Underclassic   rss   archive

Issue 25: Apple Ships 512 Gb Desk Inference

Permanent copy of issue 25. Today's issue is always the front page.

Apple just put 512GB of unified memory in a desk Mac and said it can run hundred-billion-parameter models without a token meter. Hyperscalers still spend more than they take in on cluster chips, so the money remains central.

Date: 2026-08-27
Updated: 2026-08-27 02:00
Status: LIVE

Broadcast

Four data points this week. Apple shipped 512 GB unified memory in a desktop box. Three hyperscalers spent 102 % of cloud revenue on AI chips. The United States now draws 40 % of global data-center power. Meta released an open-weight model sized for single-GPU machines. Central clusters still absorb more capital and kilowatts than they return. Local silicon and open weights are the only visible counters. Grid queues in Texas and Washington continue to lengthen. No new central capacity will come online faster than the power permits. The economic center remains inside the halls, but the inference surface is spreading to desks.

Watch

• Texas grid connection freeze through Q4
• Apple M5 Ultra volume shipping by October
• Meta Muse Glimmer adoption on consumer GPUs

Stories

Apple introduces M6 and M5 Ultra for a big leap in performance and AI compute
Apple launched M6 Mac mini and M5 Ultra Mac Studio with up to 512GB unified memory. The company claims the systems can run hundred-billion-parameter models locally.
A single desk machine now holds enough memory to run frontier-scale models without a metered API, shifting inference cost and control to the edge.
2026-08-26 / Apple / chip, efficiency / local

AI's Absurd Spending Boom? Hyperscalers Are Spending 102% of Cloud Revenue on Capex
Microsoft, Meta, and Google reported combined AI-related capital spending above cloud revenue. The gap shows central clusters still require more cash than they return.
Spending above revenue means the economic center of gravity remains inside hyperscale data centers, not at the desk.
2026-08-22 / Yahoo Finance / chip / central

The U.S. Now Uses Nearly 40% Of The World's Data Center Electricity
Data centers consumed 40% of global electricity in 2026, with the United States taking the largest share. Grid constraints are tightening around new central builds.
Power limits slow central cluster growth and raise the relative value of local silicon that does not need new grid connections.
2026-08-23 / Forbes / energy / central

Meta Releases Muse Glimmer, an Open-Weight AI Model Designed to Run on Consumer Hardware
Meta published an open-weight model sized for single-GPU or high-memory desktop machines. The release lowers the barrier for local inference without API calls.
Open weights that fit on consumer hardware shift inference away from central APIs and toward devices users already own.
2026-08-26 / Campus Technology / efficiency / local

Axis
local [################--------------] central
Score: 55   clusters still winning   -2 toward local this week
Most AI still runs in big central data centers. The new fact is a desk Mac with 512GB of unified memory, which is a local-inference box, not a campus.

Forces

Line charts (0=local 100=central 2026-08-12 to 2026-08-27)

Market overlay (2026-07-29 to 2026-08-27)

Latest: BTC $78,850 / hashrate ~888.9 EH/s / token markers avg $5.3033/M / axis 55. 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
BTC $78,850 / hashrate ~889 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 + Coinbase/CoinGecko)

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