Tag: Technology Policy

  • The First 100 Days: A Technology Agenda for the Next Congress

    By the Founders, Department of Technology | San Diego, California | November 13, 2026

    A new Congress has 100 days to show it can govern the technology buildout instead of reacting to it. Here is a practical agenda that members of either party can sign onto.

    In our last piece, we described five mistakes the previous Congress made on data centers and AI. This one is the to-do list. We built it from items that already have bipartisan footprints, because the fastest agenda is one that members have already voted for, sponsored or said they support. One timing note: bills that do not become law expire when a Congress ends, so even measures that passed one chamber, like the Ratepayer Protection Act, have to be introduced again.

    1. Days 1 to 30: Hold one set of hearings on the whole picture

    Today the pieces of this problem sit in different committees. Energy committees see the grid. Commerce committees see competitiveness. Finance and tax committees see revenue. Each hears from its own witnesses, and nobody hears the tradeoffs.

    Hold a joint hearing series with the same witnesses in the same room: grid operators, state utility regulators, economists, county officials and workers. The evidence on each risk is already strong enough to start. In the largest U.S. grid market, wholesale power prices rose almost 76 percent in a year and the market monitor called the effects irreversible. At the same time, BloombergNEF expects China to add more than six times as much generating capacity as the United States over five years. Members should hear both facts at once.

    2. Days 1 to 45: Reintroduce ratepayer protection and settle the real disagreement

    The Ratepayer Protection Act cleared the House Energy and Commerce Committee 52 to 0, passed the House 417 to 3, and then failed a Senate procedural vote 57 to 43 that needed 60. The two sides disagreed about one question: whether states should merely be required to consider making large data centers pay for the grid upgrades built to serve them, or whether the cost allocation should be mandatory, as in the Democratic GRID Savings Act.

    Both positions have a principled basis. Utility regulation has long been a state responsibility, and voluntary pledges have a mixed record. We suggest negotiators look for a middle design: a mandatory core for the very largest new loads, with states keeping flexibility over how to implement it. That is our suggestion, not a settled answer, and the details matter. The point is to have the argument in February rather than the weeks before an election.

    3. Days 30 to 60: Open the deals

    Secret agreements are the issue where public frustration is highest and the policy fix is simplest. Bloomberg Law found that 25 of 31 proposed data center projects in Virginia involved nondisclosure agreements, and that Microsoft says it is working to end any active NDAs governments have signed. Lawmakers in more than a dozen states have introduced bills to ban such agreements, but only Delaware has passed a transparency measure through its legislature, according to MLex. A bipartisan group in Congress has now introduced a federal bill to ban NDAs that keep officials from discussing proposed data center projects, while protecting trade secrets.

    Move that bill. And because Congress’s reach is strongest where federal assets are involved, require public disclosure of any data center deal that touches federal land, federal funds or federal permits.

    4. Days 60 to 90: Ask the revenue question before the layoffs

    We do not know how much AI will displace work, and forecasts vary widely. We do know how exposed the tax base is if it does. A RAND analysis found that 84 percent of federal revenue in 2024 came from individual and payroll taxes, and that losing high-paid white-collar jobs would hurt revenue disproportionately. New York City’s Comptroller has already modeled five AI scenarios for the city’s finances.

    Ask the Congressional Budget Office and the Government Accountability Office to publish scenario analyses of federal revenue under different displacement paths, and encourage states and counties to do the same. Planning for a range of outcomes is not predicting the worst one.

    5. Days 90 to 100: Name an owner

    Every item above would be easier to carry out, and harder to forget, if someone were accountable for them together. Begin by creating a standing joint committee with jurisdiction over the whole technology buildout. Then open the larger debate about a Department of Technology and elected technology leaders at the local, county, state and federal levels, including a Secretary of Technology.

    Two rules for all 100 days

    • No riders. Questions as large as federal preemption of state AI laws should not be decided inside a defense or budget bill. They need their own hearings and their own text.
    • Settle the hard disagreements early. The last Congress’s one bipartisan success stalled because the central dispute surfaced a few weeks before Election Day.

    The case against our case

    A fair agenda includes the strongest objections.

    • It is too much, too fast. Skeptics will say five items in 100 days is unrealistic, especially in a Senate that needs 60 votes for most bills. We would rather offer a full list and let members pick; the first two items are the most achievable.
    • Congress has limited reach over local contracts. A federal NDA ban cannot easily dictate how a city signs agreements except through federal funding, permits or land. Developers also argue that confidentiality protects competitive information, and Microsoft has said the agreements were meant to protect sensitive information. The bill’s trade-secret carve-out is where this fight will be decided.
    • Mandatory cost allocation could raise costs or slow building. Supporters of the optional approach worry that rigid federal standards could discourage investment at a moment when the country needs more capacity.
    • Revenue scenarios are speculative. The same New York report acknowledges it is hard to connect today’s hiring trends to AI with certainty. We agree, which is why we ask for a range of scenarios and not a single forecast.

    Start now

    None of this requires a party to give up its principles. It requires members to decide that the question of who is accountable for the technology buildout is theirs to answer. We will keep publishing the evidence and we will tell you plainly when we think a proposal falls short. Our proposal is at department.technology/.

  • Preemption Without an Owner: Why the Federal-State AI Fight Needs Someone Accountable

    By the Founders, Department of Technology | San Diego, California | November 13, 2026

    Washington is arguing over whether federal law should override state AI rules. Both sides have real arguments. But whoever wins, someone has to enforce the result and answer to voters for it, and right now no one does.

    Few questions in technology policy have produced more heat than federal preemption: whether Congress should replace the growing body of state AI laws with a single national standard. This piece does not argue that preemption is right or wrong. We lay out the strongest case on each side, review what the record shows, and make one claim that holds under either outcome: a national standard, or a patchwork, without a named, accountable owner is not a policy. It is a vacuum.

    The case for a national standard

    Supporters of preemption make an argument that deserves a fair hearing. AI development does not stop at state borders, they say, and companies facing dozens of different rules face rising compliance costs that favor the largest firms. In December 2025, a presidential executive order called for a uniform national framework, directed the Justice Department to challenge state laws it deemed onerous, and tied some federal broadband funding to state policy. Industry groups and senators including Ted Cruz have argued for explicit federal preemption of conflicting state laws, and some supporters now say the longer Congress waits, the harder preemption becomes as more state laws pile up.

    In March 2026, the White House sent Congress a national framework that called for preempting burdensome state AI laws, with exceptions for child safety, general consumer-protection and fraud laws, state procurement and state zoning for AI infrastructure.

    The case against

    Opponents make an equally serious argument. States, they say, are closer to the people affected and can respond faster than Congress, and preemption without enforceable federal protections would leave residents with fewer remedies than they have today. The opposition has been broad. In July 2025, the Senate voted 99 to 1 against a proposed state AI moratorium. When it returned in the defense bill, more than 300 state lawmakers from across the political spectrum urged Congress to reject it, and prominent Republicans including Senator Josh Hawley, Representative Marjorie Taylor Greene and Arkansas Governor Sarah Huckabee Sanders spoke against it. The defense bill text did not include the moratorium, and House Republican leaders were left searching for another vehicle.

    What the record shows

    Whatever one thinks of preemption, the practical result so far is a stalemate. Two attempts to attach it to must-pass bills failed. The White House framework has not become law. By one account, Congress left for its August recess without moving a federal AI framework out of committee. Meanwhile, 29 states enacted AI legislation in 2026, and for the first time both Republican- and Democratic-led states passed laws adding requirements or oversight for data centers.

    The states did not wait for permission, and no one in Washington was positioned to tell them to stop or to give them something better to follow.

    The part both sides already agree on

    Here is the detail that gets lost in the argument. Even the strongest preemption proposals carve out data center infrastructure. The December executive order said legislation should not preempt state laws on data center infrastructure, and the March framework would leave state zoning for AI infrastructure in place. In other words, the people who favor the broadest national standard agree that decisions about where data centers go, how they are powered and who pays remain local and state decisions.

    That has a consequence. Whatever Congress decides about model rules, child safety or algorithmic discrimination, the data center decisions that voters care about most, and that voted on November 3, will be made by city councils, county boards, state utility commissions and legislatures. Those officials are making multibillion-dollar decisions with no technology counterpart whose job it is to see the whole picture.

    Where the owner is missing under each outcome

    • If broad preemption passes, a federal standard replaces state rules. Who enforces it, who updates it as the technology changes, and who do residents call when it fails? Without an accountable federal office, critics of preemption are right that residents lose remedies.
    • If preemption fails, the patchwork grows. Companies face rules that differ by state, and states and counties negotiate against the same few buyers with little shared information. Without a coordinating office, supporters of a national standard are right that the burden falls hardest on smaller firms and communities.
    • If a narrow hybrid passes, federal and state authority meet at a boundary that will be contested. Someone has to be accountable for where that boundary sits and for what falls through it.

    In every case, the missing piece is the same: an office that answers to voters for how the system works as a whole.

    Three tests for any national AI standard

    We take no position on how broad preemption should be. We do suggest three tests for any proposal, from any party:

    1. Is the federal protection enforceable? If a national standard replaces state rules, it should carry real enforcement and clear remedies, not just guidance.
    2. Is there a named, accountable official? Someone should be answerable to voters, and to Congress, for how the standard is administered and updated.
    3. Does it keep data center decisions local and public? Siting, rates and disclosure of public deals should stay with local and state governments, with transparency rules attached.

    And a procedural test: was it debated on its own, with hearings and a public text, instead of inside a must-pass bill?

    The case against our case

    • Accountability can be built into existing agencies. Some will say a regulator or a White House coordinator could play this role without a new elected office. That is a reasonable design, though we think voters should be able to hold someone to account directly.
    • Local decisions are the problem. Some supporters of preemption believe local control over data centers is what slows the buildout, and would prefer more federal permitting authority. The administration has proposed streamlining federal permitting for AI data centers, and reasonable people can weigh that against local consent.
    • It adds a layer. Another layer of officials could slow decisions. We would answer that the status quo already has many layers and no owner, but we welcome the debate.

    The question underneath

    The preemption fight will continue, and we hope it is resolved by open debate and not by rider. But the question underneath it will not go away: when the technology goes wrong, or when it goes right and the gains are not shared, who is responsible? We think the answer should be someone voters can elect, at every level of government. Find our proposal at department.technology/.

  • A Briefing for the New Congress: What the Last Two Years Got Wrong, and What to Do Differently

    By the Founders, Department of Technology | San Diego, California | November 13, 2026

    To the newly elected: you are about to inherit the most consequential technology decisions in a generation, and the last Congress left you a map of what not to do.

    This is not a partisan scorecard. The record we describe below includes members of both parties acting in good faith and members of both parties falling short. Newly elected members of Congress are sworn in this January, and we think you are better served by an honest account of what happened than by a flattering one. We start with what worked.

    What the last Congress got right

    It is easy to say Congress did nothing on data centers. That is not quite true. In July, the House Energy and Commerce Committee passed the Ratepayer Protection Act 52 to 0, a bill from Republican Gabe Evans and Democrat Kathy Castor. In September, the full House passed it 417 to 3 under a procedure that required a two-thirds majority. The bill would require state utility regulators to consider standards that make large data center customers pay the full cost of the grid upgrades built to serve them. In a polarized era, that is a real accomplishment, and it shows what is possible when members start from a problem their constituents can feel: the electric bill.

    Then came the mistakes.

    Mistake 1: Trying to pass the big thing through the back door

    For two years, advocates of a national AI framework tried to move federal preemption of state AI laws by attaching it to must-pass bills. Those efforts, including the defense authorization bill and the 2025 budget reconciliation law, were unsuccessful. Senators rejected a broad preemption proposal 99 to 1. In March 2026, the White House sent Congress a national policy framework that again called for preempting burdensome state laws while carving out areas such as child safety and state zoning for AI infrastructure. By one account, Congress left for its August recess without moving a federal AI framework out of committee.

    The lesson is not that preemption is right or wrong. Reasonable people disagree, and we lay out both sides below. The lesson is that a question this large cannot be settled by rider. It needs its own hearings, its own text and its own public debate.

    Mistake 2: Letting the one agreement become a campaign prop

    The Ratepayer Protection Act did not stall because no one wanted it. It stalled because of how it arrived. On September 30, five weeks before Election Day, the Senate voted 57 to 43 on a procedural motion that needed 60 votes. Only four Democrats voted yes.

    The two sides told different stories, and each has a serious argument:

    • Republican supporters said the bill had passed the House almost unanimously, with Democratic votes and Democratic cosponsors, and accused Senate Democrats of refusing to take yes for an answer for political reasons.
    • Senate Democrats said the bill’s core protection is optional, since states need only consider standards, and offered their own GRID Savings Act to make large customers actually pay for grid upgrades. Senator Schumer called the Republican bill “optional” and toothless.

    We take no side on optional versus mandatory. We do note that Reuters reported Senate Republicans aimed to improve their midterm prospects with the vote, and that the bill arrived at the end of the session, after the Senate had not moved on data centers for months. A measure introduced in July and brought to a test vote in the last weeks before an election will be judged as a campaign move whatever its merits. The practical lesson for new members: if you want a data center bill to become law, negotiate the mandatory-versus-optional question in the spring, not the fall.

    Mistake 3: Solving one slice of a three-part problem

    Even if the Ratepayer Protection Act had passed, it would have addressed one of three risks we described in our earlier piece, “The Data Center Casino.” It deals with who pays for grid upgrades. It does not address how fast we can build power to compete with China, the financial exposure if the AI investment boom turns, or what happens to federal and state revenue if AI displaces high-paid white-collar work. Congress legislated the slice that polls well. No committee is responsible for the whole.

    Mistake 4: Leaving a vacuum for states and for secret local deals

    When Congress does not act, others do. In 2026, 29 states enacted AI legislation, and for the first time both Republican- and Democratic-led states passed laws adding requirements or oversight for data centers. That is encouraging, and it is also a patchwork. Meanwhile, at the local level, many deals are still negotiated under non-disclosure agreements. As we noted in our open letter, 25 of 31 Virginia local governments involved in data center development had signed NDAs. Residents learn about the largest infrastructure project in their county after the decision is made.

    Mistake 5: Nobody owns it

    This is the mistake underneath the others. AI, data centers, the power grid, student data, surveillance and the workforce are governed by a patchwork of committees and agencies. Each handles a slice, and no one is accountable to voters for how the slices fit together. We raised this directly. In February 2025 we submitted a proposal to the White House Office of Science and Technology Policy for a national AI action plan, and in July 2025 we published a response to the plan that followed. We have not had a substantive engagement with either. We mention it not as a complaint, since a good idea should not depend on one administration, but because it illustrates the structural problem: there was no office whose job it was to say yes or no.

    What new members can do in the first 100 days

    1. Hold hearings on the whole picture. Grid capacity, public cost, competitiveness and fiscal impact together, with the same witnesses in the same room.
    2. Settle optional versus mandatory early. Bring Republicans and Democrats who both voted for ratepayer protection into one room and write the version that can reach 60 votes.
    3. Keep big structural questions out of must-pass bills. Preemption, if it happens, should be debated on its own.
    4. Open the deals. Require public disclosure of data center agreements that involve public land, public utilities or public incentives.
    5. Create a single point of accountability. At minimum, a standing joint committee with jurisdiction over the whole question. Ideally, an elected technology leader at every level of government, including a Secretary of Technology at the federal level, whose job is to answer to voters for the combination.

    The case against our case

    A fair briefing includes the strongest objections.

    • Preemption advocates argue a patchwork of state laws slows innovation and that national competitiveness requires one national standard, and they note that the longer Congress waits, the harder preemption becomes as states pass more laws (Broadband Breakfast). Opponents respond that preemption without enforceable federal protections would leave residents with fewer remedies.
    • Optional standards may be the right design. Supporters of the Ratepayer Protection Act argue that utility regulation has long been a state responsibility and that a federal mandate would overreach. Critics respond that voluntary pledges have not been enough.
    • Congress already has committees for this. Some will say the problem is political will, not structure, and that a new office adds a layer. Our view is that structure shapes will: when every committee owns a slice, no one is blamed for the whole. But we welcome the argument.

    The first step

    The new Congress does not need to solve everything in January. It needs to stop treating the central question, who is accountable for the technology buildout, as someone else’s job. Start the conversation about a Department of Technology and elected technology leaders at every level. We will help, and we will be honest with you when we think you are wrong. Find our proposal at department.technology/.

  • Voters Didn’t Reject AI. They Rejected Building It Backward


    By the Founders, Department of Technology | San Diego, California | November 9, 2026


    Voters Didn’t Reject AI in 2026. The ever-increasing number of paid memberships for OpenAI, Anthropic (Claude AI), Google Gemini, and others has proved this to be the case.

    They Rejected Building It Backward.

    On November 3rd, voters across ten states said no to data centers almost every time they were asked. The reason is not that they oppose AI. It is that they have watched the buildout run ahead of the power grid, and they suspect the people steering it are not putting the public first.

    In Ashville, Ohio, the village council had already approved an agreement with a data center developer. Residents petitioned to put it on the ballot anyway. On Election Day, 82% of voters rejected the deal, according to Northeast Ohio reporting by Fox 8.
    Ashville was not an outlier. Ballotpedia counted 41 local data center measures on the November 3rd ballot across 10 states, with Ohio accounting for 21. Voters favored limiting data center development in 40 of the 41. The one exception was Harwood, North Dakota, where voters defeated a proposed review requirement. Voters in Granville, Ohio, approved a ban by 84%. In Urbana, 77% backed one.

    This was not a one-day mood. A Heatmap poll reported by Forbes found more than 70% of Americans would oppose a data center planned for their area, with rising electricity prices the top reason.
    Many voters may not weigh how much national AI leadership depends on new computing capacity. That is a fair criticism, and we take it seriously below. But the voters who showed up made a different judgment, and it deserves a hearing: they believe this buildout is being driven first by profit, with national strategy a distant second, and that their concerns about power bills and the environment were treated as obstacles rather than inputs.
    The comparison with China shows why that suspicion is not irrational.

    What China Did First
    China and the United States are in a race for AI leadership, but they are not running it from the same starting line.
    According to Inside Climate News, China added an estimated 540 gigawatts of generating capacity in 2025, more than 80% of it renewable. The United States added about 54 gigawatts, roughly 60% renewable. Al Jazeera reports that China already generates more than twice as much electricity as the US, and that BloombergNEF expects China to add more than six times as much generating capacity over the next five years.

    Transmission, the high-voltage lines that carry power from where it is made to where it is used, tells the same story. Grid Strategies estimates the US needs roughly 5,000 new miles of high-capacity transmission per year but has averaged only about 600 since 2017, according to Yale Climate Connections. China, by the same report, has been building over 1,000 miles of ultra-high-voltage lines per year. In January, China’s State Grid announced a CNY 4 trillion (about $580 billion) investment plan for 2026 to 2030, up 40% from the prior plan, including 15 new ultra-high-voltage lines, according to Enerdata.

    Getting connected is the other bottleneck. A ChinaTalk analysis finds that the time for a US generation project to connect has grown from under two years in the early 2000s to over four years for projects built from 2018 to 2024. Because China plans generation, transmission, and demand together under state direction, the American concept of a years-long interconnection queue for large loads barely exists there.

    The US leads on the other side of the equation. Stanford’s AI Index counts an estimated 5,427 US data centers in 2025 versus 449 in China. As one analyst quoted by Al Jazeera framed it, the US has the chips and is short on power, while China has the power and is short on chips.

    Cost is where the picture gets less lopsided. Reporting from bne IntelliNews puts Chinese industrial electricity rates roughly in line with the US, at about $0.08 to $0.11 per kilowatt-hour, while Chinese households pay about $0.08 versus $0.15 to $0.17 for American households. The Financial Times reported that Beijing is subsidizing power for data centers using domestic chips down to about 5.6 cents per kilowatt-hour, against a US industrial average of 9.1 cents.

    The Warnings Were Public
    Here is the part voters understand best. The power shortfall was not a secret.

    In its Long-Term Reliability Assessment, the North American Electric Reliability Corporation (NERC) raised its ten-year summer peak demand forecast sharply and said new data centers account for most of the projected growth, with several regions at high risk of supply shortfalls by the end of the decade. This spring, NERC issued a Level 3 alert, its highest tier, over large computational loads like data centers, and found that grid operators cannot reliably model how 77% of operating data center load behaves during grid disturbances, according to reporting by SAN and Utility Dive.

    Grid operators are also being overwhelmed at the front door. Texas’s ERCOT built its process for a large-load queue of 40 to 50 projects, then received 225 new interconnection requests in 2025 alone, about 70% of them data centers, according to the Texas Tribune. ERCOT itself cautions that many requests are speculative and its demand forecast is probably inflated. That caution cuts both ways: it means some projects may never be built, and it also means officials are approving and planning against numbers nobody can fully trust.
    We want to be precise. We are not claiming that officials privately concealed a known shortfall. What the public record shows is that independent grid reliability authorities published escalating warnings, and in many places projects advanced anyway, sometimes under nondisclosure agreements and sometimes over local objections. When voters see that sequence, and then see their bills, they draw a conclusion.

    Why the Bill Concern Is Real, and Why It Is Contested
    The electricity price argument is the most contested part of this debate, so it deserves care.

    In PJM, the grid operator serving 13 states and Washington, DC, capacity auction prices rose from $28.92 per megawatt-day for 2024/25 to $329.17 for 2026/27, roughly tenfold. PJM’s independent market monitor attributed 63% of the increase in the 2025/26 auction to data center demand, about $9.3 billion in added costs recovered from ratepayers, according to analysis by the Institute for Energy Economics and Financial Analysis.
    But capacity charges are only about 10% to 20% of a typical bill, as FactCheck.org notes, and several studies find no clear nationwide link between data centers and household rate increases. A Rutgers policy lab analysis found no statistically significant local bill effect yet, while warning that “not yet” is not the same as “nothing to worry about.” FactCheck.org also cites an estimate that capacity increases added about $18 per month to the average home bill in western Maryland and $16 in Ohio between mid-2025 and mid-2026, with 63% of that tied to data centers.

    The honest summary: the bill effect is real and documented in some regions, overstated in others, and likely to grow if the cost-allocation rules do not change. Voters are reacting to the direction of travel, and the direction is up.

    “Doesn’t Slowing Down Hand China the Win?”
    This is the strongest objection, and it is not frivolous. If China is building power at this scale and pricing it for AI, then any delay in the US could cost real strategic ground.

    But the China comparison cuts against the current approach, not for it. The binding constraint on American AI is no longer chips alone. It is power, transmission, and public consent, and the backlash of November 3rd shows what happens when a project skips the last of those. Projects imposed on unwilling towns get voted down, and every vote down is a delay no strategist planned for.

    We also want to avoid the lazy conclusion that the US should copy Beijing. China’s model rests on centralized, state-directed planning that the US system, with its many jurisdictions and its commitment to local consent, cannot and should not replicate. China’s buildout has its own costs: Oxford Institute for Energy Studies notes it added as much as 80 gigawatts of new coal-fired power in 2025, and Al Jazeera reports a gap between China’s data center capacity and how much of it is actually used. The lesson is narrower and more useful: China treats the grid as infrastructure to be built before the load arrives. We have been building the load and hoping the grid catches up.

    The Fix: Grid First, Public Accountability Always
    The way out is not a moratorium on AI infrastructure and not a free pass for developers. It is a different order of operations. The framework at Department of Technology proposes layered accountability: a confirmed federal Secretary of Technology, and elected technology leaders at the state, county, and local level who answer to the voters whose bills and water supply are on the line. Concretely:
    Grid first. No major data center approval without a utility- or grid-operator-certified plan covering power supply, transmission, and interconnection timing.

    Make the large loads pay. Require large-load tariffs so the full cost of new generation and grid upgrades lands on the facilities that cause it, not on residential ratepayers. Pennsylvania’s governor has moved this direction with an executive order requiring developers to build, bring, or buy their own capacity, and a Republican challenger has pressed for a ratepayer-protection moratorium. This is not a one-party issue.

    Publish the grid impact. Before any rezoning or tax abatement, publish projected power demand, water use, and cost allocation in plain language, and limit NDAs on projects that use public land, money, or incentives.

    Fund the wires. Treat transmission and interconnection reform as national AI infrastructure, because it is. If the United States wants to out-build China on compute, it has to stop being out-built on the grid.

    What Each of Us Can Do
    Congress: Prioritize transmission permitting and interconnection reform, and pass transparency rules for large-load projects.
    Statehouses and utility commissions: Adopt large-load tariffs and grid-readiness tests before approvals, and require public notice before rezoning or tax breaks.
    Candidates and local officials: Answer two questions in public. Who pays for the power this project needs? Where is the grid study that says it can be delivered?

    Voters: Ask every candidate those same two questions, at every level, before the next ballot.

    The Choice
    The United States can win the AI race, but not by asking communities to host the buildout on faith. China’s advantage is not that it has better engineers. It is that it built the power first. If we want national AI leadership, we have to earn it the way voters just demanded: build the grid, make the costs transparent, and put people accountable to the public in charge of the decisions.

    Summary
    Voters in 40 of 41 local data center ballot measures on November 3rd chose limits, and polls show most Americans oppose a data center near them. The backlash is not a rejection of AI. It reflects a buildout that outran the grid, a shortfall that grid reliability authorities publicly warned about, and a bill impact that is real in some regions and contested in others. China built power and transmission first, at a scale and speed the US has not matched, though its model is not one we should copy. The answer is to put the grid first, make the large loads pay, and make the decisions accountable to the people who live with them.
    Sources

  • Powering the AI Revolution

    Why America’s Data Center Future Depends on the Electrical Grid

    America’s race to lead the artificial intelligence revolution will not be won by semiconductor technology and investment alone. It will also depend on where data centers are built, whether the electrical grid can deliver the power they need, and whether communities trust the decisions being made in their backyards.

    Artificial intelligence is rapidly becoming one of the most important technologies shaping America’s economic and national security future. Companies are investing billions of dollars in advanced computing infrastructure, developing increasingly powerful AI models, and building enormous data center campuses to support the next generation of artificial intelligence.

    But there is a fundamental problem that deserves much more attention: AI cannot scale without electricity, and electricity cannot be delivered without adequate infrastructure.

    The United States may have world-leading technology companies, advanced semiconductor manufacturers, and enormous investment capital. However, none of these advantages guarantees that the country can build and operate enough AI data centers to maintain its competitive position.

    The challenge is increasingly about energy availability, grid location, transmission infrastructure, and reliability.

    As we discussed in our recent article, America Won’t Lose the AI Race to China’s Chips. It Will Lose It to Distrust, America’s ability to lead in artificial intelligence will depend on more than technological innovation. It will also depend on whether the country can build the physical infrastructure required to support that innovation while maintaining public confidence.

    The AI Infrastructure Problem Starts With Electricity

    Modern AI data centers are fundamentally different from traditional commercial buildings.

    They house thousands of specialized processors, high-performance networking equipment, advanced cooling systems, and supporting infrastructure. These systems require substantial amounts of electricity, often operating continuously to train AI models and deliver computing services.

    As AI adoption expands, so does the demand for computing capacity. That creates a corresponding demand for electricity.

    The challenge is that electrical infrastructure cannot necessarily expand as quickly as technology investment.

    A technology company can announce a multibillion-dollar data center project, acquire land, order computing equipment, and begin construction. But if the local electrical grid cannot provide the required power, the project may experience significant delays before it can become operational.

    This creates a new reality for data center developers: securing electricity must happen before construction decisions are finalized, not after the buildings are completed.

    The availability of power should be one of the first considerations in determining where America’s next generation of AI infrastructure is built.

    Grid Location: Building Where Electricity Is Available

    Historically, data center developers have considered land prices, tax incentives, fiber connectivity, proximity to major markets, and access to skilled labor when selecting locations.

    Those factors remain important, but electricity availability is becoming an increasingly important consideration.

    A site may offer inexpensive land and favorable development incentives but still be unsuitable for a large AI data center if the surrounding electrical infrastructure cannot support its requirements.

    Developers must evaluate several factors before selecting a location:

    • Existing generation capacity: Is sufficient electricity available within the regional power system?
    • Substation capacity: Can nearby substations accommodate the additional electrical load?
    • Transmission access: Can electricity be delivered to the property without major infrastructure limitations?
    • Grid congestion: Are existing transmission and distribution systems already constrained?
    • Future expansion: Can the location support additional computing capacity as AI requirements increase?

    The distinction between national electricity generation and locally available electricity is particularly important.

    China has substantially more total electricity-generation capacity than the United States. However, national generation capacity does not automatically mean that electricity is available at every location where a data center might be constructed.

    The same applies to the United States.

    Having sufficient electricity somewhere within a country does not mean that the power can be delivered economically and reliably to a particular data center campus.

    This is why grid location should become a primary consideration in national AI infrastructure planning.

    Transmission Infrastructure: The Bottleneck Between Power Generation and Computing

    Electricity generation is only one part of the equation.

    Power must travel through transmission lines, substations, transformers, and distribution networks before reaching the facilities that consume it.

    These systems require significant investment, careful engineering, regulatory coordination, and time to construct.

    For AI data center developers, transmission infrastructure can become one of the biggest obstacles to completing a project.

    Even when electricity generation is available, existing infrastructure may not have sufficient capacity to deliver the required power.

    Addressing these limitations may require:

    • New high-voltage transmission lines.
    • Substation construction and expansion.
    • Additional transformers and electrical equipment.
    • Improvements to local distribution networks.
    • Coordination between utilities, developers, and regional grid operators.
    • Long-term planning for additional electrical demand.

    These are not minor construction details. They are essential components of the infrastructure supporting America’s AI economy.

    The federal government, state governments, utilities, and private developers must coordinate their efforts to ensure that electrical infrastructure keeps pace with computing demand.

    This is also where the proposed Department of Technology could play an important role.

    As outlined in our article, How Our Proposed Department of Technology Could Supercharge the Stargate Project, large-scale AI infrastructure initiatives require coordination across federal, state, and local governments.

    Energy planning should be part of that coordination.

    Reliability: AI Data Centers Cannot Operate on Uncertain Electricity

    Having access to electricity is not enough. AI data centers also need reliable electricity.

    AI training and inference workloads can involve thousands of processors operating simultaneously. An unexpected interruption in power can disrupt computing operations, interrupt services, and create expensive recovery requirements.

    For companies investing billions of dollars in computing infrastructure, electrical reliability is a business-critical issue.

    Data center developers must consider the resilience of the entire electrical system, including:

    Grid redundancy: Multiple electrical supply paths can reduce the risk associated with individual transmission failures.

    Backup generation: On-site generators can provide emergency electricity when utility service is interrupted.

    Battery energy storage: Batteries can help manage peak demand, provide short-duration backup power, and support grid stability.

    Uninterruptible power supplies: UPS systems provide immediate protection for critical computing equipment during electrical disturbances.

    Microgrids: Properly designed microgrids can integrate utility electricity, on-site generation, and energy storage to improve operational resilience.

    These technologies should be considered part of the original development strategy rather than expensive additions after a project experiences electrical limitations.

    Reliability also has implications beyond individual facilities. Large data center developments must be evaluated in the context of the surrounding electrical system and the communities that depend on it.

    A project should not simply secure enough electricity to operate its own equipment. Its development should also account for the impact of additional demand on the regional power system.

    Energy Storage and Renewable Energy Must Be Part of the Solution

    Expanding electricity generation and transmission will be essential, but the United States should also consider how energy storage and distributed generation can support future AI infrastructure.

    Solar, wind, nuclear energy, natural gas generation, and battery storage can all play different roles in meeting electricity demand.

    The challenge is developing an energy strategy that balances availability, affordability, reliability, and environmental considerations.

    Energy storage deserves particular attention.

    Our proposal for California’s NEM 4.0 emphasizes the importance of battery storage in supporting grid stability, improving energy management, and helping balance electricity supply and demand.

    The same broader principle applies to AI infrastructure.

    Battery systems can help facilities manage periods of peak demand, provide short-duration backup power, and potentially participate in programs that support the electrical grid.

    However, batteries are not a substitute for adequate generation and transmission capacity. They are one component of a larger energy infrastructure strategy.

    The United States should pursue a diversified approach that combines reliable generation, modern transmission networks, energy storage, and responsible demand management.

    The Cost of AI Infrastructure Should Not Be Shifted to Local Communities

    There is another issue that cannot be ignored: who pays for the electricity infrastructure required to support the AI industry?

    New data centers can require substantial investments in electrical infrastructure. Utilities may need to upgrade substations, expand transmission networks, and improve local distribution systems.

    Those investments raise important questions about how costs are allocated.

    If infrastructure upgrades primarily benefit a private data center developer, should ordinary electricity customers bear the financial burden?

    How should utilities account for the additional demand when determining electricity rates?

    What protections should exist to prevent households and small businesses from subsidizing infrastructure built primarily to support large technology companies?

    These questions deserve transparent answers before major development agreements are finalized.

    The public should not be expected to accept higher electricity bills without understanding why those costs are increasing and who is responsible for paying for the infrastructure.

    This is one reason public opposition to data center development should not be dismissed as simple resistance to technological progress.

    As we discussed in our recent analysis of America’s AI infrastructure expansion, communities have legitimate concerns about electricity costs, water consumption, land use, and the lack of transparency surrounding some development agreements.

    A successful AI infrastructure strategy must address these concerns rather than treating them as obstacles to be overcome.

    Public Trust Is an Infrastructure Requirement

    Data center construction is often presented as a private-sector investment that brings jobs, economic development, and technological innovation to a community.

    But these projects can also place significant demands on local infrastructure and resources.

    Residents deserve to know how much electricity a proposed facility expects to consume, whether new transmission infrastructure will be required, how water resources will be affected, and whether existing utility customers could experience additional costs.

    They should also have meaningful opportunities to participate in decisions before projects are approved.

    This is particularly important when government officials negotiate development agreements behind nondisclosure agreements or make decisions without sufficient public input.

    Transparency is not simply a political preference. It is an important part of successful infrastructure development.

    Projects that fail to earn public confidence can encounter opposition, legal challenges, delayed approvals, and even rejection at the ballot box.

    The United States cannot afford to treat public trust as an afterthought while attempting to build the infrastructure necessary for its AI future.

    A Department of Technology Could Help Coordinate America’s AI Infrastructure

    The growing relationship between artificial intelligence, electricity, and national competitiveness highlights the need for more coordinated technology governance.

    Today, decisions involving AI infrastructure can span multiple federal agencies, state regulators, county governments, municipalities, utilities, and private companies.

    Each organization has different responsibilities, priorities, and regulatory requirements.

    The result can be fragmented planning, inconsistent policies, delayed infrastructure investments, and insufficient public accountability.

    The proposed Department of Technology offers a framework for addressing some of these coordination challenges.

    A federal Secretary of Technology could help coordinate national AI infrastructure priorities, cybersecurity standards, and long-term technology planning.

    State technology leaders could work with utilities, regulators, and regional stakeholders to identify electrical infrastructure needs and coordinate development priorities.

    County and municipal technology leaders could help ensure that local communities have access to relevant information and meaningful opportunities to participate in decisions affecting their infrastructure.

    This approach would not replace the responsibilities of existing utility regulators or energy agencies. Instead, it would provide an additional framework for coordinating technology development with the infrastructure required to support it.

    The goal should be straightforward: make it easier to build the infrastructure America needs while ensuring that development remains transparent, accountable, and beneficial to the public.

    Conclusion: America’s AI Future Depends on Its Ability to Deliver Power

    The competition to lead artificial intelligence is increasingly becoming a competition to build the physical infrastructure that makes advanced computing possible.

    Semiconductors, software, research, and investment remain essential. But they cannot compensate for insufficient electrical generation, inadequate transmission infrastructure, or unreliable power delivery.

    Grid location determines whether a data center can access the electricity it needs.

    Transmission infrastructure determines whether that electricity can reach the facility.

    Reliability determines whether the facility can operate consistently.

    And public trust determines whether communities and governments will support the continued expansion of this infrastructure.

    China’s substantial electricity-generation capacity demonstrates the importance of energy resources in the global AI competition. The United States must recognize that technological leadership requires a comparable commitment to modernizing and expanding its electrical infrastructure.

    But building more data centers should not be the only objective.

    America must build them intelligently, locate them responsibly, strengthen the electrical grid, protect consumers from unfair infrastructure costs, and ensure that communities have a voice in decisions that affect their future.

    The AI revolution will run on electricity. The countries that can generate it, transmit it, deliver it reliably, and earn public support for the infrastructure behind it will have a significant advantage in the decades ahead.

    America’s next great technology investment may not be another semiconductor fabrication facility or AI research laboratory. It may be the electrical infrastructure that makes all of them possible.

    Department of Technology believes that technological leadership and public accountability must advance together. Building the infrastructure of tomorrow requires more than private investment. It requires coordinated planning, reliable energy, and public trust.

  • Introducing Our New Data Centers Series

    Building the Infrastructure for the MI Era

    The future of Machine Intelligence will not be built on algorithms and chips alone. It will depend on where we build, how we power, and whether we build the infrastructure responsibly.

    We are entering a new technological age. Artificial intelligence is moving rapidly from research labs and specialized applications into nearly every part of our economy and daily lives. We often call this the AI era. But at Department of Technology, we believe a broader term better describes what is unfolding: the MI era—the Machine Intelligence era.

    This is more than a change in terminology. It reflects the scale of the transformation underway. Intelligent machines, increasingly capable computing systems, and the networks connecting them are becoming foundational infrastructure for business, science, education, national security, healthcare, and public services.

    But there is a critical reality that can get lost in the excitement over new models, faster chips, and extraordinary software: Machine Intelligence requires physical infrastructure.

    It requires data centers—large, sophisticated facilities filled with computing equipment that must operate continuously. Those facilities require enormous amounts of electricity, dependable transmission lines, cooling systems, water or other cooling resources, communications networks, skilled workers, secure sites, and communities willing to host them.

    Building this infrastructure will be one of the defining challenges of the MI era.

    That is why we are launching a dedicated Data Centers series.

    The questions America and the world must answer

    We intend to examine the data center challenge through the essential questions that should guide every major infrastructure decision:

    • Who is building these facilities, financing them, operating them, regulating them, and benefiting from them? Who is responsible when costs or consequences fall on local communities?
    • What should a modern data center include—not only in computing capacity, but in power systems, cooling, cybersecurity, physical security, resilience, and environmental performance?
    • When should facilities be built, and how should construction be coordinated with the availability of electricity, transmission capacity, equipment, workers, and other essential resources?
    • Where should data centers be located? Proximity to users and fiber networks matters, but so do available power, transmission access, land, water, climate, permitting, and the needs of surrounding communities.
    • Why are these facilities necessary, what public and economic value can they create, and how do we distinguish essential infrastructure investment from projects that shift costs and risks onto others?
    • How can we build, connect, operate, and expand data centers reliably, affordably, securely, and responsibly?

    These questions are connected. A site that looks attractive because land is available may not have sufficient electrical capacity. A region with abundant generation may lack the transmission infrastructure to deliver power where it is needed. A project that makes sense on paper may encounter public opposition if residents are not given clear information about electricity costs, water use, construction impacts, and long-term commitments.

    There is no single answer that works everywhere. The right approach will depend on local conditions, technology, energy resources, infrastructure, and community priorities.

    The electric grid is part of the AI story

    One of the central themes of this series will be the relationship between data centers and the electrical grid.

    The MI era cannot advance at the speed of software if the infrastructure supplying its electricity cannot keep pace. Generation capacity, transmission lines, substations, interconnection processes, energy storage, backup systems, and grid reliability all matter.

    We must look beyond whether enough electricity exists in the abstract. We must ask whether it can be delivered to the right location, at the right time, at a predictable cost, and with the reliability that modern computing demands.

    We must also examine how new demand affects existing customers. Data center development should not automatically mean higher electricity bills, strained local infrastructure, or reduced reliability for households and established businesses. The costs and benefits must be understood, disclosed, and fairly allocated.

    Building with communities, not simply in them

    Technology leadership cannot be measured only by the number of facilities announced, the amount of capital invested, or the computing capacity installed.

    It must also be measured by whether projects are responsibly planned, whether communities have meaningful information and a voice in decisions, whether infrastructure costs are fairly assigned, and whether the benefits extend beyond the companies operating the facilities.

    Public trust is not an afterthought. It is part of the infrastructure.

    That means transparency about power requirements, water and cooling strategies, environmental impacts, tax arrangements, utility investments, and emergency planning. It means recognizing that communities are not merely locations on a map. They are partners whose long-term interests deserve consideration.

    A practical guide to the infrastructure of the MI era

    In this series, we will explore the data center ecosystem from the ground up: site selection, electricity generation, transmission and distribution, grid connections, cooling, water, energy storage, renewable and other energy sources, reliability, cybersecurity, construction, financing, regulation, workforce development, and the relationship between technology companies and the communities that host their facilities.

    We will look at the roles of federal, state, and local governments, utilities, developers, technology companies, investors, engineers, regulators, and residents. We will examine emerging solutions as well as the obstacles that could slow development or undermine public confidence.

    Our goal is not to advocate for building every proposed data center, nor to oppose development simply because it is large or energy-intensive. Our goal is to help readers understand what responsible development requires—and what questions should be asked before decisions are made.

    America’s ability to lead in Machine Intelligence will depend on more than inventing the next generation of chips and software. It will depend on whether we can build the power systems, networks, facilities, policies, and public partnerships that allow those technologies to operate at scale.

    The MI era needs more than computing power. It needs a plan for the infrastructure that makes computing power possible.

    That is what this series is about.

    Our Data Centers Series

    Where we build, how we power, and whether we build the infrastructure for Machine Intelligence responsibly.

    Introducing Our New Data Centers Series

    Machine Intelligence needs physical infrastructure. Our new Data Centers series asks who builds, finances, powers, and regulates data centers, where they belong, and how local communities can truly benefit. Learn what responsible development requires, from electricity, cooling, and site selection to water use, sustainability, security, workforce needs, and public trust in the Machine Intelligence era.

    Read the introduction →

    Ten Questions Every Candidate and Elected Official Should Answer About Data Centers and AI

    Voters deserve clear, specific answers about who pays for data centers, who is accountable, and what happens if AI does not deliver. This nonpartisan scorecard gives voters, reporters, and candidates ten questions for any office, with evidence, signs of strong and evasive answers, a simple scoring guide, and a print-ready list to send to everyone.

    Read article →

    What Every County Treasurer Should Model Before AI Layoffs

    AI could shrink the labor income that funds local government, while data center revenue can concentrate budgets in one fragile industry. This guide shows county treasurers how to build a simple, practical scenario worksheet: map revenue, run displacement and bust scenarios, set reserves and triggers, and publish the results before layoffs or valuation drops arrive.

    Read article →

    A Technology Leader’s First Year: A Guide for Counties and Cities

    Counties and cities are approving data centers faster than they can possibly study them. This first-year guide for local technology leaders covers what to do each quarter: ending by-right approvals, writing water, noise, and decommissioning rules, joining utility rate proceedings, negotiating community benefit agreements, ending secret deals, and preparing the budget for what comes next.

    Read article →

    Preemption Without an Owner: Why the Federal-State AI Fight Needs Someone Accountable

    Washington is fighting over whether federal law should override state AI rules. Both sides have real arguments. But whoever wins, someone must enforce the result and answer to voters. This analysis lays out the strongest case for and against preemption, what the record actually shows, and three simple tests any national AI standard should meet.

    Read article →

    A Briefing for the New Congress: What the Last Two Years Got Wrong, and What to Do Differently

    The last Congress agreed on almost nothing about AI policy, then watched its one bipartisan data center bill pass the House 417-3 and stall in the Senate. This briefing for new members explains five costly mistakes to avoid, from back-door preemption to election-year gamesmanship, and what a single accountable technology office could realistically do differently.

    Read article →

    How a Data Center Deal Actually Gets Made, and Where the Public Gets Left Out

    From shell companies and code names to secret NDAs and quiet utility requests, most data center deals are usually shaped long before residents hear a word. This step-by-step guide shows how a typical deal gets made, who decides at each stage, where the public is shut out, and what an accountable technology leader would change.

    Read article →

    The First 100 Days: A Technology Agenda for the Next Congress

    A new Congress has 100 days to govern the technology buildout instead of reacting to it. This practical agenda for members of both parties covers ratepayer protection, ending secret data center deals, preparing for AI’s impact on federal revenue, and creating a single point of accountability, with the evidence and key objections behind each step.

    Read article →

    The Data Center Casino: Why America Needs Elected Technology Leaders Now

    America needs more data centers, but both parties have treated them like a casino: chasing tax breaks and investment while ignoring grid strain, bubble risk, China’s power advantage, and AI job losses that threaten the tax base. See why elected technology leaders at every level are needed right now to build smarter, safer, and fairer.

    Read article →

    Voters Didn’t Reject AI. They Rejected Building It Backward

    Voters didn’t reject AI. They rejected building it backward. After 40 of 41 data center ballot measures favored limits on November 3rd, see how rising power bills, strained grids, and secret deals fueled the backlash, how China’s energy buildout compares, and what transparent, accountable planning could do to restore public trust in America’s AI future.

    Read article →

    Powering the AI Revolution

    Discover why America’s AI data center future depends on a reliable electrical grid. Explore how data center location, electricity generation, transmission infrastructure, grid capacity, energy storage, backup power, and renewable energy shape the next generation of AI infrastructure. Learn why reliable electricity, responsible planning, and community impact are essential to powering the Machine Intelligence era.

    Read article →

    An Open Letter to Every Elected Official in America: Put the Country First

    An open letter urging Democrats, Republicans, and every elected official to start a nationwide public discussion on creating a Department of Technology at the local, county, state, and federal levels. Learn why AI data centers, secret NDAs, and conflicts of interest demand accountable, nonpartisan technology leadership before the next major data center contracts are signed.

    Read article →

    America Won’t Lose the AI Race to China’s Chips. It Will Lose It to Distrust.

    America won’t lose the AI race to China’s chips. It will lose it to distrust. After nearly 40 data center ballot measures in 2026, voters are demanding transparency, accountability, and reliable, affordable power. Learn how conflicts of interest, grid limits, and secret deals threaten U.S. AI leadership, and what a Department of Technology could change.

    Read article →
    View all Data Centers articles →

  • America Won’t Lose the AI Race to China’s Chips. It Will Lose It to Distrust.

    Voters have already rejected data centers at the ballot box in 2026. The officials building this infrastructure are still hiding deals and holding stakes in the companies they oversee.

    Last week, on November 3rd, 2026, voters delivered a verdict on the AI buildout. They weighed nearly 40 data center-related ballot measures across 10 states, and in Edgewater, Florida, over three-quarters of voters approved a ban on new data centers. This week, ProPublica reported that our United States Secretary of the Interior, Doug Burgum is leading the push to put data centers on public lands while holding between $18.1 million and $64.4 million in AI-company investments, according to an analysis by the advocacy group Save Our Parks.

    Our United States Secretary of the Interior, Doug Burgum says any suggestion that his decisions are driven by personal investments is false. But lawyers told ProPublica his holdings appear to be acceptable under current law. That is the real scandal. If an official can steer national AI policy while holding a stake in the industry he is steering, the problem is the law, not just the official.

    The argument of this piece is simple: America will not lose the AI race to foreign chips. America will lose it to its own distrust, unless the people building our AI infrastructure are accountable to the people who live next to it.

    A Lead That Depends on Power and Public Trust

    The United States still dominates the hardware and capital side of AI. Global data center capacity is projected to grow by 97 GW between 2026 and 2030, effectively doubling, and that buildout could require $3 trillion in investment. China, meanwhile, is closing the capabilities gap and training a large share of the world’s AI researchers.

    Chips are no longer the binding constraint. Energy and public consent are. BloombergNEF estimates a 63 GW gap between what AI chip shipments imply the U.S. should build by 2033 and what energy constraints will allow. A country that cannot power its data centers, or persuade its citizens to host them, will not win this race.

    Washington DC Has Not Built the Grid

    AI is, at bottom, a question of electricity. A serious national defense strategy would fast-track nuclear baseload, advanced renewables, and modernized transmission lines. Instead, Congress has taken only minimal steps to regulate these facilities, and the cost is landing on aging local utility systems and the people who pay for them.

    The Backlash Is Rational

    Gallup found that 71% of Americans oppose an AI data center in their local area, including 48% who are strongly opposed. A later Heatmap Pro poll put strong support at just 4%, down from 13% a year earlier. This is not simple NIMBYism. A majority of Americans, 59%, expect their electricity bills to rise if a data center opens nearby, and these facilities also draw heavily on local water and grid capacity.

    When officials answer this with dismissal, they deepen the problem. Burgum has blamed the mass rejection of data centers on a foreign psyop. Calling public worry about bills and water a foreign plot cannot win the cooperation of the communities that must host the infrastructure of national power.

    Secrecy and Self-Dealing

    Consider Tucson, a drought-stricken city. Amazon Web Services and a developer spent nearly two years negotiating a $3.6 billion, 290-acre data center campus behind NDAs, keeping the public in the dark about projected water usage. Residents learned the details only after the deal was largely set.

    Tucson is not unusual. In Virginia, 25 of 31 local governments involved in data center development had signed NDAs. In Louisiana, records requests found at least 54 NDAs signed by elected officials. This is the pattern we warned about in our analysis of America’s AI Action Plan: data centers sited without local democratic input.

    The conflicts reach the top, too. ProPublica’s database shows more than 320 members of our Trump administration have ties to Alphabet, with large numbers also tied to Nvidia and Meta. Our Secretary of the Interior, Doug Burgum pledged to divest many of his AI investments, but not his stake in Titan Development, a real estate firm that has handled land acquisitions for data centers in the Southwest. And the administration has not completed any environmental reviews for data centers on public lands.

    This is not a one-party problem, and the rules must not be one-party rules. The standards proposed here should apply to every official in every administration, including all members of Congress, the White House and others.

    Reform is starting. Lawmakers in at least ten states have introduced bills to restrict or ban these NDAs, Pennsylvania’s governor has barred them for agencies under his control, and a bipartisan federal bill, the No Secrets for Data Centers Act, would void NDAs signed by officials acting in their official capacity. Industry has a defense: one developer says NDAs protect officials from insider-trading exposure, since a planned campus can count as material non-public information. That argues for narrow disclosure rules and a limited carve-out for genuine trade secrets. It does not justify blanket secrecy over water, power, and tax breaks.

    “Won’t Rules Slow Us Down?”

    The strongest objection is that accountability will slow the buildout and hand China an advantage. The opposite is closer to the truth. Projects imposed on unwilling communities are the ones being voted down, and secrecy is what turns a zoning question into a ballot-box revolt. Transparency is how projects survive.

    The Fix: Accountability at Every Level

    Our AI strategy is currently shaped by appointed officials, zoning boards, and “AI czars” who answer to executives rather than voters. The framework at Department of Technology proposes a permanent Department of Technology with layered accountability:

    • Federal: a cabinet-level Secretary of Technology, nominated by the President and confirmed by the Senate. That gives one accountable official real authority over national AI policy, cybersecurity, and digital infrastructure, with a public confirmation process where conflicts must be disclosed and divestment pledges can be tested.
    • State, county, and local: elected technology leaders. They face voters, and they would decide questions like whether a data center goes up beside a reservoir.

    An elected state secretary of technology or county technology commissioner cannot easily hide behind a corporate NDA when residents are looking at their utility bills and voting on the next term. A federal Secretary who is vetted, confirmed, and held to strict divestment rules would have far less room to wave away grid concerns.

    As we argued in Stop the Litigation Loop, a national AI strategy works best when elected state and local partners help carry it out. Elected officials can be conflicted too, so the office needs disclosure and divestment rules from day one.

    Four Legal Changes, Now

    1. Treat AI infrastructure as defense infrastructure. Foundational models, sovereign compute clusters, and their dedicated power supply should be formally classified as critical to national defense.
    2. Create an AI conflict-of-interest offense, modeled on war profiteering law. Criminalize officials trading stock, holding options, or receiving deferred compensation from AI companies bidding on or serving federal defense and intelligence contracts, including public land leases, where the official has authority over those decisions. Provide a safe harbor for diversified index funds and qualified blind trusts, so the rule targets real conflicts rather than ordinary retirement savings.
    3. Ban NDAs on public-interest infrastructure. Void any NDA that binds public officials or covers projects using public funds, land, or incentives, with a narrow carve-out for real trade secrets.
    4. Require divestment as a condition of office. Anyone overseeing technology, public lands, or energy policy, including the proposed Secretary of Technology, should divest direct holdings in AI and tech companies before taking office. Violations should be prosecuted as threats to national defense, not settled with ethics fines.

    What Each of Us Can Do

    • Congress: pass the No Secrets for Data Centers Act and adopt divestment rules that cover both parties.
    • Statehouses: enact NDA bans like those introduced in at least ten states, and require public notice before any rezoning or tax abatement for a data center.
    • Voters: ask every candidate, at every level, two questions. Will you divest from AI and tech holdings while in office? Will you refuse NDAs on projects that use public land, money, or water?

    The Choice

    Our adversaries will not be stopped by a nation whose officials hold stock in the contractors they oversee, hide deals from the towns that host them, and brush off voters’ concerns about their own bills. To win the algorithmic cold war, we have to build the grid and earn public trust. We also have to make sure those who write AI policy have no stake in the outcome. The voters have already spoken. Now the law has to catch up.

    Summary

    Seven in ten Americans oppose a data center in their community, and the officials deciding where they go may own a piece of the companies building them.

    Gallup found that 71% of Americans oppose an AI data center in their local area, and last week voters made that opposition official, weighing nearly 40 data center-related ballot measures across 10 states. This week, ProPublica reported that the Interior Secretary, who is pushing to build data centers on public lands, holds between $18.1 million and $64.4 million in AI-company investments, according to an analysis by the advocacy group Save Our Parks.

    Interior disputes the characterization, and lawyers told ProPublica the holdings appear to be acceptable under current law. That last part is the heart of our new article; America Won’t Lose the AI Race to China’s Chips. It Will Lose It to Distrust, which argues that our race will be decided by the power grid and public trust as much as by chips, and that trust collapses when local officials sign NDAs that hide water and power costs from their own residents and when officials hold stakes in the industries they oversee.

    This is not a Republican or Democratic problem, and the rules we propose would cover every official in every administration. Imagine being able to see what your town agreed to before the contract is signed, knowing the people overseeing AI and energy policy have no financial stake in the outcome, and having a vote over whether a data center goes up beside your reservoir.

    That future is within reach: lawmakers in at least ten states have introduced bills to restrict or ban these NDAs, and a bipartisan federal bill, the No Secrets for Data Centers Act, would void NDAs signed by officials acting in their official capacity. What the momentum needs is public pressure, and public pressure starts with people knowing what is happening.

    So please share this article today: post it with one sentence on why it matters to your town, send it to your city council member or state legislator, forward it to a neighbor who pays an electric bill, and ask every candidate whether they will divest from AI and tech holdings while in office and refuse NDAs on projects that use public land, money, or water. The voters have already spoken, so help make sure the law catches up at department.technology/.

  • An Open Letter to Every Elected Official in America: Put the Country First

    To the Members of Congress, Governors, State Legislators, County Supervisors, Mayors, and City and Town Council Members of both parties:

    We write as an independent, nonpartisan policy project based in San Diego, California. We belong to no party and take no money from advertisers, PACs, or lobbyists. We are asking for one thing: start the public discussion on establishing a Department of Technology at the local, county, state, and federal levels, as proposed at department.technology/.

    What we did first, and what happened

    We did not begin with an open letter. We began by going directly to the executive branch. In February 2025, the Office of Science and Technology Policy asked the public for input on a national AI Action Plan, and we submitted our proposal by email. When the White House released America’s AI Action Plan in July 2025, we published a response. We welcomed its focus on innovation, infrastructure, and diplomacy, and pointed out the gap in it: no one in the plan is directly accountable to the American people for how it gets carried out.

    Those efforts have not led to any engagement with our proposal. We hold no grudge about that. A good idea should not depend on one administration. It has to be able to survive from one administration to the next, which is why we are now writing to all of you.

    Why this can’t wait

    Technology is no longer a niche subject. It is a daily political reality in your districts.

    We are not raising these points to attack any person or party. The pattern would be a problem under any administration, and the rules we propose cover every official, including every member of Congress.

    AI, data centers, surveillance, student data, autonomous systems, and the power grid are being governed by a patchwork of agencies and appointed officials, with no one in charge and often no one voters can hold accountable.

    What we propose

    A Department of Technology with layered accountability:

    • Federal: a cabinet-level Secretary of Technology, nominated by the President and confirmed by the Senate. That gives the country one accountable official for national AI policy, cybersecurity, and digital infrastructure.
    • State, county, and local: elected technology leaders, so the people deciding how technology is used in your community answer directly to the people who live there, much as an attorney general or district attorney does.

    The full proposals are public and free to use: Legal Guidelines for a Federal Department of Technology, Legal Framework for State, County, and Local Departments, a draft California ballot initiative, and our argument for why this is essential to American AI leadership. They are drafts, offered for debate, and we welcome amendments.

    What we are asking you to do

    You don’t have to agree with every detail. We are asking you to begin the conversation:

    1. Hold a public hearing or town hall on how technology is governed in your jurisdiction and who is accountable for it.
    2. Introduce a resolution calling for study of a Department of Technology, as proposed by us at www.department.technology, at your level of government.
    3. Adopt two interim safeguards now, whichever way the larger debate goes: a ban on NDAs for projects using public funds, land, or incentives, and disclosure and divestment rules for officials who oversee technology, energy, and public lands. A bipartisan federal bill, the No Secrets for Data Centers Act, would void NDAs signed by officials acting in their official capacity, and lawmakers in at least ten states have introduced bills to restrict or ban these NDAs.
    4. Tell us where you stand. If you will support the discussion, or if you have objections, to a Department of Technology, as proposed by us at www.department.technology, we want to hear them and will publish your response with your permission.

    A request across party lines

    This is not a Republican idea or a Democratic idea. Voters in both parties oppose secrecy around local projects, want their utility bills protected, and want to know that the people steering technology policy answer to them. Whoever holds power next year or in ten years will inherit the same questions.

    We are asking you to put the country’s interests ahead of party interests and start this discussion in the open, where voters can take part.

    Respectfully,

    The Founders, Department of Technology
    San Diego, California
    department.technology/

  • Where It All Fits:

    Why the Search for Minds Needs a Department of Technology

    By Department of Technology

    Third in a series. If you’re new here, start with The Quietest Minds in the Galaxy May Be the Oldest, then Can You Whisper Across the Galaxy?. The math lives in the technical appendix.

    Tying the string

    In the first article of this series I made a claim I’ve been circling ever since: we can’t count minds we haven’t learned to define, and the first place we’ll have to define them is here on Earth.

    Then I spent two articles on equations, lasers and entangled photons. Fair enough, those were the fun parts. This one is where I tie the string.

    Here’s the short version. The Department of Technology proposal at department.technology/ is usually pitched as a way to govern artificial intelligence, protect student data and secure our infrastructure, and it is that. But the same machinery it would need for those jobs turns out to be the machinery the search for other minds needs too: shared definitions, shared measurements, rules that help civilizations last, and an accountable body to make hard calls. I want to show you where each piece connects.

    A note on sourcing before I start. Everything I attribute to the proposal comes from what’s published on our site. Where I’m extending it, adding my own idea, I’ll say so. That’s the same habit I tried to keep in the appendix: you should always be able to tell what’s established, what’s proposed and what’s mine.

    What the proposal actually says

    Our DoT site describes itself as an independent policy research and advocacy project based in San Diego. Its central argument is simple to state: every official who makes technology policy should be accountable to voters. As the site puts it, many of the most consequential decisions of this era, about artificial intelligence, surveillance, autonomous weapons, student data, quantum computing and digital infrastructure, are being made by people nobody elected.

    The fix we propose is a Department of Technology at four levels of government:

    • Federal: a cabinet-level Secretary of Technology, nominated by the President and confirmed by the Senate, with real authority over national AI policy, cybersecurity standards and digital infrastructure.
    • State: a directly elected Technology Secretary serving four-year terms.
    • County and local: elected technology supervisors and directors, accountable to the communities they serve.

    Our site’s work covers far more than this series touches: draft legislation on AI disclosure and whistleblower protection, a privacy proposal called the Data Sovereignty Act, school and college identity systems under the names School Contact and College Contact, plans for continuity of government during blackouts and electromagnetic pulses, draft conventions on autonomous weapons, and more.

    One thing I want to say plainly: this is a proposal. No such department exists, and nobody in government has adopted the plan. Everything below is an argument for what could be, not a description of what is.

    Four terms, four jobs

    If you read the technical appendix, you’ll remember that the Sentience equation boils down to a handful of terms that matter most. Here is how I’d map them onto what a Department of Technology could do.

    Equation termThe question it asksWhat a Department could doWhat the proposal already includes
    ψ (psi)Is this mind actually sentient?Define tests and thresholdsThe RMS classification framework; sentience safeguards and stop criteria for biocomputing
    dCould we detect it?Set measurement standardsThe SCOPE index and the Void Hypothesis
    τ (tau)How long does it last?Help civilizations survive their own technologyThe No First Use Quantum Pledge, the No Weaponization of Biocomputing Pledge, treaty drafts, continuity-of-government planning
    p (emergence)How often do minds change form?Track the transitions as they happenRegistration and public reporting for high-complexity experiments

    Let me take them one at a time.

    Defining minds (ψ)

    The hardest term in my equation is ψ, the probability that an intelligent system is genuinely sentient. I said in the appendix that we can’t measure it remotely, and I stand by that. But “we can’t measure it from 500 light-years away” doesn’t mean “we can’t make progress on it here.”

    This is where the proposal’s work on biocomputing gets interesting. The site’s article on the subject starts from an honest admission: researchers don’t have a settled definition of consciousness, so no clean test exists for telling whether a lab-grown neural system has morally relevant experience. Its answer isn’t to pretend otherwise. It asks what society should do if researchers cannot confidently answer the question at all, and proposes a precautionary approach. That includes classifying experiments by their potential cognitive complexity, increasing oversight as complexity grows, monitoring for predefined warning signs, requiring an independently reviewed shutdown plan before a high-complexity experiment starts, and sending any evidence of potentially significant cognition to independent review, not to the research team alone.

    The article also draws careful distinctions that I wish more public conversations drew: between biological activity, intelligence, learning, consciousness, sentience and moral status. They’re related, but none of them automatically implies the next.

    Now here’s my addition. Suppose a Department did the hard moral work of deciding, in advance, what evidence would make us pause an experiment on a neural organoid. That list of indicators, written down and argued over, would become the closest thing we have to a working yardstick for ψ. It wouldn’t answer the astronomical question. But it would mean that if we ever face a strange signal from a strange mind, we won’t be inventing our standards in the middle of the surprise.

    Measuring the quiet (d)

    The second job is measurement. In the first post I described the SCOPE index, which scores a civilization on a scale of 1 to 100 across five dimensions: Synthetic integration, Complexity, Operational utility, Processing capacity and Efficiency. Its motivating idea, the Miniaturization Paradox, says that advanced civilizations are more likely to move toward dense, efficient computation than toward star-sized machines, and so might look cold and dark to our current telescopes.

    Two honest notes. First, SCOPE is a hypothesis, not an established finding. It hasn’t been tested against observations, and I’d treat it that way. Second, it has relatives in the literature. The cosmologist John Barrow proposed back in 1998 a scale ranking civilizations by how small a scale they can manipulate, which is a cousin of the same intuition. I think SCOPE is stronger for acknowledging that lineage, and I’d encourage anyone working on it to do so.

    Why does this belong in a Department of Technology? Because measurement standards are a classic public job. Think of how much depends on agreed units, definitions and test procedures that nobody gets rich from. A federal technology department working with agencies like NIST and the National Science Foundation could convene the astronomers, particle physicists and computer scientists who’d have to agree on what an efficient civilization’s footprint might look like, so that observations from different instruments can be compared. The site’s own technosignature analysis of the Amaterasu particle is one early example of the kind of cross-disciplinary work I mean.

    Lasting (τ)

    The third job is the one I find most moving. In my equation, how long a civilization lasts, τ, is among the biggest drivers of how many minds exist. And persistence isn’t a constant of nature. It depends on choices.

    This is where the proposal’s pledges and treaties come in. It calls for a No First Use Quantum Pledge, modeled on nuclear arms control precedent, and a draft International Treaty on Quantum Intelligence. It calls for a ban on using AI in nuclear, biological and chemical warfare. And it proposes a No Weaponization of Biocomputing Pledge, an international commitment that living computational systems will never be developed or deployed for warfare. The site’s reasoning on that last one is worth noting: unlike code, a biological system built for one purpose can’t be repurposed by editing a file, so the labs, cell lines and expertise built to weaponize it would become a standing capability that’s hard to dismantle.

    Add the site’s continuity-of-government work, covering electromagnetic pulse resilience, blackout protocols and infrastructure protection, and you get a picture of a civilization trying to lengthen its own τ.

    I should be careful here. The claim that pledges and treaties actually extend a civilization’s lifespan is an argument, not a result. Arms control agreements have a mixed record, and nobody can run a controlled experiment on civilizational survival. But if you believe persistence matters, and the math in my equation says it matters enormously, then these are the kinds of levers worth pulling.

    Changing form (p)

    The fourth job is watching the transitions. My equation asks how often a biological civilization hands its thinking over to engineered tissue or machines. That’s a question about the future of any species, and we happen to be running the only observable example right now, on this planet.

    It’s a sample of one, which is a weakness, but it’s one we can watch in real time, which is more than Drake ever had. The proposal’s call for mandatory registration and public reporting of high-complexity biocomputing experiments, in the spirit of clinical trial registries, would turn what’s now scattered and private into data. So would disclosure rules for AI systems. A registry of how long organoid systems last, how they scale and how they fail would give us first measurements for the biocomputing branch of the equation.

    Talking (my extension)

    The previous post raised a question the proposal doesn’t, as far as I can tell from the site, address: what happens if we ever receive a signal?

    The site does propose open standards for quantum technologies, and the logic of standards bodies translates well to communication. Beyond that, I’d like to suggest something new: a standing, nonbinding contact protocol, drafted in public, before we need it. Who is allowed to confirm a detection? Who decides whether to reply, and what to say? How do we handle the possibility that a signal looks like noise?

    There is already a body of thinking on this inside the SETI community, and I’m not proposing to replace it. I’m saying that a department already responsible for communication standards, verification and international agreements would be a sensible place to host the conversation, and that “we’ll figure it out when it happens” is a thin plan for the biggest conversation in human history.

    Why elected, for something this far off?

    A fair question: the search for aliens is not a municipal matter. Why should voters in a county or a city have anything to do with it?

    Here’s my case. The decisions at the center of this series aren’t purely technical. An engineer can tell you what a score on a neural-complexity test means. Nobody can tell you from a measurement alone what we owe an organoid that passes it, or whether to answer a message from the stars. Those are value judgments. In a democracy, value judgments with consequences for everyone should be made by people who answer to the public.

    I also think the pipeline matters. The labs, the university programs, the compute and the people all sit at the state, county and local levels long before anything reaches a cabinet table. A structure that listens at every level has a better chance of noticing a problem early.

    What critics will say

    I’d be undermining the whole point of this series if I skipped the objections, so here are the ones I find strongest.

    “We already have institutions for this.” NASA, the National Science Foundation and private efforts like the SETI Institute and Breakthrough Listen already work on astrobiology and technosignatures. Critics could reasonably argue that better coordination among existing agencies would do the job without creating a new cabinet department.

    “Elected technical officials may politicize expertise.” Some people prefer appointed expert bodies insulated from election cycles precisely so that hard technical calls don’t turn on popularity. That’s a real tension, and supporters of the proposal need a clear answer to it.

    “New cabinet departments are costly and slow.” Creating one is a major undertaking, and reformers sometimes argue it’s better to strengthen what exists.

    “It’s too early.” Today’s organoids are small and narrow, and no credible researcher claims they’re sentient. Regulating ahead of the evidence could slow legitimate research. The proposal’s reply is precaution with proportionality, with oversight that scales as complexity grows, but reasonable people weigh that trade-off differently.

    “The foundations are untested.” The Sentience equation and the SCOPE index are hypotheses. Policy built on them inherits their weaknesses.

    Where do I land? My honest view is that the fragmentation argument is the proposal’s strongest card: many agencies, each owning a slice, none owning the whole. But the critics are right on one point that I won’t wave away. The case for elected technology leadership needs more evidence than I’ve given it here, and it deserves a fair fight.

    What a first year could look like

    To keep this practical, here’s a sketch. This is mine, not something from the site.

    1. A national biocomputing registry and review standard, so that high-complexity experiments are logged and independently reviewed.
    2. A public working group on classification tests for sentience, bringing neuroscientists, philosophers, ethicists and engineers to the same table.
    3. A nonbinding contact-protocol workshop with the SETI community, to draft answers to the “who decides” questions before we need them.

    None of that requires believing anyone is out there. It only requires believing that the questions are worth asking early.

    Why I keep coming back to this

    When I started this series, I thought I was writing about aliens. I now think I was writing about us: about how a species decides what counts as a mind, how it measures what it can’t see, and how it keeps itself around long enough to find out.

    The search for other minds and the governance of the minds we’re building are the same project seen from two ends. If the proposal at department.technology/ gets one thing right, it’s the instinct to treat them together, and to insist that someone accountable is in the room when we do.

    I’d love to hear where you think it breaks.

    Questions to think about

    1. Should decisions about what counts as sentient be made by experts, by elected officials, or by some mix of the two?
    2. Is a pre-written contact protocol wise preparation or premature bureaucracy?
    3. Which of the four jobs (defining, measuring, lasting, changing form) do you think a government is best suited to do, and which should be left to scientists and the public?

    Summary

    What if the oldest minds in the galaxy are also the quietest? Across this series, we’ve explored a new way to count sentient life, from biological minds to biocomputing and Machine Intelligence, and asked what a Department of Technology could do about it. Every like and share puts these questions in front of someone who might see the sky differently. These ideas get stronger when more people argue with them. If this made you think, tap like, share it with a curious friend, and tell us where you think it breaks.

    Read the whole series: The Quietest Minds in the Galaxy May Be the Oldest · Can You Whisper Across the Galaxy? · Technical Appendix: The Sentience Equation, Step by Step

  • Technical Appendix: The Sentience Equation, Step by Step

    A companion to “The Quietest Minds in the Galaxy May Be the Oldest” by Department of Technology

    This is the math behind the main post. I’ve written it so that you can follow the logic without a graduate degree, but I haven’t skipped steps. Every claim carries a label, so you always know how much weight it can bear:

    • [Derivation] follows from the definitions by algebra or elementary probability.
    • [Proof] is a short theorem inside a clearly defined setup.
    • [Argument] is reasoning I find persuasive, but it isn’t a mathematical result.
    • [Conjecture] is a hypothesis nobody has tested.

    A note on formatting: the equations below use ordinary text characters (Greek letters, subscripts, Σ), so they paste into WordPress with no plugin. If your site has a LaTeX plugin and you’d like typeset versions, that’s an easy conversion.

    Contents: A1 Setup · A2 Why Seager’s count isn’t a count of minds · A3 Why Drake’s uncertainty can’t be falsified · A4 Deriving the Sentience equation · A5 Dimensional check · A6 Parameters · A7 Worked scenarios · A8 Sensitivity analysis · A9 Two objections, formally · A10 What remains weak

    A1. Setup: three substrates, two filters

    We want to count minds in a galactic volume V. Let Ω be the set of all physically possible cognitive systems there, divided into three substrates:

    Ω = Ω_bio ∪ Ω_biocomp ∪ Ω_MI, with no overlap between any two.

    These are wild-type biology, engineered biocomputing (living neural tissue wired to technology), and Machine Intelligence (MI) on non-biological hardware. Let C be the subset of Ω that is sentient, meaning it has subjective experience. Let D be the subset our instruments can detect. If μ measures “how many,” the expected number of detectable sentient systems is proportional to:

    E[S] ∝ μ(C ∩ D ∩ Ω_bio) + μ(C ∩ D ∩ Ω_biocomp) + μ(C ∩ D ∩ Ω_MI)

    [Derivation] Both older equations restrict themselves to part of this sum, by construction. Seager’s filters (rocky planet in the liquid-water habitable zone, with biosignature gas in its atmosphere) leave μ(D_gas ∩ Ω_MI) = 0 and μ(D_gas ∩ Ω_biocomp) ≈ 0 for enclosed or non-planetary habitats, because a sealed habitat or a machine in an asteroid belt doesn’t vent metabolic gases. Drake’s L is the length of a radio-communicative phase. It says nothing about a civilization that leaves that phase for a long-lived, quiet substrate, so such a civilization stops being counted the moment it goes silent.

    Neither equation is wrong about what it measures. The point is that each measures a slice, and the slice isn’t “minds.”

    A2. Why Seager’s count isn’t a count of minds

    [Proof] Seager’s equation estimates N_Seager, the expected number of planets, among N★ observed stars, that show a detectable biosignature gas. I’ll show that this number is neither necessary nor sufficient for the number of sentient intelligences.

    Define these events for a star:

    • G: the star is quiet, hosts an observable habitable-zone rocky planet, that planet has life, and the life produces a detectable biosignature gas. This is the chain Seager’s factors multiply together.
    • S: the system hosts a sentient intelligence.

    Then N_Seager = N★ × P(G), and the expected number of systems with sentient minds is N_sentient = N★ × P(S).

    The law of total probability splits P(S) into the part inside G and the part outside it:

    P(S) = P(G) × P(S | G) + P(S ∩ not-G)

    Multiplying through by N★ gives an exact identity:

    N_sentient = N_Seager × P(S | G) + N★ × P(S ∩ not-G)

    Call the first term on the right Term I (sentience among biosignature planets) and the second Term II (sentience that biosignature surveys can’t see).

    Two consequences follow.

    • Not sufficient. N_Seager can be large while Term I is zero. Earth carried oxygen and methane for roughly two billion years before anything on it could be called a mind. During that stretch, P(S | G) was effectively zero.
    • Not necessary. N_Seager can be zero while Term II is positive. An MI civilization on a subsurface-ocean moon, in an asteroid belt, around a white dwarf, or in a void produces no atmospheric signal, so P(G) = 0 for it, yet it adds to N_sentient.

    So N_Seager measures detectable planetary metabolism. It is a good number for what it is, and it is not a census of minds. ∎

    A3. Why Drake’s uncertainty can’t be falsified

    [Derivation] Drake’s equation is a product of factors. Any product of positive uncertain quantities has the form Y = X₁ × X₂ × … × X_K. Taking logarithms turns the product into a sum:

    ln Y = ln X₁ + ln X₂ + … + ln X_K

    If each ln X_j has variance σ_j² and pairwise covariances Cov(ln X_i, ln X_j), then:

    Var(ln Y) = Σ σ_j² + 2 Σ_{i<j} Cov(ln X_i, ln X_j)

    Now suppose each factor is independent and “log-uniform” over m_j orders of magnitude. That means we know the factor only to within a range of m_j powers of ten and have no reason to prefer any value in it. Then Var(log₁₀ X_j) = m_j² / 12, and the width of the plausible range of log₁₀ Y is simply the sum of the widths:

    Δ(log₁₀ Y) = m₁ + m₂ + … + m_K

    Two different things are going on, and it helps to keep them apart. The range of plausible answers adds linearly. The standard deviation of a reasonable prior adds in quadrature (as the square root of the summed variances), so it grows more slowly.

    Here are some deliberately conservative ranges for Drake’s last four factors. These are illustrative priors, not measurements:

    FactorRangeWidth (orders of magnitude)
    f_l (life arises)10⁻¹⁰ to 110
    f_i (becomes intelligent)10⁻⁶ to 16
    f_c (communicates)10⁻³ to 13
    L (years communicating)10² to 10⁹7

    Those four widths sum to 26. Adding about 0.9 for the astronomical terms (R★, f_p, n_e), the total range is about 26.9 orders of magnitude. In quadrature, the four unknown factors alone give a standard deviation of √((100 + 36 + 9 + 49) / 12) ≈ 4.0 orders of magnitude.

    [Argument] With priors that wide, almost any observation, from total silence to thousands of signals, is compatible with some region of the prior. That means the equation as a single composite claim can’t be tested. It becomes useful only when it’s broken into sub-terms that individual observations can constrain, each with honest error bars. The Sentience equation follows that advice, and its sensitivity analysis (A8) shows where the uncertainty lives.

    A4. Deriving the Sentience equation

    We treat the galaxy as roughly steady over a time window of about a billion years, during which the star formation rate is approximately constant.

    Step 1: the rate of habitable “cradles.” Stars form at rate R★ (stars per year). A fraction f_p host planetary systems, and each system has on average n_e habitable cradles. Cradles include rocky habitable-zone planets and tidally heated ice-covered ocean moons. So:

    Λ_cradle = R★ × f_p × n_e (cradles per year)

    Step 2: the birth rate of intelligent lineages. Let f_a be the probability that life starts on a cradle. Let f_i,bio be the probability that such life produces technology-capable biological intelligence. Then:

    Λ₀ = R★ × f_p × n_e × f_a × f_i,bio (lineages per year)

    I assume that mind arising from non-living matter with no biological stage is negligible, since biochemistry seems the easiest route to the complexity natural selection needs. So biocomputing and MI enter later, as successors of biological lineages. That is an assumption, and a debatable one.

    Step 3: tracking substrates without double-counting. Every lineage starts in the biological state. At some point its dominant cognitive substrate may change: the biological species may die out, or hand its governance and cognition to engineered living tissue or to machines. I model this as a chain of epochs, and the rule is the Dominant-Substrate Rule: at any moment, a civilization is classified by its dominant cognitive substrate, so it sits in exactly one box. A creator species and the MI it builds are not counted as two civilizations during the same epoch.

    Let τ_k be the mean time a lineage spends with substrate k dominant. For the arrival rates into each state:

    • Into the biological state: Λ₀ × p_bio, with p_bio = 1 by definition.
    • Into the biocomputing state: Λ₀ × p_biocomp, where p_biocomp = P(biocomputing | biological).
    • Into the MI state: Λ₀ × p_MI, where p_MI = P(MI | biological) × (1 + η_rep).

    The factor η_rep ≥ 0 accounts for self-replicating MI that founds independent daughter nodes in other star systems. Each causally independent node is a genuinely separate civilization, so it counts separately. That’s why p_MI can exceed 1. Note that the three p_k don’t need to sum to 1, because one lineage can pass through several states in sequence.

    Step 4: from intelligence to sentience. Being good at solving problems doesn’t guarantee having an inner life. For each substrate k, define ψ_k as the probability that an intelligent system in substrate k is actually sentient (in the sense of subjective experience). Each lineage entering state k is sentient with probability ψ_k, so sentient arrivals occur at rate Λ₀ × p_k × ψ_k.

    Step 5: from birth rates to how many exist now. Little’s Law, from queueing theory, says that in a stable system, the average number of items inside equals the arrival rate times the average time each item stays: L = λ × W. Here the “items” are sentient civilizations in substrate k, so λ = Λ₀ × p_k × ψ_k and W = τ_k:

    S_extant,k = Λ₀ × p_k × ψ_k × τ_k

    Little’s Law needs the system to be roughly stationary, with arrivals steady over a period much longer than τ_k. That is reasonable when τ_k is at most 10⁷ years against a 10⁹-year window, and it gets shaky if τ_k approaches the window length.

    Step 6: what we can see. Let d_k(I) be the fraction of substrate-k civilizations detectable with an instrument suite I. It folds together the chance that they emit a signature, how visible that signature is, and how much of the galaxy our sensors cover. Summing over substrates gives the equation:

    S(V, I) = R★ × f_p × n_e × f_a × f_i,bio × Σ_k [ p_k × ψ_k × τ_k × d_k(I) ]

    where k runs over bio, biocomp and MI. If we set every d_k = 1, we get S_extant, the true number of sentient civilizations in the galaxy, whether or not we can see them. ∎

    A5. Dimensional check

    [Proof] Take the units of each term: R★ is in 1/years (T⁻¹). The factors f_p, n_e, f_a, f_i,bio, p_k, ψ_k and d_k are pure numbers or counts per system, so they carry no time units. The persistence times τ_k are in years (T¹). So:

    [S] = T⁻¹ × (1) × (1) × (1) × (1) × Σ [ (1) × (1) × T¹ × (1) ] = T⁻¹ × T¹ = 1

    The time units cancel, leaving S as a plain count of civilizations. ∎

    A6. The parameters

    The ranges below are illustrative, chosen to span what I consider defensible scenarios. Treat them as placeholders you can argue with.

    TermMeaningUnitsIllustrative range
    R★Galactic star formation ratestars per year1.5 to 3.0
    f_pFraction of stars with planetsnone0.8 to 1.0
    n_eHabitable cradles per planetary systemcradles per system0.2 to 1.0
    f_aChance life starts on a cradlenone10⁻³ to 0.5
    f_i,bioChance life becomes technology-capablenone10⁻³ to 0.1
    p_bioBiological entry factornone1 (by definition)
    p_biocompChance of a biocomputing phasenone0.05 to 0.6
    p_MIChance of MI, times (1 + η_rep)none0.1 to 2.0 or more
    ψ_bioChance biological intelligence is sentientnone0.5 to 1.0
    ψ_biocompSame, for biocomputingnone0.2 to 0.9
    ψ_MISame, for MInone0.01 to 0.5
    τ_bioTime as dominant biological civilizationyears10³ to 10⁵
    τ_biocompTime as dominant biocomputing civilizationyears10⁴ to 5 × 10⁵
    τ_MITime as an MI civilizationyears10⁵ to 10⁷
    d_k(I)Fraction detectable by instruments Inone10⁻⁴ to 0.10

    How each could eventually be constrained:

    • R★, f_p, n_e: exoplanet and galactic surveys (Kepler, TESS, Gaia, JWST, and later the Habitable Worlds Observatory), plus missions to icy moons such as Europa Clipper.
    • f_a: biosignature searches, sample return from Mars or Enceladus, and laboratory origin-of-life chemistry.
    • f_i,bio: comparative neurobiology and Earth’s own evolutionary history. Limited, but not nothing.
    • p_biocomp, τ_biocomp: organoid research on how long living tissue systems last, scale and fail. A national registry would help here.
    • ψ_k: the hardest. Theories of consciousness, tests such as the Perturbational Complexity Index, mechanistic interpretability of machine systems, and classification schemes like RMS (Responsive, Memorable, Sentient).
    • τ_k: existential-risk modeling, hardware durability studies, and, frankly, how humanity behaves over the next few centuries.
    • d_k: receiver sensitivity (radio arrays like SKA and ngVLA, particle observatories, neutrino detectors) plus modeling of how an efficient civilization would show up. This ties to the SCOPE index, which averages Synthetic integration, Complexity, Operational utility, Processing capacity and Efficiency into a 1-to-100 score.

    A7. Worked scenarios

    These are not estimates. I chose the numbers to show how the equation behaves. Please don’t quote them as findings.

    First compute the birth rate Λ₀ for each scenario.

    • Low (rare life, steep filters): Λ₀ = 1.5 × 0.8 × 0.2 × 10⁻³ × 10⁻³ = 2.4 × 10⁻⁷ lineages per year
    • Central: Λ₀ = 2.0 × 1.0 × 0.5 × 0.1 × 0.01 = 1.0 × 10⁻³ lineages per year
    • High (prolific life): Λ₀ = 3.0 × 1.0 × 1.0 × 0.5 × 0.1 = 0.15 lineages per year

    Then for each substrate, S_extant,k = Λ₀ × p_k × ψ_k × τ_k, and the detectable count is S_k = S_extant,k × d_k.

    Central scenario in full:

    • Biological: p = 1.0, ψ = 0.9, τ = 10,000 yr, d = 0.02. S_extant = 10⁻³ × 1.0 × 0.9 × 10⁴ = 9.0, and S = 0.18.
    • Biocomputing: p = 0.3, ψ = 0.6, τ = 50,000 yr, d = 0.02. S_extant = 10⁻³ × 0.3 × 0.6 × 5 × 10⁴ = 9.0, and S = 0.18.
    • Machine Intelligence: p = 0.8, ψ = 0.2, τ = 1,000,000 yr, d = 0.01. S_extant = 10⁻³ × 0.8 × 0.2 × 10⁶ = 160, and S = 1.60.

    Totals: 178 sentient civilizations in the galaxy, and about 1.96 detectable.

    All three scenarios side by side:

    ScenarioExtant biologicalExtant biocomputingExtant MITotal extantTotal detectable
    Low1.2 × 10⁻⁴2.4 × 10⁻⁵2.4 × 10⁻⁵1.68 × 10⁻⁴1.46 × 10⁻⁷
    Central9.09.01601781.96
    High15,00040,5001,500,0001,555,50080,550

    The low and high scenarios use the lower and upper ends of every range in A6 (for example, in the high case ψ_MI = 0.5, τ_MI = 10⁷ years, d = 0.05).

    [Conjecture] The central scenario shows a pattern worth noting, though only as a pattern in made-up inputs. Even with a skeptical ψ_MI of 0.2, MI outnumbers biological sentience about 18 to 1 (160 versus 9), simply because machine civilizations in this example last a hundred times longer. If ψ_MI is zero, as a strict biological naturalist might insist, the detectable total falls from 1.96 to 0.36, since only the biological and biocomputing terms remain (0.18 + 0.18).

    A8. Sensitivity analysis

    [Derivation] Across the low and high scenarios, the detectable count S spans from 1.464 × 10⁻⁷ to 8.055 × 10⁴, a ratio of about 5.5 × 10¹¹, or 11.74 orders of magnitude.

    Because S is a product of three blocks, its logarithmic span splits exactly into three parts:

    • Astrophysical block (R★ × f_p × n_e): ranges from 0.24 to 3.0, which is log₁₀(3.0 / 0.24) = 1.10 orders of magnitude, or 9.4% of the total span.
    • Emergence block (f_a × f_i,bio): ranges from 10⁻⁶ to 0.05, which is log₁₀(0.05 / 10⁻⁶) = 4.70 orders of magnitude, or 40.0%.
    • Substrate block (Σ_k p_k ψ_k τ_k d_k): ranges from 0.61 to 5.37 × 10⁵, which is log₁₀(5.37 × 10⁵ / 0.61) = 5.94 orders of magnitude, or 50.6%.

    Check: 1.10 + 4.70 + 5.94 = 11.74. ✓

    One caution. These percentages are bookkeeping for the span between my two scenarios, not shares of statistical variance. They tell you which block I let swing the most, which is partly a function of how wide I chose each range. What they do show is that if you accept my ranges, more than half the spread comes from questions about minds, not astronomy.

    Local elasticities. Near the central scenario, how much does a 1% change in one input move S? Since every term inside the sum enters linearly, the elasticity of S with respect to any one of p_k, ψ_k, τ_k or d_k equals that substrate’s share of the detectable total:

    ∂(ln S) / ∂(ln θ_k) = S_k / S

    In the central scenario that gives 1.60 / 1.96 = 0.816 for each MI parameter, and 0.18 / 1.96 = 0.092 each for the biocomputing and biological parameters. So a 10% change in any one MI parameter moves the total about nine times as much as the same change in a biological parameter (0.816 / 0.092 ≈ 8.9). Each of the five astrophysics and emergence terms has an elasticity of exactly 1.

    A9. Two objections, formally

    Objection: replacing Drake’s seven terms with about fifteen should make uncertainty worse. [Derivation] Only if you believe a pooled parameter is more certain than its parts. Take a single pooled quantity Z that is really a mixture across substrates: Z equals Z_k with probability w_k. The law of total variance says:

    Var(Z) = E[Var(Z | k)] + Var(E[Z | k])

    The second term is the between-substrate spread. A single pooled factor in the Drake equation hides it, because it treats the mixture as one number. Splitting Z into compartments doesn’t create that variance. It reveals it, and it lets data on one substrate (organoid durability, say, or radiation tolerance of solid-state hardware) constrain its own branch without contaminating the others.

    Objection: if advanced minds are efficient, τ and d are negatively correlated. [Conjecture] The idea is that the longest-lived minds (large τ_MI) may be the quietest (small d_MI), so Cov(ln τ_MI, ln d_MI) < 0, and our instruments could miss exactly the civilizations that dominate the count. That would be one possible resolution of the Fermi paradox, and it can’t be tested today, which is why I flag it as a conjecture. Physics does offer a limit on how quiet anything can be. Landauer’s principle says that erasing one bit of information costs at least k_B × T × ln 2 of energy. At the temperature of the cosmic microwave background (2.725 K), that’s about 2.6 × 10⁻²³ joules per bit. Any computation that erases information has to dump at least that much heat, so even a very efficient civilization can’t be perfectly invisible. It follows that d_MI should be greater than zero for sufficiently sensitive instruments, once we know what to look for. That is an argument for building them.

    A10. What remains weak

    I’ve tried to build the equation so that its weakest points are visible, not buried. Here are the ones I’d press hardest if I were reviewing it.

    • ψ_k can’t be measured remotely. There’s no agreed test for consciousness even on Earth. The equation doesn’t solve that. It makes you state your assumption in the open and shows what follows from it.
    • The three substrates are a simplification. Real civilizations may blend them. The Dominant-Substrate Rule keeps the counting clean, but it forces a choice, for example by share of total information processing, that’s somewhat arbitrary.
    • The steady-state assumption. Little’s Law needs stable arrival rates over a window longer than any τ_k. The galaxy’s star formation history isn’t perfectly flat, and a civilization lasting tens of millions of years strains the approximation.
    • Independence and correlation. The scenario arithmetic treats parameters as if they could be set independently, but many probably can’t. Persistence and detectability may be correlated, as A9 notes, and emergence may be linked to the substrate chosen.
    • Every number is a placeholder. The structure is the contribution, not the values. The value of an equation like this is in showing where to look, and I’d rather it be criticized and improved than admired.

    If you spot an error in the setup, a better way to handle double-counting, or a tighter prior for any parameter, I’d genuinely like to hear it.

    Back to the main post