Underclassic   rss   archive   buyer's guide

Issue 74: TSMC Posts Record Profit on AI Chip Demand

Permanent copy of issue 74. Today's issue is always the front page.

Previous issue  |  This week  |  Next issue  |  archive

TSMC reported a record third-quarter profit of NT$763 billion on Thursday and said demand for its most advanced chips will stay strong into the fourth quarter. A day earlier, the International Energy Agency told governments to make data centers carry the cost and risk of the grid capacity they reserve. On the device side, Google and Synaptics previewed an edge AI chip that Google says uses 40% less energy, and Tether released an open speech model that transcribes audio on a phone.

Date: 2026-10-15
Updated: 2026-10-15 06:10
Status: LIVE

Broadcast

The foundry that makes the world's AI accelerators just booked its sixth straight record quarter, and the energy agency that advises rich-country governments now treats data-center grid costs as a policy choice, not a fixed fact. Both point the same way: the cluster keeps growing, and someone has to underwrite the wires. The answer from the device this week is small and specific, a speech model that fits on a phone and an edge chip meant for a watch battery.

Watch

• Synaptics' full specifications for the Google-designed edge chip on October 20
• How much of TSMC's 2-nanometer ramp goes to data-center accelerators rather than phones and laptops
• Which regulators adopt the IEA's minimum-payment and collateral rules for large loads

Stories

TSMC posts record quarterly profit as AI chip demand outruns supply

TSMC reported third-quarter net income of NT$763.06 billion, up 68.7% from a year earlier, on revenue of US$46.72 billion and a 66.4% gross margin. The company guided fourth-quarter revenue to between US$52.3 billion and US$54.3 billion and said 2-nanometer chips already account for 8% of wafer revenue.
-- A 58.7% operating margin at the foundry that builds nearly every leading AI accelerator shows where cluster spending lands first. The same leading-edge lines also make laptop and phone chips, so data-center orders compete directly with desk silicon.
2026-10-15 / TSMC / chip / central / #tsmc-posts-record-quarterly-profit-as-ai-chip-de

IEA tells governments to make data centers carry their own grid risk

The International Energy Agency published Powering Artificial Intelligence on October 14, a survey for France's G7 presidency of 134 policies in 34 jurisdictions covering about 90% of global data-center capacity. It says hyperscalers will spend more than $700 billion of capital in 2026 and that some grid operators now hold data-center connection requests larger than their entire peak demand.
-- The IEA puts the cost of renting a gigawatt of AI capacity for a year at $20 billion to $30 billion, about 100 times the cost of the same battery capacity. That is why clusters rarely switch off for the grid, and why regulators are reaching for minimum payments, collateral and exit fees to keep the bill off households.
2026-10-15 / International Energy Agency / energy / central / #iea-tells-governments-to-make-data-centers-carry

Google and Synaptics preview a low-power AI chip for glasses and watches

Google Research showed a next-generation edge AI chip, co-designed with Synaptics on Google's open-source Coral NPU design, at its Research@ event in Mountain View on October 14. Google Research vice president Avinatan Hassidim said the chip uses 40% less energy than existing parts, and Synaptics will publish full details on October 20.
-- Google builds its own TPUs for data centers, yet it is opening the design for a chip that handles voice and vision on the device. If an always-on assistant can run on a watch battery, fewer everyday requests need a round trip to the cluster.
2026-10-15 / Seoul Economic Daily / efficiency, chip / local / #google-and-synaptics-preview-a-low-power-ai-chip

Tether releases an open speech model that transcribes audio on the phone

Tether AI Research released QVAC TranscribePsy, a 254-million-parameter speech recognition model with open weights under an Apache 2.0 license and builds for phones, laptops and the iPhone Neural Engine. Tether says it transcribes ten minutes of audio in 1.65 seconds on an iPhone 17 Pro Max and scores an 8.02% word error rate on the Open ASR Leaderboard.
-- Transcribing meetings, clinic visits and interviews is a routine paid cloud service. A free model that runs offline on a phone takes that audio, and the privacy risk that comes with it, off the meter.
2026-10-15 / Tether / efficiency / local / #tether-releases-an-open-speech-model-that-transc

Axis
local [##########################----] central
Score: 85   clusters still winning   tilted central
TSMC's record margin and the IEA's estimate of more than $700 billion in hyperscaler capital spending this year show where the money still goes: into central clusters and the power to run them. Google's low-power edge chip and Tether's on-phone transcription pull some everyday work back onto the device. The score moves one point toward central.

Forces

Line charts (0=local 100=central 2026-09-28 to 2026-10-15)

Market overlay (2026-09-16 to 2026-10-15)

Latest: BTC $80,954 / hashrate ~1028.35 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 $80,954 / hashrate ~1028 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 2026-10-12; 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