The next decade of AI, mapped by what must physically exist at each stage
McKinsey, Goldman, PwC and Gartner publish the forecasts. Almost nobody maps them to the layers that must be built first. That’s the part investors can underwrite.
2026 is enterprise scale: roughly $725B of infrastructure spend, with tokens acting like refined electricity. The spend flows through $NVDA, $AVGO and $TSM, then lands on power and land underneath: $IREN, $DGXX, $TE and $EOSE.
2027 is frontier reasoning. Whatever the date, the physical requirement is fixed: bigger clusters connected by light. Interconnect scales with cluster size, so the constraint relocates from power to bandwidth: $AAOI, $CRDO, $MU and $SNDK.
2028 is physical AI. Intelligence leaves the chat window and needs eyes, ears and autonomy stacks: $ONDS, $MRLN, $AMBA and $OUST, with $TSLA and Figure assembling bodies.
2030 is GDP transformation. Value migrates from infrastructure to merchants selling finished AI by the token, including $DOCN, $NET and $AKAM. By 2035, orbital coverage and sensing add another layer through $ASTS.
Forecasts slip, CAGRs are sell-side and AGI dates are guesses. What does not slip is that every stage requires the one before it. Own the layers, size for drawdowns, and remember: past performance is not a promise. DYOR. Not FA.
