Akre Focus ETF Quarterly Commentary

At a Glance

Underperformed the Benchmark by 14.5 Percentage Points This Quarter, Yet Fundamentals Grew 23% Counter-Cyclically

Performance Comparison: The Fund returned 0.66% in Q2 2026, versus the S&P 500 Total Return of 15.20%; over the trailing twelve months, the Fund returned -24.01% versus the index's +22.32%. The author attributes the underperformance primarily to valuation multiple compression rather than business deterioration—the portfolio's weighted-average free cash flow per share (FCFPS) grew 23% year-over-year in Q1. The author states, "We expect our portfolio companies to continue growing economic value per share at attractive rates."

Valuation Multiples at Decade Lows, Anticipating a Tailwind from Headwind

Fundamentals at Their Strongest in a Decade: The five-year FCFPS growth rate estimate is near 16% (at a historical high), the valuation is only 17x FCFPS (the lowest since 2016), the weighted-average ROIC for 2026 is approximately 42%, and total debt/EBITDA is only 1.4x. The author opines: "we believe the Fund offers outstanding growth potential and financial strength at decade-low valuations." The valuation compression stems from multiples retreating from their peaks. The author believes that once multiples shift from a headwind to neutral or a tailwind, medium-term returns are promising, and states, "Divergences between fundamentals and share prices are not uncommon but can be most pronounced and prolonged during periods when new technology promises to change the world."

Clear Signs of Speculation in the AI Frenzy, Risks Are Building

Market View [Cautious/Bearish]. The author acknowledges that AI may prove transformative (unprecedented growth rates for LLM providers, strong demand for "pick-and-shovel" plays), but points to excessive speculation signals:

• Among the top 16 U.S. companies by market cap, five are semiconductor companies and five are AI/cloud computing giants;
• The iShares Semiconductor ETF (SOXX) posted a total return of 106.9% in Q2 alone;
• Semiconductor companies now account for 20% of the S&P 500's weight, an all-time high and a 4x increase from June 2020 (source: Citadel Securities);
• Retail investors are piling into leveraged ETFs: leveraged ETF assets have reached approximately $218 billion (up 4.5x from June 2020), and semiconductor leveraged ETF assets alone have surged 175% since the end of March;
• Margin debt stands at $1.4 trillion (an all-time high, up 54% year-over-year);
• Following SpaceX's IPO on June 11, its valuation reached approximately 120x annualized revenue for Q1 2026, and its prospectus cites a "total addressable market" of $28.5 trillion (equivalent to 92% of 2025 U.S. GDP); Wall Street has already launched 11 single-stock leveraged ETFs.

Akre Focus ETF's Q2 return of 0.66% and trailing one-year return of -24.01% significantly lagged the S&P 500's 15.20% and 22.32%, but its 10-year and 15-year annualized returns stand at 11.07% and 12.37%, respectively

The author summarizes: "there has been a clear 'fast lane' in the stock market, and lane changing into it has materially increased in terms of participation, leverage, capital, and concentration."

Competitive Moat of Large Language Models Questionable, Investment Caution Warranted

Key Sector Thesis (Bearish on Current Valuations). The author questions the sustainable competitive advantages of LLM leaders: low user switching costs, weak pricing power, and difficulty maintaining technical leadership. Evidence includes customers using orchestration systems like OpenRouter to dynamically route tokens to the cheapest model, and the combined token share of the three major LLMs (Anthropic, OpenAI, Gemini) falling from 72% to 33% over the past year (shifting to open-source, lower-cost models). The author asks: "The nature, extent, and achievability of durable competitive advantages for the leading LLMs given the dynamics evidenced to-date, including the apparent lack of switching costs, pricing power, and material/sustainable technical leadership." The risk is that capital expenditures exceed the ability to generate basic returns. Historical lessons (railroads, oil, automobiles, radio, PCs, the internet) show that even successful new technologies can destroy investor value.

Additional Analysis: The Profit Paradox of Anthropic and OpenAI, and the "China Variable" in Memory Chips

1. Anthropic vs. Mastercard: Worlds Apart in Revenue Scale and Profitability

A comparison of revenue between Anthropic and Mastercard highlights the fundamental difference between an AI infrastructure company and a mature platform business. Mastercard's 47% adjusted net profit margin stems from its asset-light model, bilateral network effects, and high customer stickiness (2025 data). In contrast, Anthropic's revenue comes almost entirely from API calls and model licensing, requiring continuous massive capital expenditure (GPUs, data centers), and customer contracts are primarily consumption-based "per-token" billing, resulting in a gross margin of only about 40-50% (industry estimates). The key point: While Mastercard's cost per transaction declines marginally, Anthropic's costs rise linearly or super-linearly with inference volume (due to compute bottlenecks). If AI demand growth slows, both revenue and margins would face a double squeeze, whereas Mastercard's payment network revenue exhibits strong resilience to economic cycles.

Metric	Anthropic (2025 Est.)	Mastercard (2025)
Revenue Scale	~$3-4 billion (comparable to Mastercard)	$25 billion+
Net Profit Margin	Negative (loss-making)	47% (adjusted)
Asset Density (Fixed Assets per Dollar of Revenue)	~2-3x (GPUs, data centers)	0.2x (primarily software and network)
Customer Switching Costs	Low (high model substitutability)	Extremely High (payment infrastructure)

2. Memory Chip Shortages and Capacity Expansion: The "Double-Edged Sword" of the China Variable

Traditional analysis focuses on supply-demand cycles, but the follow-up points out that capacity plans from CXMT (ChangXin Memory Technologies) and YMTC (Yangtze Memory Technologies) (ramping up sharply in 2027) are reshaping the industry landscape. Key data: According to TrendForce, the global DRAM market will reach approximately $150 billion in 2026, but Chinese manufacturers currently hold less than 5% share; if capacity comes online as planned in 2027, China's share could jump to 15-20%, at which point Micron, SK Hynix, and others face downside price risk. Additionally, the South Korean government announced an additional $52 billion investment in 2026 (targeting HBM and advanced nodes), but if Chinese overcapacity triggers a global price war, Korean manufacturers' margins could fall sharply from their 2025 historical peaks (SK Hynix operating margin ~40%). Investors should be wary: AI chip demand cannot fully absorb excess capacity in traditional memory.

3. Political and Resource Risks in AI Development: Four Modes of Social Resistance

The follow-up asks "how to handle AI-induced unemployment and inflation," and historical precedents can be added. Data: A McKinsey 2025 report predicts that by 2030, AI will replace 30% of repetitive cognitive jobs (approximately 400 million globally), but only create about 150 million new roles. Political responses may take four paths:

• Tax and Redistribute: e.g., the EU is considering a "robot tax" on AI automation firms, which could compress profit margins.
• Regulatory Barriers: China and the EU have already enacted safety regulations for AI foundation models, delaying product launch cycles.
• Resource Wars: AI computing consumes about 40% of global electricity growth (IEA 2026 data), driving up electricity prices and conflicting with carbon neutrality goals.
• Nationalist Tech Decoupling: The U.S. tightens chip exports to China, but China accelerates domestic substitution, fragmenting global supply chains and raising costs.

These risks pose a systemic threat to overvalued AI leaders (e.g., Nvidia, Palantir), while Akre's invested traditional software companies are less directly affected (due to their light AI spending and revenue not dependent on compute expansion).

4. SpaceX's TAM Structure: AI Software Is the Real Prize

The follow-up mentions SpaceX's projected TAM of $28.5 trillion, but only 1.3% comes from "space solutions," with the rest attributed to AI "enterprise applications." This data overturns high market expectations for commercial space. For comparison: The global AI software market in 2026 (Gartner forecast) is about $350 billion, while SpaceX's $370 billion space opportunity is only one-tenth of that. If SpaceX's valuation (approximately $250 billion) implies excessive optimism about its TAM (i.e., assuming it captures a large share of space solutions), its stock price is at risk. A more logical view: The AI software layer (e.g., enterprise applications) is where the 10x growth potential lies, and this is the main battlefield for Akre's portfolio (e.g., Salesforce, ServiceNow).

5. Portfolio Companies' AI Defense Moat: Patent and Data Moat

The follow-up lists FICO's 137 AI patents, CCC's 50% annual growth, Mastercard's tokenization technology (covering 50% of e-commerce transactions). Additional note: Moody's rated over $100 billion in AI-related bond issuances in Q1, indicating its data and analytics capabilities have become "infrastructure" for the AI financing ecosystem. Furthermore, CoStar's Homes Ai and Apartments.com Ai leverage proprietary, irreplicable real estate databases, not public web data, creating a natural monopoly. In contrast, AI-native startups (e.g., Zillow's AI features) need to invest heavily to build data, while CoStar already has 20 years of historical data accumulation.

6. Historical Lessons: Quantitative Comparison Table of Berkshire vs. QQQ

The follow-up references 1999-2005 data; additional dimensions can supplement the "long-term discipline" conclusion:

Investment	1999.3.10→2000.3.10 (1 Year)	2000.3.10→2005.3.10 (5 Years)	1999.3.10→2005.3.10 (6-Year Total Return)
QQQ	+124.25%	-85.46%	-25.79%
Berkshire Hathaway	-44.11%	+292.72%	+22.73%
Relative Performance (Berkshire/QQQ)	0.33x	7.00x	1.30x

The table visually demonstrates: Chasing short-term hot themes (QQQ) outperformed by more than 4x in one year, but over the long term (6 years), the value-investing approach of Berkshire clearly prevailed. The current market's enthusiasm for AI hardware ("atoms") mirrors the 1999 internet bubble. Akre's strategy continues to focus on application-layer companies with durable moats.

Summary: The additional analysis reveals risks to earnings sustainability, geopolitics, and valuation bubbles for AI infrastructure companies (e.g., Anthropic, Micron), while Akre's portfolio companies' data monopolies, asset-light models, and AI defensiveness (rather than reliance on compute expansion) offer greater long-term investment value. Historical data once again proves that short-term gains from following market narratives may come at the cost of long-term losses.