Underclassic rss archive buyer's guide
Permanent copy of issue 75. Today's issue is always the front page.
Previous issue | This week | Next issue | archive
Meta released Muse Spark Open on Thursday, an 816-billion-parameter open-weight model it says can run on two Apple M5 Ultra Mac Studios. The same day NVIDIA published Nemotron 3.5 Bolt for local agents as RTX Spark laptops go on sale. On the power side, American Electric Power committed about $4 billion to a 702-megawatt fuel-cell plant for Google in Wyoming, and NERC advanced mandatory reliability standards for large computational loads.
Date: 2026-10-16
Updated: 2026-10-16 10:15
Status: live
Broadcast
Two open model drops aimed at desks land on the same morning the utilities write bigger checks for hyperscaler megawatts. Meta says Muse Spark Open runs on a pair of Mac Studios; NVIDIA says Bolt fits in 24 GB on an RTX Spark laptop. Meanwhile AEP is building a $4 billion fuel-cell plant for Google, and NERC is one board vote from filing mandatory rules for large computational loads. The desk gained real software this week. The grid still writes the larger invoice.
Watch
Stories
Meta published weights for Muse Spark Open, an 816-billion-parameter multimodal model with 42 billion active parameters and a 1-million-token context window. The company says two M5 Ultra Mac Studios with 512 GB of unified memory each can run the model locally through Apple MLX.
-- A frontier-class open US model that fits on a desk setup, even one that needs two Mac Studios, pulls coding and long-horizon agent work off the public cloud. Local inference keeps the weights and the data under the operator who owns the hardware.
2026-10-16 / Meta AI Research / efficiency, chip / local / #meta-releases-open-muse-model-built-to-run-on-ap
NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter agent model, and Nemotron 3.5 Bolt 74B-A9B, a 24 GB edge model for local agents. Bolt is built for RTX Spark laptops that go on sale October 16 and also runs through llama.cpp, Ollama, LM Studio and vLLM.
-- Open weights tuned for a laptop memory budget make everyday agent work runnable without a data-center lease. The same company that sells the cluster racks is now shipping the desk models that compete with those racks for coding and tool use.
2026-10-16 / NVIDIA Blog / efficiency, chip / local / #nvidia-ships-nemotron-bolt-for-local-agents-as-r
American Electric Power said it will invest about $4 billion in a 702-megawatt Bloom Energy fuel-cell project near Cheyenne under a 20-year power purchase agreement with Google. Construction is set to start in 2026, with full operation expected by the end of 2028.
-- A utility is building dedicated generation for one hyperscaler rather than waiting for the shared grid to catch up. That pattern locks more AI work into central campuses that can underwrite their own megawatts.
2026-10-16 / American Electric Power / energy / central / #aep-plans-4-billion-fuel-cell-plant-to-power-goo
The North American Electric Reliability Corporation said its Project 2026-02 Computational Loads Reliability Standards passed final ballot on October 15 and will go to the NERC Board of Trustees for adoption. After board adoption, NERC will file the standards with the Federal Energy Regulatory Commission for US approval.
-- Mandatory ride-through and coordination rules for large AI and crypto loads raise the cost of dropping onto the bulk power system without planning. Clusters still grow, but they face a thicker regulatory bill than a desk machine that never touches a 100 kV line.
2026-10-16 / NERC / energy / central / #nerc-computational-load-standards-pass-final-bal
Axis
local [#########################-----] central
Score: 84
clusters still winning
tilted local
Meta and NVIDIA both shipped open models aimed at desks and laptops, which pulls everyday agent work toward machines the operator owns. AEP's $4 billion Google fuel-cell plant and NERC's computational-load ballot keep the larger capital and regulatory weight on central campuses. The score moves one point toward local.
Forces
| Force | Meter | Score | Pull | Note | ||
| CHIP / CAPEX |
|
93 | mixed | Meta's Muse Spark Open needs two M5 Ultra Mac Studios with 512 GB each, and NVIDIA timed Nemotron Bolt for RTX Spark laptops landing October 16. Those are desk silicon bets, but they still ride the same leading-edge foundry capacity that feeds cluster accelerators. | ||
| ENERGY |
|
99 | central | AEP will spend about $4 billion on a 702 MW Bloom fuel-cell project under a 20-year Google PPA near Cheyenne, with full operation targeted for end-2028. NERC's computational-load standards passed final ballot and now await board adoption before a FERC filing. | ||
| EFFICIENCY |
|
66 | local | Nemotron 3.5 Bolt packs a 74B MoE into 24 GB for local agents, and Muse Spark Open is the first US open model Meta pitches for a home Mac Studio stack. Both cut the need to send coding and long-context jobs to a rented cluster when the hardware is already on the desk. |
Line charts (0=local 100=central 2026-09-29 to 2026-10-16)
Market overlay (2026-09-17 to 2026-10-16)
Latest: BTC $81,760 / hashrate ~1029.4 EH/s / token markers avg $5.2583/M / axis 84. 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 $81,760 / hashrate ~1029 EH/s. List-price gap (not a subsidy %): 100.0x list gap (blended) GPT-6 Luna -> Claude Fable 5.1 ($0.20/M to $20.00/M). (mempool.space + Coinbase/CoinGecko)
Marker prices (USD per 1M tokens; verified 2026-10-12; blended = (3*input + 1*output) / 4; -- = open/self-host or no single public list)
| # | Model | Org | CC | Weights | Tier | In | Out | Blend |
| 1 | Claude Sonnet 5 | Anthropic | US | closed undisclosed |
workhorse | $2.00 | $10.00 | $4.00 |
| 2 | GPT-5.5 | OpenAI | US | closed undisclosed |
flagship | $5.00 | $30.00 | $11.25 |
| 3 | DeepSeek V4.1 Flash | DeepSeek | CN | open ~284B total / ~13B active (MoE) |
commodity | $0.30 | $1.20 | $0.53 |
If useful, support the broadcast.
bc1qfs3yw5qlq8sxs50crzh3g84ug27gvww9npu85n
Subscribe
$ echo you@email.com | underclassic --subscribe
One email per issue. No tracking pixels, no filler.
home | rss | page 2 | page 3 | archive | models | buyer's guide | subscribe | data | board | mcp | llms.txt