Forty-six years after Desert One, business still refuses to run the one calculation that matters.
In April 1980, the United States launched eight helicopters into Iran on a mission that required six, believing two spares constituted a margin of safety. The binomial math — never computed by anyone in the chain of command — says the mission had roughly a one-in-three chance of reaching its start line. Eight men died, and the failure reshaped the American military.
The lesson of Operation Eagle Claw is not a military lesson. It is a lesson about a specific, teachable analytical failure: substituting a point estimate and a gut-feel margin for a probability distribution and a confidence threshold. The planners knew the requirement (six helicopters). They knew the fleet had failures (that’s why they added spares). What they never did was state the confidence level the mission demanded and compute the fleet size that achieved it. Had they done so, the answer — seventeen aircraft — would have forced a redesign of the entire operation.
Even the Pentagon’s own Holloway Commission, which reviewed the failure and concluded ten to twelve helicopters should have launched, understated the problem — because it sized the fleet against the first leg of the mission, as if the requirement ended at Desert One. It did not. The demonstrated failure rate that night — three of eight aircraft lost before a single hostage was approached — was generated by the easiest portion of the profile. The six surviving helicopters still had to sit through a full day parked in talcum-fine desert dust on the outskirts of Tehran, extract the hostages from a soccer stadium under fire from students and Iranian military, and survive the flight out. Every one of those phases carried its own probability of mechanical failure or combat loss, and none of them appears in the Holloway arithmetic. Price the full mission profile honestly and seventeen is not the ceiling of the estimate. It is the floor.
I call this the eight-helicopter problem: committing capital or lives to a plan that requires N things to go right, without ever computing the probability that N things go right. And nearly half a century after Desert One, it remains one of the most common — and most expensive — failures in corporate strategy. Two current examples, one already concluded and one still unfolding, show the pattern operating at billion-dollar and trillion-dollar scale.
Kroger and Ocado: A $2.6 Billion Impairment That a Spreadsheet Could Have Prevented
In 2018, Kroger partnered with Ocado, the U.K. automation company, to build a national network of automated customer fulfillment centers…massive robotic sheds, some approaching 375,000 square feet, designed to assemble online grocery orders with swarming robots on high-density grids. The ambition at launch: as many as 20 CFCs across America. Kroger built eight.
In November 2025, based on independent reporting, the bill came due. Kroger announced it would close three of the eight located in Pleasant Prairie, Wisconsin; Frederick, Maryland; and Groveland, Florida, taking a $2.6B impairment charge. In December it canceled the planned Charlotte CFC outright and agreed to pay Ocado $350 million to walk away, while shuttering spoke facilities from Nashville to Florida and laying off well over a thousand workers. Ocado’s stock fell to levels last seen at its IPO fifteen years earlier. Kroger’s own explanation was devastating in its simplicity: after a “full site-by-site analysis,” three of the CFCs did not meet the demand-density thresholds required to justify their fixed costs.
Read that sentence again. The site-by-site demand analysis that killed the facilities in 2025 is the same analysis that should have preceded the groundbreaking in 2019. Nothing Kroger learned was unknowable in advance. It was simply never computed to a confidence standard. What makes this so damming to Kroger is that they own a company called 84.51, staffed with mathematicians and statisticians…none of them did the math.
Here is the calculation that was owed to Kroger’s shareholders before a single shovel of dirt was turned, for every proposed CFC site:
The requirement. Each CFC carries enormous fixed costs: the building, the Ocado technology fees, the robotic grid, the delivery fleet. Those costs imply a breakeven volume: some number of orders per week below which the facility loses money no matter how efficiently the robots swarm. That number was known, or knowable, to the dollar.
The distribution. Weekly order volume in a given catchment is not a number; it is a probability distribution driven by e-grocery adoption rates, competitive intensity, delivery-speed expectations, and critically for Kroger, brand presence. A CFC is not a store; it cannot create demand, only serve it.
The confidence threshold. The question that was never asked: What is the probability that demand in this catchment reaches breakeven utilization within three years, and is that probability at least 75 percent or higher? Not the base case. Not the board-deck forecast. The probability.
Run that math honestly and the portfolio sorts itself. The CFCs Kroger is keeping in Ohio, Texas, Georgia, Colorado, and Michigan, sit in markets dense with Kroger-banner stores, where the demand distribution was narrow and its center was high. The CFCs Kroger is closing were built in Florida and the Mid-Atlantic, markets where Kroger had thin or no store presence and the demand distribution was wide, flat, and centered too low. The company built its most speculative facilities precisely where its demand uncertainty was greatest…this is the the statistical equivalent of launching eight helicopters and calling two of them a margin.
There is a second, crueler parallel to Desert One. A CFC, like the hostage rescue, is a threshold system: it does not degrade gracefully below breakeven. At 95 percent of required volume, the mission still fails; the fixed costs still crush the P&L. When your plan has a hard minimum, “close” is worth nothing which is exactly why the confidence level, not the expected value, is the number that matters. Kroger managed to the expected value. The distribution collected $2.6B.
An honest concession: in 2018, Kroger was staring at Amazon’s acquisition of Whole Foods and a consensus forecast that online grocery would swallow the industry. The pandemic then produced two years of demand data that validated the bet before reverting and stranding it. But that concession sharpens the argument rather than blunting it. High uncertainty is precisely when probability-of-sufficiency analysis matters most, a wide distribution is an argument for staging capital and buying information, not for committing eight facilities’ worth of fixed costs against a point forecast at the distribution’s optimistic edge.
The Data Center Buildout: The Eight-Helicopter Problem at Trillion-Dollar Scale
Which brings us to the largest capital deployment in corporate history, happening right now.
Based on independent reporting, the four largest hyperscalers — Amazon, Microsoft, Alphabet, and Meta — plan to spend on the order of $700B on capital expenditures in 2026, up roughly 77 percent from an already record 2025, with the overwhelming majority directed at AI data centers, GPUs, and power. Multi-year projections run into the trillions of dollars through 2030. Amazon alone is guiding toward $200B in a single year…capital intensity at levels no industrial enterprise has ever sustained.
I am not predicting this ends like Desert One. I am observing that, as far as any public disclosure reveals, nobody building it has published the Desert One math. Three distributions matter, and all three are being discussed, when they are discussed at all, as point estimates:
The demand distribution. The entire pure-play AI vendor economy, the primary tenant class for this infrastructure, generates revenue that amounts to a single-digit percentage of one year’s buildout cost. The reported revenue of the largest AI model company represents roughly 3 percent of projected 2026 hyperscaler capex. The buildout is therefore a bet not on current demand but on a demand distribution five to ten years out. What is the probability, not the forecast, the probability that AI revenue reaches the level required to earn a return on trillions of dollars of deployed capital? No hyperscaler has shown that number.
The asset-life distribution. A grocery CFC depreciates over decades. A GPU cluster is economically obsolete in roughly three to six years, and where the true number falls within that range is itself a distribution, driven by the pace of chip improvement. If useful life lands at the short end, the annual revenue required to clear the cost of capital doesn’t rise incrementally; it balloons. Fleet sizing against MTBF and fleet sizing against depreciation schedules are the same calculation wearing different clothes.
The correlation problem. This is the haboob. At Desert One, the dust cloud was dangerous because it threatened every helicopter simultaneously, a correlated failure mode that redundancy cannot fix. Every hyperscaler is making the same bet, on the same demand curve, with the same asset class, on the same timeline. If AI demand disappoints, it will not disappoint one builder; it will disappoint all of them at once, and the “spare capacity” across the industry becomes a synchronized glut. Independent redundancy protects you. Correlated redundancy is just a bigger bet.
The steelman deserves its due, and it is stronger than Kroger’s was. Compute is fungible in a way a Florida grocery shed never can be, an AI data center that misses on model-training demand can serve inference, cloud workloads, or enterprise hosting, which widens the effective demand distribution considerably. The hyperscalers report being supply-constrained today, and unmet present demand is real evidence, not hope. And for a trillion-dollar company, some of this spending is rationally priced as an option on the biggest technology transition in a generation, where the cost of underbuilding may genuinely exceed the cost of overbuilding.
All of that may be true. None of it is a substitute for the calculation. “Supply-constrained today” is a data point on the current draw from the distribution, not a probability statement about 2030 — every fiber company in 1999 was supply-constrained too. The option-value argument is precisely the kind of claim that should be computed: at what utilization, at what revenue per gigawatt, over what asset life, does this buildout clear the cost of capital with 75 percent probability? If the hyperscalers have run that math, they have not shown it. If they have not run it, they are launching eight helicopters with better branding.
The Discipline: Five Steps That Fit on One Page
The remedy has not changed since 1980, and it fits on a single page:
State the requirement as a threshold, not a target. Six helicopters. Breakeven orders per week. Revenue per gigawatt above cost of capital. If the plan has a hard minimum, name it.
Price the per-unit probability from data, not conviction. The RH-53D’s maintenance record existed; nobody pulled it. E-grocery adoption curves by market density existed; Kroger built anyway. Historical technology-adoption distributions exist today. Use the demonstrated rate, not the aspired one.
Set the confidence level before you see the answer. A mission carrying hostages’ lives demands 99 percent. A billion-dollar facility deserves at least 75. Choosing the confidence level after computing the result is how base cases become business cases.
Compute the probability of sufficiency — and believe it. Not the expected value. Not the base case. The probability that the threshold is met. If the answer says seventeen helicopters and the deck holds twelve, the answer is not “launch twelve and hope.”
Write the kill conditions in advance. Eagle Claw got this one right…fewer than six flyable meant abort, decided months early, honored under crushing pressure. Kroger’s kill condition arrived seven years and $2.6B late. The data center builders should publish theirs now: the utilization rate, the revenue-per-gigawatt, the date, below which the buildout stops.
Statistics is the cheapest tool in the strategy toolkit. It requires no consultants, no software, no capital, only the willingness to state a requirement, price a probability honestly, and accept an answer you did not want. The planners of Eagle Claw were not stupid, and neither are the executives at Kroger or the hyperscalers. They are something more common: confident. And confidence, unpriced, is how eight helicopters keep getting launched in the desert, in the grocery aisle, and now across a trillion dollars of silicon and concrete.
The math was available in 1980. It is available now. The only question, every time, is whether anyone in the room insists on running it.