A PC downturn frees leading-edge wafers, but not every AI bottleneck. Packaging and power cap accelerator upside; server CPUs benefit more because they use less silicon and no HBM. thediligencestack.com/p/titl…

AI hardware is becoming more tool-intensive even when wafer volume slows. Gate-all-around, backside power, HBM and advanced packaging each add steps, pushing our 2026 WFE base case to $153B. thediligencestack.com/p/chip…

In a power-capped AI market, uptime becomes capacity. Ten points of training goodput at a gigawatt-class site can be worth roughly 50–75 MW of effective capacity—years faster than adding new power. thediligencestack.com/p/prom…

A security agent can contain an attack and still harm the business if it acts incorrectly. The gating layer is authority: permissions, approval for sensitive changes, recovery paths and verified closure. thediligencestack.com/p/agen…

Credo’s advantage is owning SerDes, DSP, firmware and photonics. That lets it trade power, reach, reliability and qualification inside one design—but revenue still has to convert into gross profit and cash. thediligencestack.com/p/cred…

Enterprise agents create less storage than expected. The bigger bill is access: making data reachable, governed, permission-aware and current, then paying the system that authorizes each action. thediligencestack.com/p/the-…

Peak-cycle multiples hide different assumptions. Run current values backward: the market implies the deepest downturn for SK hynix despite its highest HBM exposure—the opposite of a higher-floor thesis. thediligencestack.com/p/bet-…

There is no universal CPU:GPU ratio. Host CPUs, dedicated agent racks, context length, scale-up topology and manufacturing limits all change the denominator. Useful sizing starts with the workload. thediligencestack.com/p/cpug…

Power is a small share of AI-campus capex, but an outage can strand the much larger compute investment. That makes redundancy, fuel delivery and contract length the real commercial constraints. thediligencestack.com/p/expe…

Can AI Pay for Itself? When the revenue covers the bill, why a rented gigawatt is not the problem, and what still has to happen thediligencestack.com/p/can-…

Optical demand is not the constraint. The gating path is qualified supply: 300mm laser integration could expand capacity, but volume yield, test and field reliability must clear before NPO and CPO scale. thediligencestack.com/p/exec…

Chip-equipment demand no longer tracks wafer volume alone. GAA, HBM and advanced packaging add process steps per usable wafer, while cleanroom space and tool output stretch the cycle across more years. thediligencestack.com/p/chip…

Compute becomes fungible only when software can place work across alternatives without giving back savings in deployment friction, networking or idle capacity. Economics follow the layers that keep control. thediligencestack.com/p/the-…

GPUs do the thinking; CPUs carry out the actions. More cores help parallel tools, faster cores shorten dependent steps, and memory bandwidth determines whether either advantage is usable. thediligencestack.com/p/agen…

A weaker PC market frees leading-edge wafers, but not every bottleneck moves with them. Advanced packaging caps accelerator upside, while server CPUs can absorb redirected wafer supply more directly. thediligencestack.com/p/titl…

Advanced packages multiply the cost of a late defect. Wafer-level burn-in moves screening ahead of HBM and package assembly, protecting more value before weak die become expensive system failures. thediligencestack.com/p/the-…

AI infrastructure grows only as fast as its most constrained layer. Capacity added in 2027 still needs qualification and yield improvement, so relief requires several layers to clear those gates together in 2028. thediligencestack.com/p/2027…

Similar peak-cycle multiples hide very different implied downturns. The widest discount sits with the supplier carrying the most HBM—the opposite of what a higher memory floor should produce. thediligencestack.com/p/bet-…

Abundant natural gas does not mean fuel at a data center. Pipeline extensions, compression and customer commitments must be financed and timed with generation, delivery and the compute opening date. thediligencestack.com/p/expe…

Better models can investigate threats, but authority sits with security platforms. The proof comes when governed agents complete remediation and verify exposure is closed without breaking the system. thediligencestack.com/p/agen…