UnitedHealth, Salesforce, and the AI Value-Chain Debate
Takeaways from Our Latest Member Call
We held our third bi-monthly member call on Thursday. Several instructors from the just-completed Wide-Moat Investing Summit delivered “elevator pitch” versions of ideas they had presented at the summit. A single theme ran through most of the pitches: high-quality businesses the market has marked down on an artificial-intelligence narrative the presenters consider either overblown or simply irrelevant. The call closed with a candid roundtable on who in the AI value chain gets squeezed.
For those who couldn’t join us, here is a synthesis of the key insights and specific investment thesis highlights shared by our community of intelligent capital allocators.
UnitedHealth: two presenters, one thesis
Dave Sather of Sather Financial Group framed UnitedHealth as misunderstood. The market lumps it with HCA, Centene, and Aetna and sold it off on fears that Washington would cut Medicare reimbursement. Dave’s rebuttal: health insurance is short-tail risk, repriced every twelve months, so a pricing mistake is not a thirty-year problem, and Washington has historically come back with reasonable rates. The more interesting asset, he argued, is Optum, which he called “a technology company with a healthcare wrapper,” sitting on roughly 330 million identified data points it can mine and sell back to the industry. Dave sees about 50% upside if management delivers its own 13 to 16% growth guidance.
Jonathon Fite of KMF Investments reinforced the case with the economics of vertical integration. Because UNH owns provider practices through Optum, which helps explain why its margins, returns, and EV-to-sales multiple of 1.25 to 1.5 times sit above an industry that trades at 0.5 to 0.7 times. Jonathon walked through the post-COVID stumble, when a new team mispriced policies and spiked the loss ratio, and the housecleaning that followed: the ex-CEO returned and dismissed half of the top 100 managers, which he likened to the Dollar General turnaround. With CEO compensation tied almost entirely to stock through 2030, Jonathon expects buybacks until the shares clear $500 and a multi-year fair value north of $600.
Salesforce: the moat is distribution, not the model
Amit Nath, Director of Research at Montaka Investments, relayed the Salesforce thesis his colleague Andy Macken presented at the summit. The “SaaS-pocalypse” fear, that agents make seat-based software obsolete and hand the economics to the model providers, has in Montaka’s view priced Salesforce for roughly 1% revenue growth as compared to guidance for 10% organic growth. Amit argued the narrative is backwards. Building agents is the easy part; deploying them into large, regulated enterprises that demand governance, security, auditability, reliability, and increasingly cost control is the hard part, and that is where Salesforce’s 250 petabytes of proprietary data, customer metadata, and decades of trusted distribution form the real moat. The models themselves are becoming the most substitutable layer in the stack, with open-source Chinese models running at a fraction of frontier cost and only months behind. Customers report 5 to 20 times ROI on Agentforce deployments, agentic usage is up several hundred percent year over year, and fee holidays at large customers roll off in 2027, aligning with management’s guidance for a second-half 2026 acceleration. Amit likened the optics to Simpson’s paradox: a flat surface over violent underlying change. The company carries a $155 billion market cap, a $50 billion buyback (half already executed), and a double-digit free-cash-flow yield.
4imprint: a marketing engine hiding in plain sight
Jim and Abby Zimmerman of Lowell Capital screen for fortress balance sheets and low enterprise value to operating cash flow, and they put 4imprint squarely in that bucket. The company generates almost all revenue in North America but is listed in London, which keeps it off most US investors’ radar. Abby reframed it: not a promotional-products distributor but a customer-acquisition and marketing platform, built on a proprietary database and decades of disciplined direct marketing. The drop-ship model carries no inventory, driving high returns on capital. Jim noted the firm spends $150 to $175 million a year on marketing, more than anyone in its industry, treats that spend as capex, and draws about three-quarters of sales from repeat customers. Tariff fears hit the stock on the assumption suppliers would raise prices, yet margins have held. With roughly $100 million a year returned through regular and special dividends on an enterprise value near $1.3 billion, Jim and Abby think the stock can at least double, and invoked their Celestica win as the template for buying growth at a value price.
Toast: another AI fear, another mispricing
Chris Crawford of Crawford Fund Management pitched Toast as a rare combination of growth, moat, and value, dislocated by AI fears he considers misplaced. The restaurant operating and payments platform is growing revenue above 20% and taking share from legacy systems like NCR and Oracle’s Micros, and from modern peers Square and Clover (owned by Fiserv). Chris sees Toast as an AI beneficiary, not a victim. It is founder-led, debt-free, and holds about $1.7 billion in cash against a $13 billion market cap, and it began its first-ever buyback in Q1 at depressed prices. His valuation methods converge near $48 a share, about 80% above today’s price. Revenue has scaled from roughly $700 million to about $7 billion in five to six years. The moat is switching costs: the system is embedded across front and back of house, so migrating would mean retraining staff and rebuilding menus and historical data. Two-thirds of revenue is payments and one-third the higher-margin SaaS layer, which management wants to grow; the main risk Chris flagged is payment competitors such as Fiserv using price to pry the business loose.
Midea: a compounder priced like a commodity
Rodrigo Lopez Buenrostro of KuE Capital pitched Midea Group, the Chinese industrial conglomerate listed in mainland China and Hong Kong. Midea built its moat in HVAC, where it holds roughly 45% of the global market for compressors, the most expensive component in an air conditioner. Two-thirds of revenue still comes from home appliances growing 7 to 8% a year, but 25% now comes from a commercial and industrial arm growing around 15%, spanning robotics through its KUKA subsidiary, elevators, and building and industrial technology. At 15 times earnings with a 7% FCF yield, against Honeywell, Siemens, and Carrier at 25 to 30 times, Rodrigo sees a re-rating candidate that has historically generated returns on equity above 20%. He put intrinsic value near $19 a share against about $11 today, for 45 to 50% upside in dollar terms.
The AI roundtable: who holds the hot potato?
The liveliest stretch was a debate on where AI economics ultimately settle. Bryan Lawrence of Oakcliff Capital, a longtime Alphabet holder, called Google’s most recent quarter one of the most remarkable he has seen: backlog doubling to roughly $460 billion, incremental cloud margins above 65%, and reaccelerating search. He described running two AI “analysts,” an investment analyst he calls Marvin and an information-gatherer named Vicky who digests hours of podcasts daily.
James Emanuel framed the question that ran through the rest of the hour. If intelligence is commoditizing, with cheap open-source and Chinese models good enough for most enterprise work, who absorbs the cost as it moves down the funnel from hyperscaler capex, to the software vendors paying for tokens, to the end customer? Someone is left holding the hot potato, and it was not clear whether it is the hyperscaler or the application company in the middle.
Amit Nath argued that intelligence is becoming the most substitutable layer in the stack. The open question is whether a frontier model is interchangeable with a far cheaper Chinese or open-source model dropped into the same harness, and for ordinary enterprise work the answer increasingly looks like yes. Microsoft is preparing to run DeepSeek inside Copilot, ring-fenced within Azure, and model capability has begun to outrun real-world usefulness. Once a model is good enough, price becomes the binding constraint, which is why the frontier labs are pushing into material science and drug discovery, and why Google no longer feels compelled to sit at the bleeding edge.
Bryan sharpened the point with a contrast. A model that runs a call center for Pizza Hut need not be Claude, and the restaurant will not care if a cheap open-source model runs on its own hardware, so the deflationary force in that middle layer is enormous and, in his view, not yet in the price of the model labs. Drug discovery is the opposite, where the highest-order intelligence is worth almost any price, a point he credited to Gavin Baker. He also noted that his own Tesla self-drives on roughly seven gigabytes of on-device software with no network connection, and asked why so much inference will not migrate to cheap models at the edge.