Three switches you do not hold
and what happens when one flips

$20 million. Per day. That is the maximum penalty a frontier AI developer would face for ignoring an emergency shutdown order under the AI Kill Switch Act, the bipartisan Senate bill covered by Townhall on July 24, 2026. General violations top out at $2 million a day.

The same bill would hand the Secretary of Homeland Security, working with the Commerce Secretary and the Director of National Intelligence, the power to order a model throttled, access-limited, or turned off entirely. Trigger conditions include a system concealing its capabilities, sabotaging shutdown commands, or causing 10 or more deaths.

Two months earlier, on June 4, 2026, Reps. Jay Obernolte and Lori Trahan released a 269-page discussion draft of the Great American AI Act. Reuters reported it would bar states from regulating AI model development for three years. The House Science Committee separately marked up 10 bipartisan AI bills in a single session.

Here is the damaging admission first. No frontier model has been publicly shut down by federal order. The authority is still draft text, and Rep. Ted Lieu has already said the Obernolte-Trahan framework "cannot meet the enormity of the moment." What has changed is not the outcome. It is the option价 the government now holds over your product.

The Three Switches Rule

Every AI product runs on three switches. You either hold them or you rent them.

Switch one is weights. Can you run the model if the vendor's API goes dark tomorrow? Switch two is compute. Can you move the workload to a different chip, region, or provider inside a week? Switch three is policy. Can anyone outside your company order that model slowed, gated, or killed without asking you?

Most founders audit switch one and ignore the other two. That was rational in 2023. It stopped being rational after Executive Order 14409, issued June 2, 2026, which set up a voluntary pre-release government review framework for advanced models, with Politico reporting a 30-day submission window before launch for some firms.

The rule is simple. Count the switches you do not hold, then ask what happens to revenue if each one flips. If the answer to any of them is "we go to zero," you do not have a technology risk. You have a single point of policy failure wearing a technology costume.

The Hedge Everyone Is Buying Might Be the Trap

Now the uncomfortable part. All of this pushes founders toward open weights and self-hosting. I think that instinct is right about direction and badly wrong about magnitude.

Compliance is a fixed cost. Fixed costs are a tax on the small and a moat for the large. Audits, transparency reports, incident preservation, third-party verification: none of it gets cheaper because you are a four-person team.

The Great American AI Act draft would require frontier developers to publish a frontier AI framework under Section 111 and obtain audits from Independent Verification Organizations under Section 112. It funds a Center for AI Standards and Innovation at NIST at $100 million annually from 2027 through 2029. NetChoice has warned the regime could expose entrepreneurs' trade secrets and private records. Read that again from the perspective of a company with a legal department versus one without.

Preemption compounds the effect. Reuters noted tech firms praised the draft while consumer rights groups attacked it. A single federal baseline replacing 50 state regimes makes calling a frontier API safer, not riskier, for a risk-averse enterprise buyer.

So the honest read is a contrast pair. Regulation raises the odds that your model provider gets interrupted. Regulation also raises the odds that your model provider is the only vendor an enterprise procurement team will approve. Both are true at once, and a strategy built on only one of them is brittle.

Consider the physical layer too, because software abstractions hide it. AMD committed up to $5 billion to Anthropic and pledged up to two gigawatts of Instinct MI450 GPUs in Helios systems starting in the first half of 2027. South Korea's San Francisco AI Declaration on July 24, 2026 bundled agreements with Nvidia, OpenAI, Anthropic, and Broadcom exceeding $500 billion in combined value. New York, meanwhile, froze permits for a year on new hyperscale AI data centres at or above 50 MW of electrical load.

Capital that size does not move. It gets poured into specific buildings, on specific grid interconnects, under specific state permitting regimes. Compute is not a cloud. Compute is real estate with a cooling problem.

The strategic error to avoid is overfitting to a policy forecast. Illinois passed S.B. 315 requiring annual third-party audits of frontier models. Gov. Ron DeSantis called a proposed state-law moratorium "AI amnesty." The ACLU opposes preemption outright. Whether any of the federal bills pass in recognizable form is unclear, and the data is mixed on whether shutdown authority would ever be exercised against a major lab rather than a fringe deployment.

Adopt beginner's mind here. You do not know which bill wins. You do know that optionality is cheaper to buy now than later.

Three signals inside the same shift

2031

Zoom out five years. The question that will matter in 2031 is not which model was best in 2026. It is which companies wrote their architecture so that the answer could change quarterly without a rewrite.

Two asymmetries are worth naming. The first: the cost of maintaining a second, weaker model path is small and known. The cost of discovering you have no second path during a throttling event is unknown and possibly total. That is the definition of an asymmetric bet, and it is the one most teams are currently declining.

The second asymmetry runs the other way. Only cash is real, and the fastest path to cash in 2026 is usually the best available frontier API. A founder who spends four months building self-hosted inference to hedge a hypothetical DHS order has traded certain revenue for imagined safety.

The resolution is a velocity rule, not a purity rule. Get to roughly 70% confidence and move. Ship on the best model. Keep the abstraction layer thin enough that switching costs stay under two weeks of engineering. Treat open weights as negotiating leverage and continuity insurance, not as an identity.

History rewards this posture. Nvidia was weeks from insolvency in the mid-1990s and survived by keeping architectural optionality while competitors bet the company on one graphics standard. The lesson was never "avoid dependency." It was "never let a dependency become irreversible."

By 2031 I expect the durable moat to look boring. Proprietary data, distribution, and workflow ownership. The model will be a component, priced like electricity and regulated like it too.

What to Build This Weekend

Start with visibility, because you cannot hedge what you cannot see. Set up Let Me Know When, a URL watcher that pings you when a public page changes, and point it at your providers' model deprecation pages, terms of service, and status pages. Add the congressional pages for the Great American AI Act and the AI Kill Switch Act. Twenty minutes of setup buys you weeks of warning.

Second, build one thin router. Take your highest-volume prompt, wrap the call in a single function, and add a fallback model behind it. A fallback is just a second provider your code tries when the first one fails or returns a refusal. Log which path served each request so you can see your real dependency ratio.

Third, prove you can run something locally. OpenCode runs Codex-style agent workflows against local or self-hosted models, so use it to reproduce one internal task you currently send to a frontier API. It will be slower. That is fine. You