← All notes

Research note · Follow the money · 16 Oct 2026

AI keeps getting cheaper. Big Tech keeps spending more on it.

Running AI costs far less than it did two years ago, yet the biggest tech companies are spending more than ever on buildings and equipment. Here is where the money goes, in plain English, with every source.

Illustration: a huge data center campus under construction with cranes and a large orange sun
Cover: illustration made for YosQuant.
$91.4BAlphabet spending on buildings and equipment, 2025 (+74%)
300×+more AI use at Google in about two years (Google’s count)
$20 → $0.07cost per million tokens for the same skill, 2022 → 2024

1. Spending on buildings and equipment

Company20242025Change
Alphabet$52.5B$91.4B+74%
Meta$39.2B$72.2B+84%
Microsoft$44.5B$64.6B+45%

Alphabet and Meta: calendar years (Meta includes finance-lease principal). Microsoft: fiscal years ending in June. Different bases, so we do not add them up. [1] [2] [3]

2. Cheaper did not mean less

DeepSeek released its R1 model on 20 January 2025, showing strong AI could be built for less. Spending still went up in 2025. That is consistent with “cheaper means more use”, but it does not prove cause and effect. [4]

3. Where the money goes

Into data centers: chips, memory, buildings, networks and the power to run them. That is why memory and electricity keep showing up in our notes.

4. How much more AI is used

Google said its products handled about 9.7 trillion tokens (pieces of words) a month a year before May 2025, more than 480 trillion in May 2025, and more than 3.2 quadrillion by May 2026. That is Google’s own count across its products, not the whole industry. [6] [5]

5. The test

The Stanford AI Index found the cost of a GPT-3.5-level answer fell from $20 to $0.07 per million tokens between November 2022 and October 2024, before DeepSeek. Cheaper often means more use, but nobody has proven it causes more spending, and spending can fall. [7]

Sources

  1. Alphabet Q4 and full-year 2025 results (Exhibit 99.1). sec.gov
  2. Meta Q4 and full-year 2025 results (Exhibit 99.1). sec.gov
  3. Microsoft 2025 Annual Report. microsoft.com
  4. DeepSeek release log, Jan 2025. api-docs.deepseek.com
  5. Google I/O 2026 keynote notes (token volume). blog.google
  6. Google I/O 2025 keynote (token volume). blog.google
  7. Stanford HAI, AI Index 2025 in 10 charts. hai.stanford.edu

Image credits

  1. Cover and carousel slide 1: illustration made for YosQuant.
  2. “Front of server racks at NERSC”, Derrick Coetzee from Berkeley, CA, USA, CC0. Wikimedia Commons
  3. “One Wilshire Building (Unsplash)”, Owen CL owencavlys, CC0. Wikimedia Commons
  4. “BT Italia Settimo Milanese”, Davide Alotto, CC0. Wikimedia Commons