Enterprise Data Platform

The question

What does the market have to believe about enterprise data infrastructure for today's prices to make sense, and is that imputed belief defensible against the base rate for data-plane vendors in prior technology cycles? Model weights are commoditizing on a roughly six-month half-life. Inference margins are compressing. The only durable economic rent in the stack sits where data is stored, moved, governed, and operationalized — the substrate beneath the model. This idea is a focused bet that the substrate layer captures a disproportionate share of the remaining value, and that a hard one-year-versus-SPY filter has already done most of the selection work.

Base rates

The reference class here is data-plane infrastructure vendors across three prior cycles: the relational-database buildout (1995–2005), the web-scale NoSQL and cloud-data wave (2010–2018), and the current AI-era data fabric. In each cycle, the distribution of outcomes was right-skewed and narrow: roughly one in five named vendors delivered a ten-year return in excess of the broad market, and the winners took multiples of the losers' value. Gross retention for surviving platform vendors clusters between 108% and 125% net revenue retention at scale. The unconditional base rate for a category-leading data vendor to be acquired or to IPO into durable compounding is ordinary — somewhere near 30%. Conditional on already having beaten SPY over a trailing year while the cohort as a whole was being rerated, the base rate for continued outperformance is meaningfully higher. That conditioning is the core of the margin-of-safety selection here.

Why the consensus view is wrong (or incomplete)

Consensus treats enterprise software as a single pool and marks it to the median multiple in that pool. This conflates two businesses with very different economics. Application-layer SaaS is being priced for a world where AI agents compress seat counts and workflow surface area. That fear is rational at the application layer and is probably still under-discounted. The data plane is the inverse: every new agent, every new model, every new workflow generates more writes, more reads, more governance load, more streaming throughput, and more developer artifacts. The information-efficiency argument cuts the opposite way from what most analysts are applying. The market is reading the whole sector through the application-layer lens and leaving the substrate layer trading at a discount to its probability-weighted cash-flow trajectory.

Position construction — name each holding's role

Palantir (PLTR, 20%) is the decisioning and ontology layer; it is the position closest to the application layer in the book and carries the highest expectations premium. We size it at the cap because further upside is now mostly a multiple story. Cloudflare (NET, 20%) is the edge data plane and the most credible second network-effect business in the portfolio; the developer-platform attach rate is the underappreciated driver. MongoDB (MDB, 20%) is the operational document store for AI-native applications, sized for the vector-plus-operational convergence. Twilio (TWLO, 19.6%) is the communications data exhaust — a customer-data asset the market is still valuing as a messaging toll road. Confluent (CFLT, 9.3%) is the streaming backbone; weight reflects tighter reference-class base rates and a narrower customer concentration. JFrog (FROG, 6.6%) is software-supply-chain governance — a small, asymmetric position on a secular requirement. DigitalOcean (DOCN, 4.5%) is the developer-SMB complement, a low-beta toehold with a different customer profile than the rest of the book.

Asymmetric payoff — rough EV framing

We model three scenarios over a three-year horizon. In the base case (55% probability), the book compounds at roughly 14% annually as the data plane continues to grow faster than its revenue multiples. In the upside case (25%), at least two holdings re-rate as durable compounders with NRR above 120% and the book delivers 25%+ annualized. In the downside case (20%), AI monetization stalls, customer consolidation intensifies, and the book tracks or lags SPY by 400–600 bps. Weighted expected value sits in the mid-teens, with a right-skewed payoff distribution that justifies the concentration.

Three things that would change our mind

First, sustained net revenue retention below 110% across three or more holdings for two consecutive quarters. Second, evidence that hyperscaler-native data services are taking share from independent substrate vendors at an accelerating rate rather than the observed steady state. Third, a compression of gross margin at the data-plane layer toward the inference-layer average — the sign that the substrate is no longer a rent-extracting position.

What we are explicitly NOT betting on

We are not betting on application-layer SaaS multiples recovering. We are not betting on any single model provider. We are not betting on generative AI revenue inflecting at a specific calendar quarter. We are not betting on the broader software index; the filter deliberately excluded most of the cohort. We are not betting on any specific AI-feature monetization inside these vendors' own product catalogs. We are not betting on a hyperscaler-native data services retreat; we are betting only that independent substrate vendors continue to take a share of incremental workload proportional to their existing footprint. This is a focused bet on the substrate, priced against the conditional base rate of names that already cleared a one-year performance hurdle against SPY.