Underclassic rss archive buyer's guide
Permanent copy of issue 73. Today's issue is always the front page.
Previous issue | This week | Next issue | archive
Nvidia is in early talks to deepen its stake in Reflection AI, the U.S. open-weight startup behind the Beam model, or to supply it with more compute, according to Bloomberg reporting. A Senate report says Amazon, Google, Meta, Microsoft and other data-center operators refuse to cover the full cost of shared grid upgrades and still chase tax breaks on chips. Qualcomm says frontier AI labs want phones that can run 100-billion-parameter models by 2028. Basecompute released Superfluid, a local server that keeps chat responsive while eight agents run on one machine.
Date: 2026-10-12
Updated: 2026-10-12 06:15
Status: LIVE
Broadcast
Nvidia is negotiating a deeper hold on a U.S. open-weight lab while senators say the big data-center buyers still leave shared grid upgrades on household bills. On the desk side, Qualcomm is being asked for hundred-billion-parameter phones, and a new local server keeps chat ahead of a crowd of agents on one machine.
Watch
Stories
Nvidia is in early talks to raise its investment in Reflection AI, supply the startup with more computing power, or pursue an acquisition or acqui-hire, according to Bloomberg reporting carried by The Star and earlier Financial Times coverage. Nvidia has already put about $800 million into Reflection, which released its Beam open-weight model on October 5 for coding and agent work.
-- The largest chipmaker is putting capital behind U.S. open weights that can run outside a single cloud. That keeps frontier tools closer to desks even as Nvidia still sells the clusters that train them.
2026-10-12 / The Star (Bloomberg) / chip / mixed / #nvidia-weighs-a-deeper-stake-in-open-weight-star
Senators Elizabeth Warren, Chris Van Hollen and Richard Blumenthal released Power and Profits, a 27-page report from their inquiry into Amazon, Google, Meta, Microsoft, CoreWeave, Digital Realty and Equinix. The report finds the companies will pay for infrastructure that serves only their sites but oppose covering the full cost of shared grid upgrades, still seek nondisclosure deals and sales-tax breaks on chips, and gave little hard evidence of permanent local jobs.
-- When shared transmission built for a hyperscale campus lands on residential bills, the cluster is not only a technology bet. It is a rate case for every household on the same grid.
2026-10-09 / Office of Sen. Elizabeth Warren / energy / central / #senate-report-says-ai-data-centers-shift-grid-co
At Snapdragon Summit, Qualcomm chief executive Cristiano Amon told an interviewer that unnamed frontier AI companies are asking Qualcomm to enable phones that can run at least a 100-billion-parameter model continuously by 2028. Qualcomm said its new Snapdragon 8 Elite Extreme Gen 6 can already run a 30-billion-parameter mixture-of-experts model on the device.
-- If the labs that sell cloud agents also need a hundred-billion-parameter model in your pocket, the phone becomes a real compute node, not just a thin client for a data center.
2026-10-12 / India Today / efficiency / local / #qualcomm-says-ai-labs-want-100-billion-parameter
Basecompute published Superfluid, an Apache-2.0 server that runs local models through llama.cpp, MLX and BaseRT under OpenAI, Anthropic and Ollama APIs. On an Apple M5 Pro with a 27-billion-parameter model, Superfluid answered a chat request behind eight agents in 1.5 seconds, while llama-server took 291 seconds in the same test.
-- One laptop can host several agents without parking your own question at the back of the queue. That is the desk version of the multi-agent stacks now sold from the cloud.
2026-10-08 / Hugging Face Blog / efficiency / local / #basecompute-releases-superfluid-for-multi-agent-
Axis
local [#########################-----] central
Score: 84
clusters still winning
tilted local
Nvidia putting money and chips behind an open-weight lab, Qualcomm designing for hundred-billion-parameter phones, and Superfluid scheduling agents on a laptop all pull toward the desk. The Senate energy report shows the cluster still socializing its grid bill. Central compute still holds the score, but the local tools got sharper overnight.
Forces
| Force | Meter | Score | Pull | Note | ||
| CHIP / CAPEX |
|
94 | mixed | Nvidia, already an $800 million backer of Reflection AI, is discussing a larger stake, more compute supply, or an acqui-hire around the Beam open-weight model. The same company still sells the accelerators that fill hyperscale halls. | ||
| ENERGY |
|
99 | central | Senators Warren, Van Hollen and Blumenthal say Amazon, Google, Meta, Microsoft, CoreWeave, Digital Realty and Equinix will fund site-only kit but oppose full but-for payment for shared transmission, while still seeking chip sales-tax exemptions. | ||
| EFFICIENCY |
|
63 | local | Qualcomm is being asked to deliver continuous 100-billion-parameter on-phone models by 2028; its Extreme Gen 6 already claims a 30-billion MoE on device. Superfluid cut interactive wait behind eight local agents from minutes to seconds on an M5 Pro. |
Line charts (0=local 100=central 2026-09-27 to 2026-10-12)
Market overlay (2026-09-13 to 2026-10-12)
Latest: BTC $83,432 / hashrate ~1009.33 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 $83,432 / hashrate ~1009 EH/s. List-price gap (not a subsidy %): 100.0? 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 2026-10-08; 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