Local compute wins: fast disks, local indexes, local models
"Postgres isn't slow, your storage is" struck a nerve. One commenter said network-attached storage only made sense with HDDs and that paying a 10x penalty on IOPS and latency for operational flexibility is ridiculous. Another moved RDS instances from r8g to r8gd, which adds local NVMe, and saw average I/O latency drop from about 1.0ms to 0.1ms.
The same instinct shows up in AI tooling. A local-indexing post argues that agents like Claude Code and Codex waste tokens by grepping, and commenters said they've seen about 20 similar tools, with new ones monthly. Someone shared Hugging Face's funes as their go-to. In the Ask HN about replacing Claude/GPT with local models, results were mixed: Gemma 4 on an M4 was too slow, while a pair of RTX Pro 6000 cards running DeepSeek V4 Flash hit 160 tok/s. Fun side signals: PS2 games statically recompiled to WebAssembly, and an unmodified Linux amdgpu driver running on an iPad.
The pattern: people are pulling work closer to the hardware to cut latency and cost.
So what?
Cheap wins exist in your stack today: check whether your database storage is the bottleneck before you scale up. In AI tooling, indexing and context management are crowded and fragile as a standalone product. If you build there, expect the big agent vendors to absorb it.
Read these
Postgres isn't slow, your storage is
Local Indexing to reduce tool calling and token usage
Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?
Show HN: PS2 games statically recompiled to WebAssembly
Running the unmodified Linux amdgpu driver on an iPad