Future Technology[Daily Briefing] Nvidia buys Groq silicon 🧠+ a 2nm Mac mini for 899 💻
Rubin CPX was supposed to be the Nvidia answer to long-context inference: a purpose-built class of GPU for jobs where the model has to hold hundreds of thousands of tokens at once. At GTC 2026 it quietly disappeared from the public roadmap. It never reached production silicon.
What replaced it is the interesting part. Nvidia paid roughly 20 billion dollars to license technology from Groq, and the Vera Rubin rack-scale system now ships with the Groq 3 LPX rack, which entered full production in late August.
The swap is architectural rather than cosmetic. CPX was a conventional GPU approach, streaming model weights out of HBM as they were needed. Groq designs keep the weights in SRAM, on the chip itself. You give up capacity and you get latency back, which is exactly the trade you want when the job is serving tokens rather than training a model from scratch.
Nvidia defined training hardware. Spending 20 billion dollars to license somebody else instead of shipping its own part is a real admission about where the serving half of this market landed.
Why it matters: inference is now most of the compute bill, not training. The memory architecture that wins the serving half of AI decides what every model you use costs to run, and Nvidia just told you which one it thinks that is.
Quick hits
- Apple put a 2nm chip in an 899 dollar Mac mini. The M6 has a 12-core CPU, a 12-core GPU and two 16-core Neural Engines, while the 1,699 dollar M5 Pro version only has one. Full specs and the current listings.
- A perfect 10.0 Oracle flaw is being exploited right now. CVE-2026-21962 hits Oracle HTTP Server and the WebLogic Proxy Plug-in with no authentication required, and CISA added it to the exploited catalogue on Monday. What to patch.
- Amazon drone delivery jumps from 11 metros to nearly 500 towns. Packages up to 5 pounds, delivery in as little as 30 minutes, free for Prime members over 50 dollars. The city list.
- A 96 percent lunar eclipse happens early Friday. Partial phase runs 02:33 to 05:52 UTC on 28 August, peaking at 04:12. No telescope needed. Visibility map.
Tool of the day: Ollama
Ollama is the least painful way to run a language model on your own machine. One command to pull a model, one to run it, and it handles the quantisation choices you would otherwise spend an evening reading about. The honest limitation: it is a wrapper, and when something goes wrong you end up in llama.cpp territory anyway. It also will not save you from the real constraint, which is unified memory rather than clever software.
Rough rule: model size in billions of parameters, times 0.6, gives the gigabytes you need at 4-bit quantisation. Our full walkthrough is here, and if you are wondering which chip in your machine actually does the work, we explained NPU vs GPU vs CPU.
Question for you: if a 899 dollar desktop can run a decent model privately at home, would you actually cancel a cloud AI subscription, or is the frontier model still worth the monthly fee? Hit reply and tell me which one you would drop first.
Nath, Future Technology
Forwarded? Get your own free briefing at futuretechnologyhq.com/newsletter
Some links may be affiliate links.