This piece explains a new way to invest in AI: a company's edge comes from a 'data feedback loop'—more users generate more data, which improves the model, which attracts more users—not from traditional 'moats' like hard-to-build software. Ash Fontana favors B2B AI startups built from scratch over big companies pivoting later. He highlights three holdings: Tractable (AI for car insurance claims, now superhuman accuracy), Lilt (enterprise translation where users correct translations to improve the model in real time), and Focal Systems (real-time inventory management using cameras and human labels to spot discrepancies).
Ash Fontana (Managing Partner at Zetta) discussed AI investment strategies in Invest Like the Best EP.90. The core argument: AI companies' competitive advantages come from "loops" (data feedback loops), not traditional moats. Fontana emphasizes that data itself is "dumb," but high-quality datasets c
Ash Fontana (Managing Partner at Zetta) discusses AI investment strategies in Invest Like the Best EP.90. Core thesis: The competitive advantage of AI companies comes from "loops" (data feedback loops), not traditional moats. Fontana emphasizes that data itself is "dumb," but high-quality datasets can fuel great companies. He argues that the key to an AI company's framework lies in feedback mechanisms, and that companies building AI from scratch (rather than later-stage pivots) will hold a significant advantage. He focuses on B2B AI enterprises, evaluating datasets based on their uniqueness and recyclability. The report also discusses AI regulation, how to prevent being overtaken by competitors, and the valuation challenges of AI models as intangible assets. Fontana is bullish on AI applications in finance and lists AI companies he favors.
Ash Fontana argues that the sustained competitive advantage of AI companies does not stem from traditional moats (such as the difficulty of building software), but from a "virtuous loop"—a closed cycle where data, models, and user feedback mutually reinforce one another.
Fontana points out that the moat of traditional SaaS is disappearing: "Software used to be hard to build... but now there are too many Lego bricks (APIs, cloud services) to piece together, and you no longer need the skills of an advanced computer scientist." (In other words: the barrier to building software has been significantly lowered.)
He proposes that the true moat for AI companies is the "loop"—a self-reinforcing feedback system. Specifically:
Fontana emphasizes that a single dataset plus a single model is unsustainable, as data may become outdated, replaceable, or insufficient in dimensionality. "If you only have one dataset and one model, sooner or later someone will come and beat you."
Fontana proposes five dimensions for evaluating datasets to assess their long-term value:
| Dimension | Definition | Ideal Characteristics |
|---|---|---|
| Uniqueness | Difficulty of acquisition | Hard to obtain, hard to replicate |
| Substitutability | Whether similar datasets can be used as substitutes | Non-substitutable |
| Dimensionality | Number of variables | Multi-variable (e.g., age + gender + income + ...) |
| Breadth | Coverage of population/scenarios | Broad and representative |
| Timeliness | Speed of data obsolescence | Depends on acquisition method: exclusively continuously obtained data with high timeliness is valuable; for one-time acquisitions, low timeliness is preferred |
Fontana adds: "When we look at companies, we record their rankings across each dimension. The real challenge is assessing a dataset's predictive power for a given problem and how that prediction translates into a competitive advantage."
Fontana proposes four key elements for evaluating AI models, which determine whether they can form a "runaway advantage":
1. Accuracy Threshold: Does the model reach or surpass human-level performance? Is there a path for improvement?
2. Data Critical Mass: How much data is needed to achieve effective predictions? If only a small amount of data is required, competitors can easily catch up.
3. Payoff Shape: Is it convex (large gains from correct predictions, low cost of errors) or concave (catastrophic cost of errors)?
4. Stability: Over time, will the model exhibit "erratic behavior" due to data bias? (e.g., Microsoft's Twitter bot turning racist)
Fontana emphasizes: "Once all four elements are in place, you have a runaway advantage, and fast followers have no chance at all."
Fontana argues that companies designed around AI from day one possess three structural advantages over incumbents undergoing later-stage transformation:
1. Product Design: AI-native companies design interfaces from day one to "collect user feedback data," rather than merely providing functionality
2. Data Rights: AI-native companies establish data-sharing agreements with clients from day one, whereas legacy firms like Salesforce once promised "we won't touch your data" and now need to renegotiate
3. Market Positioning: AI-native companies tell clients from the outset, "We are in this together, and you benefit from everyone's data"
Fontana uses Salesforce as an example: "When Salesforce entered the market, they convinced clients to put their data in the cloud, promising 'we won't touch your data.' Twenty years later, when data became extremely valuable, they had to go back and renegotiate."
Fontana divides AI technology adoption into four stages, each with distinct risk-return characteristics:
| Stage | Description | Example | Risk Characteristics |
|---|---|---|---|
| Consumer AI | Low-risk recommendations (started 10 years ago) | Netflix recommendations, Google Search | Extremely low cost of errors |
| Workplace AI Enhancement | Decision support (started 2010–2012) | Inside Sales (CRM recommendations) | Low cost of errors, high upside from correct decisions |
| AI Empowerment | Changes workflows, AI makes decisions with human oversight | Tractable (auto insurance claims) | Errors have real-world impact |
| AI Autonomy | AI makes decisions humans cannot (next 3–5 years) | Energy grid optimization, medical diagnosis | Errors can be catastrophic |
Fontana states: "Our job is timing. If the market is not ready, the company will not succeed. Currently, we spend most of our time on the third stage while preparing for the fourth."
Fontana explains why Zetta only invests in B2B AI companies, citing two core reasons:
1. Demand is Deductible: "B2B investing is more like science—you can call GE or Pfizer and ask, 'Would you use this product? What's your budget? Who approves it?' Then you can gauge demand. Consumer investing, on the other hand, requires predicting people's desires and trends, which is very difficult."
2. Data Ownership: "Most B2C datasets are already owned by giants like Facebook, Google, and Netflix. It is extremely challenging to obtain a sufficiently large consumer dataset."
Fontana adds: "What's most surprising is that the primary reason tech companies fail is not the difficulty of building the technology, but misjudging demand. B2B allows us to deduce demand more accurately."
Fontana acknowledges that the valuation of AI models as intangible assets is still in its early stages, currently relying more on talent value than the models themselves.
"Today, most AI companies are valued based on talent—a machine learning PhD from a top research institution is worth between $5 million and $20 million. This is not foolish; these individuals can indeed generate far more value than that for large companies."
Fontana is researching how to quantify the "residual value" of AI models: "Just last week, I was reading Michael Mauboussin's paper on measuring moats and ROIC, trying to apply it to data and AI. Currently, buyers in due diligence will try to apply the model to their own data to see if it can solve a specific problem—for example, an autonomous driving company tests whether a startup's pedestrian recognition model can integrate into its own model stack."
Fontana predicts: "In the future, metrics similar to SaaS's LTV/CAC will emerge to evaluate AI models. At that point, I will stop investing in this space—because it will mean it has become a commodity."
Fontana emphasizes that the key to fending off "fast followers" lies in identifying and waiting for the right moment when a "virtuous loop" takes shape.
He proposes a risk-adjusted investment timing framework:
Fontana cites Tractable as an example: "It took nearly three years to reach human-level accuracy, but only a few months to surpass it, and just weeks to achieve the final 5% of perfect accuracy. At that moment, I knew—the game was over. I stopped paying attention to competitors."
| Position | Guest Stance | Key Data |
|---|---|---|
| Tractable | Bullish (Invested) | AI for auto insurance claims; 3 years to reach human-level accuracy, months to surpass it, last 5% in weeks |
| Lilt | Bullish (Invested) | Enterprise-level translation; users can correct translations, model recalculates in real time |
| Focal Systems | Bullish (Invested) | Real-time inventory management; shopping cart cameras + human labeling + discrepancy detection |
| Inside Sales | Bullish (Invested) | CRM recommendation system; hundreds of billions of data points, predicts optimal call timing |
| Numerai | Bullish (Invested) | Financial data + crowdsourced prediction models + token incentives |
| ClearBit | Bullish (Invested) | Gmail extension collects data, provides company information via API |
| Cambridge Cancer Genomics | Bullish (Invested) | Mid-chemotherapy efficacy prediction; objective function: weekly treatment effectiveness |
| Invenia | Bullish (Invested) | Optimizes day-ahead supply-demand forecasts for the five largest U.S. power grids |
| Salesforce | Risk Warning (as a cautionary case) | Early promise of "not touching customer data" now makes AI transition difficult |
| Google/Amazon/Facebook/Netflix | Neutral (as background) | Own most B2C datasets; not interested in small vertical markets |
1. "Competitive advantage shifts from 'software is hard to build' to 'data loops'" (Ash Fontana) — The traditional SaaS moat is disappearing because building software has become easy; the new moat is a "virtuous loop": data → model → user feedback → more data → better model.
2. "Four factors for evaluating AI companies: accuracy threshold, data critical mass, revenue shape, stability" (Ash Fontana) — These four factors determine whether an AI model can form a "runaway advantage"; once all are in place, fast followers "have no chance at all."
3. "AI-native companies vs. transforming companies: three structural advantages" (Ash Fontana) — Product design (collecting feedback data from day one), data rights (sharing data from day one), and market positioning ("we are in this together") — all three are indispensable.
4. "Five-dimensional framework for evaluating datasets: uniqueness, substitutability, dimensionality, breadth, timeliness" (Ash Fontana) — The first four dimensions have clear direction (stronger is better), but timeliness depends on the acquisition method: exclusive, continuously acquired high-timeliness data is valuable, while one-time acquisition favors low timeliness.
5. "Four stages of AI adoption: consumer → workplace augmentation → AI-enabled → AI-autonomous" (Ash Fontana) — Risk ranges from low to high; currently, most opportunities lie in the third stage (AI makes decisions + human oversight), while the fourth stage (AI makes decisions humans cannot) is still 3-5 years away.
6. "B2B investing is more like science, consumer investing is more like art" (Ash Fontana) — B2B allows demand validation by directly asking customers ("Would you use it? What's your budget?"), while consumer demand is hard to deduce; moreover, B2C datasets are already monopolized by giants.
7. "AI model valuations are currently based on talent rather than the model itself" (Ash Fontana) — A top machine learning PhD is worth $5–20 million; model valuations are still in early stages, and metrics similar to SaaS LTV/CAC will emerge in the future.
8. "Investment timing: wait before the virtuous loop forms, but don't wait until competitors enter" (Ash Fontana) — A few weeks' difference in accuracy can be waited out, a million-point gap in data critical mass requires longer waiting, and a 12-month gap in value validation carries greater risk; investment must be completed before competitors enter.