type: pattern tags: [gross-margin, ai-inference, cloud-infrastructure, mix-shift, gpu, platform, margin-compression] confidence: medium created: 2026-04-01 source: DOCN stock-analysis Q4_FY25 persona: atlas provenance: legacy source_analysis_path: null source_paragraph_quote: null source_transcript_span: null source_loss_log_path: null

AI Inference Mix Shift Structurally Compresses Cloud Platform Gross Margins

Cloud infrastructure platforms pivoting to AI inference face a predictable, durable gross margin headwind as the inference revenue share grows. The mechanism: GPU-intensive inference workloads carry materially lower gross margins than core cloud services (Droplets, managed databases, storage) because GPU CapEx amortization, power, and cooling are large per-unit COGS line items. As inference revenue grows from a small share to a significant percentage of revenue, blended gross margins compress even if underlying unit economics are stable.

This is distinct from the SaaS-plus-hardware-device GM compression pattern (AXON). In that pattern, a software company adds a physical device attach. Here, a cloud infrastructure company is adding a higher-COGS compute product — both are cloud, but inference is structurally heavier on COGS than general-purpose cloud.

Evidence

• DOCN: Gross margin declined from 63% peak (FY22) to 58.7% (Q4 FY25) as AI inference grew to 12% of total ARR. Management explicitly confirmed: "AI margins are lower than core cloud margins, so you will have a little bit of a mix impact." FY26 EBITDA guidance (36-38%) is down from FY25 (42%) — confirming the compression is structural and frontloaded as AI mix grows. AI customer ARR at $120M growing 150% YoY will continue to be a larger share of revenue each quarter.
• The pattern is predictable: at 12% AI mix and ~58.7% blended GM, continued AI growth toward 20-25% of revenue implies further GM compression toward 55-57% unless inference-specific margins improve at scale.

Implication

When evaluating cloud platform companies pivoting to AI inference, build a GM bridge model: (1) estimate steady-state inference GM vs. core cloud GM, (2) project inference revenue mix at each quarter, (3) calculate the blended GM trajectory. Do not apply the legacy GM as a stable assumption — the mix shift is predictable and quantifiable. For valuation screens, be cautious applying the 60% gross margin threshold rigidly to cloud platforms mid-pivot; the relevant question is whether inference-specific unit economics are improving and whether scale reduces per-unit GPU costs over time.