Cross-subsidy: which part of the business should pay for the rest

The finding is in the records that failed

Twenty-two companies in this corpus made a decision about which part of the business would pay for the rest. Fourteen of the twenty settled decisions were never reversed. That rate is not the interesting number.

The interesting number is eight. Eight further attempts to record a cross-subsidy were rejected before they reached the corpus, every one for the same reason: the story described a company that had a subsidy and never described a company that decided to have one. No other decision type on this site fails that way. A subsidy is usually a condition — one unit earns, another spends, and the arithmetic has been true for years before anyone writes it down.

That is the practitioner's problem in one line. The subsidies that held were built. The ones that failed were discovered, or bought in the hope that they would appear.

Built means there is a mechanism

In 1956 McDonald's set up a separate company to buy the land its future restaurants would stand on, lease it, and sublease the package to the franchisee. That is a mechanism: a lease with a markup, signed before the restaurant exists. By 2009 the company collected $4,841.0 million in rent against $2,379.8 million in royalties, it was the lessee at 13,620 locations, and the property on its books had grown to a net $24,160.0 million by the end of 2019. Rent has run roughly two to one over royalties for decades because rent was the design and the brand license was the byproduct.

Ryanair started charging €2.50 for a checked bag in March 2006 and cut every fare by exactly €2.50 in the same announcement, calling the pair revenue neutral. Nothing was earned on day one. What was built was a line item that could move independently of the fare. By the year ended March 2012 ancillary revenue was €886 million against €3,439 million of scheduled revenue; in the most recent reported year it was €4.99 billion against €15.54 billion of group revenue, and its share has sat near a third — 35.7%, 32.0%, 33.8% across the last three years — rather than drifting.

Berkshire Hathaway bought an insurer in March 1967 for $8.6 million, not for the underwriting but for the gap between collecting premiums and paying claims. Float was $19.4 million then and about $171 billion at the end of 2024, alongside a $9.02 billion underwriting gain. Amazon's cloud business started as a data-center competency the company decided to sell rather than an accident of spare capacity, and in 2024 it returned $39.8 billion of operating income on $107.6 billion of revenue at a 37.04% margin while North American retail earned $25.0 billion at 6.44%.

Microsoft priced the Xbox below cost on 15 November 2001 and said so; the Home and Entertainment division lost roughly $348 million in the December 2002 quarter against $180 million a year earlier. Disney put $6.99 on the screen in April 2019 to buy reach, and the direct-to-consumer loss widened to $693 million in the quarter the service launched while the cable networks it was supposed to replace grew. In each case the size of the loss was a decision before it was a result.

Bought means there is a hope

Sears is the cleanest case of the other kind. In October 1981 it agreed to buy Coldwell Banker and Dean Witter Reynolds, paying $607 million for the brokerage, on the argument that 36 million regular customer families would buy stocks where they bought socks. There was no mechanism — no lease, no transfer price, no fee that existed whether or not the synergy did. Four years in, the financial centers sitting inside 312 stores were still unprofitable overall. Dean Witter Discover was cut loose in March 1993, Coldwell Banker Residential that May, and Allstate in 1995.

CVS Health bought Aetna in 2018 to make the insurer the engine of an integrated health company. By 2024 the insurance arm's adjusted operating income had collapsed to $307 million from $5,577 million the year before, and group operating income fell 38% to $8.5 billion. The passenger was bought at the price of an engine.

The pattern is not that acquisitions are bad. It is that a subsidy acquired is a forecast, and a subsidy built is a contract. McDonald's rent arrives whether the franchisee's synergy thesis works.

The corpus's other reversal runs the opposite way and is worth holding next to this one. Philip Morris was not shopping for a payer; it had one. It used cigarette cash to buy Kraft for roughly $13.1 billion in October 1988, raising an all-cash offer to $106 a share, and the stated reason was to diversify away from a declining, regulated market. The passenger was the point. Nineteen years later, with tobacco litigation pressing on the payer, Altria distributed about 88.9% of Kraft to its own shareholders and the arrangement ended. Buying a passenger with a healthy payer's cash is a sound thing to do and it does not make the passenger permanent — it makes the passenger the first thing available to sell when the payer needs the money back.

Which way the money runs, and who can see it

A transfer price is the mechanism at its most explicit, and it can be pointed either way. Samsung's memory division nearly doubled the internal price it charges Samsung's own phone and appliance arm during the 2026 upcycle. The chip division then accounted for 93.8% of group operating profit while the consumer division's profit fell 38% and appliances slipped to a loss, on group operating profit of KRW 57.23 trillion. Nothing was sold differently. Margin was moved to the part of the company the market pays a higher multiple for.

Alphabet went the other way in August 2015 and restructured specifically so shareholders could see the cash leaving the advertising business. Other Bets had lost about $24.3 billion cumulatively by mid-2021; in 2024 Google Services earned $121.3 billion of operating income while Other Bets lost about $4.4 billion. The structure did not reduce the subsidy. It made arguing about the subsidy possible, which is why it has survived a decade.

UnitedHealth ran the same logic in April 2011 when it put its services businesses under one Optum brand. Optum went from 14.8% of group operating earnings in 2011 to 44% by 2017, crossed $100 billion of revenue in 2018, and by 2024 booked $150.9 billion from affiliated customers — close to 60% of its own reported revenue, most of it billed to its sister insurer.

What a subsidy buys is time, which can be spent or wasted

GE reached for the FDIC's debt guarantee in December 2008 because GE Capital could not roll its funding. The rescue held. What it bought was years in which the finance arm's earnings covered an industrial business that was declining underneath them; GE Capital was designated a systemically important nonbank in July 2013 and the designation was rescinded in June 2016 only after GE exited most lending and cut the unit's assets by more than 80%.

The Washington Post Company spent three decades funding a newspaper with a test-prep business. When Kaplan's enrollment fell the subsidy broke, and the paper was sold in August 2013 to leave a company where education would be nearly two thirds of remaining revenue, renamed Graham Holdings. The subsidy was real for thirty years and it was never permanent.

Samsung's counter-cyclical build through the 2022 downturn is the same risk in a different shape. The chip division lost roughly KRW 14.87 trillion in 2023 against group profit of KRW 6.57 trillion. It financed a genuine rebound, and it cost Samsung the DRAM revenue lead it had held since 1992, with SK Hynix ahead 36% to 34% by 2025. Lockheed Martin's F-35 sustainment franchise is worth $1.58 trillion over the program's life, up 44% from $1.1 trillion, and full mission capable rates fell from 38% in FY2021 to 25% in FY2025 until the Pentagon launched a $13.7 billion restructuring of the arrangement. A payer that is doing well enough to cover the passenger can hide the fact that it is being cut open to do it.