The Investable AI Stack, Part II: Own the Hard Assets

Entry 25 of 32 · C177 · AI CAPEX · April 13, 2026
Power and grid delays persisted, supplier backlog grew faster than revenue, and Ciena backlog reached about $8.5B by fiscal Q3 2026 against $1.67B quarterly revenue.

[On the track record · Article · Original post on X]

Entry 26 of 32 · C178 · AI CAPEX · April 13, 2026
April and May 2026 data put 30% to 50% of planned 2026 United States data center capacity in delayed or canceled status, with the shortfall slipping past 2027.

[On the track record · Article · Original post on X]

Claim 1 · thesis
Independent general-purpose model companies without proprietary chips, unique data, or direct distribution face pressure from both ends of the market: hyperscalers are better positioned for expensive, high-liability frontier workloads, and smaller local models can take on repetitive, privacy-sensitive, and cost-conscious tasks; without a defensible advantage such as exclusive data, embedded distribution, a regulated vertical, superior workflow integration, or a hardware channel, the model risks becoming a replaceable component.

What would break it: Independent mid-tier general-purpose model vendors lacking proprietary chips, unique data, direct distribution, or another named defensible advantage sustain durable pricing power and margins while both the hyperscale frontier and local small-model ends of the market grow.

Claim 2 · thesis
The AI market is developing a barbell structure: hyperscalers continue to invest heavily in frontier systems, while enterprises are beginning to deploy smaller models for narrow, repetitive, privacy-sensitive, and cost-conscious workloads.

What would break it: AI deployment concentrates in only one mode at scale, with the other side of the barbell failing to materialize: frontier hyperscale spending collapses, or enterprise deployment of smaller local models remains negligible.

Claim 3 · empirical
High-bandwidth memory is a critical component of advanced AI accelerators, and memory remains a cyclical business: new supply and normalized pricing can compress margins even when long-term AI demand remains intact.

What would break it: High-bandwidth memory pricing and margins remain permanently elevated through a full new-supply cycle with no compression, or high-bandwidth memory ceases to be a critical component of advanced AI accelerators.

Claim 4 · mechanism
TSMC is largely neutral about which chip architecture wins: it manufactures frontier accelerators, custom AI chips, and edge processors while providing increasingly important packaging capabilities, so greater manufacturing complexity can benefit TSMC whether the market favors Nvidia GPUs, custom accelerators, or edge chips.

What would break it: TSMC's advanced-node and packaging business fails to benefit as AI chip complexity grows, for example because a competing foundry captures the advanced AI manufacturing share or AI chip volumes shift to less complex processes.

Claim 5 · mechanism
As autonomous and semi-autonomous agents spread across enterprise environments, companies will need to authenticate activity, segment networks, control access, monitor behavior, detect anomalies, and preserve audit trails, so distributed AI increases the importance of the infrastructure surrounding the models.

What would break it: Enterprise agent deployment scales widely without a corresponding increase in networking, security, and observability requirements or spending.

Claim 6 · thesis
The strongest choke points in the AI stack remain advanced compute, semiconductor manufacturing, networking, enterprise permissions, operational data, and deployment infrastructure, and the more durable opportunity lies in the infrastructure and control layers surrounding model competition rather than in any single model provider.

What would break it: Durable value accrual concentrates in a single model provider while suppliers of advanced compute, manufacturing, networking, permissions, operational data, and deployment infrastructure fail to capture durable economics.

Claim 7 · mechanism
Both sides of the barbell require substantial physical infrastructure, so infrastructure offers exposure to AI growth without requiring investors to predict which individual model will dominate.

What would break it: AI infrastructure suppliers' growth turns out to depend on a single model provider's success, so that the physical layer fails to benefit across different model architectures.

Claim 8 · mechanism
Broadcom benefits when hyperscalers seek alternatives to infrastructure built entirely around Nvidia, because custom chips can improve cost, performance, and power efficiency for stable workloads running at enormous scale.

What would break it: Hyperscalers abandon custom accelerator programs and consolidate their infrastructure on Nvidia's standard GPUs, eliminating the demand Broadcom's custom-chip exposure depends on.

Claim 9 · empirical-forecast
As AI clusters become larger, the connections between accelerators and data centers become more demanding, and Arista benefits from the growth of Ethernet-based scale-out networking.

What would break it: AI cluster growth does not increase demand for Ethernet-based scale-out networking, for example because proprietary interconnects dominate cluster buildouts or interconnect spending stagnates as clusters grow.

Claim 10 · mechanism
Nvidia's advantage increasingly lies in the platform rather than the GPU alone: customers are buying an ecosystem that includes software, optimized libraries, networking, engineering support, and a large developer base.

What would break it: Custom accelerators, alternative architectures, and customer concentration weaken Nvidia's pricing power at scale, with customers switching away despite the ecosystem, showing the platform confers no durable advantage.

Claim 11 · mechanism
Dell is a practical beneficiary of private, local, and sovereign AI deployment: it sells the servers, storage, workstations, integration, and support needed by organizations that want AI capacity under their own control, and the requirements of on-premises processing, data sovereignty, predictable inference costs, and long-term support favor established vendors with enterprise relationships and service capabilities.

What would break it: Private, local, and sovereign AI deployment fails to grow materially, or enterprises assemble and maintain such systems without established vendors' integration and support, leaving Dell no deployment-complexity business to capture.