Rulya (@Rulyaxd)

There's a subculture of Chinese hardware modders quietly turning the Mac Mini into the Raspberry Pi of local AI. Portable battery rigs. Custom cooling shrouds. 3D-printed docks with LED status widgets. English AI-Twitter hasn't noticed yet, iykyk. Frame one. A Mac Mini strapped into a hand-built rig sitting in front of a wall screen running Openclaw, a Chinese Claude Code gateway most people over here have never opened. Frame two. A different Mac Mini docked into a 20,000mAh 150W portable power station. Full mobile inference box for under $700 with the enclosure printed on a Bambu at home. Frame three. A Mac Mini clipped into a mini status display module. The screen reports GPU wattage in real time so you can watch the model actually thinking through the glass. Frame four. A Mac Mini shoved into a full-tower 3D-printed shroud with NASA-style venting and carrying handles. Apple silicon inside, community-designed chassis outside. Nothing here is for sale. The scene runs on shared STL files, Xiaohongshu build logs, and modded firmware repos on Chinese Github mirrors. Local AI is already a hardware-modding culture that ships weekly if you can read Chinese build logs. That's the whole gap. Everything else is discourse.

Apple runs its own AI models, its own silicon, its own map platform. And it still needs a guy in a red polo walking Malaysian sidewalks with a $50,000 LIDAR backpack strapped to his shoulders. Five states in Malaysia. Pedestrian-only zones where the Apple Maps cars can't drive. Areas around markets, courtyards, mosques where the roads stop but the sidewalks keep going. Somebody has to physically walk it. What's actually on the backpack. A 360 spherical camera on the top pole. A LIDAR unit for depth. GPS at the base. Combined cost of the rig runs around $50,000 in sensors. Contractors get hired through outsourced recruiting sites, not the Apple careers page. Apple doesn't want its logo on this part. This is the layer at the bottom of the AI stack that nobody writes threads about. Foundation models don't wake up knowing Kuala Lumpur's alleys. Somebody feeds them the alleys. That somebody is a guy in a red polo stacking bag as an Apple contractor while everyone else fights over the newest Claude release. There's a whole layer of AI-adjacent jobs that pay quietly while the discourse stays loud somewhere else. Pedestrian mapping. RLHF review. Data annotation. Voice recording for TTS training. And one more that most of AI-Twitter would refuse to admit even exists.

In March 1980, a Texas oil family owed $1.7 billion they couldn't pay by Thursday. Two months earlier, silver had hit $50 an ounce. The Hunt brothers owned a third of the world's supply. Their paper worth: $9.8 billion. Then COMEX changed the margin rules. New long positions banned mid-trade. Silver dropped to $10 in two weeks. Wild collapse. The corner didn't fail because it was stupid. It failed because the exchange rewrote the rulebook while the position was still open. Look at where AI compute sits in 2026. Nvidia owns roughly 90% of training silicon. TSMC is the only place it's manufactured. Claude, GPT, Gemini, Grok all run on the same pipe. Every AI operator right now is running a version of that same trade. The bet is that the pipe stays the way it is. That export controls hold. That custom silicon from Google and Amazon underperforms. That no open model quietly makes the pipe optional. The corner-lasts-a-decade crowd and the crack-is-already-here crowd. Both of them look right. Until one of them isn't.

Mac Mini on Metal: 55.64 t/s. NVIDIA rig, two cards: 83.65 t/s. Same model. Same quant. Guess where that gap actually comes from. Qwen3-Coder 30B (A3B active), Q4_K_M on all three boxes. The Mac isn't dying. It's ~33% behind a real GPU stack on token generation. On a chip that fits in a lunchbox. Where NVIDIA runs away is prefill. pp512: 2107 vs 563 t/s. That's raw compute. Prompt processing is compute-bound. Token generation isn't. tg is bandwidth-bound. M-series unified memory sits around 400 GB/s. RTX 3090 sits at 936. That's your ratio. tg128 confirms it almost line for line. Which is the whole reason a used 3090 stomps a $4k DGX Spark at 273 GB/s. And why the guys dropping $2–4k on Ryzen AI Max+ 395 keep asking where the tokens went. ngl the "best rig" everyone chases is the wrong axis. What's actually in your box?