Teams sizing "Mac for AI dev" often hit two traps: wait for M5 before building the pipeline, or run Runner + Ollama together on a laptop. This post follows the main causal chain and covers how to split local Mac and Cloud Mac—breaking the degradation path on the right of the diagram, not a rent-vs-buy false choice.
Vs Cloud Mac vs local Mac workstation: here we focus AI-dev "rent first vs wait for M5"—not CodeGraph / agent architecture.
Two common traps
Trap 1: "M5 will fix everything." If the bottleneck is swap and CI drift (main guide formula), same-RAM M5 cannot replace scheduling or 24GB.
Trap 2: "Cloud Mac = remote desktop." For AI dev it is a dedicated macOS node for GitHub Runner and Ollama—not mirroring your laptop screen.
How to split local and cloud
| Where | Typical tasks |
|---|---|
| Local Mac | Code, review, Claude Code |
| Cloud Mac | Runner, xcodebuild, signing |
| Cloud Mac or off-peak | Ollama embedding, batch inference |
After the laptop lid closes, Runner and Ollama should keep running in the cloud—that is why many AI teams add Cloud Mac.
When memory is tight—do this first
Using the main guide's upgrade pressure estimate, when result is clearly > 0:
- Schedule —
ollama stopbefore CI (runbook) - Partition — Runner on Cloud Mac, Ollama on nightly batch
- Upgrade RAM — 24GB hardware or long-term larger Cloud Mac tier
"Wait for M5" is not in the top three—unless headroom is fine and only tok/s is low (pure compute bound).
When renting Cloud Mac beats waiting for M5
- Pipeline gap — 6–9 months until keynote, but iOS CI and agent experiments must continue; daily Cloud Mac runs Runner execution now.
- Purchase uncertainty — validate on M4 24GB Cloud Mac with Ollama benchmark and pressure formula before choosing 16 vs 24GB hardware.
- Peak load — release-week CI doubles; temporary Cloud Mac nodes beat replacing the whole machine.
When to still buy a physical M4
Low-latency local IDE for daily interaction; or 24GB always full and Cloud Mac monthly exceeds depreciation. Even with owned hardware, keep Runner + inference partitioned—avoid "one Mac runs everything."
Series navigation
ZavCloud
Code locally, build and infer on Cloud Mac
Dedicated Mac mini M4, native macOS, static IPv4—validate Runner and Ollama daily without waiting for M5.
View plans and pricing