, so everyone knows what outcome would count as wrong, and when. And the maker is exposed: reachable, on the record, correctable in public. An estimate with those properties can embarrass its maker, and the possibility of embarrassment is the feature, because it’s what makes the estimate informative.
Judged that way, a fair amount of futures work qualifies. A regulator’s migration deadline is a decision rule rather than a forecast, auditable to the day, which is exactly why a budget can be built against it and why I’ve argued for years that PQC budgets should run on regulators’ deadlines rather than on Q-Day predictions.
My own Quantum Utility Map asks the feasibility question before the value question – what they will be able to do in finance, checked against published resource estimates with DOIs – so that every claim in it can be attacked by anyone with a library card. And the five questions I put to McKinsey at the end of the teardown are a small experiment in manufacturing the accountability the genre lacks. The questions are specific, answerable, and public, and their answer will be data, and their silence will be data too.
The same test runs against me, and it should. Every recomputation in the teardown can be redone with a pocket calculator, every resource estimate resolves to a paper, and corrections on this site are dated, additive, and permanent. If my arithmetic is wrong, it will be wrong in public within a week of someone checking. That exposure is the whole difference. An estimate that can be wrong is information. An estimate that can’t be wrong is comfort.
The oracle’s defense
In Herodotus’s telling, Croesus survived the fall of Sardis, and what he did next completes the story. He sent his chains to Delphi with a bitter question, whether ingratitude was the custom of Greek gods. The Pythia’s reply has outlived almost everything else in the story. The god answered truly, she said, and a great empire did fall. If Croesus wanted to know which one, he should have sent again and asked. The failure was the client’s, for not asking the follow-up question.
Twenty-five centuries before a footnote first said approximate rather than definitive, the genre’s defense already existed in final form. The caveats were available, and you chose the headline. I read the Pythia’s answer the way I’d read a modern PR statement, which is what it is, and I note that it worked. Delphi stayed in business for roughly nine centuries more.
So I won’t end by proposing a regulator for consulting forecasts, because the demand side is the side we control. The next time a branded number arrives in your deck offering relief, recognize the relief as the product.
Then go one step past Croesus, whose test was sound and whose error came after it. Test the forecast itself, the checkable thing on the table. Ask for the model, rerun the arithmetic, trace an assumption to its source, and extend trust no further than the checks. It took me an afternoon and a calculator.
