THE NEW AI INFRASTRUCTURE STACK
Piper Sandler mapped the new AI infrastructure stack across 8 layers from the power grid all the way to the agent as AI scaling creates bottlenecks across power, cooling, networking, storage, cloud & inference:
Layer 1 | Power, Grid & Real Estate
•
$NEE building renewable power backbone hyperscalers need to scale AI
•
$CEG monetizing 24/7 nuclear power AI data centers increasingly require
•
$TLN building behind the meter nuclear power model for AI campuses
•
$IREN turning scarce power & interconnects into hyperscale AI capacity
•
$DLR owning physical real estate underneath global AI infrastructure
•
$EQIX owning interconnection layer where AI networks & clouds meet
Layer 2 | Physical Datacenter, Shell & Cooling
•
$VRT building 800V DC powertrain needed for next jump in AI rack density
•
$SMCI assembling liquid cooled rack systems that turn GPUs into usable compute
Layer 3 | Silicon & Hardware
•
$NVDA building compute platform underneath AI economy
•
$AMD building second full stack accelerator platform at hyperscale
•
$AVGO designing custom AI silicon behind Google TPUs & Meta MTIA as hyperscalers move deeper into proprietary compute
•
$GOOGL turning TPUs from an internal advantage into a hardware business sold directly into customer data centers
•
$INTC turning 14A into a real external foundry business with
$TSLA as an anchor customer and its largest reported foundry deal yet
•
$ARM moving into full server CPUs & capturing the full value of the chip instead of dollars per core
Layer 4 | Scale Up & Scale Out Networking
•
$ANET building Ethernet backbone connecting hyperscale AI clusters
•
$CSCO extending enterprise networking into AI data center fabric
•
$CLS manufacturing physical networking hardware connecting AI clusters
Layer 5 | High Performance Storage
•
$PSTG using DirectFlash to reclaim power & rack space for AI infrastructure
•
$DELL bundling servers & storage into an integrated enterprise AI stack
•
$NTAP connecting enterprise data estates to cloud AI workloads
•
$HPE bringing enterprise storage into an infrastructure as a service model
Layer 6 | Orchestration & Workloads
•
$IBM turning Red Hat OpenShift into enterprise control plane for private & hybrid AI deployments
•
$NTNX bringing AI infrastructure stack into private & on prem environments
Layer 7 | GPUaaS & Cloud
•
$CRWV proving GPU residual value with fully priced A100 capacity contracted through 2029
•
$NBIS building an AI cloud around 5GW of contracted power with customers prepaying 50% of capex
•
$DOCN bringing AI compute & inference below hyperscale cloud layer
•
$MSFT turning Azure into one of largest distribution layers for AI compute
•
$AKAM pushing AI inference closer to where users consume it
•
$AMZN monetizing AI through hyperscale cloud infrastructure & its own silicon
•
$ORCL building AI cloud capacity against a heavily concentrated backlog of large customer commitments
Layer 8 | Model Execution, Tokens & Agents
•
$NET becoming distribution layer between AI agents and the internet as non human traffic crosses more than 50% of its network
•
$FSLY moving inference closer to the end user through edge compute