Build American AI: when compute needs a PR campaign
AI's biggest funders are spending millions on data center ads in Kansas, Ohio and Wisconsin. That ad buy is a bottleneck report, and it prices the compute you rent.
Data center ads are about to run in Kansas, Ohio and Wisconsin, paid for by a group funded by Marc Andreessen, Ben Horowitz and OpenAI president Greg Brockman. TechCrunch reported on 31 August 2026 that the group, Build American AI, will spend millions telling voters in those states that data centers are good for them.
I read that as a bottleneck report. Nobody is buying television time to explain that GPUs are good. They are buying it to talk about land, power and neighbours.
🗳️ Who is spending, and on what
The money is not hiding. Here is what the reporting lays out:
| Entity | Behind it | What it is doing |
|---|---|---|
| Leading the Future (super PAC) | Andreessen, Horowitz, Brockman | Launched with over $50 million |
| Build American AI | LTF-affiliated | Millions in ads for data centers in Kansas, Ohio, Wisconsin |
| Public First Action | Backed by Anthropic | Argues the pro-regulation side |
| Fairshake | Supported by Andreessen Horowitz | The crypto PAC that proved the playbook works |
Andreessen Horowitz has put $115 million or more into disclosed federal contributions for these midterms, which the New York Times has described as the largest of any donor so far.
OpenAI has put distance between itself and the PAC. In a statement quoted by TechCrunch, the company said it "does not direct the activities of LTF, or have visibility into their operations," and added that groups advocating on AI should "be honest about whom they represent, and not use astroturfing tactics."
Two of the largest AI labs are now funding opposite sides of the same political argument. That is not a detail. That is the industry telling you it has no shared position on its own regulation.
⚡ You do not buy ads for things that are easy
Capital does not spend millions of dollars across three states persuading the public about an uncontroversial building. If a warehouse full of servers were an easy sell, it would go up quietly, the way warehouses do.
So the ad spend is a map of the friction. What actually gets contested at a county planning meeting:
- Electricity. A large site is a new baseload on a grid that was sized for something else, and someone has to pay for the upgrade.
- Water. Evaporative cooling competes with agriculture and with taps, and that fight is local and loud.
- Property tax abatements. The jobs number after construction is usually small relative to the incentive.
- Land and noise. Permanent, and the people living next to it get a vote.
None of that is solved by a better chip. It is solved by consent, and consent is what the ads are shopping for. That is why I think this story matters more than the average PAC story: the constraint on AI capacity has visibly moved from silicon supply to local politics.
If you want the physical side of this made concrete instead of abstract, our AI Energy & Carbon Footprint Calculator turns your own daily prompt volume into electricity, CO₂e and water, with every figure traced to a published source. The per-prompt number is small. The point is what it looks like multiplied by a few hundred million users.
🌐 What a compute importer should read into this
Sri Lanka is on the consuming end of this. We rent tokens and rent GPUs. Nobody here votes in Kansas, and no amount of local engineering skill changes whether a substation gets approved in Ohio.
What that leaves us exposed to is narrow but real:
| We control | We do not control |
|---|---|
| Which model we call, and how often | Whether new capacity gets built |
| Tokens per request, caching, batching | The price per token next quarter |
| How portable our stack is | Rate limits and regional availability |
| Our own cost visibility | Which states say yes |
Key takeaway: Your AI bill is a downstream effect of zoning fights you cannot attend. Build as if the price can move against you, because the people who set it are spending $50 million-plus to keep their options open.
There is a mirror image here that is worth sitting with. Sri Lanka has had its own arguments about who pays for grid capacity and what a big industrial load does to tariffs. Anyone who has read a CEB bill through a tariff revision already understands the Kansas argument. The scale differs. The question does not.
🛠️ Five hedges that cost you nothing today
None of this requires predicting the politics. It requires not being fragile to it.
- Know your unit cost before you scale. Cost per request, not cost per month. If you cannot state it in one number, you cannot notice when it moves.
- Keep the model layer swappable. One interface, one config value. If switching providers means a refactor, you have written the price into your architecture.
- Cache like you mean it. Prompt caching and plain result caching are the cheapest performance work available, and they are also price insurance.
- Know your self-host break-even. Our self-hosting vs API calculator shows the monthly token volume where renting a cloud GPU for an open model beats paying per token. For most small teams that crossover is far away. Knowing where it is still tells you how much room you have.
- Do not build a product that only works at one price point. If your margin dies on a 2× API increase, that is a business risk, not a technical one.
Point 5 is the one people skip. A free-tier side project can absorb a price change. A paid product built on a thin token margin cannot, and the funders currently buying advertising have every incentive to build capacity that eventually makes tokens cheaper. They also have every incentive to be the ones who own it.
💡 What this means for you
If you are a student or a small-team builder here, this news changes nothing about what you should do tomorrow morning. It changes how confidently you should extrapolate.
Cheap inference is not a law of nature. It is the output of an industry that is currently spending real money to secure the land, power and permission it needs, in specific counties, against real opposition. Sometimes that opposition wins. When it does, capacity slips, and slipped capacity shows up as rate limits and prices, months later, on a bill in Colombo.
So measure your usage, keep your stack portable, and treat today's token price as a market rate rather than a fact about the world. The people advertising on your behalf are not doing it for your margins.