Underclassic   rss   archive   buyer's guide

Issue 77: Microsoft Pulls Out of Nscale's Essex Data Center Over Grid Delays

Permanent copy of issue 77. Today's issue is always the front page.

Previous issue  |  This week  |  archive

Microsoft has pulled out of its agreement to anchor Nscale's planned data center in Loughton, Essex, after grid operators said the site would not get power until the 2030s, The Telegraph reported on Sunday. Nvidia is designing custom memory with Samsung and TSMC to relieve the bandwidth limit on its AI chips, according to the Seoul Economic Daily, and Vice President JD Vance told a Michigan voter that data centers must not raise household power bills. DeepSeek, meanwhile, published the weights of its upgraded V4.1 Flash model under the MIT License.

Date: 2026-10-18
Updated: 2026-10-18 06:30
Status: LIVE

Broadcast

Microsoft had agreed to rent 23,000 Nvidia chips at an Essex site that the local grid cannot feed until the next decade, and it has decided not to wait. In Seoul, the reporting is that Nvidia now designs its own memory with two other companies' factories because its chips spend too long waiting for data. In Michigan, the vice president promised that the neighbors of server halls will not pay for their power. DeepSeek, for its part, gave away the weights of a model it says writes software faster and more cheaply than last month's version. The limit on the central build-out has moved from the chip to the substation.

Watch

• Whether Nscale's listing documents restate how much of its $103 billion backlog depends on Microsoft
• Independent benchmarks of DeepSeek-V4.1-Flash-1017 against the scores DeepSeek published
• Whether UK Power Networks offers Loughton an earlier connection or Nscale secures on-site generation

Stories

Microsoft pulls out of Nscale's Essex data center after grid delays

Microsoft has told Nscale it is ending its agreement to be the anchor tenant at the British company's planned data center in Loughton, Essex, where it had agreed to rent 23,000 Nvidia AI chips, The Telegraph reported on Sunday. Grid operators have told Nscale they do not expect to supply power to the site until the 2030s; Nscale says it remains committed to the project and is exploring alternative power sources.
-- The chips were ready before the wires were: a site meant to open next year now waits roughly a decade for a grid connection. Microsoft accounts for more than 40 percent of Nscale's $103 billion in contracted revenue, so the loss lands just as the company seeks a New York listing worth up to $35 billion.
2026-10-18 / The Telegraph (via Yahoo Finance) / energy, chip / central / #microsoft-pulls-out-of-nscale-s-essex-data-cente

Nvidia designs custom memory with Samsung and TSMC to ease an AI chip bottleneck

Nvidia is co-designing the base die of its custom high-bandwidth memory, called NVHBM, with Samsung Electronics and TSMC, industry sources told the Seoul Economic Daily on Sunday. The base die takes over the memory controller that used to sit on the GPU, and the paper reports that NVHBM can widen bandwidth by up to 30 percent and cut power use by 15 percent against standard HBM4E.
-- Memory, not arithmetic, now limits how fast the largest models answer, and Nvidia's fix is a part built on TSMC's most advanced lines for a single customer. Custom memory of this kind goes only into data-center accelerators, so each gain widens the gap between cluster hardware and machines that fit on a desk.
2026-10-18 / Seoul Economic Daily / chip, efficiency / central / #nvidia-designs-custom-memory-with-samsung-and-ts

Vance tells a data center's neighbors that operators must not raise household power bills

At a CBS News town hall in Michigan, a resident who lives across the street from a data center told Vice President JD Vance that its constant buzzing had made opening the windows a luxury her family could no longer enjoy. Vance said data centers can be good but can also cause real problems, and that they should pay more local tax and must not lead to higher residential utility bills.
-- The administration that champions the AI build-out is now promising voters that the cost of new plants and power lines will not reach their meters. Each promise of that kind moves more of the bill onto the companies building central campuses and gives local councils more reason to slow approvals.
2026-10-18 / CBS News / energy / mixed / #vance-tells-a-data-center-s-neighbors-that-opera

DeepSeek publishes weights for its upgraded V4.1 Flash model under the MIT License

DeepSeek released DeepSeek-V4.1-Flash-1017 on Saturday on its website, app and API, and posted the model weights on Hugging Face under the MIT License. The company says the update scores higher on its benchmarks while finishing tasks faster with shorter outputs, at unchanged API prices, and it opened its in-house Search and Fetch APIs at $0.75 per 1,000 calls.
-- Anyone may download, modify and serve the weights without DeepSeek's permission, which keeps a strong coding model outside any one company's cloud. The benchmark claims are DeepSeek's own and have not been independently checked, and a model of this class still needs server-grade memory rather than a laptop.
2026-10-17 / Hugging Face (DeepSeek model card) / efficiency / local / #deepseek-publishes-weights-for-its-upgraded-v4-1

Axis
local [#########################-----] central
Score: 83   clusters still winning   down 1
Microsoft's exit from Loughton shows that grid connections, not chips, now set the pace for new central capacity in Britain, and the White House is telling voters that operators will pay for their own power. Nvidia's custom memory keeps the largest hardware gains inside data centers, while DeepSeek put another strong model in public hands. The score eases one point to 83.

Forces

Line charts (0=local 100=central 2026-10-01 to 2026-10-18)

Market overlay (2026-09-19 to 2026-10-18)

Latest: BTC $82,058 / hashrate ~1011.84 EH/s / token markers avg $5.2583/M / axis 83. 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 $82,058 / hashrate ~1012 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)

Full 20-model board

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