First, here's how much money is donated to charities and nonprofits focused on AI safety.
$1.3B2026 projectedThe majority of donor funding comes from Coefficient Giving. Together with Longview Philanthropy, Schmidt Sciences, and the Survival and Flourishing Fund, they make up 78% of all the funding in 2025.
Scroll to see more
- Headline: Independent donors are giving a sizeable and increasing amount
- Money from independent donors is growing fast
- As recently as a few years ago, the amounts were much lower
- ?hover over segments for more information, click for an infobox (i will build infobox)?
Governments aren't responding fast enough
$0.3B2026 projectedGovernment funding didn't arrive until the public started to understand the crisis - and even now, it can't move fast. While it has been increasing since 2023, it's still only about a quarter of what private donors give.
What about the labs' own safety efforts? ↓“Safety” teams at AI companies
$2–6B2026 projected, estimateLabs' safety teams exist to serve the labs. The majority of their work goes to making sure models don't cost the company money or reputation by creating explicit or illegal content, ignoring users' instructions or causing lawsuits. A small fraction goes towards the serious dangers that Nobel prize winners are warning about - like human extinction.
So just how much more is spent on making AI more powerful before we've made it safe? ↓Before you scroll, ask yourself how much you'd spend. Twice as much? Half? 10%?Safety doesn't make money for the labs. Power does.
So here's how much OpenAI alone spends on capabilities.
Add in Google DeepMind.
And Meta Facebook.
And Anthropic.
Once you add in everyone else - including xAI, Microsoft, Scale, Surge, Mercor and other contractors - all the safety spending is barely visible on our graph.
$100–150B2026 projected, estimateIn 2026, capabilities spending is projected to reach ninety-eight billion dollars - $98,000,000,000.
- From Fabian (code changes, 3 Nov): Claude changed the chart code, not just the copy. Capabilities now builds up one company at a time (OpenAI, DeepMind, Meta, Anthropic, xAI, then everyone else), with a pause slide and a bigger scroll gap before it. New STEPS entries are added in the CO-STEPS block, so the build-out and Apollo slides moved from data-step 4 and 5 to 10 and 11. If anything else refers to those step numbers, it needs updating. The "$98B, 2026 projected" big number moved to this last capabilities slide. "Add in DeepMind" includes Google Brain in early years. The "ninety-eight billion" sentence above is typed in by hand, so it won't update if the estimate changes.
- Link to next slide: Now let's add the full spend: data centres, chip fabs, power ↓
- Name the flattening: everything about safety is now a sliver at the bottom.
- ?Flag the overlap with the next slide (labs rent Big Tech data centres), so readers don’t add the two.? Now let's add the full spend on AI
Then the build-out
~$520Babove trend, 2026 projectedThe full spend on datacenters, power, and chip fabs dwarfs even the capabilities spend, and makes all safety spending invisible.
- If you include the extra spend on the three main inputs to AI: Data centers, power supply, and chip fabs, the combined spend on safety becomes virtually invisible at this scale
23 Apollo peak years, in one year
23×Apollo's peak yearEven compared with the peak spending of the program that put a man on the moon, this is an extreme amount of capital to deploy.
- History's most expensive projects, like the Apollo moon landings and the nuclear bomb, pale in comparison to the spend tech companies are deploying. They are spending this much because they expect to be able to automate ~all jobs in the near future.