Texas has spent three years courting the artificial intelligence buildout. On 21 September 2026, Governor Greg Abbott ordered the Texas Commission on Environmental Quality to halt every permit sought by a data centre project, extending a moratorium that previously covered only grid connections. The directive freezes approvals until state agencies finish auditing a connection queue that stood at more than 438,000 megawatts of proposed demand in mid-2026, nearly nine in ten of them data centres. That queue is more than five times the state’s all-time peak electricity demand of 85,508 megawatts. Texas is not retreating from AI. It is discovering that the physical terms on which AI is built are now a matter of public sovereignty, and that those terms had been negotiated by default rather than by design.
Featured image: a protest against data centre construction. Credit: Getty Images via CNBC, 21 September 2026.
What the original report says
The Texas Tribune reported on 21 September 2026, in a story by Paul Cobler, that Governor Greg Abbott ordered the Texas Commission on Environmental Quality to stop issuing new permits for data centres. Abbott’s letter to TCEQ Executive Director Kelly Keel directs the agency to pause all data centre permits until the Electric Reliability Council of Texas and the Texas Water Development Board complete an audit of facilities seeking to be built in the state. The audit seeks information on electricity use, water use, community impact, ownership and tax breaks. The Tribune notes that the order appears designed to close a gap for developers who build “behind the meter” to bypass grid approval, and that it follows an August moratorium on new grid connections and a September directive on water reporting. TCEQ must report compliance to the governor’s office by 19 October 2026. The report also notes that rapid data centre development has angered Texas communities ahead of the November election.
THE SCENARIO: A board approves an eighteen-month AI deployment plan. The models, the training data and the inference capacity all sit with two vendors in one foreign jurisdiction. The chief risk officer reports on cybersecurity, vendor concentration and regulatory compliance. Nobody reports on what happens if the grid that serves those vendors is curtailed, or if the state hosting them changes its permitting rules, or if a single political decision closes the queue the workloads depend on. The plan passes unanimously. It is a good plan on the terms the board was given. Those terms were assembled by people whose incentive was speed, not control.
The Question
When the physical foundations of artificial intelligence become subject to political permission, who actually controls an institution’s AI capability? The question is not whether governments should regulate data centres. The question is who holds the power of interruption over the systems an organisation has already made load-bearing.
Texas has made the answer visible. A governor can freeze permits. A grid operator can defer connection. A water board can demand reporting. A legislature can strip incentives. Each of those levers sits between an AI provider and the electricity that provider needs. An enterprise that has purchased capability without mapping those levers has purchased an option it does not control.
That matters because the sovereignty debate has spent two years focused on the wrong layer. Boards have been briefed on data residency, on the EU AI Act, on model provenance. Comparatively few have been briefed on interconnection queues, water rights, behind-the-meter generation, or the permitting calendars that determine whether the compute they have been promised exists on the date the business case assumes.
For leaders, the practical issue is this. AI dependency is no longer a software procurement decision. It is an infrastructure dependency with a political interface. The question this article examines is whether institutions can convert that dependency into a governed relationship, or whether they will continue to discover the terms of their own capability in someone else’s regulatory diary.
What Happened and Why It Matters
The sequence in Texas is worth reading in full because it shows how quickly a permissive posture can invert. In November 2025, Abbott described Texas as an emerging epicentre of AI development, citing a large Google investment. By June 2026 he had directed the Public Utility Commission of Texas and ERCOT to require data centres to fully fund the electric infrastructure needed to serve them. On 3 August he ordered a moratorium on new data centre grid connections and instructed ERCOT to audit developers on tax breaks, power use, generation, water use, cooling, community impact and facility ownership. Projects that failed what his office called a comprehensive verification and audit were to be denied connection.
On 14 September, after as few as 28 per cent of data centres responded to a state-mandated water usage survey, Abbott directed the Texas Water Development Board to penalise non-reporters and to join the ERCOT audit. Then came the 21 September letter halting TCEQ permits entirely. CNBC reported that the pause covers environmental approvals and that TCEQ is expected to update the governor’s office on compliance by 19 October. CBS Texas reported the same day that the audit covers electrical and water use, grid reliability and security, and impact on local communities.
The governor’s letter is direct on intent. “Simply put, Texans must come first,” Abbott wrote. “Data centres must pay their own way, protect our grid and water, and complete the ERCOT and TWDB audits.” The letter also commits to working with the Legislature to eliminate financial incentives for data centres, and states that facilities “must not use water needed by local communities.”
The scale explains the pressure. ERCOT’s own documentation records that large users totalling more than 438,000 megawatts of proposed demand had entered its connection queue by mid-2026, with nearly 90 per cent of requests from data centres. ERCOT’s all-time peak demand record is 85,508 megawatts, set in August 2023. ERCOT also reported receiving 198 new large load interconnection requests in the first quarter of 2026 alone.
Queue figures are requests, not commitments, and ERCOT has repeatedly cautioned that many projects may never be built or may need far less power than requested. That caution cuts both ways. It means the headline number overstates near-term load. It also means the permitting system was making consequential decisions on the basis of data that was, by the state’s own account, incomplete. Fewer than three in ten data centres answered the water survey. A planning regime cannot allocate water and transmission on a 28 per cent response rate.
That matters because the politics have moved faster than the infrastructure. Brookings analysis records that public opinion has shifted from cautious optimism to outright opposition, with 64 per cent of Reuters/Ipsos respondents saying rapid data centre construction is not a good thing, 77 per cent worried about electricity rates, and only 14 per cent willing to live near a facility. CNN reported that more than $31 million had been spent on political advertising mentioning data centres by late August, with over 99 per cent of that spending opposing AI infrastructure. NPR put the figure above $45 million by early September. A June Reuters/Ipsos poll found 77 per cent concerned about rate increases and a Gallup poll in March found 71 per cent opposed to local construction.
None of this is a Texan anomaly. Bisnow reported that the US House passed the Ratepayer Protection Act 417-3, requiring states to “consider” making data centres of 100 megawatts or more bear the full cost of the electricity serving them, before the bill stalled in the Senate. Customers on PJM Interconnection, the largest US grid, have already seen bills rise largely on data centre demand. FERC has issued show-cause orders to all six of its jurisdictional grid operators, directing them to justify or reform large-load connection rules. Ohio, Virginia, New York and Pennsylvania have all moved toward cost-shifting protections, moratoriums or guardrails.
The sovereignty signal is not that one governor changed his mind. It is that the physical preconditions of AI capability have become electorally salient, which means they are now permanently contestable. Permit regimes can tighten within a single political cycle. Interconnection rules can be rewritten. Water allocations can be revoked. Tax incentives can be repealed. For an institution whose operating model assumes those conditions are stable, that is an unhedged exposure.
The Sovereignty Risk
The sovereignty risk in this story is not that Texas might slow down AI. It is that the audit revealed how much of the AI buildout was never governed by anyone with a duty of care to the institutions depending on it.
Consider the power shift first. A hyperscaler that has secured a grid connection, a water allocation and a tax abatement holds something a state cannot easily replicate. The state controls permitting, but the operator controls the capacity. When permitting tightens, the operator’s existing assets appreciate and the state’s leverage over incumbent operators diminishes. Every additional restriction on new entrants is, in effect, a barrier to entry that protects whoever already connected. The practical consequence is concentration: fewer, larger, more entrenched providers, each with more negotiating power than before.
That is the first sovereignty test an enterprise should apply. Does the regulatory tightening you are reading about increase or reduce your number of viable suppliers? In Texas, the answer is likely to reduce it, at least in the near term, because projects in the queue that cannot demonstrate water and power viability will not be built. The institutions that contracted with those projects now inherit a delay, and delays in compute provision translate directly into delayed capability.
The dependency has a second dimension that is less visible: information asymmetry. Fewer than 30 per cent of data centres answered a state water survey. A planning authority cannot govern what it cannot measure, and a regulator cannot protect citizens from a cost it cannot see. That is a governance failure, but it is also a market opportunity for whichever operator chooses transparency first. An operator that volunteers water intensity, power draw, ownership structure and community impact data is not merely a good corporate citizen. It is positioning itself as the low-risk counterparty in a market where regulatory risk is now the dominant cost of AI infrastructure.
The accountability gap is the third dimension. Abbott’s letter instructs TCEQ to pause permits until ERCOT completes its review, but Texas grid regulators have estimated the audit could last until late next year. TCEQ must report on compliance by 19 October, but the substantive question of how much load the grid can carry does not resolve on that timeline. The result is a period during which permitting authority is suspended and the underlying facts remain unknown. Projects sit in limbo. Capital is committed but unproductive. Nobody is accountable for the delay because no single agency owns the decision.
For institutional buyers, the implication is that the AI capacity they have been sold is contingent on decisions taken by agencies they have no relationship with, in jurisdictions where they have no standing and where the political calendar runs on a different clock to their own planning cycle. A hospital group, a bank, a university or a ministry that has moved critical workloads onto rented capacity has transferred its operational continuity to a permitting process it cannot influence and did not examine.
The institutional capability risk is the most quietly damaging of the four. If the decision to adopt AI was made without an assessment of power, water, permitting and political exposure, then the organisation has no internal capability to reassess it now. There is no team that owns infrastructure dependency mapping. There is no register that tracks which workloads depend on which facilities in which jurisdictions. There is no threshold that would trigger a re-evaluation if a jurisdiction changed its rules. The audit that Texas has just forced on data centre developers is the same audit that most enterprises have never run on their own AI supply chain.
That matters because the cost of the gap compounds. An organisation that discovers its dependency during a disruption pays the disruption cost plus the remediation premium. An organisation that maps the same dependency in advance pays only the mapping cost. The difference between those two numbers is the entire business case for sovereignty work in the AI era.
The TEE Method Response
The TEE Method organises sovereignty work into three connected moves: Talent, Enterprise and Ecosystem. Each has a specific application to infrastructure dependency.
Talent is the capability to ask the question in the first place. Most organisations do not have a person whose remit includes knowing which regulator controls the power supply to the facilities running their models. That is not a technical skill and it is not a legal skill. It is a governance skill: the ability to trace a business dependency through an infrastructure chain and into a political system. Building it requires deliberate hiring or deliberate secondment, not a training module. The practical first step is to name one accountable owner for infrastructure sovereignty and give them authority to ask vendors questions the procurement process has never asked.
Enterprise is where the dependency becomes contractual. Every AI vendor agreement should now disclose the physical facilities serving the contracted workloads, the jurisdictions those facilities sit in, the ownership structure of the facility operator, and the interconnection status of any capacity not yet operational. An agreement that promises compute without naming the facility is an agreement that has not been priced for regulatory risk. The enterprise move is to convert that disclosure into terms: service credits or exit rights tied to material permitting delay, a contractual right to relocate workloads between named regions, and a requirement to notify on regulatory events affecting named facilities. Texas did not change its law overnight; it escalated over four months. A vendor with a notification obligation would have given its customers four months of warning.
Ecosystem is the recognition that no single institution can hold this position alone. A bank, a hospital, a university and a public agency in the same jurisdiction face identical exposures and are individually too small to influence permitting outcomes. Pooled, they become a constituency with standing. That can take the form of shared infrastructure requirements, common disclosure standards demanded of vendors, regional compute arrangements, or simply a shared evidence base that lets a regulator make decisions on real data rather than a 28 per cent survey response. The ecosystem move is the one most often skipped, because it requires an organisation to accept that its individual sovereignty is partly a collective property.
Applied to this news, the TEE response reads as follows. Talent asks who in the organisation owns infrastructure dependency and whether that person has ever read a permitting calendar. Enterprise asks what the vendor contracts say about facilities, jurisdictions and delay, and answers honestly that they say very little. Ecosystem asks whether the organisation’s peers have the same problem and whether a joint position would be stronger than four separate ones. None of the three requires new technology. All three require a decision to treat AI capability as governed infrastructure rather than purchased service.
Sovereignty Test Matrix
The following matrix scores five domains against the infrastructure dependency revealed by the Texas moratorium. A score of one means the organisation has effectively outsourced control. A score of five means it holds meaningful, tested agency. Most institutions assessing themselves honestly will score between one and three on four of the five domains.
| Domain | Score | Weak signal | Stronger signal |
|---|---|---|---|
| Political | 2 | No map of which regulators control the power, water or permits serving contracted AI capacity; no standing in those jurisdictions. | A named register of governing authorities per facility, with a monitoring routine and a defined escalation path on regulatory change. |
| Economic | 2 | Compute cost modelled on contract price only, with no allowance for permitting delay, curtailment or re-siting. | Total cost model includes regulatory delay scenarios, alternative siting costs and the price of redundancy. |
| Cultural | 2 | Infrastructure dependency treated as a technical matter owned by IT and unworthy of board attention. | Board receives a standing infrastructure sovereignty report alongside cyber and regulatory risk. |
| Intellectual | 1 | No internal capability to trace a business workload through to the physical facility and its governing jurisdiction. | A documented dependency graph linking critical workloads to facilities, operators, jurisdictions and permits. |
| Technological | 2 | Single-region, single-vendor inference with no tested migration path and no portability evidence. | At least one exercised failover to a second region or provider, with measured migration time and cost. |
The political domain is scored low because of the asymmetry that Texas exposed. Institutions make adoption decisions on commercial timelines and discover regulatory constraints on political timelines. The economic domain is scored low because permitting risk almost never appears in a compute business case, even though it can add years to delivery. The cultural domain matters because the issue does not reach the board. The intellectual domain scores lowest because the dependency graph simply does not exist in most organisations, which means every later assessment is guesswork. The technological domain is where most organisations feel confident and are usually wrong, because an untested failover is not a failover.
The sovereignty test is not whether an organisation can operate without foreign or third-party infrastructure. No institution of any size can. The test is whether it can describe, today, which of its critical functions would stop if a specific facility in a specific jurisdiction were curtailed, delayed or de-permitted, and how long it would take to move. If that description cannot be produced, the dependency is real and unmanaged.
What Leaders Should Do This Week
The Texas directive gives every institution a useful external prompt. The state has demonstrated that permitting regimes can reverse inside four months. The following actions are ordered so that each one is assignable this week and produces an artifact a board can read.
Build the dependency graph. Assign one accountable owner to list every AI workload the organisation considers critical, the model provider serving it, the cloud or hosting provider beneath that, the physical regions involved, and the facility operators in those regions. The output is a one-page register, not a project. This is the artifact whose absence made every other question unanswerable. Time to first draft: five working days.
Ask vendors four specific questions in writing. Which facilities serve our contracted workloads? In which jurisdictions do those facilities sit? Who owns and operates them? What is the interconnection or permitting status of any capacity we have been promised but not yet received? Vendors that cannot answer are disclosing material information regardless. Send the questions to every provider carrying critical workloads and record the responses.
Price the delay scenario. Ask finance to model what happens to the business case if a named region’s capacity slips by twelve months. Texas grid regulators have estimated the audit could run until late next year, so twelve months is not a stress test; it is the observed timeline in a comparable jurisdiction. If the business case fails on a twelve-month slip, the business case was never load-bearing.
Test one failover before you need it. Select one non-critical workload and move it to a second region or provider. Measure the time, the cost and the number of things that broke. An untested migration path is an assumption, and assumptions are the specific failure mode this story exposes. Report the measured numbers to the board rather than the intended design.
Add a regulatory notification clause to new contracts. Require vendors to notify within a defined period of any material permit suspension, interconnection delay, moratorium or change in ownership affecting facilities serving your workloads. This converts a public political event into a private commercial trigger. Texas escalated across June, August, 14 September and 21 September. A notification clause turns four months of warning into four months of preparation.
Set a decision threshold in advance. Define, now, the conditions under which the organisation will re-site, dual-source or pause an AI deployment. A workable threshold names the trigger and the required response: for example, a permit suspension exceeding ninety days affecting a facility carrying tier-one workloads requires activation of the alternate-region plan within thirty days. The value of a threshold is that it removes the decision from the moment of crisis, when the incentives all point toward waiting.
Take the collective position. Identify three peer institutions facing the same exposure and propose a shared minimum disclosure standard for AI infrastructure vendors. This is the ecosystem move, and it is the cheapest lever available. Individually, an institution is a customer. Collectively, a group of regulated buyers in the same jurisdiction is a constituency whose evidence base regulators need. Texas ran its water survey and received responses from fewer than three in ten facilities. Institutions that hold the missing information have leverage they are not using.
The Question Revisited
The question was who actually controls an institution’s AI capability when its physical foundations become subject to political permission. The answer, on the evidence from Texas, is that control currently sits with whoever holds the permits, the interconnection queue position and the water allocation, and that most institutions have never identified those parties or established any relationship with them.
That is the sovereignty failure, and it is not caused by regulation. It is caused by adoption that outpaced governance. Texas is not the first jurisdiction to reverse course on data centre development and will not be the last. FERC has instructed six grid operators to justify or reform their large-load rules. The United States Congress passed a cost-shifting bill through the House and stalled in the Senate. What happened in Texas is a preview of a permanent condition: the physical preconditions of AI are contested, and the contest runs on political time.
Through the TEE Method, the resolution is not withdrawal. Talent builds the capability to trace a dependency into a jurisdiction. Enterprise converts that trace into contractual terms with notification, portability and exit rights. Ecosystem turns individual exposure into collective standing so that the organisation is a participant in the rules rather than a recipient of them. Each of the three is available this quarter and none requires new technology.
The honest conclusion is that most institutions have already accepted a dependency they cannot describe. The work ahead is not to reverse that decision but to govern it, beginning with a register that names the facilities, the operators and the jurisdictions their critical functions depend on. That register is the difference between a strategy and a hope. For boards assessing where to begin, the diagnostic questions are already written and the first artifact takes five days.
[tt_consultation_booking]
Editorial note: This is a news-led Sovereignty Briefing analysis produced with Hermes-assisted research and source checks.
Sources
Paul Cobler, “Gov. Greg Abbott broadens moratorium on data center approvals to include environmental permits”, The Texas Tribune, 21 September 2026.
Office of the Governor of Texas, letter to TCEQ Executive Director Kelly Keel, 21 September 2026.
“Texas Gov. Abbott orders data center permit halt for electricity audit”, CNBC, 21 September 2026.
“Gov. Abbott pauses new data center permits until water, power audit complete”, CBS Texas, 21 September 2026.
ERCOT, “ERCOT’s New Batch Connection Process for Large Electricity Users”, June 2026 (438,000 MW large load queue; 85,508 MW peak demand record of 10 August 2023).
ERCOT, Interconnection and Grid Analysis Update, April 2026 (198 new large load requests in Q1 2026).
“Why data centers are a top issue in the 2026 midterms”, Brookings Institution.
“Data Center Energy Bill Stalls In Congress As Opponents Say It Lacked ‘Real Teeth'”, Bisnow.
“Voters are fed up with data centers. Both parties are trying to cash in for midterms”, GPB News, 5 September 2026.
“Can ERCOT’s new ‘Batch Zero’ process alleviate the burden of 438,000 MW of large load requests?”, Renewable Energy World.
“The midterms’ unlikely villain: Data centers”, CNN Politics, 31 August 2026.
Image credit: protest against data centre construction, Getty Images via CNBC, 21 September 2026 (CNBC).
This article draws on the TEE Method™ framework from SOVEREIGN: Who Owns the Future? For the complete framework and advisory support, see tonishatagoe.com.
Sign in to save