The Investable AI Stack, Part I: The Great Split and Who Captures the Value

Claim 1 · empirical
Local models, often under 10 billion parameters, now account for more than 90% of open-source downloads.

What would break it: Open-source download statistics show models under 10 billion parameters accounting for 90% or less of downloads.

Claim 2 · mechanism
Value shifts away from undifferentiated model access and toward the surrounding stack: chips and advanced packaging, networking and memory, enterprise servers and private infrastructure, data preparation and governance, orchestration and workflow control, and security, permissions, and observability.

What would break it: Undifferentiated model-access providers capture the dominant share of AI value-chain profit while the surrounding stack layers commoditize and lose pricing power.

Claim 3 · mechanism
Alphabet's advantage is structural integration, not model quality alone: proprietary frontier models, DeepMind research, custom TPUs, Google Cloud, search distribution, consumer applications, proprietary data, and advertising monetization connect directly, reducing its dependence on third-party accelerators.

What would break it: Alphabet's AI deployment remains materially dependent on third-party accelerators, or its distribution and advertising surfaces fail to convert AI capability into cloud and ad revenue while its infrastructure spending produces no new revenue or profit-pool protection.

Claim 4 · empirical
Palantir's supplied figures show first-quarter 2026 revenue growth of 85% and a 112% increase in U.S. commercial remaining deal value.

What would break it: Palantir's reported first-quarter 2026 results show revenue growth other than 85% or a U.S. commercial remaining deal value increase other than 112%.

Claim 5 · empirical-forecast
Undifferentiated standalone frontier-model startups and mid-tier proprietary model vendors are the likely frontier-layer losers: they face enormous capital requirements, rising certification costs, and limited distribution versus the hyperscalers, while remaining too expensive for narrow tasks that can move to smaller local models.

What would break it: Standalone or mid-tier proprietary model vendors secure distribution and capital on terms comparable to hyperscalers and hold premium pricing and share in both frontier and narrow-task workloads.

Claim 6 · mechanism
AI increases the importance of data gravity: for many companies, moving a model closer to protected enterprise data is easier and safer than repeatedly sending sensitive data to an outside model, which favors Oracle's installed base of databases, regulated workloads, and systems of record.

What would break it: Enterprises predominantly migrate regulated data and systems of record off incumbent platforms like Oracle's toward external model providers, and Oracle's position in enterprise AI workloads shrinks.

Claim 7 · empirical-forecast
The market will segment: the largest models become premium reasoning utilities, while smaller models increasingly operate as embedded machinery, specialized and closer to the workflow.

What would break it: Frontier models absorb routine narrow tasks (invoice reading, document classification, ticket routing) at prices that make local embedded deployment stagnate or shrink.

Claim 8 · thesis
The AI market is forming a barbell: capital-intensive frontier platforms run by hyperscalers at one end and compact local models at the other, with the middle (general-purpose model vendors without frontier-level performance, proprietary hardware, enterprise distribution, or a defensible path to deployment) vulnerable.

What would break it: Mid-tier general-purpose model vendors without frontier performance, proprietary hardware, or enterprise distribution durably gain enterprise share and sustain premium pricing against both hyperscalers and local models.

Claim 9 · mechanism
Meta's open-weight models can weaken the pricing power of closed-model providers while improving Meta's own advertising, recommendation, content, and engagement systems; Meta's largest benefit is internal value capture, not direct model licensing revenue.

What would break it: Closed-model providers' pricing power is unaffected or rises as Meta's open-weight models spread, or Meta's AI value shows up primarily as standalone model licensing revenue rather than improved core-platform economics.

Claim 10 · mechanism
Microsoft is less dependent on the long-term dominance of any single model provider than is often assumed, because enterprises consume AI through its existing software contracts, identity, permissions, security controls, and workflows (Azure, Microsoft 365, GitHub); the model can change while the surrounding control system remains in place.

What would break it: Microsoft's AI revenue and margins prove tightly coupled to one model provider's dominance, or enterprises predominantly bypass Microsoft's stack to consume standalone models directly.

Claim 11 · mechanism
The durable enterprise product is the control plane, not the model: the systems that connect models to corporate data, permissions, applications, and auditable workflows (Microsoft's agent-control environment, Palantir's governed orchestration, ServiceNow's workflow ownership) are harder to dislodge than models, which become cheaper and easier to replace.

What would break it: Enterprises routinely replace their workflow, permissions, and orchestration platforms when switching models, and durable profit accrues to model providers rather than to the owners of workflow and systems of record.