Five outlets ran the same weekend comparison, led by Nasdaq: Salesforce (CRM) against Dell Technologies (DELL), software-as-a-service against physical AI infrastructure. The framing treats the two as competing bets, but they occupy different floors of the same stack, and the margin structure of that stack is set by whoever controls the scarcest component. The bill of materials tells the real story here. Every claim Salesforce or Dell makes about AI growth resolves into a question about memory supply, power delivery, and cost per inference.

The Spec

Dell's AI business assembles accelerators, high-bandwidth memory, networking silicon, and liquid cooling into rack-scale systems it ships to enterprises. That makes Dell a systems integrator whose revenue per rack is large but whose margin depends on component costs it does not set.

Salesforce sits at the opposite end of the deployment pipeline, consuming compute through cloud contracts and selling recurring software seats on top. Its gross margin per seat is structurally high, but every AI feature it ships adds an inference cost line that scales with usage.

The model layer between them keeps getting cheaper. Alibaba (BABA) launched its flagship Qwen3.8 Max this week while Moonshot widened developer access to its own models, per Yahoo Finance. Each capable open or low-cost model compresses the price of a token of output. Compression at the model layer lowers Salesforce's input cost and simultaneously raises total inference volume, which is demand for Dell's racks.

The Bottleneck

This cycle's constraint is not accelerator logic. It is the memory bonded to it and the power feeding it. High-bandwidth memory requires stacking DRAM dies with through-silicon vias, a process concentrated among only three manufacturers worldwide. That concentration hands the memory layer pricing power that neither the integrator below it nor the software vendor above it can escape.

The International Business Times framed SK Hynix this weekend as the direct expression of AI memory demand, set against SpaceX as the speculative alternative. Memory ships into every accelerator architecture regardless of which logic vendor wins, which makes it the least substitutable line item on the bill of materials.

The second constraint is electrical. Vera Silva, chief strategy and technology officer at GE Vernova's electrification unit, told The Financial Express that "the power grid must get smarter before it gets bigger," citing AI data center load. Rack power density rises with every accelerator generation, and liquid cooling adds its own draw. A rack that cannot be energized generates zero revenue for Dell and zero inference capacity for anyone renting it.

The Unit Economics

Run the Salesforce-Dell comparison on per-unit cost rather than growth narrative. Dell earns integration margin on each system, but that margin is squeezed between concentrated component suppliers and price-sensitive enterprise buyers. Volume growth in AI servers can therefore expand Dell's revenue faster than its profit.

Salesforce's unit is the seat, and its AI economics hinge on inference cost per query falling faster than feature usage rises. The Qwen3.8 Max release and Moonshot's developer push both work in Salesforce's favor on that axis. Cheaper models mean the software layer captures the falling input cost as gross margin, provided customers keep paying the same seat price.

The memory supplier faces neither squeeze. It sells a component with three producers and no near-term substitute into a market where every architecture needs more of it. Pricing power on the bill of materials flows to the node with the fewest alternatives, and right now that node is stacked DRAM.

The Inflection

The next inflection is the migration of AI spending from training clusters to enterprise deployment, and each company proves its claim differently. Dell's proof point is its next earnings cycle: AI server backlog converting into shipped systems at stable or improving margin. If backlog converts at thin margin, the integrator thesis weakens even as revenue grows.

Salesforce's proof point is disclosure that AI features lift revenue per seat without a matching rise in cost of revenue. Falling model prices make that outcome more plausible each quarter, but it remains unproven until the margin line confirms it.

SK Hynix's proof point is the simplest: sustained memory allocation and pricing through the coming reporting cycle. If the thesis is that enterprise AI deployment accelerates through 2026, the cleanest expression lives in the component and power layers — memory suppliers and grid equipment vendors — with Dell as the volume play and Salesforce as the margin play one step removed.

Follow the wafer allocation and the substitution risk at each layer. Software competitors multiply monthly, server integrators number in the dozens, and stacked-DRAM producers number three. On that cost curve, SK Hynix owns this node, and the Salesforce-versus-Dell debate is a downstream argument about who pays it.