Research

Digital Surface Labs

AI Stack Value-Accrual — Where Value Accrues, and Why We Own the Consent Rail

Why intelligence commoditizes, the customer relationship + consented data is the scarce complement, and the box/custom-model are moat-deepeners not businesses
2026-06-18 strategyai-stackvalue-accrualaggregation-theorydata-sovereigntytrustmoats

The decision this informs

Whether DigitalSurface / Wintermute is standing in the right place in the AI stack, and how aggressively to invest up the stack (customer relationship, consent rail, data market) vs down the stack (a custom model, hardware, the box). The recommendation: climb up, rent the commodity layers, and own the two scarce assets — trusted users and the only consented, attested path to them. Do not train a foundation model. Treat the box as a moat-deepener, not a hardware business.

The verdict, up front

Intelligence is commoditizing. The capability gap between the best closed and best open model collapsed from 17.5 points in late 2023 to 3–5 points by early 2026, at roughly 10× lower price (open-vs-closed 2026). An input that cheap, that fast-falling, and available from five vendors plus open weights is not a moat — it's a rental.

Aggregation theory has been right for fifteen years: power flows to whoever owns the customer relationship and commoditizes suppliers (Stratechery). The AI era adds a second scarce asset next to the customer relationship that almost nobody owns: verified, consented, rights-cleared access to real humans and their data. Models get cheaper and smarter every quarter; what they starve for is trusted, permissioned, real-world data and the right to act for a real person. That is the trust-MD layer.

Layer-by-layer (case for → case against → verdict)

1. Open source vs closed source. For closed: frontier reasoning, rent the best. For open: gap is single digits at ~10× lower cost, and open weights are required for "data never leaves home." Verdict: closed as a bridge, open as the destination, model-agnostic always — never marry a model.

2. Dedicated-model company vs cross-model orchestrator (Harvey). Harvey (~$8B) went multi-model in 2025; its moat is the vertical data/workflow/trust stack, not the model — "foundation models are commodities; applied intelligence is a product" (Harvey deep dive; Harvey model offerings). Verdict: be the consumer-side Harvey — orchestrate models, own a consent/data moat for individuals.

3. Application layer. Thin wrappers are no longer funded; same revenue fetches ~8× (wrapper) vs ~25× (workflow + data moat); the prize is becoming the system of record (vertical-AI comp sheet; thin-wrapper collapse). Verdict: become the system of record — the consented vault + trust manifest where the data and the right to act live.

4. Consumer vs business. Business = higher ACV. Consumer = aggregation theory's prize, gated by CAC/trust — which the settlement wedge detonates: "the state owes you $347" is verifiable, zero-risk, viral. Verdict: consumer, because the settlement wedge is a CAC hack — buy trust + acquisition with the user's own found money, not ad spend.

5. Trust MD / the consent rail. As intelligence commoditizes, the bottleneck becomes trust between agents. trust.md + /.well-known/agent.json already describe agent-to-agent consent. The data-union literature says coordinated, consent-managed data selling is welfare-maximizing (Bergemann, Yale). Risk: protocols are hard to own. Verdict: highest ceiling / highest variance — play "open protocol, owned network": open the spec, own the graph of verified humans and the attestation/reputation layer that prices the data.

6. Climbing up vs down / the box / a custom model. Down = models, silicon, hardware — capital-intensive, commoditizing, contested by hyperscalers + NVIDIA. A custom foundation model is the worst use of capital on the list — do not build it. The box does three real jobs: removes the Gmail Limited-Use-Policy revenue cap (data-sovereign-email), is the ultimate trust signal, and is sovereign-tier lock-in — but our own economics work already concluded "the business has to be in the value the box generates, not the box itself… avoid the Helm trap." Verdict: climb up; the box is a moat-deepener priced to retain.

Two sharpenings of the current framing

  1. "Sell a box later" undersells the box. The monetization event is the consent-gated data marketplace (user paid ~90% per trust.md, we take a platform fee); the box unlocks its full margin by removing the Gmail/platform dependency. Lead with the marketplace, treat the box as the sovereignty upgrade.
  2. "Own the customer relationship" must name both scarce sides. Own the demand side (trusted users) and the supply side (the only consented, attested path to them). The moat is being the sole low-friction bridge between data buyers and rights-cleared real humans — a take-rate on a two-sided network where we own the scarce side, not a SaaS seat.

Value-accrual ranking (for us, highest → lowest)

  1. Aggregated demand of users who trust us (customer relationship).
  2. Proprietary consented + attested data asset reachable only through our consent rail.
  3. The trust/consent protocol if it standardizes (high variance, optionality).
  4. The box (retention / trust signal / removes Gmail cap; low-to-negative hardware margin).
  5. A custom model (value-destroying; do not build).

Shape of the thesis: rent the commodity (intelligence, compute), own the scarce complement (verified consented humans + the trust to act for them).

Sequenced roadmap

  • Phase 1 — Settlements as CAC + trust primer (now). Found money is the front door. (≈ "Year 1: cloud relay + KYC" in sovereign-stack.)
  • Phase 2 — Consent-gated data marketplace (the real business). User paid, we take a published fee. Build DP/aggregation + attestation rails now. (≈ "Year 2: hybrid mailstore.")
  • Phase 3 — Box as sovereign tier. Removes the Gmail cap, hardens marketplace margin, sells the premium "data never leaves home" trust tier. Priced to retain. (≈ "Year 3: full sovereignty.")

Sources and prior research