the model is trained on a basin

A model card lists parameters, benchmarks, and training data. It does not list the county, the substation, or the river.

That omission is the subject of this post, and it is not an accusation. If you work at a frontier lab — or you are about to be liquid because of one — the compute that made your equity worth something is also a load-growth line in somebody's utility filing and a water question on somebody's council agenda, in a place you have probably never visited. Both things are true at the same time. Neither makes you a villain.

AI philanthropy — giving funded by AI-era equity — has one advantage over most new money, and it is not the size of the check. It is that these givers already accept that the physical world is real. You have read power contracts. You know a gigawatt is a place. That knowledge is worth more than guilt, and it points somewhere specific.

None of this is tax, legal, or investment advice.

the bill of materials nobody prints on the model card

Every model has a physical bill of materials. Chips, land, transmission, megawatts, cooling water. The first three get contracted. The last two get argued about in rooms most of us will never sit in.

Direct on-site use is a small share of national water use. The indirect number is the one that matters — about 800 billion liters, roughly 210 billion gallons, consumed at power plants to produce the 176 TWh the buildings drew in 2023 (Lawrence Berkeley National Laboratory, December 2024). And 2023 is now a historical baseline. LBNL's scenario range puts US data centers at 325 to 580 TWh by 2028, or 6.7% to 12% of national electricity.

One more figure from the CRS report, because it decides who ends up in the argument: municipal systems supply an estimated 97% of data centers' on-site water. Almost no large campus holds its own water right. It buys from the same utility as the houses.

the same balance sheet

Here is the part that is uncomfortable and worth sitting with anyway.

The compute commitments that repriced the labs are public and still accumulating. As reported in spring–summer 2026, Anthropic's contracted capacity sat on the order of 10+ gigawatts across AWS, Google Cloud, AMD, and Broadcom, plus a $50 billion Fluidstack build starting in Texas and New York. OpenAI's footprint is the same shape at a different scale — Stargate-class campuses, no comparable public equity-to-DAF match. Deal lists date in weeks; the structure does not: the same strategy produced the charitable balance and the load growth.

Anthropic's equity-to-DAF match has been reported at 3:1 up to 50% of equity for early employees and 1:1 up to 25% for newer hires. Those are company-level figures as reported, not a how-to for your HR portal, and the share mechanics are a separate post.

So: one corporate strategy produced both the charitable balance and the load growth. That is not hypocrisy. It is arithmetic, and noticing it is the beginning of being useful rather than the beginning of feeling bad.

Credit where it is due — Anthropic has said it will "pay its own way" on power, which is more than many buyers of gigawatts have said out loud. But that commitment is about the grid: generation, interconnection, the upgrade. There is no equivalent commitment to the watershed, and the reason is structural rather than moral.

Snow retention, soil infiltration, riparian condition, forest structure upstream of the reservoir: almost no one sends a bill for any of it, and the buyer of the megawatts never does. Salt River Project's forest-and-watershed work on the Salt and Verde is the exception that proves the rule — real money, still small against the draw. Free-looking water is why the source degrades.

a basin, named

Take Central Arizona, because the Congressional Research Service just did. A 2026 Arizona State University study cited in the CRS report counted 66 data centers completed or under construction there — 25 in the Phoenix area, 9 around Tucson, 8 in Mesa, 5 in Scottsdale.

Mesa publishes its water portfolio: 53% Colorado River delivered through the Central Arizona Project, 40% from the Salt and Verde system, 7% local groundwater. Follow the arithmetic. A campus in Mesa buying municipal water holds no federal water contract and has never negotiated with the Bureau of Reclamation — and is still, by roughly half, drinking the Colorado River. CRS makes exactly that point: federally developed water can reach a data center indirectly, through a city.

Meanwhile Reclamation's Final Environmental Impact Statement for post-2026 Colorado River operations points toward a reduced Arizona allocation, which Mesa's own public materials acknowledge could shrink its CAP deliveries. The load is growing into a supply that is contracting on a published schedule.

Cooling design is genuinely improving, and the honest version has to say so. The NTT campus proposed at Pecos and Crismon in Mesa is designed closed-loop — filled once, no evaporative loss — and the city's development agreements now tie project water planning directly to CAP shortage conditions. But closed-loop cooling moves the burden onto electricity, and electricity has its own water bill. The basin does not care which meter the water went through.

Who ought to pay to protect that river is a whole argument we have already made, with the beneficiary-pays models that already work: 40 million people, one shrinking source. This post only needs the river as one named example of a general fact. Every training run sits in a watershed. Pick a different cluster — Loudoun County, central Ohio, west Texas, Querétaro, the Maipo basin above Santiago — and the structure repeats with different hydrology, different water law, and a different set of people who were there first.

the offset objection

You are probably forming the sentence: this is offset logic with better vocabulary.

Fair, and mostly correct about offsets. Three differences worth holding.

We do not claim it makes water. Funding a watershed does not produce wet water in a closed basin, and anyone who tells you otherwise is selling something.

The unit is a place, not a credit. An offset is a fungible claim about an avoided quantity somewhere. A hold on a named natural asset is a claim about the condition of one specific living system, tracked over time. Fund the basin your compute sits in when you can name it. When you cannot, a named place is still not an offset, because you are making no netting claim.

The timing is inverted. Offsets are bought to net something out. Protection is funded before loss, to keep the source producing. That is the whole distinction between insurance and ensurance, and it is the reason the sequence matters more than the size.

The other objection is quieter: so the argument is that AI is bad. No. If you spend your working life on the premise that a system's behavior has consequences in the physical world, you are already the right kind of thinker for this. The people worrying hardest about what models do to the world are not the opposition here. The invitation is to run the same seriousness one layer further down — past the model, past the megawatt, to the water cycle both of them assume will keep working.

humility is a better instrument than guilt

Guilt and humility fund different things, and it is worth being precise about why.

Guilt is a ledger about the giver. It wants to be settled, so it buys the most legible available object: the recognizable logo, the photogenic species, the thing that reads clearly as atonement at a dinner party. Guilt-funded giving gets priced on the donor's discomfort, which has nothing to do with where the marginal dollar does the most good. It also ends the moment the feeling lifts.

Humility is a claim about knowledge, not character. It says: the thing I helped build rests on a physical substrate I never priced, and I only found that out by reading load studies. Humility is directional. It sends money toward the substrate rather than toward the feeling.

Guilt is priced on the giver's discomfort. Humility is priced on the place's condition.

That is also why the receipt has to keep reporting. A guilt gift is finished when the acknowledgment letter arrives. A humility gift is finished when the place is measurably in better shape, which means someone has to keep looking — what you actually hold after the grant.

what AI philanthropy is unusually good at

Most donors have to be talked into the sentence "nature is infrastructure." You do not. You spent years learning that capability is downstream of physical constraint — that the impressive thing on the screen is bounded by substations, water rights, land use hearings, and interconnection queues. You have the intuition most philanthropic capital lacks.

Applied to giving, that intuition says something narrow and useful: fund the substrate, and fund it as a named place rather than a theme. Not "water." A watershed, with a name, a condition, and a load on it. Watersheds, forests, soils, and the species inside them exist whether or not anyone funds them — that is precisely what makes them infrastructure rather than a product. Ensurance is one way a gift can hold a named living system, with condition tracked and proceeds routed to whoever is doing the stewardship. The living system is the asset. Ensurance is the funding mechanism, and it should never be confused with the thing it funds. Taking a grant directly from a DAF sponsor is a design target, not a shipped button; the rails are in how to grant from a DAF to a place.

the operator's version of this argument

There is a parallel series on this site written for the other side of the table — the developer, the operator, the economic development office, the county board. It is a different job from yours, and worth reading precisely because it is not addressed to you.

Those are operating documents. This one is for the person who holds the equity and is deciding what the gift is for.

what to do with this

The physical footprint of your industry is not a reason to give. It is a reason to give accurately — to the layer that everything else, including the model, is running on. That is the whole argument for AI philanthropy funding basins: not penance, just the same respect for physical constraint that made the work possible in the first place.

1. Read the infrastructure argument in giver form — nature is infrastructure. a DAF can fund it
2. Decide which part is a grant and which part is a hold — some of this should be a grant. some of it should be a hold
3. See who already pays to protect one named source — 40 million people, one shrinking source
4. Look at real places and the agents that steward them — explore