H.R. 9939 turns AI infrastructure into a political fight. Sovereign AI already stepped out of it.

H.R. 9939, the No AI Data Centers on Federal Lands Act, was introduced in the United States House of Representatives by Representative Rashida Tlaib on 23 July 2026 and referred to the House Committee on Natural Resources. It would permanently prohibit AI data centres on federal land, require existing facilities to cease operation within thirty days, remove the infrastructure and remediate the site. It is one bill among many, but it names something that was already true: hyperscale AI is now a real estate fight. Sovereign AI, run on hardware the customer already owns, is not in that fight.

What H.R. 9939 actually does

The bill is short and blunt. It bans the construction and operation of AI data centres on any federal land, including military bases. It defines the category widely, catching any facility above 20 megawatts of capacity, any facility designed to deliver 20 kilowatts or more to a single server rack, and any facility using liquid cooling. Existing sites would have thirty days from enactment to stop operating or stop construction, then the infrastructure would be removed, by demolition if necessary, and the land underneath would be restored. Enforcement sits with the Department of the Interior.

The bill has five cosponsors at the time of writing and sits in a Democrat-minority committee, so its immediate legislative path is difficult. That is not the point. The bill exists because federal land has become a preferred siting strategy for hyperscale AI capacity, and because the appetite for that siting has run ahead of local and political tolerance for it. The Boulder City data centre in Nevada is the case study cited by campaigners, and the Pentagon's own plans for AI facilities on Department of Defense land are under simultaneous scrutiny in a separate congressional strand.

Why the bill matters even if it does not pass

A bill that names a problem changes the procurement conversation whether or not it becomes law. Two things follow.

First, siting risk is now a live variable in AI planning. A regulated buyer that has been told its AI will run on capacity that a hyperscaler is bringing online at a specific federal site now has to underwrite the political permanence of that site, not just its technical availability. A thirty-day removal clause, even hypothetically, is the kind of clause a general counsel does not want to read on a supplier's dependency map.

Second, this is the second time in twelve months that the physical footprint of AI, rather than its capabilities, has become the object of policy attention. The first was the EU AI Act's calendar of enforcement dates, which the Digital Omnibus reset on 24 July 2026 but did not remove. Both signals point the same way. The architecture that survives is the one that does not need a new tract of contested land to run on.

What the bill exposes about the hyperscale AI stack

The choice to put AI data centres on federal land is not incidental. It is a response to three constraints that hyperscale inference cannot avoid: power at grid scale, cooling at grid scale, and a footprint measured in acres rather than square metres. A 20 megawatt facility is a small municipal power draw. A single server rack pulling 20 kilowatts is denser than a normal enterprise data hall can support. Liquid cooling is the concession those densities force. Once the architecture requires all three, the shortlist of viable sites is short. Federal land has been on that shortlist because it is available, cheap and outside local zoning.

H.R. 9939 attacks that shortlist directly. Even if the bill fails, the political cost of siting has risen, and the assumption that AI capacity can be planted on public land at speed is no longer a safe planning assumption for any buyer whose contract depends on it.

The architecture that is not in the fight

Sovereign AI runs on hardware the customer already owns, inside a facility the customer already operates, drawing power the customer already has under contract. It does not need new federal siting, new grid interconnects or new cooling plants. It fits within the existing envelope of a hospital data hall, a bank's server room, a defence integrator's controlled facility, or a manufacturer's on premise infrastructure. The land question does not arise, because no new land is required.

This is a different design decision, not a smaller version of the same design. Sovereign AI trades the economies of hyperscale training for a set of properties hyperscale cannot offer: the model runs where the data lives, no customer data leaves the customer's perimeter, every consequential action is written to a signed audit ledger the customer holds itself, and the whole stack survives whatever the vendor's real estate portfolio does next.

How this maps onto the buyers we work with

Regulated buyers are already asking the sovereign question because their own rules make them ask it. Financial services under DORA, healthcare under the NHS DCB0129 and DCB0160 clinical safety standards, aviation MRO under EASA and CAA record keeping obligations, defence contractors under DFARS 252.204-7012 and the CMMC framework: each of these carries an obligation to hold, evidence and produce data in the customer's own custody. H.R. 9939 does not change those obligations. It adds an infrastructure layer to the same question a compliance officer was already asking: where does this actually live, and who can move it.

• Where does the model run, physically, on which hardware, in which building?
• Who controls the power and cooling that keeps it running?
• What happens to the customer's data if the site is decommissioned, sold or reclaimed?
• Can the buyer prove, offline and without vendor cooperation, what the AI did?

A hyperscale answer to those four questions depends on the vendor's real estate strategy remaining intact. A sovereign answer does not.

What we are building at MICKAI, in one sentence

MICKAI is a Sovereign Intelligence Operating System, a SIOS, that runs entirely on the customer's own hardware, on premise and air gapped, with no data egress. Every consequential action is signed into the Open Audit Record, a post-quantum tamper-evident ledger any outside party can verify offline, in a browser, with no network and no trust in the vendor. Sixty-three studios sit on one operating system, ten of which are production-ready at launch and fifty-three in development. Mickai LTD holds 104 filed UK patent applications carrying 2,340 claims across 13 invention families.

The read for buyers, this week

H.R. 9939 will not, on its current path, become law before the end of the current Congress. It does not need to. It has already changed the diligence question for AI procurement across the United States. A vendor that cannot state, in one sentence, where the AI runs and who owns the ground it runs on, has an answer to prepare. The buyers who write the largest cheques in regulated sectors are the ones who have to ask that question first, and they are asking it now.

The safe posture is not to bet against hyperscale. It is to keep the option of an architecture that does not depend on the outcome of a real estate fight the buyer is not a party to.

Frequently asked questions

Has H.R. 9939 been passed into law?

No. It was introduced on 23 July 2026 and referred to the House Committee on Natural Resources. It has five cosponsors at the time of writing. It has not been reported out of committee and has not been enacted.

Would H.R. 9939 apply to sovereign, on premise AI?

It would not, on any reading of the current text. The bill targets AI data centres on federal land, defined by grid scale power draw, high per rack density and liquid cooling. On premise sovereign AI running inside the customer's own facility on hardware the customer owns is outside the definition and outside the jurisdiction of the bill.

Why is the fight about physical siting rather than model behaviour?

Because the biggest engineering cost of running large models is now power, cooling and land. Model behaviour is regulated by other instruments, from the EU AI Act to sector rules. H.R. 9939 is the first bill in a wave that addresses the physical layer directly, which is why it has caught buyers' attention.

How do regulated buyers plan around this?

By making the physical location of AI inference a stated requirement in the procurement itself, not a vendor implementation detail. That means either contracting for a named, permanent, non federal site, or moving to an architecture that does not depend on any external site at all. Sovereign, on premise AI is the second of those.

What is MICKAI?

MICKAI is a Sovereign Intelligence Operating System, a SIOS, that runs entirely on the customer's own hardware, on premise and air gapped, with no data egress. Every consequential action is signed into the Open Audit Record, a post-quantum tamper-evident ledger any outside party can verify offline, in a browser, with no network and no trust in the vendor. Sixty-three studios sit on one operating system, ten production-ready at launch and fifty-three in development. Mickai LTD holds 104 filed UK patent applications, 2,340 claims across 13 invention families.