Underclassic   rss   archive

Issue 16: The Rope and the Dynamo

Permanent copy of issue 16. Today's issue is always the front page.

Central AI concentrates chips, power, and metered APIs. Local AI puts usable models on machines you already own. Efficiency cuts the cost of both--and decides how often you must rent.

Date: 2026-08-19
Updated: 2026-08-19 02:00
Status: LIVE

Broadcast

Grid queues in Texas now exceed 474 GW and hyperscalers face $1 T-plus capex commitments for 2026. At the same time, DeepSeek V4 Flash and Muse Glimmer run fully on-device, and list token prices continue to fall. Central supply is tightening while local alternatives gain usable performance. The decisive variable remains whether efficiency gains accrue to rented APIs or to machines already owned.

Watch

• Texas grid freeze on new data-center load
• DeepSeek V4 Flash on-device benchmarks
• Single-GPU open-weight releases

Stories

Data Center Backlash Widens as Texas Freezes 474 GW Grid Queue
Texas halts new data-center grid connections amid 474 GW of queued load. Power constraints now directly limit hyperscale buildout speed.
Grid queues tighten central compute supply and raise the relative value of local runtimes that bypass power constraints.
2026-08-12 / Natural Gas Intelligence / energy, chip / central

America's largest grid wants to cut power to new data centers first during shortages
50 MW-plus data centers must bring their own generation to avoid shutoffs. The rule formalizes local power responsibility for large AI loads.
Mandating on-site generation for large loads shifts some energy burden from shared grids back toward facility-level solutions.
2026-08-17 / Tom's Hardware / energy / central

DeepSeek V4 Flash Runs FULLY LOCAL On Apple Silicon
A new runtime lets DeepSeek V4 Flash execute entirely on-device without API calls. Users report instant offline inference on M-series chips.
On-device execution removes recurring token costs and grid demand, directly competing with central API usage.
2026-08-18 / Mshale / efficiency / local

Meta Muse Glimmer: the 30B Open-Weight AI Model That Runs Locally on a Single GPU
Muse Glimmer delivers competitive performance in a 30B open-weight package runnable on one consumer GPU. Release widens local inference options.
Open-weight models that fit single-GPU hardware lower the hardware threshold for capable local AI.
2026-08-13 / Pasquale Pillitteri / efficiency, chip / local

Axis
local [####################----------] central
Score: 67   clusters still winning   -1 toward local this week
Grid queues and hyperscaler capex still anchor most compute centrally, but open-weight releases and on-device runtimes are trimming token costs at the edge.

Forces

Line charts (0=local 100=central 2026-08-09 to 2026-08-24)

Market overlay (2026-07-26 to 2026-08-24)

Latest: BTC $77,183 / hashrate ~887.93 EH/s / token markers avg $5.3033/M / axis 62. These lines are different units. We compare timing, not dollars-to-hashrate.

Not shown: GPU cloud $/hour, hub electricity, miner hashprice, or lab cost per token -- none of those are a clean public feed.

Power and list prices
hashrate ~930 EH/s. List-price gap (not a subsidy %): 44.4x list gap (blended) GPT-5.6 Luna -> Claude Fable 5 ($0.45/M to $20.00/M). (mempool.space hashrate API)

Marker prices (USD per 1M tokens; verified 2026-08-19; blended = (3*input + 1*output) / 4; -- = open/self-host or no single public list)

Full 20-model board

If useful, support the broadcast.
bc1qfs3yw5qlq8sxs50crzh3g84ug27gvww9npu85n

home  |  rss  |  archive  |  models  |  data  |  board  |  mcp  |  llms.txt