10 posts on SDAD Writes.
Generator step-up transformer lead times passed 160 weeks by Q1 2026, up from about 52 weeks in 2020. Johns Hopkins projects unmet demand of 14.1 GVA for data centre transformers and 22.1 GVA for UPS by 2027, and Kimmeridge says as many as half of planned US data centre projects are at risk of delay or cancellation. FERC is fast-tracking grid connections. A faster queue does not conjure a transformer.
One side says 519ml per query. The other says 0.3ml. Meta's indirect water use alone was 19 billion gallons in 2024, more than 20 times its direct use, and one Meta facility in Georgia drinks 10 percent of an entire county's supply. Both camps are measuring a different quantity and both are correct, because nobody publishes the only number that matters.
A high-end desktop is Rs 1.5-7 lakh behind a 230 V wall socket nobody controls. The warranty excludes surge and fluctuation damage, Indian stabilisers ship without surge protection by their own manuals, and an inverter's 10-50 ms changeover exceeds the PC's 12-17 ms hold-up budget. This is the full technical and financial case for all three layers.
Two developments are breaking NVIDIA's CUDA monopoly: the Vulkan compute backend for llama.cpp (cross-vendor LLM inference on any GPU) and the Mojo programming language (a CUDA alternative that compiles to any accelerator).
Real Qwen3.8-27B numbers on the DGX Spark GB10: the memory wall caps single-stream decode at 11-12 tok/s, and speculative decoding is the only way out. SGLang with DSpark hits 34 tok/s, veloGB10 reaches about 40 tok/s, and NVFP4 keeps quality within noise of FP8.
VRAM is everything for local AI agents: the 4-bit cheat sheet runs 7B at about 5GB, 14B at 10GB, 30B at 20GB, and 70B at 40GB. Three builds cover it, from a $1,200-1,500 RTX 4060 Ti 16GB starter to the used 24GB-card sweet spot that runs 32B models at near-cloud quality for daily work.
Claude Code ranked the best local AI model for each VRAM tier from 4GB to 384GB - Nanbeige 3B at 4GB to Motif-3 314B at 384GB. Fact-checked July 2026: all five picks are real, available, and the 5.9GB Bonsai 27B even runs on an iPhone.
The memory wall ends here: NVIDIA's DGX Station pairs 252GB of HBM3e at 7.1TB/s with 496GB of LPDDR5X into one 748GB coherent pool on GB300 Grace Blackwell Ultra, adds NVFP4 quantization and 800Gbps ConnectX-8 networking (two boxes pool about 1.5TB), at roughly $94,930.
AMD's $1500 Strix Halo runs the same 120B model as Nvidia's $4700 DGX Spark, only 13% slower. But DGX Spark wins prefill by 3x and compute by 13x. Full benchmarks, India pricing, and a hands-on check on 2x DGX Spark GB10 with a 30B Q4 model.
Every AI system runs on one of five chip types — CPU, GPU, TPU, LPU, or NPU. Each makes a different tradeoff between flexibility and speed. This guide explains all five in plain English, with enough depth to actually help you decide which one matters for your work.
