Underclassic   rss   archive   buyer's guide

Issue 69: Samsung Forecasts Record Profit on AI Memory

Permanent copy of issue 69. Today's issue is always the front page.

Previous issue  |  This week  |  Next issue  |  archive

Samsung Electronics said Thursday it expects a record 107.4 trillion won, about $80 billion, of operating profit for the third quarter, nearly nine times its profit a year earlier, and analysts credit memory chips sold for AI servers. CoreWeave said it will fill three buildings near Mumbai, 240 megawatts in all, with Nvidia systems from mid-2028. Duke Energy agreed with Amazon, Google, Meta, Microsoft and North Carolina's consumer advocate that new data centers should pay up front for the grid equipment built to serve them. Liquid AI released two small open models that make yes-or-no decisions on a laptop or an edge computer board.

Date: 2026-10-08
Updated: 2026-10-08 06:12
Status: LIVE

Broadcast

The memory in an Nvidia rack and the memory in a family laptop come off the same wafers, and Samsung is now earning more from that scarcity than any Korean company has in a single quarter. In Charlotte, the utility is drawing a line between the household meter and the server campus, and asking the campus to pay first. The small open models keep improving, but they are being built for sorting work, while the frontier stays in the rack.

Watch

• Whether the North Carolina Utilities Commission approves Duke Energy's data center settlement by mid-November, as Duke expects
• Whether Samsung's full results on Oct. 29 show how much of the profit came from HBM4 memory for Nvidia
• Whether Mistral posts the Large 4 weights by the end of October, as it promised

Stories

Samsung forecasts a record $80 billion quarterly profit on AI memory

Samsung Electronics said Thursday it expects operating profit of about 107.4 trillion won ($80 billion) for the third quarter on sales of about 195 trillion won, up from 12.17 trillion won of profit a year earlier. The company gave no division figures before its full results on Oct. 29, but analysts credit memory chips, including HBM4 sold for Nvidia's Vera Rubin AI systems.
-- Meritz Securities estimates Samsung's DRAM prices rose about 20 percent in the quarter while shipments grew about 3 percent, so the record came from scarcity rather than volume. The same chip makers are steering wafers to AI-server memory, leaving less ordinary memory for the laptops and phones that run AI at home.
2026-10-08 / Samsung Newsroom / chip / central / #samsung-forecasts-a-record-80-billion-quarterly-

CoreWeave plans 240 megawatts of Nvidia capacity near Mumbai

CoreWeave said it will take all three 80-megawatt buildings at AdaniConneX's Taloja campus in Navi Mumbai and fill them with Nvidia Vera Rubin systems, with the first phase due in mid-2028. The company holds an option to add another 240 megawatts at the site and will open an office in India.
-- A rented AI cloud from New Jersey is booking buildings and power in India two years before the first rack switches on. Capacity for 2028 is being claimed now by the companies that rent compute, not by the people who will use it.
2026-10-07 / CoreWeave / chip, energy / central / #coreweave-plans-240-megawatts-of-nvidia-capacity

Duke Energy makes data centers pay up front for grid connections

Duke Energy said Wednesday it reached a settlement with North Carolina's Public Staff, Amazon, Google, Meta, Microsoft and others that would make new large customers pay up front for substations built only for them and post deposits for shared upgrades such as transmission lines. The terms would cover loads of 50 megawatts or more that signed service agreements after June 1, 2026, and put them on a separate high-load rate, pending a state commission ruling expected by mid-November.
-- A large utility is now pricing the data center as a different kind of customer from the household. If regulators approve, the bill for wiring a server campus in the Carolinas lands on its owner rather than on the households sharing the grid.
2026-10-07 / Duke Energy / energy / central / #duke-energy-makes-data-centers-pay-up-front-for-

Liquid AI releases open-weight decision models for laptops and edge boards

Liquid AI released d1-3B and d1-omni-600M, two open-weight models that return a yes or no, a choice or a score in one pass instead of writing text. The company says d1-3B answers a single question in 30 milliseconds on an Apple M5 Pro laptop chip and 50 milliseconds on an Nvidia Jetson Orin Nano, with llama.cpp support from launch.
-- Much of what businesses pay cloud models to do is sorting: routing a ticket, flagging a message, checking a form. A 3-billion-parameter model that does that on a desk machine takes a slice of the metered work off the API.
2026-10-07 / Liquid AI / efficiency / local / #liquid-ai-releases-open-weight-decision-models-f

Axis
local [#########################-----] central
Score: 84   clusters still winning   up 1
Samsung's record profit forecast and CoreWeave's 240-megawatt lease near Mumbai show the cluster still buying memory, chips and power years ahead. Duke Energy's settlement would make those clusters pay first for their grid connections in North Carolina. Liquid AI's open decision models move routine sorting work onto local hardware. The money and the megawatts still go to the cluster.

Forces

Line charts (0=local 100=central 2026-09-23 to 2026-10-08)

Market overlay (2026-09-09 to 2026-10-08)

Latest: BTC $83,009 / hashrate ~989.58 EH/s / token markers avg $5.2583/M / axis 84. 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 $83,009 / hashrate ~990 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