4090, M5 Max Still Trail Cloud

The blunt take: consumer hardware can cover a lot of real developer work, but it does not erase the gap to Sonnet, Opus, or Codex on hard coding tasks. The 4090 is the best single-box value for fast local inference; the MacBook buys portability and memory headroom, not magical model quality.

• –A 4090’s 24GB VRAM is enough for a surprising amount of daily coding work, especially with quantized 20B-35B-class models and agentic helpers
• –M5 Pro 64GB is the practical portability option; M5 Max 128GB is for fitting bigger models and longer context, not for outpacing a desktop GPU on raw throughput
• –Local models shine on privacy, offline use, log analysis, repo scanning, boilerplate, and sub-agent workloads where “good enough” matters more than perfect reasoning
• –For hard refactors, nuanced debugging, and highest-confidence code generation, cloud subscriptions still look like the rational default
• –The smartest spend is probably the hardware you already own plus selective cloud usage, unless your real need is always-on local experimentation or laptop-based inference