13 Pieces Latest · 2026-07-01
Static archive fallback
Meta's reported plan to resell excess AI compute reclassifies AI infrastructure economics: the same GPU capacity that was a capital burden on Meta's books becomes contracted revenue on Nebius's and CoreWeave's books — until Meta can supply it too.
Micron's FQ3 2026 was not just a record earnings quarter. The stronger signal was contractual: multi-year Strategic Customer Agreements, take-or-pay commitments, pricing floors, customer deposits, and RPO disclosure suggest that AI demand is turning part of memory supply from a spot-cycle input into contracted scarcity.
Hyperscaler capex is growing faster than the cloud revenue it funds, and depreciation is catching up — a filing-based look at whether cloud CapEx is translating into demand or into balance-sheet burden across Microsoft, Amazon, Alphabet, and Meta.
A framework essay examining why inference efficiency gains reshape rather than eliminate AI infrastructure demand — tracing how lower cost per token expands deployment, shifts workload mix from training toward inference, and changes which parts of the stack face sustained load.
A framework essay arguing that 'AI exposure' is a category mistake: AI demand changes economic form as it moves through the value chain — becoming platform leverage at NVIDIA, capital obligation at TSMC, cycle risk at Micron, and volume infrastructure at Vertiv. Treating these as the same economic category produces bad analysis.
Part III extends the series into market sorting: once the AI constraint stack is visible, markets begin to separate bottleneck control from thematic exposure.
A series essay on how frontier model bottlenecks propagate outward into GPUs, HBM, advanced packaging, interconnect, optical links, and data center infrastructure.
A framework essay on why frontier model progress is increasingly constrained by system-level scaling limits rather than algorithmic improvement alone. This essay traces the path from Transformer compute shape to memory, bandwidth, interconnect, and heavier inference budgets.
Recent volatility may reflect a timing mismatch: shipment timing is being read as demand erosion. This note uses a constraint-based framing to explore two physical constraints on revenue recognition—advanced packaging ramp dynamics and downstream cooling retrofit lag.
A public-data note using public filings and operating metrics to analyze Palantir's business model and operating efficiency.
A public-data note on Oracle's cloud investment, capital plan, and RPO/revenue timing.
A public-data note on Oklo's power-market narrative, regulatory timeline, and construction constraints.
Personal 2026 analytical notes for tracking how AI exposure converts into physical capacity, capital intensity, revenue quality, and issuer-level financial evidence.