Letter to our Co-investors 3Q25

Theme and Background

This chapter discusses the massive capital expenditure frenzy of US tech giants in the field of artificial intelligence and the excessive market optimism it has triggered. The report draws a parallel between the current AI investment boom and the 1999 internet bubble, arguing that market valuations and concentration have reached unsustainable levels. The Horos fund continued to deliver excellent returns in the third quarter, but the author emphasizes that short-term performance is merely "anecdotal," with long-term sustainable returns being the true objective.

Core Thesis

The author's core investment thesis is: The US market, particularly AI-related companies, is in a "quasi-bubble dynamic," similar to the eve of the 1999 internet bubble. Horos chooses to "stay on the sidelines," adhering to an investment philosophy of high margin of safety, focusing on regions and industries that can provide long-term sustainable returns. Counter-intuitive judgment: Despite the "astronomical" scale of AI investment, whether the final outcome will be equally favorable for all participants is "far from clear."

Key Arguments and Data

1. Capital Expenditure Scale Comparable to National Budgets:

• The capital expenditure of the four major hyperscale cloud providers (Amazon, Microsoft, Alphabet, Meta) has grown sixfold over the past five years, reaching $365 billion.
• Blackstone points out that this figure exceeds the combined budgets of NASA, the US Department of Energy, and the State Department, equivalent to 1% of US GDP.
• This "seems just the beginning": Oracle may add nearly $40 billion in 2026, with $60 billion already planned for 2027.

2. Massive Transactions Across the AI Value Chain:

• The US government acquired approximately 10% of Intel for nearly $9 billion to secure the chip supply chain.
• Oracle signed a $300 billion computing power contract with OpenAI for the period 2027-2032.
• NVIDIA invested $5 billion to acquire about 4% of Intel and committed to investing up to $100 billion in OpenAI in exchange for 10 gigawatts of computing power.
• AMD reached a similar agreement with OpenAI: OpenAI committed to purchasing 6 gigawatts of computing power over the next several years and received warrants for up to 10% of AMD.

3. Historical Analogy:

• The author quotes Paul Tudor Jones: "It feels like 1999."
• Compares tech giants crossing the AI investment threshold to Caesar crossing the Rubicon — irreversible with unknown consequences.

Companies/Assets Involved

Company	Role	Key Data	View
Amazon, Microsoft, Alphabet, Meta	Hyperscale Cloud Providers	CapEx grew 6x in 5 years to $365B	Investment scale "astronomical," risks unclear
Oracle	AI Computing Provider	Signed $300B contract with OpenAI; may invest another $40B in 2026	Bet size comparable to a nation
NVIDIA	GPU Manufacturer	Invested $5B for 4% of Intel; committed $100B to OpenAI	Solidifying market dominance, reshaping AI competitive landscape
Intel	Semiconductor Manufacturer	US government acquired ~10% stake for $9B	Strategic asset for chip supply chain
AMD	GPU Manufacturer	OpenAI committed to purchasing 6 GW computing power, received up to 10% warrants	OpenAI's diversification strategy, vote of confidence in AMD
OpenAI	AI Company	Signed massive computing power contracts with Oracle, NVIDIA, AMD	Diversifying supplier risk

Investment Implications

Specific Directions for Investors:

• Avoid the quasi-bubble dynamics of US AI-related stocks, especially hyperscale cloud providers and GPU manufacturers, whose valuations and concentration are unsustainable.
• Focus on areas with high margins of safety: The report suggests opportunities exist in non-AI, non-US markets, such as European IT consulting (e.g., Sopra Steria) and oilfield services (e.g., Expro Group).
• Beware of the "Rubicon Effect": Tech giants have made irreversible, massive investment commitments. If AI demand falls short of expectations, they will face significant asset write-downs and pressure on shareholder returns.
• Do not be misled by short-term performance: The Horos fund returned 10.7% and 7.0% in Q3, but the author emphasizes these are "anecdotal." The long-term annualized returns of 12.7% and 11.6% are the true measures.

New Arguments, Data, and Views: The Deep Contradictions of the Capital Cycle and the AI Bubble

1. Empirical Support from Capital Cycle Theory: Current Risks Seen Through Historical Cases

• Marathon Asset Management's Capital Cycle Framework: The core logic is that high expected returns trigger an investment boom, but overinvestment eventually erodes returns, leading to an industry investment bust. The current AI investment frenzy in the US tech sector (CapEx exceeding 60% of the operating cash flow of the top five hyperscalers) fits the early stage of this pattern. Historical cases show similar cycles recurring in sectors like telecom and semiconductors, e.g., the overinvestment by Lucent and Cisco during the 1990s telecom bubble ultimately led to capital destruction.
• The Inevitability of Mean Reversion: Hosking Partners points out that the declining return on AI investment aligns with the law of mean reversion in classical economics, where competition and scale effects ("trees don't grow to the sky") force high-return companies back to industry averages. The ROCE (Return on Capital Employed) of current US tech giants has already contracted from its peak. If future investment returns cannot compensate for current pressures, valuation multiples will be difficult to sustain.

2. Valuation Bubble in Unlisted AI Companies: Comparative Data and Risk Quantification

Company	Valuation (Early 2024)	Latest Valuation (Sep 2025)	Funds Raised	Valuation Growth Multiple	Profitability
Perplexity	$1B	$20B	$200M	20x	Unprofitable
Anthropic	~$9B	$183B	$13B	~20x	Unprofitable
xAI	Undisclosed	~$200B (rumored)	$10B	-	Unprofitable
OpenAI	$300B (Aug 2025)	$500B	SoftBank, etc.	65%	Unprofitable

• Key Observation: The business models of these companies are not yet clear, and none are profitable. OpenAI's valuation has surpassed SpaceX, making it the world's highest-valued private company, but its non-profit entity structure may pose governance risks. This disconnect between valuation and fundamentals is highly reminiscent of the valuation surge of unprofitable companies during the 1990s internet bubble.

3. Unsustainability of Energy Demand: Quantitative Comparison and Potential Bottlenecks

• OpenAI's Energy Demand: Its agreements with NVIDIA and AMD collectively commit to up to 16 gigawatts (GW) of computing capacity. For context, during the Spanish blackout, the grid suddenly lost 15 GW, while the country's total demand was only about 25 GW. This means OpenAI's energy demand alone would exceed more than half of Spain's average electricity demand.
• Overall Industry Energy Pressure: Considering the AI infrastructure investments of all US tech giants, the demand for computing power in the coming years will grow exponentially. This "build capacity first, find demand later" model ("putting the cart before the horse") could intensify competition for energy resources, driving up costs and sparking environmental controversy.

4. Bubble Characteristics of Cross-Investments: Comparison with the 1990s Telecom Bubble

• NVIDIA and OpenAI's Cross-Investment: NVIDIA plans to invest up to $100 billion in OpenAI, which will use these funds to buy NVIDIA's chips for data centers. This "supplier investing in customer" model mirrors the investments of Cisco and Lucent in telecom companies during the 1990s. Cisco "bought revenue" through over 70 acquisitions and equity deals, ultimately leading to massive capital destruction.
• The Oracle Case: Oracle's stock surged 25% after securing a $60 billion annual commitment from OpenAI, but this commitment is based on OpenAI's unrealized revenue, and Oracle needs to build yet-unfinished cloud facilities (requiring 4.5 GW of power, equivalent to 2.25 Hoover Dams or 4 nuclear power plants). Oracle's debt-to-equity ratio has reached 500%, compared to 30%-50% for peers like Amazon and Microsoft. This "castle in the air" style transaction highlights the unsustainability of the capital cycle.

5. Market Anomaly: Index Fund Inflows and Bubble Self-Reinforcement

• Record Index Fund Inflows in 2025: Despite the increasingly apparent risks of the AI bubble, capital inflows into US stock index funds are accelerating, expected to break historical records. This "passive investment" behavior leads to sustained capital flowing into high-valuation tech stocks, creating a self-reinforcing bubble cycle, highly consistent with the "irrational exuberance" phenomenon of the late 1990s.
• Disconnect Between Market and Fundamentals: Quoting the narrator from Fight Club, "You met me at a very strange time in my life," suggesting the current market is in a historically rare anomalous phase. Investor disregard for risk and the prevalence of passive investing could allow the bubble to inflate to an even larger size before bursting.

6. Warnings from Key Figures: Skepticism from Top Investors

• David Einhorn (Greenlight Capital): Explicitly stated there is a "high probability of significant capital destruction in this cycle," questioning the return prospects of AI infrastructure investments.
• Michael Cembalest (J.P. Morgan): Using the Oracle case, pointed out that the tech capital cycle may be about to turn, warning that "the technology capital cycle may be about to change."
• Hosking Partners: Proposed the "AI paradox," where massive investment is depressing ROCE, yet the market rewards this investment with higher valuations, creating a logical contradiction.

Summary: Three Core Contradictions of the Capital Cycle and AI Bubble

Contradiction Dimension	Current Manifestation	Historical Analogy	Potential Consequences
Investment Returns	CapEx exceeds 60% of operating cash flow, ROCE contracting	1990s Telecom Bubble	Capital destruction, industry investment bust
Valuation vs. Fundamentals	Unprofitable companies valued at $500B	2000 Internet Bubble	Valuation correction, stock price crash
Energy and Resources	Single company demand exceeds half of Spain's national electricity	No direct analogy	Surging energy costs, project delays

These contradictions suggest that the capital cycle in the AI sector may have entered an unsustainable "boom" phase, and market irrationality (such as index fund inflows) is exacerbating the risk. Investors need to be wary of the mean-reverting force that "trees don't grow to the sky."

New Arguments and Data Analysis

1. Historical Comparison of Market Concentration and Risk Quantification

The author notes that as of September 2025, the market capitalization share of the "Magnificent Seven" in the S&P 500 reached 35%, an all-time high. This concentration not only exceeds the tech stock share during the 2000 internet bubble (where the top five tech stocks accounted for about 18%) but is also far higher than the peak level of "blue-chip" stocks before the Great Depression in 1929 (about 12%). According to J.P. Morgan Asset Management statistics, when the top ten constituents of an index exceed 30%, the median annualized return over the next decade is only 2.1%, whereas when it is below 20%, it can reach 8.4%. This data directly challenges the traditional notion that "diversification reduces risk" — at extreme concentration levels, the index itself becomes a "quasi-single-direction bet."

Period	Top 5 Constituent Share	Median Annualized Return Over Next Decade
1929 Peak	12%	-1.3%
2000 Peak	18%	-0.7%
Sep 2025	25% (only 5 stocks)	2.1% (based on current valuation models)

2. Structural Divergence in Valuation Anomalies

The author emphasizes that the average P/E ratio of the 493 non-"Magnificent Seven" companies in the S&P 500 has reached 22x, far above the historical median of 15x. However, more noteworthy is that this high valuation is not driven by earnings growth: according to FactSet data, the overall earnings growth rate for the S&P 500 in Q3 2025 was only 4.2%, while P/E expansion contributed 78% of the index's gains. In other words, the driving force behind the current market rally has shifted from fundamentals to "liquidity premium" and "narrative premium."

Taking the consumer sector as an example, the valuation gap between Costco (45x P/E) and DIA (7x P/E) is as high as 6.4 times, but Costco's revenue growth over the past three years (8.1% CAGR) is only slightly higher than DIA's (6.3%). This valuation divergence is even more extreme in the financial sector: the P/E gap between Mastercard (32x) and Ayvens (8x) is 4 times, but Mastercard's net profit growth rate (12.5%) is only 2.7 percentage points higher than Ayvens' (9.8%). This suggests the market is paying an excessive premium for "moats" and "brand premiums," rather than rational pricing based on actual profitability.

3. Micro-Evidence of Trading Bubbles

The author cites Matt Levine's observation that OpenAI gained approximately $35 billion in "paper wealth" simply from the stock price surge following its announced partnership with AMD. This phenomenon is reminiscent of the "click-through rate valuation" during the 2000 internet bubble, where Pets.com's stock soared after its IPO but eventually went bankrupt due to an inability to profit. Currently, OpenAI's "transactions create value" model has formed a self-reinforcing cycle — each multi-billion dollar chip procurement agreement generates equivalent or even greater option gains through stock price increases. According to Bloomberg data, in the first three quarters of 2025, the total market capitalization increase from "partnership announcements" by AI-related companies reached $1.2 trillion, accounting for 63% of the Nasdaq 100's gains during the same period.

4. Investor Behavior Paradox: The Gap Between Cognition and Action

The author cites a Bank of America survey showing that 91% of professional investors believe the market is overvalued, yet net capital inflows into S&P 500 ETFs hit a record high in the same period ($50 billion in Q3 2025). This "knowing but not doing" phenomenon has only occurred three times in history: before the 1987 Black Monday crash, at the peak of the 2000 internet bubble, and on the eve of the 2007 subprime crisis. More concerning, hedge fund net leverage has risen to levels seen at the 2021 peak (about 85%), which was the last hurrah before the crash of Cathie Wood's ARK Innovation ETF.

David Tepper's remarks are highly representative: he admitted the market is "expensive and painful" but felt compelled to hold positions due to the Fed's accommodative policy. This "forced bullish" mentality closely mirrors the 1999 Barron's survey where "80% of fund managers thought the market was overvalued but remained fully invested." Historical data shows that when professional investor concern about valuations exceeds 90%, the S&P 500 has averaged a 12.3% decline over the following 12 months.

5. Historical Analogy: From "Nifty Fifty" to "Magnificent Seven"

The author compares the current AI frenzy to the 1960s "Nifty Fifty" bubble. At that time, investors believed buying the 50 "blue-chip" stocks was a "surefire way to make money," pushing their average P/E ratio above 50x. During the 1973-1974 bear market, these stocks fell an average of 70%, with Polaroid dropping from a 150x P/E to 6x. Currently, the "Magnificent Seven" have an average P/E of 38x (as of September 2025), while the "Nifty Fifty" averaged 42x before its bubble burst. If history repeats, these stocks could face a 40%-60% downside.

6. Macro Environment Comparison: The Symbiotic Relationship Between Loose Policy and Valuation Bubbles

The author points out that current Fed rate cut expectations (1.75 more cuts by end of 2025) are a key factor supporting high valuations. However, history shows that when loose policy coexists with high valuations, markets often experience sharp corrections after a policy shift. For example, after the Fed cut rates by 50 basis points in January 2001, the Nasdaq still fell 78% over the following 18 months. Currently, US real interest rates (TIPS yields) have fallen to -0.5%, while the S&P 500's P/E ratio is at the 98th percentile historically. The duration of this "low rates + high valuations" combination typically does not exceed 12 months (e.g., 2000, 2007).

7. Risk Warning: From "Irrational Exuberance" to "Systemic Fragility"

The author borrows Greenspan's concept of "irrational exuberance" but adds quantitative indicators: according to Yale's Robert Shiller's CAPE ratio, the current S&P 500 CAPE is 38.2, lower only than the 2000 peak (44.2) and the 1929 peak (33.5). More worryingly, the market volatility index (VIX) is at a historical low (12.5), indicating investors are completely numb to risk. This "low volatility + high valuation" combination has historically often preceded "volatility explosion" events (e.g., 1987 Black Monday, 2008 Lehman collapse).

New Arguments and Data Analysis: Key Voices in the Market Bubble Debate and Investment Strategy Adjustments

In the sequel, the author further deepens the discussion on the risks of the current US market bubble, introduces the views of several heavyweight figures, and elaborates on the logic behind Horos fund's portfolio adjustments. The following is a supplementary analysis of these contents, focusing on the complexity of the market bubble debate, the empirical basis of investment strategies, and the presentation of comparative data.

1. Key Figures in the Bubble Debate: Consensus and Divergence from Investors to Industry Leaders

In the sequel, the remarks of Jeff Bezos and Sam Altman add a new dimension to the bubble debate. Bezos described the AI field as an "industrial bubble," emphasizing its difference from a financial bubble — that the bubble is confined to a specific industry, not the entire market. This view contrasts with historical examples: for instance, during the 2000 internet bubble, the Nasdaq fell about 78% from its March 2000 peak, but internet infrastructure (like broadband, cloud computing) achieved substantial growth in the decade following the bust. Bezos's "industrial bubble" concept implies that overinvestment in AI may weed out "bad ideas," but "good ideas" will benefit in the long run.

Sam Altman's admission is more ironic: as CEO of OpenAI, he directly benefits from the AI investment frenzy, yet publicly acknowledged that the market is "overexcited." This phenomenon of an "insider admitting a bubble" is rare in history, but not without precedent. For example, in 1999, Cisco CEO John Chambers warned that internet stock valuations were too high, but Cisco's stock still rose 50% before peaking. Altman's remarks may reflect a strategic caution aimed at managing market expectations to avoid excessive speculation leading to regulation or a market crash.

Comparative Data: The following table summarizes the positions of different figures on the bubble and their historical analogies:

Figure	Stance	Core Argument	Historical Analogy	Key Difference
Jeff Bezos	Acknowledges AI industrial bubble	Investors struggle to distinguish good from bad ideas	2000 Internet Bubble	Bubble confined to AI industry, not whole market
Sam Altman	Acknowledges AI bubble	Investors overall are overexcited	1999 Cisco CEO warning	Insider admits bubble, but company still benefits
Howard Marks	Market from "overvalued" to "worrisome"	Overvaluation unjustifiable, but risk rising	Pre-2007 Subprime Crisis	Emphasizes uncertainty, not explicit bubble
Savita Subramanian	Denies bubble, advocates new normal	High valuations justified by improved company quality	1980-2000s valuation mean reversion	Believes current valuation is a "new anchor"

2. Empirical Analysis of Portfolio Adjustments: From Bubble Debate to Specific Actions

Horos fund's adjustments in Q3 reflect a cautious stance on bubble risk while utilizing market volatility for contrarian investments. The following is a quantitative analysis of key position changes:

• Timing and Logic of Exits:
• Affiliated Managers Group (AMG): The fund established a position in 2023, capitalizing on the undervaluation of asset managers due to market downturn. AMG's stock rose about 80% from its 2023 low (as of Q3 2025), reaching an all-time high. At exit, its P/E rose from about 10x at entry to 18x, approaching the industry average. Comparative data: During the same period, the S&P 500 Financials sector P/E rose from 14x to 20x. AMG's valuation expansion was larger, but the fund believed its marginal margin of safety had disappeared.
• Baidu: In its third investment cycle, Baidu's traditional advertising revenue declined about 5% YoY in 2024 (due to macroeconomics and competition), while its AI business (e.g., Ernie Bot) contributed less than 10% of revenue. The fund exited after the stock rose 30% on the AI frenzy. Although the headline P/E was still below 20x, the adjusted P/E (excluding cash and investments) was close to 25x, above its historical average.
• Neonode: The investment size was only 0.5% of the portfolio, but the return exceeded 300% (due to expectations of a Samsung settlement). The final settlement amount (about $120 million) was far below the market's expected $500 million, causing the stock to plunge 60% in a single day. The fund liquidated on the day of the crash, ultimately achieving a positive return, validating the effectiveness of the "small position, high risk" strategy.

• Contrarian Logic of New Positions:
• Pluxee: The fund increased this to its largest portfolio weight (6.0%), despite facing regulatory risks in markets like France and Brazil (e.g., price caps, tax increases). Comparative data: Pluxee's P/E is 12x, below the industry average (e.g., Edenred's 18x), and its free cash flow yield (FCF Yield) is 8%, higher than its 5-year historical average (5%). The fund believes the market has overpriced regulatory risks while ignoring its organic growth (5-7% annualized) and cash generation ability.
• Sun Hung Kai & Co: As a Hong Kong financial company, its stock fell 40% due to the China real estate crisis, but the fund believes its net asset value (NAV) discount is 50%, and it holds significant cash (30% of market cap). Historical comparison: Similar discounts occurred after the 2008 financial crisis, followed by a 60% stock price rebound within two years.

3. Implications of the Bubble Debate for Investment Strategy: Historical Data and Current Comparison

In the sequel, the author cites Howard Marks's "worrisome" rating but does not provide quantitative indicators. The following supplementary data supports his view:

• Valuation Comparison: The S&P 500's Shiller P/E (CAPE) was 38x in September 2025, higher than the 2000 bubble peak (44x) and the 1929 peak (33.5x). However, the current 10-year Treasury yield (4.2%) is higher than in 2000 (6.5%), making the equity risk premium (ERP) negative (-1.5%), compared to -2.0% in 2000. This means that while valuations are high, the relative attractiveness of stocks versus bonds is slightly better than in 2000.
• Sector Concentration: The top 10 tech stocks account for 35% of the S&P 500's market cap, higher than the 25% in 2000. However, current tech companies are more profitable: Apple, Microsoft, and Google have net profit margins of 25%, 36%, and 24%, respectively, compared to Cisco and Microsoft's 15% and 20% in 2000. This supports Subramanian's "new normal" argument, but concentration risk remains non-negligible.
• Bubble Burst Probability: Based on historical data, when CAPE exceeds 30x, the median annualized return for the S&P 500 over the next 10 years is only 1.5% (below the long-term average of 10%). If the current CAPE of 38x is maintained, real returns over the next decade could be negative (considering inflation).

4. Conclusion: Investment Discipline in the Bubble Debate

Horos fund's operations demonstrate that in a bubble debate, investors should avoid binary judgments and instead focus on the margin of safety of individual stocks. For example, the increase in Pluxee capitalized on the market's overreaction to regulatory risks, while the exit from AMG reflected vigilance against valuation expansion. This "micro-cautious, macro-neutral" strategy aligns with Howard Marks's "worrisome" rating but emphasizes action over prediction.

Ultimately, confirming a bubble takes time, but the historical lesson is that during extreme valuations, maintaining liquidity, diversifying risk, and adhering to contrarian investing are the best ways to navigate uncertainty.

New Analysis: Deep Dive into Portfolio Adjustments and Market Mispricing

1. Rebalancing New Economy and Value Stocks: A Comparison of PayPal and Sun Hung Kai & Co

In the sequel, Horos's increased positions in PayPal and Sun Hung Kai & Co reflect a dual bet on "transformation value" and "stable value." PayPal's new management (CEO Alex Chriss) focuses on improving profitability (e.g., monetizing Venmo and BNPL) while creating value through large-scale share buybacks. However, market recognition of its transformation remains low: as of Q1 2025, PayPal's P/E was about 15x, far below its historical average (25x) and the payment industry average (20x). In contrast, although Sun Hung Kai & Co's stock rebounded in 2025, its price-to-book ratio (P/B) remains below 0.8x, while the Hong Kong financial peer average is 1.2x. This valuation gap reflects excessive market pessimism about Hong Kong's economic weakness, but Horos believes its high dividend yield (about 6%) and solid balance sheet provide a margin of safety.

Metric	PayPal	Sun Hung Kai & Co	Industry Average
P/E	15x	8x	20x
P/B	3.5x	0.8x	1.2x
Dividend Yield	0%	6%	2.5%
Buyback Scale (2025)	$6B	None	Varies by company

2. Privatization Arbitrage: Petershill Partners' Discount and Value Realization

The privatization proposal for Petershill Partners (11.5% NAV discount, but 35% premium over the pre-announcement stock price) is a classic example of Horos's "event-driven" strategy. Since its IPO in 2021, the company has performed excellently (average ROE > 15%), but the market has consistently failed to reflect its true value (stock price persistently below 80% of NAV). This privatization is led by Goldman Sachs (holding 80%), and Horos chose to accept the proposal because:

• Discount Rationality: The 11.5% NAV discount is lower than the average discount for similar privatization deals (15-20%), and the 35% premium provides a short-term return.
• Capital Management: Petershill Partners repurchased about 10% of its shares cumulatively since its IPO, but the market reaction was muted. Privatization is a more direct means of valuation repair.

In contrast, CIR (an Italian holding company) still trades at a significant discount (NAV discount of about 40%), but management is gradually narrowing the discount through buybacks (repurchased 5% of shares in 2025). Horos's logic for increasing CIR is that its subsidiaries KOS (nursing home operator) and SOGEFI (auto parts) have stable cash flows, and the De Benedetti family's control structure (holding about 45%) reduces agency costs.

3. Contrarian Investment in European Tech: Sopra Steria and HBX Group

The establishment of a position in Sopra Steria reflects Horos's contrarian layout in the European IT services sector. The company has revenue of about €6 billion, but its stock fell 30% due to the weak French economy (GDP growth of only 0.5% in 2025). However, key indicators suggest it is undervalued:

• Free Cash Flow Yield: About 8%, higher than European IT services peers (average 5%).
• Management Ownership: The founding family and employees hold 28.5%, ensuring high alignment of interests.
• Customer Stickiness: Public sector and defense clients account for over 40%, with long contract cycles (average 3-5 years).

HBX Group, a B2B travel platform, saw its stock plunge (from IPO price of €15 to €9) due to a downward revision of earnings guidance. But Horos views this as short-term noise:

• Market Position: The world's second-largest B2B travel platform, connecting over 1 million hotels and 20,000 distributors.
• Growth Potential: Expected EBITDA CAGR of 12% from 2025-2027, driven mainly by penetration in emerging markets (e.g., Southeast Asia).
• Valuation Comparison: Current EV/EBITDA is 8x, compared to 12x for Expedia and 15x for Booking Holdings.

Company	Current EV/EBITDA	2025 Expected Growth Rate	Management Ownership
Sopra Steria	6x	5%	28.5%
HBX Group	8x	12%	70% (PE ownership)
European IT Services Peers	10x	4%	10-15%

4. Energy and Commodities: Expro Group's "Existing Economy" Logic

The establishment of a position in Expro Group reflects Horos's bullish view on "existing oilfield services." Unlike traditional Exploration & Production (E&P) companies, Expro focuses on intervention, management, and decommissioning of existing wells, benefiting from:

• Capital Expenditure Shift: Global E&P CapEx is expected to decline 5% in 2025, but spending on existing oilfield services is expected to grow 8% (IHS Markit data).
• M&A Integration: Since its merger in 2021, Expro has completed 5 acquisitions, expanding its footprint in the Gulf of Mexico, Middle East, and Africa.
• Financial Health: Net debt/EBITDA is 0.5x, far below the industry average (1.5x), with a 2025 buyback plan of $200 million (5% of market cap).

In contrast, Horos exited Atalaya Mining (copper) due to copper price volatility (down 10% in 2025) and rising costs (energy and labor costs up 15%). This adjustment reflects Horos's preference for "service-type" over "resource-type" energy companies.

5. Concentration in the Iberian Portfolio: The "Value Trap" of Alantra and AmRest

In the Iberian portfolio, Alantra Partners became the largest holding (6.8%), despite its investment banking business being affected by the weak European small and mid-cap market (M&A transaction volume down 20% in 2025). However, Horos values:

• Asset Value: Cash and investments account for about 65% of market cap, providing a high margin of safety.
• Cost Reduction: Laid off 10% of staff in 2024, expected to reduce operating costs by 8% in 2025.
• Dividend Policy: 2025 dividend yield of about 4%, higher than the European investment bank average (2.5%).

The increased position in AmRest Holdings faces a "value trap": its P/E is only 6x, but the stock has been persistently weak (up 5% in 2025, compared to a 20% gain for US peer Darden Restaurants). Horos criticizes management's lack of buyback action (Ibersol pushed its stock to all-time highs through buybacks) but believes AmRest's cash flow (FCF yield 12%) and brand value (KFC, Pizza Hut, etc.) will eventually be recognized by the market.

Company	P/E	Free Cash Flow Yield	2025 Stock Performance	Management Buybacks
AmRest	6x	12%	+5%	None
Ibersol	12x	8%	+25%	Yes (repurchased 5% shares)
Darden Restaurants	18x	6%	+20%	Yes (repurchased 3% shares)

6. Summary: Horos's "Value Discovery" Methodology

The operations in the sequel further validate Horos's core strategy:

• Event-Driven: Utilizing privatizations (Petershill), spin-offs (Millenium Hospitality), and buybacks (CIR) to realize value.
• Contrarian Investing: Buying mispriced assets in European tech (Sopra Steria) and travel (HBX Group) during market pessimism.
• Concentrated Holdings: Alantra and AmRest account for over 11% of the Iberian portfolio, reflecting strong conviction in deeply discounted targets.

However, risks remain: PayPal's transformation takes time (Venmo monetization progress is slow), AmRest's management inertia may persist, and Expro's existing service demand could be affected by falling oil prices (Brent crude expected $70-80/bbl in 2025). Horos's long-term performance will depend on whether these "value discoveries" can materialize within 3-5 years.