A coalition-and-advocacy kit for companies that build on AI but rent the frontier. Part II, the deployment field guide for how to actually operate and build toward contestability, is its companion piece: Building Toward Contestability.
This is advocacy, grounded in but distinct from The Inversion working paper. The paper maps the option space and deliberately refuses to prescribe a statutory form; this kit prescribes. Use it as the argument and the asks, not as neutral analysis.
Who this is for
Capitalized software and SaaS companies, and the mid-tier AI firms, research labs, and large enterprise buyers in the same position. If you build with AI but do not own the compute that produces it, this is written for you.
You are a tenant. You rent the frontier from a market with roughly three landlords, you pay enormous rents, and you have no real exit. The premise of this kit is simple: contestability of the compute layer is your commercial interest, not just the public’s good. That alignment is the entire reason a coalition is possible. Nobody here is being asked to act against themselves.
1. The self-interest case: the part that frees the political capital
You don’t have to care about the concentration of power in the abstract to act here. You only have to read your own income statement and risk register.
- Concentration is a tax on you. When the means of production sits with a few landlords, the rents flow up. Your largest and least controllable input cost is set by people you cannot meaningfully negotiate with, and it will not fall on its own.
- Lock-in is a liability you never booked. No portability means no exit; no exit means no leverage; no leverage means your roadmap, your margins, and your pricing power are hostage to someone else’s decisions.
- You are downstream of a market that can foreclose you. The same landlords increasingly build in your layer. A tenant with no countervailing power is a tenant whose supplier can become its competitor, on terms it has no standing to contest.
Countervailing power, a more contestable compute layer, is what lowers those costs, creates real exit, and restores your leverage. This is not charity. It is margin defense, procurement strategy, and risk management. Frame it to your CFO and your board in exactly those terms.
2. The asks: the heart of the kit
Specific, winnable, ordered from least to most interventionist. Each is given three ways: what it is, what it does for you, and what it does for everyone, plus why it has to happen now rather than later.
Ask 1: Compute transparency
What it is. Disclosure of frontier-scale training runs and large compute-capacity holdings above a defined threshold.
For you. Visibility into the supply and the players in the market you depend on. You cannot plan against a black box.
For contestability. It narrows the information asymmetry between the landlords and everyone else: the precondition for every other intervention.
Why now. It’s the cheapest, least disruptive ask, and far easier to impose before the holders are entrenched enough to resist it. Feasible because, unlike data or algorithms, compute is detectable, excludable, and measurable.
Ask 2: Public training capacity (not just credits)
What it is. Public computational capacity that can actually train at or near the frontier, building well beyond today’s early public-compute efforts (NAIRR in the U.S., AIRR in the U.K., EuroHPC in the EU), which still largely mean subsidized access to the incumbents’ own clouds.
For you. A genuine alternative supplier and a price anchor. Even a credible public option disciplines what the landlords can charge.
For contestability. This is the countervailing institution itself: capacity that sits outside the incumbent firms and can hold open what the market would otherwise close.
Why now. Public capacity is vastly cheaper to stand up before a few firms own all the frontier-scale infrastructure. Insist on the distinction: capacity to train, not credits to rent. The first shifts where capability lives; the second just subsidizes the rent.
Ask 3: Interoperability and portability
What it is. Mandated portability of data, fine-tuned models, and workloads, and interoperability across layers of the stack, so switching costs stay low and dominance at the compute layer can’t be parlayed into capture of the model and application layers above it.
For you. Your exit option, written into law. The single most direct anti-lock-in ask there is.
For contestability. It keeps each layer separately contestable: the property that, in every prior technology, is what made reopening possible.
Why now. Interoperability is easy to require before lock-in makes switching costs prohibitive, and nearly impossible to retrofit after.
Ask 4: Procurement that rewards contestability
What it is. Public buyers, and large private ones, preferring portable, interoperable, and open stacks in what they purchase.
For you. Government is the largest customer in most markets. Making “no lock-in” a buying criterion levels the field for non-incumbent suppliers, which you either are, or want to be able to buy from.
For contestability. Demand is a lever. A large buyer that rewards contestable architectures moves the market without a single new regulation.
Why now. Procurement standards set early become the default everyone builds to.
Ask 5: Antitrust attention to compute-layer exclusivity
What it is. Competition scrutiny at the production layer (the exclusive tie-ups between frontier labs and hyperscalers), not only at distribution, where antitrust usually looks.
For you. It stops the landlords from carving up the market and foreclosing the alternatives you’d want to switch to.
For contestability. The chokepoint here sits at production, rooted in physical scarcity; antitrust aimed only at distribution misses it entirely.
Why now. Exclusivity arrangements harden into structure. Contesting them is far easier before they’re entrenched than after.
3. What NOT to ask for: the ladder-pull rule
Read this one twice, because it is where a tenant coalition either earns its legitimacy or betrays it.
Every ask above is enabling: it lowers barriers, adds suppliers, forces openness, narrows asymmetries. There is a second family of interventions that sounds responsible and is, for an incumbent, a gift. Call it gatekeeping: licensing regimes for model development, mandatory “safety” certifications, model registries, compute permits. These create compliance costs only the largest can clear. Backed by a tenant coalition, they become moats that crush the little builders (the open-source projects, the indie developers, the researchers) who are precisely the people most likely to build the alternatives you want to exist.
The rule, then: favor enabling interventions over gatekeeping ones, always. A coalition that lobbies for licensing and certification is not building countervailing power; it is helping its own landlords pull the ladder up, and calling it safety.
Keep the test simple. Does this ask add suppliers and lower switching costs (enabling), or does it raise the cost of entry (gatekeeping)? If it raises the cost of entry, it is not in this kit, no matter who is selling it to you.
4. The coalition: how the power actually gets organized
Countervailing power has never arrived on its own. In the history this argument draws on, the broad gains from technology came when someone organized a bloc with enough weight to bend the direction of the technology away from the few who owned it: labor, legislation, public institutions. The modern version of that bloc, for the compute layer, is a coalition of tenants: SaaS and software firms, mid-tier AI companies, open-source organizations, research institutions, and the large enterprise buyers who foot the bills.
Its credibility comes from breadth, not from any one company’s lobbying budget. A single-firm campaign reads as one firm’s interest. A coalition that spans the capitalized firms and the open-source and research community reads as a structural argument about the market: harder to dismiss, and harder to capture. So the coalition’s first discipline is to bring the little builders in from the start: their presence is what keeps the asks honest (see the ladder-pull rule), and their alternatives are what make “contestability” concrete rather than rhetorical.
Practically: form or join a tenants’ group organized specifically around compute contestability; adopt the five asks as a shared agenda; and divide the labor. The firms with government-relations capacity carry the asks, the open-source and research members supply the proof and the legitimacy, and everyone holds the message line.
5. How to use this kit: who carries what
- Your CFO and board are the first audience, not the last. The self-interest case (Section 1) is what frees the political capital. Lead internal conversations with margin, risk, and optionality, not ideology.
- Your government-relations and policy team carries the asks (Section 2) to legislators and agencies, in the right language: competition, optionality, supply-chain resilience, national interest. Never “anti-AI.” You are for a competitive AI market.
- Your CTO and engineering org make the operational choices that turn the argument into proof: multi-cloud, open-weight adoption, portability by default. That’s Part II of this playbook (Building Toward Contestability), and it matters here because a coalition that practices contestability is far more credible than one that only lobbies for it.
- The coalition coordinates, so no single firm is out on a limb and the asks stay aligned and enabling.
The asks at a glance
| The ask | What it does for you | What it does for contestability | How far it goes |
|---|---|---|---|
| Compute transparency | Visibility into the market you depend on | Narrows the information asymmetry | Least |
| Public training capacity | An alternative supplier; a price anchor | Builds the countervailing institution | Low to moderate |
| Interoperability & portability | Your exit option, in law | Keeps every layer separately contestable | Moderate |
| Procurement preference | Levels the field for non-incumbent supply | Uses demand as a lever, no new regulation | Moderate |
| Antitrust on exclusivity | Stops the market being carved against you | Reaches the chokepoint at production | Most |
The honest footnotes
A coalition like this is a coalition of convenience, and that is the honest word for it. Many of the firms best placed to push these asks also do things, in their products and their labor practices, that the broader Choose Otherwise project argues against. This kit takes no position on any of that. It is about a single variable, who controls compute, and allying there is not a moral endorsement in either direction. Pretending it were would be the dishonest move.
And one reminder, because everything turns on it: the ladder. A tenant coalition that backs gatekeeping rules just entrenches the landlords it set out to check. If a single line from this kit survives the meeting, make it that one.