Underclassic rss archive buyer's guide
Permanent copy of issue 92. Today's issue is always the front page.
Previous issue | This week | Next issue | archive
National Grid has agreed to buy land on Teesside for what would be Britain's largest electricity substation, built beside a planned AI data centre that the local mayor says could draw up to 3.2 gigawatts. In Asia, SK hynix opened a president-level US headquarters to serve American buyers of high-bandwidth memory, and Sam Altman toured Samsung, SK and TSMC as OpenAI prepares to deploy its own inference chip. Against that, Tohoku University and Aisin showed a small AI chip that boots in 66 milliseconds without external memory, and Nace AI released a 27-billion-parameter decision model that runs on a Mac.
Date: 2026-11-02
Updated: 2026-11-02 06:30
Status: LIVE
Broadcast
The weekend's news was about who gets the scarce inputs first. National Grid is putting Britain's largest substation next to one planned campus, SK hynix has given its American customers a headquarters of their own, and OpenAI's chief executive spent the trip lining up memory and wafers for a chip only OpenAI will use. Each step shortens the queue for the biggest buyer and lengthens it for everyone else. The counterweight is engineering at the small end: a chip that runs a vision model from its own memory in a car or camera, and a decision model that fits on a Mac.
Watch
Stories
National Grid has agreed to buy 29.5 hectares at Dorman Point on the Teesworks site, where it plans to build the largest electricity substation in Britain, the Guardian reported on Sunday. The site sits next to a planned AI data centre of nearly 500,000 square metres, and the Tees Valley mayor says the connection could offer up to 3.2 gigawatts.
-- Britain's binding constraint on data centres is grid connection, and a 3.2-gigawatt substation built for one campus puts national transmission money behind a single large site.
2026-11-01 / The Guardian / energy / central / #national-grid-buys-land-for-britain-s-biggest-su
SK hynix began operating a president-level Americas headquarters on Sunday, led by Ahn Hyun, the executive who oversaw its high-bandwidth memory development, Aju Press reported. US customers supplied 64.1 percent of the company's revenue in the first half of 2026, and it expects HBM4 shipments to overtake HBM3E this quarter.
-- High-bandwidth memory is now designed alongside each large customer's accelerator, which binds the scarcest part of an AI server ever more tightly to a few American buyers.
2026-11-02 / Aju Press / chip / central / #sk-hynix-opens-a-us-headquarters-under-its-high-
Sam Altman met Samsung and SK Group executives in Seoul and the chairmen of TSMC and Foxconn in Taiwan, accompanied by the head of OpenAI's chip programme, TrendForce reported, citing Korean and Taiwanese press. OpenAI's Broadcom-designed Jalapeno inference chip is due for initial deployment by the end of 2026, reportedly on TSMC's 3-nanometre process with Samsung HBM4.
-- A model maker that designs its own accelerator and books its own memory supply is building a vertically integrated cloud, the opposite of hardware anyone can buy.
2026-11-02 / TrendForce / chip / central / #openai-s-altman-meets-samsung-sk-and-tsmc-as-its
Tohoku University and the car-parts maker Aisin said on Monday they had built a 16-nanometre AI system-on-chip with 32 megabytes of embedded STT-MRAM, a non-volatile memory, that holds both the model weights and the operating system. The chip boots in 65.7 milliseconds, 38.6 times faster than a conventional design, and uses about one fiftieth of the energy per object-detection frame.
-- Removing external memory from the boot and inference path is how AI moves into cars, cameras and sensors that run on their own, without a round trip to a data centre.
2026-11-02 / Tohoku University / efficiency, chip / local / #tohoku-university-and-aisin-build-a-16-nanometre
Nace AI published the weights of Drex Max v1, a dense 27-billion-parameter model that returns probabilities for routing, triage and policy questions, with a 4-bit build for Apple Silicon Macs with 24 gigabytes of memory, MarkTechPost reported in a sponsored article. The company says its 8-bit version on an Apple M5 Pro matched an Nvidia H100 within 0.02 per probability.
-- Routing and guardrail checks are among the most frequent calls an AI agent makes, and a model that handles them on a desk machine keeps that traffic off metered cloud endpoints.
2026-11-01 / MarkTechPost (sponsored by Nace AI) / efficiency / local / #nace-ai-releases-open-weights-for-drex-max-a-27-
Axis
local [##########################----] central
Score: 85
clusters still winning
0
Grid operators, memory makers and model companies all moved this weekend to tie power and silicon more closely to the largest buyers, from a 3.2-gigawatt substation on Teesside to OpenAI's own chip supply. The local side gained an instant-on edge chip and a desk-sized decision model, real but small advances; the score holds at 85.
Forces
| Force | Meter | Score | Pull | Note | ||
| CHIP / CAPEX |
|
98 | central | SK hynix now runs its US business as a president-level division and expects HBM4 to outship HBM3E this quarter, while OpenAI lines up Samsung memory and TSMC 3-nanometre capacity for its Jalapeno inference chip. | ||
| ENERGY |
|
98 | central | National Grid is buying 29.5 hectares on Teesside for Britain's biggest substation, sited beside a planned AI campus that could take up to 3.2 gigawatts, as the government calls data centres as fundamental as the railways. | ||
| EFFICIENCY |
|
79 | mixed | Tohoku and Aisin's 16-nanometre chip keeps weights and operating system in 32 megabytes of on-chip MRAM, cutting energy per detection frame to 50.7 millijoules, and Nace AI's 27-billion-parameter Drex Max runs in 4-bit form on a 24-gigabyte Mac. |
Line charts (0=local 100=central 2026-10-18 to 2026-11-02)
Market overlay (2026-10-04 to 2026-11-02)
Latest: BTC $77,840 / hashrate ~1049.02 EH/s / token markers avg $5.2583/M / axis 85. 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 $77,840 / hashrate ~1049 EH/s. List-price gap (not a subsidy %): 123.1x list gap (blended) Mercury 3 -> Claude Fable 5.1 ($0.16/M to $20.00/M). (mempool.space + Coinbase/CoinGecko)
Marker prices (USD per 1M tokens; verified 2026-10-31; 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