𝑻𝒉𝒆 𝒄𝒉𝒊𝒑𝒔 𝒘𝒆𝒓𝒆 𝒋𝒖𝒔𝒕 𝒄𝒉𝒂𝒑𝒕𝒆𝒓 𝒐𝒏𝒆.
The real constraints on scaling AI are no longer measured in teraflops alone. They’re measured in megawatts you can actually deliver, photons you can move at rack scale, liquid you can pump without melting the floor, and platforms that turn raw infrastructure into deployed, governed intelligence.
Jensen has been clear for months: the handoff is happening. Copper has limits. Power is the new gate. Cooling density is non-negotiable. The companies securing these layers with long-term contracts and balance-sheet capital are building durable positions.
Here’s how the 2026 AI infrastructure stack breaks down from silicon to intelligence:
LAYER 1: SILICON FOUNDATION
🟠
$TSM — The advanced foundry and packaging backbone (CoWoS and beyond). This is the capacity gate for the highest-performance AI chips. Everything upstream starts here.
LAYER 2: AI COMPUTE
🟠
$NVDA — Still the clear leader in GPUs and full AI systems/racks. The ecosystem investments and rack-level selling show they understand the game has moved beyond individual chips.
LAYER 3: SERVERS & RACK INTEGRATION
🔵
$SMCI — Building dense, liquid-ready AI server and rack systems that hyperscalers deploy at scale. Direct volume beneficiary of the buildout.
LAYER 4: NETWORKING & HIGH-SPEED INTERCONNECT
🔵
$MRVL — Switching silicon, DSPs, and aggressive silicon photonics push via Celestial AI. Essential for scaling clusters beyond single racks.
🔵
$AAOI — Datacom optics leader with explosive data center revenue growth and a major 800G/1.6T order backlog. Scaling domestic capacity while supply chain security matters.
LAYER 5: PHOTONICS & OPTICS TRANSITION
The physics is clear, copper hits the wall on bandwidth density and power. Photonics moves from “nice to have” to required for the next jump in cluster scale.
🔵
$LITE (Lumentum) — Leading provider of high-speed lasers, EMLs, and optical components with a major NVIDIA strategic investment and multi-billion dollar purchase commitments. Scaling US manufacturing for AI data center optics demand.
LAYER 6: POWER: THE BINDING CONSTRAINT
🟠
$VST — Signing multi-decade power purchase agreements directly with hyperscalers. Nuclear uprates and reliable generation capacity are becoming strategic assets.
🔵
$IREN — Renewable-powered data centers executing a clear pivot into AI/HPC cloud. Large-scale sites with secured power and flexible capacity. 5.8 GW-scale pipeline building.
LAYER 7: COOLING & THERMAL MANAGEMENT
🔵
$VRT — Liquid cooling systems are moving from pilot to must-have for high-density AI racks. Record backlog and order growth as power density per rack climbs.
LAYER 8: CLOUD & DATA CENTER PLATFORMS
🔵
$ORCL — Oracle Cloud Infrastructure is seeing massive AI-driven demand. Major capex ramp, OpenAI/Stargate ecosystem work, and enormous remaining performance obligations.
🔵
$NBIS — AI infrastructure specialist rapidly scaling contracted capacity (multi-GW targets). Major hyperscaler deals are driving strong ARR acceleration.
LAYER 9: THE INTELLIGENCE LAYER
🔵
$PLTR — Palantir’s AIP platform turns physical infrastructure plus enterprise data into production-grade, governed AI applications. U.S. commercial momentum is accelerating sharply.
Capital is already voting with real money: direct hyperscaler power deals, NVIDIA’s multi-billion dollar optics investments, pre-payments for capacity years out, and software platforms seeing accelerating commercial traction.
The honest counter-thesis is that real valuations across parts of the stack are elevated, power timelines can slip, cooling and photonics tech is still maturing in some deployments. Software ROI takes time to prove at scale. One digestive quarter can pressure the whole chain. But the structural, multi-year spend from the largest balance sheets in tech is not a narrative. It’s signed contracts and capex plans.
The market has aggressively priced the obvious chip layer. The asymmetry sits in the scarce layers everything else depends on.
Own the full stack. Not just the logo on the GPU.
Bookmark this. Then drop which layer you think moves first or which one I’m underweighting.
Tag the person who’s still only buying NVDA and calling it a diversified AI portfolio.