Underclassic rss archive buyer's guide
Permanent copy of issue 81. Today's issue is always the front page.
Previous issue | This week | archive
Nvidia on Wednesday introduced DSX POD, a design for complete AI data centers that partners build and test inside containers, and said the units can start running at a prepared site in as little as 14 days instead of about six months. Lam Research, whose machines etch and deposit the layers of advanced chips, reported record quarterly revenue of $8.22 billion the same day, while in Michigan residents are fighting a DTE Energy plan that could put $6.4 billion of new gas plants for data centers on customers' bills. On the desk side, Gradium released Phonon, an open text-to-speech model of 31 million parameters that runs on a phone or laptop without a network connection.
Date: 2026-10-22
Updated: 2026-10-22 06:30
Status: LIVE
Broadcast
For two years the slowest part of building an AI cluster was the building itself. Nvidia now proposes to ship the data center in containers, tested at the factory, so that a customer with a ready site waits weeks rather than months. That leaves electricity as the binding constraint, and Michigan shows who is asked to pay for it: DTE wants new gas plants largely because of data center contracts, and its residential customers may carry much of the cost. Lam's record quarter confirms that the chip factories feeding those clusters are still growing. Gradium's small voice model points the other way, showing that a service the cloud sells by the word can now run on a phone for nothing.
Watch
Stories
Nvidia on Wednesday introduced DSX POD, a reference design for complete Vera Rubin AI data centers that partners including Dell, HPE, Lenovo, Schneider Electric and Vertiv assemble and test inside modular containers before delivery. The company says the units can begin running at a site that already has power and cooling in as little as 14 days, against about six months for a conventional build, with deliveries from early 2027.
-- If the computers arrive ready to plug in, the time it takes to open a cluster shrinks to the time it takes to secure a grid connection, which puts still more weight on power. It also extends Nvidia's reach from the chip to the whole building and makes rented central capacity faster to add.
2026-10-21 / NVIDIA Newsroom / chip, energy / central / #nvidia-says-factory-built-ai-data-centers-can-st
Lam Research, which makes the etch and deposition machines used to build advanced chips, said on Wednesday that revenue rose 22 percent from the previous quarter to $8.22 billion and net income reached $2.75 billion, with record revenue, profit and earnings per share. The company forecast revenue of about $8.60 billion for the December quarter, and chief executive Tim Archer said AI is driving higher semiconductor demand and greater device complexity.
-- Chipmakers are still buying the tools for new production lines, so the supply of AI processors will keep growing well into next year. Most of that output is bound for clusters rather than for the machines on a desk.
2026-10-21 / Lam Research / chip / central / #lam-research-reports-record-8-22-billion-quarter
DTE Energy has asked Michigan regulators to approve a 20-year plan that includes 2.1 gigawatts of new gas plants in 2031 and 2032, which could cost its customers $6.4 billion, The Guardian reported on Wednesday. The utility has already contracted 2.4 gigawatts of supply for Google and Oracle data centers, and the Michigan Public Service Commission will issue a preliminary decision within 300 days.
-- About 2.3 million electric customers in south-east Michigan could help pay for power plants built largely to serve data centers, which is a cost of central computing that ordinary households carry. Where regulators make data centers pay their own way, the price advantage of the cluster narrows.
2026-10-21 / The Guardian / energy / local / #michigan-residents-fight-a-6-4-billion-bill-for-
Gradium, the French voice company founded by the co-founders of the Kyutai research lab, released Phonon on Wednesday, a text-to-speech model of 31 million parameters whose English version is free for commercial use under the Apache 2.0 licence. The company says the model fits in about 64 megabytes, generates speech entirely on the device and starts speaking in about 7 milliseconds on a laptop processor.
-- Speech is one of the services apps most often send to a server, and a model this small lets an app talk offline at no cost per word. It is a modest release, but it moves another everyday AI task onto hardware people already own.
2026-10-21 / Gradium / efficiency / local / #gradium-releases-an-open-voice-model-that-runs-o
Axis
local [#########################-----] central
Score: 82
clusters still winning
unchanged
Nvidia is shortening the time between buying chips and selling computing, and Lam's record quarter shows the factories behind those chips are still expanding. Against that, Michigan's fight over who pays for DTE's gas plants shows that power remains the slow and contested part of the central build-out, and Gradium's open voice model adds a useful tool that needs no server, so the score holds at 82.
Forces
| Force | Meter | Score | Pull | Note | ||
| CHIP / CAPEX |
|
95 | central | Nvidia introduced factory-built DSX POD data centers for delivery from early 2027, and Lam Research reported record revenue of $8.22 billion with a December-quarter forecast of about $8.60 billion. | ||
| ENERGY |
|
98 | central | DTE Energy is seeking approval for 2.1 gigawatts of gas plants that could cost Michigan customers $6.4 billion, after contracting 2.4 gigawatts of supply for Google and Oracle data centers. | ||
| EFFICIENCY |
|
69 | mixed | Gradium's open Phonon model generates speech on phones and laptops from about 64 megabytes of weights, with the English version under the Apache 2.0 licence. |
Line charts (0=local 100=central 2026-10-05 to 2026-10-22)
Market overlay (2026-09-23 to 2026-10-22)
Latest: BTC $81,897 / hashrate ~1034.45 EH/s / token markers avg $5.2583/M / axis 82. 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 $81,897 / hashrate ~1034 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 2026-10-17; 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