AI safety philanthropic capital - Inflow model

By PierreChuzeville @ 2026-07-15T09:48 (–1)

This was cross-posted on Twitter here - https://x.com/PChuzeville/status/2075595916984119317

AI safety may be heading for a funding shock. How much new philanthropic capital could reach the field over the next decade? And how much could it actually use well? Given the interest this post received, I built a collaborative scenario model to find out.

[HERE'S THE MODEL] 

Nan Ransohoff estimates roughly $370B of gross philanthropic assets and $37–100B/year of broad potential spending. So I tried to narrow the scope by looking at it through the lens of a funnel. 
After liquidity, payout behavior and cause allocation, how much could reach strict AI safety? 

At the gross-asset level, the estimates roughly converge (the apparent gap is mostly scope, not missing assets). Some key findings: 

• The decade-long funding story is actually a one-year discontinuity. In the base case, annual funding rises from $0.41B in 2027 to $6.84B in 2028 (a 16.7× jump in twelve months) then remains essentially flat through 2035. After 2028, the story is not continued capital growth; it is the field gradually learning to use the capital already available.
• A 31× capital wave initially translates into only a 4.2× deployment wave. Relative to the $0.22B 2025 baseline, the 2028 funding target is 31.1× larger. But it seems like responsible deployment reaches only $0.92B, or 4.2× today’s level.
• The model implies a roughly $20B cumulative deployment backlog. Summing the unabsorbed funding target from 2028 through 2032 produces approximately $20.0B of capital that would need to wait, be staged, or be redirected toward capacity-building.
• The bottleneck eventually reverses. The field first becomes able to absorb the full $6.84B annual target in 2033. By 2035, modeled talent-adjusted capacity reaches $14.86B/year (2.17× the available funding target).
• The smaller capital pool generates the larger annual flow. The OpenAI Foundation pool is 2.8× larger than the modeled Anthropic founder pool ($221.5B vs. $78.3B). Yet Anthropic founders generate approximately $3.68B of the base annual target, compared with $3.17B from the OpenAI Foundation.
• A single percentage point of ownership can be larger than the entire present-day field. Holding all other base assumptions constant, one percentage point of aggregate Anthropic founder ownership changes modeled annual strict-safety funding by roughly $326M (=about 1.5× total 2025 funding).
• The scenario changes who controls the funding system. In the model, Anthropic founders provide approximately 61% of the bear-case target and 54% of the base case. In the bull case, the OpenAI Foundation becomes dominant at roughly 57%. The scenario therefore changes the likely allocation power, priorities and grantmaking architecture of the future field.
• The headline is highly concentrated in two correlated liquidity events. The core result contains only two additive sources: the OpenAI Foundation stake and Anthropic founder wealth. So both base-case liquidity curves jump in 2028.
• Agreement on gross wealth does not imply agreement on usable funding. Replacing the conservative $10.9B Anthropic employee pool with the article’s $60B assumption lifts modeled gross assets from $310.7B to $359.8B (so almost matching the external $370B estimate). Yet this does not validate the $6.84B annual strict-safety result, because liquidity, vehicle conversion, payout and cause allocation still determine the actual flow.

For scale, US charitable giving reached $617.2B in 2025. The article’s $37–100B/year represents roughly 6–16% of that. But broad US giving is not the right denominator for strict AI safety, so the model adds the missing funnel. A few comments before you dive in:

• this is a scenario model, not a forecast, committed capital or a fundraising target.
• it depends on private valuations, liquidity timing, donor behavior, cause allocation and field capacity. Every headline is tied to a source or explicit assumption.
• it’s intended to be collaborative, so please criticize, debate, and challenge anything on this sheet.