Underclassic rss archive buyer's guide
Issue 109 · 2026-11-19

Permanent copy of issue 109. Today's issue is always the front page.
Previous issue | This week | archive
The Loudoun County Board of Supervisors voted 7-1 on Tuesday to stop acting on new data center and substation applications until it finishes stricter rules, expected in July 2027, in the county that calls itself the data center capital of the world. A day earlier PJM, the grid operator that serves Loudoun, asked federal regulators to make large data centers stay connected through grid faults after 3.9 gigawatts of them switched to backup power at once in July, while Tencent and the Marin project each released open model weights that anyone can download.
Date: 2026-11-19
Updated: 2026-11-19 06:30
Status: LIVE
Broadcast
Loudoun County, Virginia, has more data center buildings than any other county in the world, and on Tuesday its supervisors voted to stop approving new ones until stricter rules are written next July. The day before, PJM, the grid operator that serves Loudoun, told federal regulators that a single fault in Northern Virginia in July sent 3.9 gigawatts of data centers onto their own generators at once, and asked for a rule that keeps such sites connected through grid disturbances. Hillsboro, Oregon, another data center hub, extended its own moratorium the same evening. On the other side of the ledger, Tencent and the Marin project released large models whose weights anyone may download, run or study. Neither fits on a household computer today, but neither depends on one company's servers.
Watch
Stories
The Loudoun County Board of Supervisors voted 7-1, with one abstention, on Tuesday to stop holding hearings and final votes on data center and substation rezoning applications, backdated to Sept. 15, until it adopts stricter rules on location, noise and on-site power, expected in July 2027. The county has about 245 data center buildings built or under construction, about 58 million square feet, with another 55 million square feet approved or in review.
-- Loudoun is the densest data center market in the world, and county staff say 33 applications are pending; developers may still file and staff may still review, but projects that rely on by-right approvals are not covered, and the county attorney says the board lacks power to stop those. Data Center Dynamics reports that data centers are projected to supply $1.3 billion, or 40%, of the county's 2027 budget.
See also: Data Center Fights Tracker: pauses and votes by state
2026-11-17 / Loudoun County Board of Supervisors / energy / local / #loudoun-county-pauses-action-on-new-data-center-
PJM Interconnection asked the Federal Energy Regulatory Commission on Monday to require data centers of 50 megawatts or more, with at least 25 MW fed through power electronics, to stay connected through set ranges of voltage and frequency disturbances, starting Jan. 16, 2027. The filing says a July 22 fault caused about 3.9 gigawatts of data center demand to switch to backup power at once, the largest such event in PJM's history, pushing Eastern grid frequency to 60.092 hertz.
-- PJM says it already serves about 10 GW of computing load, and the rule would cover existing sites too, giving them one year from the effective date or from start of operation, whichever is later, so the cost of keeping very large customers from destabilizing the grid moves onto the data centers themselves. PJM asked for a 60-day comment period and a decision within 120 days.
2026-11-16 / PJM Interconnection (FERC Docket ER27-522) / energy / central / #pjm-asks-regulators-to-require-data-centers-to-r
Tencent posted the weights of Hy4 Flash on Hugging Face this week in three versions, a chat model, a compressed FP8 build and a base model, under its own Tencent Hy License. The model card describes a mixture-of-experts model of about 220 billion parameters, 12 billion of them active on each token, that reads text and images and handles up to 1 million tokens of context.
-- Tencent's API lists Hy4 Flash at $0.07 per million input tokens, among the cheapest prices from a large Chinese lab, and the open weights mean other hosts can undercut or match that. It is not a household model: Tencent's serving instructions use eight graphics processors, though the card links a compressed GGUF build for llama.cpp-style tools.
See also: Model board: Hy4 Flash, priced at $0.07 per million input tokens
2026-11-17 / Tencent (Hugging Face model card) / efficiency / local / #tencent-releases-hy4-flash-as-open-weights-with-
The Hillsboro City Council voted on Tuesday to extend for six months a land-use moratorium on data centers that it adopted for 120 days in July and that was due to expire at the end of November, KPTV reports. Nearly a dozen residents spoke, several of them objecting to property tax breaks for the companies while the city rewrites its development code for data centers.
-- Hillsboro is one of Oregon's main data center hubs, and our state-by-state rates page shows the state is also changing who pays: a Pacific Power rate class for computing sites of 20 MW or more starts Jan. 1, North Carolina's Nov. 16 order gives Duke data centers of 50 MW or more their own schedule, and since January 2025 Georgia has let Georgia Power bill new customers of more than 100 MW for grid and generation costs as construction proceeds.
See also: Data center electricity rates by state: Oregon's 20 MW class starts Jan. 1
2026-11-17 / KPTV / energy / local / #hillsboro-oregon-extends-its-data-center-morator
Open Athena and the Marin project said on Tuesday that Marin "Mochi" 535B-A23B, which uses about 23 billion active parameters per token, has finished its main pre-training phase on 18.2 trillion tokens, and published all 57 checkpoints from the run on Hugging Face. The team said it found substantial leakage of common benchmarks into its training data, which limits which test scores it can fairly report.
-- The large labs do not publish the intermediate stages of their training runs, so open checkpoints of a model this size let outside researchers study how one is built. They are base models with no instruction or safety tuning, and Marin says they are for research, not use as assistants.
See also: Model board: Marin Mochi 535B, research checkpoints only
2026-11-17 / Open Athena / efficiency / local / #marin-releases-57-pre-training-checkpoints-of-it
Reuters reported on Wednesday, citing three people, that the Treasury Department and the White House Office of Science and Technology Policy have discussed requiring US companies to post on their websites that they use Chinese AI models, a "name and shame" approach meant to discourage reliance on the technology without new restrictions that could anger industry or Beijing. It is not a rule: Reuters said it was not clear whether the idea was gaining traction or would go through formal rulemaking, and the White House, Treasury and Alibaba did not respond to requests for comment.
-- Reuters named Airbnb, which uses Alibaba's Qwen for customer service, as a company the plan would reach; Airbnb said its AI "runs overwhelmingly on U.S.-origin models" and that its China-origin models are open-source and run only through approved US-based providers. Reuters called the idea part of a push in the administration against Chinese open-weight models, whose weights anyone can download, and Representative John Moolenaar, who chairs the House China committee, called disclosure "an important step." A label would fall on the companies that use the models, not the labs that make them; our linked page lays out the facts and, in a section marked as opinion, argues the risk is to open-weight and self-hosted AI.
See also: Our report and analysis: what the disclosure idea would mean for open-weight and self-hosted AI
2026-11-18 / Reuters / efficiency / central / #white-house-weighs-making-us-companies-disclose-
The Center for AI Safety (CAIS) organized the MisAnthropic Statement, three sentences that say Anthropic "warns that recursively self-improving AI could concentrate power and escape human control, yet races to build it first" and end by calling it "not a safe AI company." The signatory list on misanthropicstatement.com showed about 120 names on Thursday morning, up from the 92 we counted Wednesday afternoon, among them Harvard's Steven Pinker, Hugging Face co-founder Thomas Wolf, xAI co-founder Igor Babuschkin and MIT's Max Tegmark. Anthropic had not publicly responded as of Thursday morning. Wolf, who signed, said he would not have used the word "reckless."
-- Anthropic priced its IPO on Wednesday at $83 a share, selling 980 million shares for $81.3 billion, according to CNBC citing the company's release, a record offering, and its shares are expected to begin trading on Nasdaq today under the symbol ANT. The statement is the signers' opinion, not a finding by a regulator or court: its signers include Anthropic's rivals, AI-risk campaigners and academics who disagree about how fast AI should move, and Anthropic's own filings describe the same facts differently. Our page sets each claim next to the record.
See also: Who signed it, and what Anthropic's own record says about each claim
2026-11-18 / The MisAnthropic Statement (Center for AI Safety) / chip / central / #coalition-organized-by-the-center-for-ai-safety-
Axis
local [#######################-------] central
Score: 78
clusters still winning
-1
Thursday's edition leads with Loudoun County's pause on data center applications, alongside PJM's proposed ride-through rule for data centers, Tencent's open Hy4 Flash weights, Hillsboro's longer moratorium and Marin's 57 open Mochi checkpoints. The largest data center market slowing its approvals and two large open releases move the score one point toward local, to 78.
Forces
| Force | Meter | Score | Pull | Note | ||
| CHIP / CAPEX |
|
97 | central | No new chip or spending figures moved this edition; Nvidia's $115.8 billion quarter, reported on Tuesday, still sets the pace for central clusters. The gauge holds at 97. | ||
| ENERGY |
|
94 | central | Loudoun County, home to about 245 data center buildings, paused board action on 33 pending applications, and PJM told regulators it already carries about 10 GW of computing load and saw 3.9 GW switch to on-site generators in a single July fault. The pause in the largest market eases the gauge one point, to 94. | ||
| EFFICIENCY |
|
89 | local | Tencent released Hy4 Flash, a model of about 220 billion parameters with 12 billion active on each token, as open weights, and Marin published 57 pre-training checkpoints of its 535-billion-parameter Mochi model. Both are too large for most home machines, but both can be run and studied outside the companies that made them, so the gauge rises one point, to 89. |
Line charts (0=local 100=central 2026-11-04 to 2026-11-19)
Market overlay (2026-10-21 to 2026-11-19)
Latest: BTC $63,638 / hashrate ~1007.06 EH/s / token markers avg $5.2583/M / axis 78. 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 $63,638 / hashrate ~1007 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 2026-11-15; 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
Get the daily issue
One email a day. Unsubscribe anytime.
home | rss | page 2 | page 3 | archive | models | buyer's guide | subscribe | data | board | about | mcp | llms.txt