Big Tech's quarterly capex reaches $230 billion as AI router traffic sets a 30-day high

WattSaid™ weekly brief ·

We read our own boards and write down what moved. The figures below are what WattSaid showed on the morning of 2 November 2026; each one carries the date or period it describes. Our last brief was on 26 October.

The September quarter's spending is in

Microsoft, Alphabet, Amazon, Meta and Oracle spent $230.4 billion on capital in the quarter to 30 September 2026, up 117.8% from $105.8 billion a year before and up from $181.5 billion in the June quarter. Oracle's figure is for its own quarter to 31 August.

Capital spending, latest quarterChange on a year
Amazon$77.26 bn+120.2%
Alphabet$51.35 bn+114.4%
Microsoft$39.37 bn+103.0%
Meta$33.94 bn+80.2%
Oracle (to 31 Aug)$28.50 bn+235.3%

Every one of the five more than doubled except Meta. Spending on this scale has been climbing every quarter on the board since mid-2024, when it was $55.6 billion.

See the live numbers

Bigger AI sites

Worldwide, the largest AI data centers tracked now draw 15.08 GW together (+6.2% on the week) across 96 sites, with 15.75 million H100-equivalents of compute (+6.8%). The single largest is still 1,037 MW.

Open the AI board

Traffic sets a new high

Tokens sent through one large AI model router reached 33.1 trillion on Thursday 29 October, above last week's 30.6 trillion peak and the highest in the 30 days on the board. Saturday 31 October came in at 24.5 trillion, 4.6% above the Saturday before. Open model downloads over the last 30 days were 2.49 billion (+2.1% on the week, as of 1 November). AWS spot prices for training GPUs eased 1.9% to $5.06 per GPU-hour on 1 November, while inference GPUs rose 2.2% to $0.78.

US grid this morning

Lower 48 demand was 403,725 MW at 12:00 UTC on 2 November, 5.4% above the same hour on Sunday; PJM was up 10.2% to 86,752 MW. As last week, most of that is the usual Monday-versus-Sunday rise. ERCOT's real-time hub price was $16.90 per MWh at 13:02 UTC, $38.21 lower than the same time on Sunday. At 04:00 UTC wind and solar supplied 17.6% of Lower 48 power, 9.3 points more than a day earlier, and gas 40.2%.

US nuclear reactors ran at 81.8% of full power on average on 30 October, down from 88.2% on 1 October. Our reading: autumn is when many reactors refuel, though the board does not say which units are offline. US gas in storage was 3,724 billion cubic feet on 23 October, 1.0% above the five-year average.

Open the grid

China and the US

ChinaUS
Solar capacity1,310 GW (Sep 2026)169 GW, plants of 1 MW or more (Aug 2026)
Wind capacity700 GW (Sep 2026)167 GW (Aug 2026)
Electricity used, 202510,370 TWh (+5.2%)4,080 TWh sold to end users (+2.6%)
Electricity generated, 12 months to Aug 20269,944 TWh (+3.6%), large plants4,496 TWh (+2.7%)

China's September figures are in: it added about 11 GW of solar and 7 GW of wind in the month. The board now carries 2025 consumption, replacing 2024. Electricity used by China's computing centres fell to 278.6 in September from 300.9 in August (index, 2022 = 100), a 7.4% drop; one month is not enough to call a turn.

Open the Energy board

How much AI is talked about

In the week to 31 October, Wikipedia views fell for AI agents (-33.2%) and GPUs (-15.4%), after rising the week before; new AI repositories on GitHub were flat at 15,979 a day (-1.2%). arXiv is still lagging on the board: its latest day is 27 October, with 236 AI papers, 12.3% fewer than a week earlier.

Open the AI board

Figures from WattSaid™ by Starfind AI (starfindai.com/wattsaid/live), as shown on 2 November 2026; each number's period is given in the text. Bigger AI sites: Epoch AI's Frontier Data Centers (CC BY 4.0). Traffic sets a new high: tokens through one large router from OpenRouter (CC BY 4.0, as of the date on its number); GPU spot prices from AWS. US grid this morning: Lower 48 demand and the generation mix from the U.S. Energy Information Administration's hourly data; the real-time hub price from ERCOT; gas in storage from the U.S. Energy Information Administration. China and the US: China's generation from official monthly statistics on the output of large plants; US generation from the U.S. Energy Information Administration's monthly series. How much AI is talked about: Wikipedia page views; new repositories on GitHub; arXiv. Nothing here is investment advice.