FIG. 55MOAT, UNRENTABLE

Buffett on moats when cognition is a commodity

More than 80% of businesses in software segments serving multiple industries qualify as having a weak vertical moat. AlixPartners reported that result in July after assessing more than 500 privately held software businesses for exposure to AI disruption. Those segments comprise DevOps, middleware and workflow automation, tools that all industries purchase and none owns. Just 14% of the full sample had strong data and vertical moats.

Those businesses were protected because they were expensive to build, and most of that expense was cognition. Suppliers now meter cognition. A moat’s value depends on what competitors cannot rent. When anyone can purchase thinking per token, the lasting advantages are those beyond metered pricing: customers’ affection, information held exclusively by the incumbent, and physical cost position.

What Buffett bought in 1972

Warren Buffett popularised the idea of a moat, and his clearest practical example is a chocolate manufacturer. His 1983 letter to Berkshire Hathaway shareholders sets out the figures. Blue Chip Stamps, which subsequently merged into Berkshire, acquired See’s Candies in early 1972 for $25 million. See’s then held roughly $8 million in net tangible assets and made around $2 million after tax. By 1983, it made $13 million after tax with roughly $20 million in net tangible assets.

Buffett attributed the premium to reputation. He wrote, “Consumer franchises are a prime source of economic Goodwill” (1983 letter). With a franchise, the buyer’s valuation of the box determines its price, not the expense of producing it. He identified the source too. “We regard product quality as sacred” (1983 letter).

Customers remembered the box, and the kitchen maintained its standards. Those were the moat. Cognition accounted for none of it.

The meter drains moats made of work

Suppliers now meter cognition, and the price per unit drops quickly enough to upend a traditional budget. Buyers benefit from that price curve in their budgets; sellers lose protection around their businesses.

Robert Almeida at MFS identified the most vulnerable businesses in February. They earn money through information asymmetry, charge for human expertise, depend on friction in processes, or incur substantial marginal costs through labour-intensive services. Each relies on the expense of paying someone to think, and horizontal software falls squarely within that group. Workflow software encodes knowledge of a process. A competitor using a model can already recreate much of that software, as can its customer.

Four pillars stopped predicting

Morningstar identifies five sources of advantage in its moat framework (switching costs, network effects, intangible assets, efficient scale and cost advantage). Adrian Helfert, chief investment officer of multi-asset strategies at Westwood, rated 460 S&P 500 businesses for the strength of their moats and their exposure to AI. Speaking to Morningstar in March, he said four of those five pillars “have almost no predictive power in today’s AI environment”.

“The only moat that still clearly matters is physical cost advantage” (Morningstar), he said, citing supply chains, factories, mineral reserves and regulated infrastructure. He has yet to publish the study on SSRN, so I regard this as one manager’s interpretation.

Switching costs illustrate the reasoning. Customers leaving a software provider spend most of their exit costs on migration, work a model performs at token rates. The lock remains. The key gets cheaper.

Two kinds of data

The sources disagree about data. Helfert says investors are pricing in AI wearing away proprietary data advantages. Almeida writes that “AI can be both a moat destroyer and a moat amplifier” (MFS), adding that systems of record could become more valuable.

Both views can hold, since each concerns a different kind of data. Some proprietary information amounts to cognition after a delay: research, benchmarks, curated content. A competitor using a model can put together an alternative. Other information exists solely because the incumbent processed a transaction. A bank has the customer’s payment history because it transferred the funds. Nobody can hire that history at token rates, since nobody else participated when the transaction occurred.

Why the chocolate survives

Helfert’s screen appears to condemn See’s because it had an intangible asset for a moat. That classification groups too much together. Businesses selling expertise through advice belong in Almeida’s most vulnerable category. Businesses earning affection for chocolates produced to a consistent standard fall outside it. A model can produce a competitor’s advertising. It cannot cause customers to feel affection for the box or reduce cocoa prices.

The distinction therefore cuts through individual pillars, not the gaps between them. Brands based on affection retain their advantage; brands based on expertise lose theirs. Transaction-generated data retains its advantage; research-generated data loses it.

A test for the finance team

Boards want to know whether the business has a moat. I would ask which pillars depend on something competitors can purchase per token. Four checks cover most of the assessment.

  • Switching costs: Price the exit at token rates. If leaving is mostly migration work, a model will do it for a fraction of today’s cost.
  • Data: Find out how the data came into being. Data generated by processing customers’ transactions holds; data a determined rival could compile buys time and no more.
  • Brand: Separate affection from expertise. Customers who love the product stay. Customers who pay for know-how will test a cheaper source of it.
  • Cost position: Count the physical assets, contracts and licences behind the margin. A meter cannot price those.

The findings affect goodwill on software acquisitions, where teams often base impairment assumptions on switching costs. They also affect supplier renewals, where suppliers with weak moats have less pricing power than they suggest in their renewal letters.

Blue Chip Stamps spent $25 million to acquire roughly $8 million in net tangible assets, and eleven years later See’s made $13 million after tax. Metered cognition has left those sums unchanged. It has altered which businesses qualify.