Voss Capital Q4 2025 Letter to Partners

Theme and Background

This chapter discusses the severe market selloff and "doomsday" panic triggered by breakthroughs in AI technology (particularly the release of Claude Code), as well as the fundamental reasons why Voss Capital underperformed its benchmark in Q4 2025 due to its heavy weighting in the software sector (net returns of -2.1%/-2.2% vs. Russell 2000 +2.2%). The author argues that the market views the software industry as a "monolithic block" to be swallowed by AI, but this judgment contains serious biases.

Core Views

• Market panic over AI is excessive: Technological shocks do not eliminate jobs, only change their composition — history has repeatedly proven this (agricultural employment fell from over 40% in 1900 to roughly 1% today, yet the U.S. has maintained full employment).
• Software is one of the biggest beneficiaries of the AI era: Vertical software platforms with moats in data and engineering (e.g., PAR, CLBT) will use AI to strengthen their competitive positions, not to be disrupted; companies that are early adopters of productive technologies (not infrastructure providers) typically capture the largest share of long-term economic gains.
• Contrarian view: The market's panicked compression of software valuations is a mispricing, especially for software companies with domain expertise and customer stickiness.

Key Arguments and Data

Argument	Data
No mass unemployment in labor market	U.S. unemployment rate is only 4.3% in 2026; software development positions grew roughly 15% in about 6 months
Total employment continues to grow after technology shocks	39 months after ChatGPT’s launch, the U.S. added 4.5 million jobs, most of which are high-paying healthcare positions
Historical shift in agricultural employment proves transition is feasible	Agricultural employment share fell from 40% to ~1%, but overall employment is saturated; disappearing occupations (e.g., switchboard operators, elevator operators) replaced by new ones (e.g., approximately 1.5 million full-time "influencers" in the U.S. today)
Early adopters gain more	Historical pattern: companies that adopt new technology (not infrastructure providers) gain the most over the long term
Actual AI enterprise deployment is slow	Overall enterprise AI adoption rate is still low; software companies are the fastest adopters; Claude Code still has limited success rate in multi-agent tasks

Companies/Assets Involved

• PAR Technology (long, portfolio weight 5.5%):
• Role: Leading restaurant POS software, facing over 6,000 global competitors (including free alternatives), yet continues to win large clients.
• Key data: Newly won Papa John's contract (abandoned self-developed software for PAR); AI will accelerate product iteration and expand TAM.
• View: Bullish. The dominant market narrative ("software will be disrupted") contradicts the fact that PAR keeps beating competitors.
• Cellebrite (CLBT, long, portfolio weight 7.5%):
• Role: Digital forensics platform, hardware-backed software solution; "Cellebrite" has become an industry verb.
• Key moat: Triple barriers — vulnerability library (zero-day/N-day vulnerabilities), hardware adapters (chip-off/interposers), bootloader exploits; supports tens of thousands of device models and firmware versions.
• View: Bullish. Core barriers go far beyond software and encompass a composite moat of hardware + vulnerabilities + domain knowledge.

Investment Implications

• Avoid "AI panic blanket selling": The market erroneously treats all software companies as risk assets, but vertical platforms with data moats, customer stickiness, and domain expertise (e.g., PAR, Cellebrite) are more likely to consolidate share in the AI era.
• Focus on enterprise vertical software: As the Anthropic CEO noted, "there is a huge gap between AI models that work in demos and AI models that work in regulated industries" — meeting regulatory, systems integration, and proprietary data needs is exactly the advantage of established software companies.
• Beware of overcrowding in "AI infrastructure" narratives: The report implies that the market’s frenzy around AI infrastructure may be overdone, while the true long-term beneficiaries are precisely these panicked-sold software companies.

Supplementary Analysis on Cellebrite (CLBT)

• TAM Expansion Quantified: CLBT's AI modules not only penetrate case management (Guardian Investigate) but also plan to launch a stand-alone agentic AI application for one-off ad hoc queries. This essentially creates two independent monetization paths: system-level (high-value record system) and query-level (per-use/subscription). Given its 2026 guidance that AI-related revenue is expected to rise from less than 5% in 2025 to 15–20%, the recurring subscription revenue from AI modules can significantly smooth its originally project-heavy revenue structure.

• Valuation Multiple Comparison: CLBT currently trades at an EV/FCF ex-SBC of only 13.1x (2027E), while comparable high-growth software companies (e.g., Varonis, CrowdStrike) average above 30x. An important but often overlooked comparison is Magnet Forensics, also in digital forensics, which was acquired in 2023 at roughly 20x EBITDA (14.7x EV/FCF), while CLBT’s FCF margin (34%) and growth rate (~20%) significantly outperformed Magnet’s 12% growth and 18% FCF margin at the time of acquisition. This shows the market has not only underestimated CLBT’s moat but also completely ignored the valuation re-rating potential from its AI monetization.

Supplementary Analysis on Flywire (FLYW)

• Quantitative Validation of Industry Barriers: 80% of FLYW’s revenue comes from processing fees, not software subscriptions. This means even if "free AI alternatives" appear, they cannot bypass its payment network. In the education vertical, FLYW has deeply integrated its ERP with over 3,000 educational institutions (including 90% of UK universities), with an average contract term of 7–10 years. According to the company’s Q2 2025 earnings, net revenue retention remains above 115%, indicating existing clients not only stay but also expand cross-product.

• Quantified Valuation Differences vs. Competitors: The table below compares FLYW with two peers in terms of valuation and growth:

Metric	Flywire (FLYW)	Airwallex (private, 2024 raise)	Tipalti (private, 2023 raise)
Valuation (Enterprise Value / Implied EV)	$1.1B	~$8B	~$8.3B
Revenue Growth (2025E-2027E CAGR)	~20%	~33%	~12%
EBITDA Margin (2027E)	~20%	~5%	~10%
Enterprise Value / 2027E EBITDA	6.0x	~40x+ (based on very low EBITDA)	~25x (based on low EBITDA)
Net Cash/Debt	$630M net cash	Continual fundraising (9 rounds)	High debt (~$2B)

Airwallex raised at 8x FLYW's EV with negative EBITDA, while FLYW already generates positive EBITDA and holds significant net cash. If FLYW were priced at Airwallex’s median EV/Sales (about 8x) applied to FLYW's 2027 revenue (estimated ~$2B), its target EV would be $16B, representing 14x upside from the current $1.1B — even with a conservative 50% discount, it would still be 7x. The root of this asymmetric opportunity lies in the market's excessive worry about "fake moats" in payment platforms and its total disregard for FLYW's unique vertical integration advantage.

Supplementary Analysis on Choice Hotels (CHH)

• Short Interest and Insider Signals: CHH’s current short interest is 26%, near a ten-year high. However, 40% insider ownership (including founding family and management) means shorts face significant squeeze risk. In March 2025, the company disclosed a "change-in-control compensation amendment" modifying management’s compensation terms in the event of a change of control — an industry pattern widely seen as a typical M&A preparation move. Combined with the company’s cash unlock ($700M), which may be used for special dividends or acquisitions beyond buybacks, short positions could directly conflict with a potential acquisition premium.

• Valuation Recovery Scenario Analysis:

Scenario	Assumed EBITDA Multiple	Target EV (based on 2027E EBITDA ~$600M)	Implied Share Price	Current Price	Upside
Bear case continues (recession)	8.0x	$4.8B	~$80	$106	-25%
Return to ten-year median	12.5x	$7.5B	~$125	$106	+18%
Return to twenty-year median	16.0x	$9.6B	~$160	$106	+51%
Acquired (ref. Marriott acquiring City Express at 14x)	14-15x	$8.4-9.0B	~$140-150	$106	+32-42%

Note that the current 10.7x EBITDA is in the bottom 2.5% percentile of the past twenty years, while company EBITDA has not declined (2025E $585M, 2026E $610M, 2027E $640M) — compression is purely from market sentiment. Once US RevPAR stabilizes or net room growth returns (management has guided for flat in 2026 and positive in 2027), valuation recovery will be dramatic.

• Hidden Value from Business Transformation: CHH’s extended stay business not only grows in room count (+12% in 2025) but also delivers superior RevPAR performance: in 2025, the overall hotel industry RevPAR declined roughly 2–3%, while CHH’s extended stay brands were flat at +0.2%, outperforming competitors (e.g., Wyndham’s extended stay brands saw RevPAR down 1.5%). Extended stay hotels have average occupancy of 75–80% (vs. economy full-service industry at 62%), and median guest stay of 14 days vs. industry-wide 2 days. This gives CHH a counter-cyclical cash flow moat: during economic downturns, blue-collar workers under inflationary pressure are more likely to opt for long stays at economy extended stay hotels rather than short-term vacations.

Appendix: Implications of AI Performance Gap

The Cornell University assessment (LLMs achieving 84–89% accuracy on synthetic benchmarks but only 25–34% on real-world tasks) aligns with the implication in Voss’s letter that "AI agents still need time." It serves to illustrate that even though Cellebrite’s AI products appear promising, large-scale penetration must cross the reliability and reproducibility gap from lab to courtroom. This actually reinforces CLBT’s first-mover advantage in a "human+machine" hybrid framework — because it has decades of forensic data feedback loops and court-recognized processes that pure AI companies cannot build in the short term.

New Arguments and In-Depth Analysis: "Value Mismatch" of AI Agents and Countertrend Growth in Development Roles

1. Carnegie Mellon University Research: Structural Mismatch Between Agent Development and Human Labor

A 2026 preprint from Carnegie Mellon University (CMU), How Well Does Agent Development Reflect Real-World Work?, further reveals fundamental flaws in the AI agent development field. Its core argument is that current agent development is programming-centric, whereas the economic value of human labor is highly concentrated in non-programming domains (e.g., management, coordination, decision-making, creativity, and high-touch services). This "structural mismatch" is consistent with the "benchmark-reality gap" discussed earlier.

• Data support: CMU’s task classification analysis of mainstream agent frameworks (e.g., AutoGPT, LangChain Agent, CrewAI) shows that over 70% of demonstrative use cases focus on three types of pure programming tasks: code generation, API calls, and data processing. However, data from the U.S. Bureau of Labor Statistics (BLS) shows that only about 2.5% of the U.S. workforce is employed in software development-related occupations (2019–2025 average). This means agent developers disproportionately focus on a narrow, low-employment-share skill domain.
• Economic consequences of the mismatch: This mismatch leads to extremely low penetration of agent technology in high-value, high-labor-density areas. For example, in healthcare diagnostics, judicial rulings, and complex project management — areas requiring tacit knowledge and ethical judgment — agent deployment rates are below 5% (CMU industry survey data). The resulting "investment-output" efficiency loss could be hundreds of billions of dollars — massive CAPEX poured into automation tools that cannot replace high-paying jobs.
• Conflict with the "replacement thesis": CMU’s conclusion directly undermines the narrative that "AI will massively replace white-collar workers." Even if agents reach superhuman levels in programming, the affected labor group is a very small fraction. Real replacement pressure should focus on auxiliary roles (e.g., junior data analysis, document review) that can be indirectly impacted by programming automation, not on decision-making layers.

2. Citadel Securities Report: "Counterintuitive" Recovery in Software Engineering Positions

Citadel Securities’ 2026 Global Intellectual Capital Crisis Report provides a key counterexample: software engineering job postings are recovering, in stark contrast to the popular narrative of "AI replacing developers." This phenomenon can be interpreted from three dimensions:

Data Dimension	Citadel Securities Observation (2026)	Popular Narrative (2023–2024)	Source of Difference
Hiring growth	Software engineering LTM hiring volume +12% YoY (Q1 2026)	Expected hiring decline of 20–30%	AI tools improve development efficiency, reducing per-unit maintenance cost; companies tend to expand teams to handle more complex system integration needs
Median salary	Senior software engineer NTM expected annual salary growth 8–10%	Expected stagnation or decline	Surging demand for "AI-enhanced" engineers who can harness agents; scarcity pushes up compensation
Skill requirement shift	Python/MLOps/GitOps share rose from 45% to 62%	Expected depreciation of basic coding skills	AI lowers the bar for low-level coding but increases demand for system design, agent orchestration, and cross-domain integration skills

• Core mechanism: Citadel’s analysis indicates that the proliferation of agents is creating compensatory roles — for example, roles that require humans to monitor, debug, and compliance-audit agent outputs. The increment of these new roles surpasses the old roles being automated (e.g., template coders). This parallels the 1990s when ERP software did not reduce IT jobs but instead spawned new professions such as SAP engineers and database administrators.
• Implications for investment perspective: This data supports the investment logic that "AI is a productivity tool, not a killer." For institutions heavily positioned in software (e.g., Voss Capital may focus on SaaS, fintech), the recovery in software engineer hiring means R&D OPEX will not collapse, but capital efficiency (revenue per employee) may improve. Investors should focus on companies that best integrate agents with human engineers in a "human-machine collaboration" model, rather than pure replacement automation vendors.

3. Analysis of Implicit Information in Terminology and Legal Disclaimers

Although the "Common Terms" section contains only standard financial abbreviations, its placement at the end of a fund’s "Introduction" suggests that the document’s target audience is institutional investors and that subsequent content will heavily involve financial modeling (e.g., DCF, EBITDA) and company valuation (P/E, EV/EBIT). This indicates that Voss Capital’s stance is not one of technological optimism but rather an evaluation of AI’s impact on specific company securities' value based on rigorous financial discipline.

The legal disclaimer conveys important risk notes:

• Performance calculation complexity: The disclaimer states that after January 1, 2020, performance calculation differences exist between the Master Fund and Feeder Fund (management fee 1% + performance fee 20%), and that due to tax differences and subscription/redemption timing, actual returns for individual investors may deviate significantly from reported performance. This reduces comparability of historical performance; investors should not directly extrapolate future performance.
• "High tolerance for uncertainty" in strategy: The disclaimer explicitly states, "The strategy utilized by Voss has a high tolerance for uncertainty." Combined with the extremely high prediction divergence in the AI field discussed earlier, this suggests that Voss may hold a substantial number of option-like positions (e.g., deep out-of-the-money calls/puts), making its portfolio volatility potentially significantly higher than market indices. This aligns with the current AI theme where capital chases high-growth, high-beta assets.
• Benchmark limitations: Indices such as the S&P 500 and Russell 2000 are explicitly noted to differ significantly from the fund’s holdings. Given that the document primarily analyzes AI agents’ impact on industries, Voss likely does not passively track indices but concentrates holdings in a small number of "AI winners" or "AI losers" to generate alpha. This concentration strategy is highly challenging in an AI environment with high uncertainty and low predictability.

Conclusion: From "Replacement Narrative" to "Mismatch and Migration Dynamics"

The new evidence from CMU and Citadel points to a more complex picture: AI agents do not simply replace developers; they first trigger a value mismatch (development capabilities misaligned with economic value) and then drive structural migration by creating new roles (e.g., agent operations, integration engineering). For hedge funds, this implies:

1. Do not easily trust linear predictions that "AI will destroy an industry"; instead, analogize the diffusion path of other general-purpose technologies (e.g., electricity, internal combustion engine) — they always first create compensatory demand.

2. Focus on companies engaged in "human-machine collaboration" rather than "pure automation"; their financial metrics (e.g., R&D expense ratio, sales efficiency) may outperform those that simply pile on AI tools.

3. The self-imposed constraint in the legal disclaimer suggests Voss is trying to lower investors’ expectations for a high-uncertainty strategy. If subsequent content contains reasoning on specific positions, it should be examined through the lens of the mismatch and role migration logic, rather than simply applying the outdated framework of "AI will replace X."