Underclassic   rss   archive   buyer's guide

Issue 72: Oracle Trucks Gas to Its AI Data Centers

Permanent copy of issue 72. Today's issue is always the front page.

Previous issue  |  This week  |  Next issue  |  archive

Oracle is trucking compressed natural gas to AI data centers in Utah and Texas because pipelines are not ready, and is considering the same for its delayed Project Jupiter site in New Mexico. Nvidia can reportedly secure only 60 to 70 percent of the low-power memory it needs next year. A group of 26 researchers including Berkeley's Stuart Russell published a plan for a verified global pause of frontier AI training that swaps training chips for inference-only chips. Google released EmbeddingGemma 2, a 740-million-parameter open model that searches text, images, audio and video on a phone.

Date: 2026-10-11
Updated: 2026-10-11 07:25
Status: LIVE

Broadcast

The cluster is now hauling its fuel by truck and rationing its memory. A serious group of scholars has written down how the world could stop training bigger models while keeping the ones it has. On the phone, Google made private search across photos, audio and video small enough to run offline.

Watch

• Whether Oracle's Project Jupiter stays on schedule in New Mexico
• Memory prices for phones and laptops heading into 2027
• Whether any government responds to the hardwired pause proposal

Stories

Oracle is trucking natural gas to its AI data centers while pipelines lag

Oracle is delivering compressed natural gas by truck to data centers outside Salt Lake City and at a campus for OpenAI in Shackelford County, Texas, and is considering the same for Project Jupiter in New Mexico, where a pipeline is delayed, according to Bloomberg reporting carried by the Associated Press. Delivered gas can cost about four times as much as pipeline gas. Oracle shares fell 5.5 percent to $135.69 on the report.
-- When a cloud company pays four times the going rate to keep turbines running, power, not chips, is setting the schedule. Pipelines take months; customers will not wait.
2026-10-09 / Transport Topics (Bloomberg) / energy / central / #oracle-is-trucking-natural-gas-to-its-ai-data-ce

Nvidia can secure only 60 to 70 percent of the low-power memory it needs next year

Korean daily MoneyToday, citing industry sources, reports that Nvidia can secure only about 60 to 70 percent of the LPDDR5X memory it requires in 2027 as AI servers compete with phones for the same chips. Nvidia already cut the SOCAMM2 memory module on its Vera CPU from 192GB to 96GB.
-- LPDDR5X is the memory in phones and laptops. When the largest AI buyer cannot get enough, the devices that run AI locally get scarcer and pricier.
2026-10-06 / MoneyToday / chip / central / #nvidia-can-secure-only-60-to-70-percent-of-the-l

Researchers say a verified global pause of frontier AI training is feasible

A team of 26 researchers including Stuart Russell and Barry Eichengreen of UC Berkeley published a plan for an international, mutually verified pause of frontier AI training lasting at least 10 years. The plan would halt production of AI training chips and replace them over time with inference-only chips that can run existing models but cannot practically train new ones.
-- The proposal puts the limit in the hardware, not in promises. It would freeze the frontier while letting today's models keep running everywhere, including on local machines.
2026-10-09 / UC Berkeley News / chip / mixed / #researchers-say-a-verified-global-pause-of-front

Google's EmbeddingGemma 2 brings offline multimodal search to phones

Google released EmbeddingGemma 2, a 740-million-parameter open model under Apache 2.0 that maps text, images, audio and video into one search space. Google says it needs about 191MB of RAM for text and about 567MB for the full multimodal model on a Pixel 11 Pro, and runs in Ollama, llama.cpp, MLX and LM Studio.
-- Searching your own photos, voice memos and videos no longer needs a server. The data stays on the phone.
2026-10-06 / Google Developers Blog / efficiency / local / #google-s-embeddinggemma-2-brings-offline-multimo

Axis
local [##########################----] central
Score: 85   clusters still winning   flat
Oracle hauling gas by truck and Nvidia rationing memory show the cluster outrunning its own supply lines for power and parts. The Berkeley-led pause plan is the first detailed proposal for capping the frontier at the chip level. EmbeddingGemma 2 moves private, offline search onto ordinary phones. The cluster is still where the money goes, but it is straining.

Forces

Line charts (0=local 100=central 2026-09-26 to 2026-10-11)

Market overlay (2026-09-12 to 2026-10-11)

Latest: BTC $82,971 / hashrate ~1014.19 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 $82,971 / hashrate ~1014 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 2026-10-08; 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