Six AI Infrastructure Roles, One Data Flow: How MU, AVGO, MRVL, WDC, STX, NTAP Differ

Same AI cycle, six different jobs

The AI memory super-cycle gets bundled into one trade — Micron, Broadcom, Marvell, Western Digital, Seagate, NetApp. Each company actually sits in a different spot in the data flow. Think of an AI datacenter as a city: these six are fuel, highways, bridges, engine, warehouse, and traffic control. Before deciding where to put capital, you have to know what each one actually does.

1. Micron (MU): the fuel — HBM4 and DRAM

A GPU is useless if memory can't keep up. Micron makes that memory. HBM4 in particular is the backbone of next-gen AI accelerators from NVIDIA and AMD.

194.1% revenue growth tells you how valuable this seat is right now. Supply is constrained, demand is exploding. The 4-1 scorecard win versus the rest of the group makes more sense once you see this position from the inside.

2. Broadcom (AVGO): the highways — custom AI accelerators and networking silicon

Inside a datacenter, between GPU and memory, server to server, rack to rack — data needs highways. Broadcom builds them. Custom AI silicon (Google's TPU among them), Ethernet, InfiniBand-class networking — all Broadcom territory.

42.3% levered FCF margin shows how profitable this seat is. Best cash conversion in the group. AI revenue inside Broadcom is among the fastest-growing categories in semis, and it falls straight to free cash flow.

3. Marvell (MRVL): the bridges — custom silicon, CXL, optical interconnects

If Broadcom is the highways, Marvell is the bridges crossing them. CXL connecting memory and compute. Optical interconnects between datacenters. Hyperscaler-specific custom chips. Connectivity plus intelligence.

The catch is that this story has been priced ahead of the numbers. Marvell didn't win a round in the six-round scorecard. 7.6% CROIC and 32.6% margin trail the rest of the group. The seat is good — Marvell just hasn't fully monetized it yet.

4. Western Digital (WDC): the engine — active-tier HDD storage

AI training data isn't only in GPU memory. Frequently accessed petabytes live on HDDs. Western Digital, after spinning off SanDisk, is now a pure-play HDD company serving the engine room of active data.

35.6% margin is real evidence the spin-off is working. Capital efficiency and balance sheet still need a couple of quarters to validate under the new structure. The next two or three reports will determine how the market values this seat.

5. Seagate (STX): the warehouse — high-capacity HAMR storage

AI training corpus, backups, cold storage. Low access frequency, massive capacity. Seagate's HAMR (heat-assisted magnetic recording) lets each disk hold 30TB+ — making the company a key supplier for this seat.

27.1% revenue growth and 34% CROIC show the business is executing. The shadow over everything is 1,046.6% debt-to-equity. Warehouse businesses are capital-heavy by nature, but leverage at this level requires its own evaluation before sizing into a position.

6. NetApp (NTAP): the traffic control — data management software

Deciding where data should go and how to get there efficiently. NetApp is mid-transition from hardware company to data management software company, and 45.1% CROIC reflects how that pivot is showing up in capital efficiency. Best in the group on that metric.

The flip side: 236.1% debt-to-equity and 4.2% revenue growth. Best efficiency, weakest growth and balance sheet. Belief in the software pivot is required to size NetApp as a core position. Without that belief, treat it as a complementary one.

Which seat is actually the best?

Take the city analogy further: the most valuable seat is the one with the worst bottleneck. Right now that bottleneck is memory — Micron's seat. HBM4 supply is the limiting factor on GPU shipments, and that constraint isn't unwinding fast.

Broadcom's highways seat is attractive too, but highways have more room for new entrants — Marvell is targeting that exact territory. Memory has higher entry barriers. Stack all six seats up and the safest bet converges on the narrowest, deepest one. Which is to say: Micron.