SKYWAY SUMMIT
Investments · Topic Brief
The Intelligent Stranger
Two commentaries, one conclusion. A founder-investor's note tracing JEPA and the enterprise data layer, and 13D's Knowledge Monoculture, arrive from opposite directions at the same place: as foundation models commoditize into infrastructure, durable value migrates to what a model cannot copy. Proprietary data, tacit and relational knowledge, and expert judgment that carries liability. The Burry framework already in the brain calls the same boundary Ballard's Test. This is a lens for the source diet, not an entry screen.
Models commoditize. The edge moves to what they cannot reach.
The convergence is the finding. Three independent sources, written for different audiences, describe the same shift in where value sits as models get stronger.
The data note (JEPA to Enterprise Data). Every company now reaches the same small set of frontier models, so the model layer is becoming infrastructure, like electricity or cloud. A model is, in the author's phrase, an extraordinarily intelligent stranger that knows almost everything except you. What turns the stranger into a collaborator is proprietary context, and context comes from data. Value migrates to the data layer and the application layer.
The markets note (13D, Knowledge Monoculture). AI debases explicit knowledge, the codifiable and transferable kind, the way central banks debase currency. So value migrates to tacit knowledge: relational, embodied, context-dependent. The warning is structural, not philosophical, and it is covered in Section 03.
The framework already in the brain (Burry, Ballard's Test). Workflows whose output is language get commoditized; workflows resting on expert judgment that carries real-world liability resist substitution, because a model cannot assume the liability. Same boundary, drawn from the bottom up.
The three even share a casualty. Intuit, whose business was codifying tax and accounting expertise into software, is the example 13D reaches for to show explicit knowledge being commoditized, and it is the same name Burry rates a liability-anchored Castle. One company, three lenses, one conclusion: the durable edge is whatever the model cannot replicate.
JEPA, and why robotics may stay vertical without it
The first half of the data note is a bet on how AI comes to understand the physical world. Today's leading vision systems predict the next pixel; they learn the appearance of reality rather than its structure. Gravity, inertia, and collision do not exist at the pixel level, they exist in the relationships between abstract entities. JEPA (Joint Embedding Predictive Architecture) tries to predict the world at that abstract level instead, learning what matters and discarding what does not.
The stakes for the thesis are concrete. If the field stays on the current paradigm, robotics still succeeds, but mostly in vertical, structured domains: warehouses, factories, task-specific workflows with stable distributions. General-purpose robots that adapt to unfamiliar environments may require a world-model breakthrough. So the question is not whether robots get useful, it is whether they get general.
The cleanest expression of the world-models bet is AMI Labs, Yann LeCun's post-Meta venture built around JEPA. It is private: roughly a $1.03B seed at a $3.5B valuation, backed by Nvidia, Samsung, and Bezos Expeditions among others. There is no direct listed exposure. Public-market access to this leg of the thesis is indirect, through the compute layer and the listed backers, which is itself a tell about how early this is.
When everyone runs the same model on the same data
13D's market-structure argument is the sharper of the two for a trading framework. As AI adoption saturates, convergence happens beneath the model layer: even where the architectures differ, the inputs are the same earnings transcripts, the same news feeds, the same price histories, increasingly AI-filtered themselves. Outputs converge regardless of model choice, and dispersion across strategies compresses. The result is a monoculture, efficient under normal conditions and fragile under stress, the way a genetically uniform crop falls to a single pathogen.
The casualties are already measurable: 13D cites Goldman's prime brokerage estimating quant equity managers lost over 4% during otherwise benign mid-2025 conditions, a crowding signature. The conclusion, echoed by practitioners from Citadel to Goldman's quant desk, is that edge migrates to the frontier, to private information, experiential judgment, networks, and relationships, the domains AI cannot train on. Explicit processing becomes table stakes; tacit judgment becomes the differentiator.
Taken seriously, this thesis does not point at a stock to buy. It points at the brain itself. If the durable edge is proprietary synthesis plus the trader's own tacit judgment and network, then the moat is the process, the curated source diet, the documented framework, the relationships, not exposure to the commoditizing software names. The strongest application of this source is a reason to trust the system Skyway is building, and to be skeptical of buying the very tools that are converging.
A watchlist, with the thesis arguing against the obvious names
If the data and application layers capture value, the listed candidates cluster in three groups. The honest tension: the most obvious names are both richly valued and themselves exposed to the convergence and moat-erosion the thesis describes. This is a watchlist for the source diet, not a buy list. All figures are as of the source dates and require verification before any action.
| Name | Role in the stack | Current read | Access / vehicle |
|---|---|---|---|
| Data layer | |||
| SNOW Snowflake | Governed data cloud; agent control plane | Q4 FY26 product revenue ~$1.23B, +30%, NRR ~125%, RPO ~$9.77B. But the "open by default" shift (Iceberg, open catalogs) is stripping the storage lock-in that was the moat, the convergence risk in microcosm. | US-listed; CFD 5x OK |
| MDB MongoDB | Flexible DB + vector search for AI apps | ~27% revenue growth, Atlas ~29%, vector-search customers roughly doubling, but flagged Atlas softness on macro. Earlier-innings, higher-beta data play. | US-listed; CFD 5x OK |
| CFLT Confluent | Real-time data streaming (Kafka) for agents | Now inside IBM following acquisition; the streaming-for-agents exposure is consumed by a far larger, slower vehicle. | Via IBM (US-listed) |
| Application layer | |||
| PLTR Palantir | AIP; intelligence-to-decision workflows | Q1 2026 revenue ~$1.63B, +85%, US commercial +133%, NRR ~150%, ~57% adj FCF margin. Best-in-class growth, but ~55-67x forward sales prices in perfection. Extreme volatility. | US-listed; CFD 5x, but AP-7 vehicle risk acute |
| NOW / CRM / SAP ServiceNow, Salesforce, SAP | Agentic enterprise workflows | The incumbents embedding agents into finance, HR, procurement, service. SAP's "Autonomous Enterprise" is the reference case; its CEO's point, that agents create value only with organizational transformation, is the application-layer thesis stated plainly. | US / EU-listed; CFD 5x OK |
| Physical-world AI | |||
| AMI Labs LeCun, world models | Purest JEPA / world-models bet | Private. ~$1.03B seed at ~$3.5B. No listed exposure. | Not accessible |
| NVDA Nvidia | Compute under every layer; AMI backer | The picks-and-shovels expression of all three layers at once, and the only liquid way to touch the world-models leg today. Already heavily owned by the market. | US-listed; CFD 5x OK |
Samsung is also an AMI backer but is Korea-listed and not cleanly CFD-accessible from the current setup, the same access constraint that applies to mainland Chinese names. Confluent figures are pre-acquisition; the exposure now lives inside IBM.
Off-edge, and that is the point
None of these names sits in Skyway's confirmed edge zone, which is structural scarcity plus a Western policy catalyst. This is an AI-strategy lens, useful for reading the source diet and for sorting any AI name that drifts onto the watch list, not a screen that produces entries. Two anti-patterns govern any temptation to act.
AP-4, No-Thesis Trades. "The thesis says the data layer wins" is not a thesis for any single name. Each of these would need its own written structural driver, named catalyst, timeframe, and invalidation before becoming a trade. The convergence argument actively warns against buying the crowded, obvious expression.
AP-7, Right Thesis, Wrong Vehicle. These are long-duration, high-multiple, high-volatility accumulation names. PLTR at 55-67x sales is the clearest case. The fixed-leverage 5x CFD book is the wrong vehicle for any of them; cash equity or nothing.
The strongest signal in both pieces is not a name, it is a confirmation of the edge thesis itself. As models commoditize, value accrues to proprietary data, tacit judgment, and liability-bearing expertise, the things a model cannot copy. For a discretionary book that is a reason to lean harder on the curated source diet and the documented framework, and to treat the listed AI names as a lens rather than a watch list to chase. If any name does earn a thesis, it belongs in cash equity, not the CFD account, and the physical-world leg (AMI) is not accessible at all today.
Source ageing. The JEPA commentary references events from late April and May 2026 (AMI Labs, SAP Autonomous Enterprise, Anthropic Founder Playbook, Gemini, Snowflake Summit 1-4 June). The 13D piece is dated 21 May 2026. All company figures are as of their reporting dates and must be verified before any action. Private-company valuations are point-in-time and illiquid. This brief is research and analysis only, not an instruction to trade. The final decision is the trader's.