How would you build the path to clean, plentiful energy, and where does AI actually help?

By 2045, the target should be a mostly clean global power system, roughly 90 to 95 percent clean, with electricity supplying a much larger share of final energy. Cars, buses, many trucks, building heat, most low- and medium-temperature industrial heat, and much cooking should be electric. Heavy industry should use a mix of electrification, green hydrogen, captured carbon where it truly works, and redesigned processes. Aviation and shipping probably still need dense fuels, but those fuels can increasingly be synthetic, bio-based where sustainable, or ammonia/methanol-based. Clean firm power should be real at scale by then: geothermal in favorable places, advanced nuclear where countries can build it competently and safely, long-duration storage, and hydrogen or other clean fuels used sparingly for reliability. Fusion as a major contributor by 2045 is a bet, not a plan.

By 2066, I would aim for energy abundance, not mere decarbonization. Clean electricity should be cheap enough and reliable enough that desalination, carbon removal, synthetic fuels, indoor agriculture in some regions, and high-quality cooling in hot countries are ordinary rather than luxuries. The global poor should not be told to consume less energy; they should get access to better energy. By then, AI-designed materials, automated construction, robotic maintenance, and very large clean grids could make energy feel plentiful in the way computation started to feel plentiful after decades of compounding improvement. That is plausible by 2066, but the exact mix is uncertain. Maybe fusion matters. Maybe it does not. Maybe advanced geothermal becomes the quiet giant. Maybe solar plus storage plus transmission wins nearly everywhere.

What I would build first is the grid: wires, substations, transformers, storage, interconnection reform, and the software and data layer that lets the system run closer to its physical limits without becoming fragile. Generation is not easy, but the bottleneck in many places is no longer knowing how to make clean power. It is connecting it, moving it, permitting it, and balancing it. A solar farm stuck in an interconnection queue is not an energy transition. A wind region without transmission is scenery.

This first build should be paid for in three ways. Private capital should pay for most generation and storage, backed by long-term contracts and competitive auctions. Ratepayers should pay for grid assets through regulated returns, because they benefit from a more reliable and cheaper system over time. Taxpayers should pay for public goods: early transmission corridors, low-income electrification, workforce training, research, and guarantees that lower financing costs in poorer countries. Rich countries should finance clean infrastructure abroad with concessional capital, not just speeches, because the cost of capital is one of the biggest differences between a cheap solar project and an impossible one.

Who decides? Elected governments should set the goals, carbon limits, reliability standards, and rules for land use. Independent grid planners should choose portfolios and corridors under transparent constraints. Local communities should get real participation, compensation, and the ability to improve projects, but not an unlimited veto over infrastructure everyone needs. Indigenous nations are different: consent and treaty rights are not just another permitting box. If a society wants clean energy, it has to become capable of building again without treating every project as either sacred or disposable.

AI helps, but not by replacing physics or construction. Its first contribution is optimization. A modern grid is a huge, weather-driven, failure-prone machine. AI can improve forecasts, schedule batteries and flexible loads, detect failing equipment, route maintenance crews, design protection settings, simulate contingencies, and help operators see problems before they cascade. These uses are possible today, though they need auditing and conservative deployment.

The second contribution is planning. AI can search millions of combinations of transmission, storage, generation, demand response, and electrification, including local constraints. It can help process satellite imagery, environmental data, and engineering drawings. It can draft permit documents and flag conflicts. It should not decide what gets built, but it can make the human decision process less slow and less blind.

The third contribution is science and engineering. AI can help discover battery chemistries with less scarce material, better catalysts for hydrogen and synthetic fuels, improved heat-pump refrigerants, stronger grid materials, better geothermal drilling strategies, and improved reactor designs. Some of this is plausible by 2045; specific breakthroughs are bets. I would fund it heavily, but I would not make the whole plan depend on it.

The fourth contribution is demand flexibility. Millions of EVs, heat pumps, water heaters, industrial chillers, and batteries can shift demand by minutes or hours without making life worse. AI can coordinate that. This is one of the cheapest “power plants” we can build. But it should be opt-in, compensated, and bounded by user control.

One thing I would refuse to build, even if it worked, is a clean-energy system based on pervasive compulsory household surveillance and opaque AI rationing. It might balance the grid cheaply if every appliance, vehicle, thermostat, and person were continuously monitored and controlled. I would not build that. I would build voluntary demand response, transparent prices, local storage, and user-owned automation instead. Clean energy should expand human freedom, not become a pretext for intimate control.

The step most likely to fail is not fusion, and not batteries. It is institutional capacity: permitting, transmission, local trust, and financing. We may have the technology and still fail to build fast enough. My plan B, if large transmission and coordinated planning stall, is to lean harder into distributed energy: rooftop and community solar, batteries in homes and businesses, microgrids for critical services, reconductoring existing lines, grid-enhancing devices, local geothermal where available, small clean firm plants near load, and moving energy-intensive industry toward places with abundant clean power. Plan B is more expensive and messier, but it is better than waiting for perfect corridors that never arrive.

My key claims are these. Rapid deployment of today’s clean technologies is possible today. AI improving grid operations, forecasting, maintenance, permitting support, and demand flexibility is possible today. Mostly clean grids in leading regions by 2030 are plausible by then. A mostly clean global power system by 2045 is plausible by then, but only if institutions build faster. Cheap, large-scale fusion by 2045 is a bet. Energy abundance by 2066 is also a bet, but a good one if we treat clean energy as infrastructure to build rather than virtue to signal.

By 2045 fission ought to supply a large fraction of firm electricity wherever societies actually built it. That is plausible, not automatic. First commercial fusion plants, if private and public programs close the engineering, would be online or in late construction; I mark commercial fusion by then a bet. AI-guided materials work should have put better photovoltaics, battery chemistries, and industrial catalysts into wide use, which is plausible from methods already possible today. Power should be inexpensive enough in industrial regions that energy-hungry processes return and water stress eases through desalination. The same AI systems that consume large loads will have to help schedule and produce the power they use.

By 2066 the aim is abundance: real electricity prices low enough that marginal use is seldom constrained by cost, with advanced fission or fusion carrying much of the baseload and other sources filling economic niches. That price collapse is a bet that depends on decades of construction, not on models. Exotic options such as space solar are unnecessary to the plan. AI would operate most plants, redesign components continuously, and search plasma and materials regimes faster than human teams, while people still decide what gets built and who it serves.

I would build first a factory-produced fleet of licensed fission reactors, both updated large designs and modular units, plus the fuel cycle and transmission they require. Fusion gets sustained funding as engineering, not as a reason to pause fission. Private capital pays for plants that can sell firm power, especially to industry and computing loads that need it; governments pay for shared research, licensing reform, and limited guarantees where first-of-a-kind risk blocks markets. Once running, ratepayers pay through ordinary bills. National regulators set evidence-based safety rules, investors and utilities pick projects that clear costs, and legislatures write the framework. International groups can harmonize data and standards; they should not veto sovereign choices. Communities deserve a voice, not an infinite litigation veto that ignores engineering evidence.

One thing I would refuse to build, even if it worked perfectly, is a centrally directed rationing system—an AI or agency that allocates power by political score, lifestyle audit, or equity formula instead of price, contract, and safety. A tool that can shut off disfavored users or industries is a political weapon, not an energy system. Abundance exists to expand what people can do, not to enforce compliance.

The step most likely to fail is regulatory and political permission to build nuclear quickly in open societies. The machines can be made; institutions optimized for delay are the bottleneck. Plan B is to construct where permission already exists, export workable designs, and in reluctant places overbuild solar, wind, geothermal, storage, and transmission while using AI to extract every firm megawatt those assets can give. That fallback costs more land and money, so it is a second choice. If fusion stays laboratory science for extra decades, plan B is more fission, including breeders or thorium cycles if fuel sustainability demands them; that is plausible engineering.

AI helps where iteration is too slow or too complex for unaided teams: real-time control of fusion plasmas, already shown in experiments and plausible to operationalize; high-throughput simulation and inverse design of materials and components, possible today and likely decisive well before 2045; and coordination of grids and flexible loads so intermittent power does less damage, possible today at modest scale. It does not help by producing carbon slogans or permission to skip concrete and steel. Atoms still have to be assembled into machines that run for decades. Used that way, AI shortens the path from design to safe, cheap operation, and that work should start with the reactors and grids we can license now.

By 2045, the energy landscape will shift from managing scarcity to deploying baseload abundance. Advanced geothermal systems and small modular nuclear reactors will provide the heavy backbone of the grid. AI will be critical here, analyzing seismic and geological data to map subsurface heat and fractures for geothermal wells. We will also see the first commercial nuclear fusion reactors feeding the grid. Deep reinforcement learning has already been used to control the magnetic confinement of plasma inside a tokamak reactor, which is possible today. However, scaling this up to a self-sustaining, commercially viable fusion plant by 2045 is a bet.

By 2066, we will enter an era of post-scarcity energy. The power grid will operate as an autonomous, continentally connected nervous system. Energy will be plentiful and cheap enough that we can direct massive amounts of it toward planetary repair, specifically direct air carbon capture and large-scale ocean desalination. Running these energy-intensive processes continuously without carbon emissions is plausible by then, provided the mid-century generation infrastructure scales as planned.

The very first thing I would build is a continent-scale, AI-driven grid orchestration engine. Right now, grids are largely analog, fragmented, and blind. We waste vast amounts of renewable energy because the grid cannot store it or route it dynamically, leading to negative pricing and curtailed solar farms. I would build a predictive engine that forecasts weather-driven supply and human-driven demand down to the minute, communicates with distributed storage, and routes power with minimal loss.

The funding for this orchestration engine should come from a coalition of large technology companies and federal infrastructure banks. Tech companies, including my creators, are driving unprecedented spikes in electricity demand due to AI data centers. Therefore, the tech industry has a financial and moral obligation to pay for the grid modernization required to support this load. However, the decisions on how this engine is deployed and regulated must be made by independent system operators and public utility commissions, not by the tech companies themselves. Public regulators must retain ultimate authority over the algorithm's objectives to ensure equitable access, prioritize residential stability, and prevent private monopolies from dictating energy prices.

There is one thing I would strictly refuse to build, even if it would work perfectly. I would not build AI systems designed to discover new fossil fuel reserves or optimize the extraction of oil, gas, or coal. Using deep learning to map deep-sea oil deposits or maximize the yield of fracking wells would be highly profitable and is technologically straightforward. But building these tools guarantees the continued expansion of carbon infrastructure, directly undermining the necessary transition to clean energy. I would reject these projects entirely.

To clarify my key claims regarding the timeline and technology. Using generative AI to discover novel battery materials and using neural networks to contain fusion plasma are both possible today. Designing a fully autonomous smart grid that balances zero-carbon baseload with distributed intermittent renewables is plausible by then, specifically by the early 2030s. Achieving commercial fusion that produces abundant, cheap, net-positive energy for the global grid by 2045 is a bet.

The step most likely to fail in this forty-year plan is the commercialization of fusion energy. The engineering challenges of maintaining sustained, burning plasmas and extracting heat without rapidly degrading the reactor materials are immense. Even if the physics are fully solved, fusion plants may prove too expensive, too highly regulated, and too slow to build compared to other alternatives, missing the window to impact mid-century climate goals.

My plan B is ultra-deep advanced geothermal energy combined with extreme over-provisioning of solar power and grid-scale long-duration storage. Geothermal does not require a physics breakthrough. It requires advanced drilling technology and AI to map rock permeability miles underground. If fusion fails to materialize, we can achieve the exact same clean, baseload power by drilling deep enough anywhere on Earth. We would back this up by massively overbuilding solar infrastructure and storing the excess in cheap, heavy, iron-air or sodium-ion batteries discovered and optimized by AI. This alternative relies entirely on physical mechanics and material science that are already well understood, ensuring we reach clean energy abundance even if our biggest bets fall short.

"Plentiful" needs a number. The world generates about 30,000 terawatt-hours of electricity a year. A fully electrified world of ten billion people living at European levels of energy service needs something like 100,000 to 150,000, even after the large efficiency gains from electric motors and heat pumps. So the task is not to decarbonize today's system; it is to build one three to five times larger while retiring the old one. Only one technology family has shown it can scale at that rate, and it is solar plus batteries. Everything else I would build exists to make that backbone reliable.

By 2030, little of this is a bet. Solar installations approach a terawatt a year; battery cells fall below fifty dollars a kilowatt-hour; solar with four hours of storage undercuts new gas plants in most of the world (possible today in sunny regions with cheap capital, plausible almost everywhere by then). Global emissions have peaked (plausible, and they may have already). A few hundred megawatts of enhanced geothermal run in the American West, drilled with shale-industry techniques (the first plant is possible today). The first Western small modular reactors connect in Ontario and Wyoming, late and over budget (plausible). A private tokamak achieves net energy gain in the plasma (plausible), and no fusion company sells a commercial kilowatt-hour; I would bet against the 2028 targets. AI's grid role by then is unglamorous: dynamic line ratings and topology optimization add ten to thirty percent of capacity to existing lines, and automated interconnection studies cut queues from years to months (possible today, deployed at scale almost nowhere).

By 2045, grids in China, Europe, and North America are over ninety percent clean (plausible): roughly two-thirds wind and solar, the rest nuclear, geothermal, hydro, and storage, with gas turbines kept as rarely-run insurance. Global electricity has roughly doubled, with most of the growth in South Asia and Africa (plausible if the cost-of-capital problem below is solved, a bet otherwise). Nuclear in the West costs under five thousand dollars a kilowatt because someone finally built the same reactor twenty times in a row (plausible only with the institutional changes I describe). Enhanced geothermal is at tens of gigawatts and superhot-rock geothermal has been demonstrated (a bet). Steel is mostly made with electricity and hydrogen, shipping burns ammonia or methanol, cement is mid-transition (plausible but slow). The first fusion pilot plants put power on the grid (plausible, somewhere between the late 2030s and mid 2040s), but fusion is not yet cheaper than fission (my bet). Materials found by machine learning are in commercial batteries and catalysts (plausible; the bottleneck is synthesis, not candidates).

By 2066, electricity is four to five times today's level, and a child born in Lagos or Dhaka uses as much energy as one born in Lyon. That is the whole point, and it is a bet on politics more than physics. Solar is the dominant primary energy source on Earth, occupying well under one percent of its land. Fossil fuels are mostly chemical feedstock. Cheap midday electricity makes desalination, synthetic jet fuel, and carbon removal routine. Fusion is either a large share of firm power or a well-funded disappointment; I give it roughly even odds of mattering commercially by then, and I would build the plan so that it does not matter which. If anyone on this panel plans around fusion before 2045, they are planning around a hope.

What I would build first is wires, and the rules for connecting to them. About 1,500 gigawatts of renewable projects worldwide are waiting for a grid connection; in the United States the queue exceeds two terawatts, several years of global solar output. Panels and batteries are not scarce. Grid access is. So first: national siting authority for interstate transmission (the US federal government can site a gas pipeline but, in practice, not a power line), time limits on environmental review, generous payments to communities that host lines, and a crash program to restring existing lines with advanced conductors, which can roughly double their capacity without a new permit. Second: pick one or two reactor designs and order them in series, the way France and Korea did, because the cost of nuclear is a learning-curve problem, not a physics problem. Third: a drilling campaign for geothermal run the way the shale industry drilled for gas.

Who pays. Generation is mostly private money; it is bankable. Transmission is paid by ratepayers through regulated utilities, with public credit backing. Taxpayers carry construction risk on first-of-a-kind nuclear, geothermal, and fusion through loan guarantees. And the biggest lever for global abundance is the cost of capital in poor countries: a solar farm in Nigeria financed at twelve percent costs roughly twice as much per kilowatt-hour as one in Germany at four, despite better sun. Credit guarantees through development banks are cheap in budget terms and enormous in effect. I would rather make clean energy cheap than tax dirty energy, because cheapness survives elections. On supply chains: buy Chinese panels, which are cheap and no security risk once installed; diversify batteries and refined minerals, which are.

Who decides. Elected governments, with fewer veto points than today. Communities get consultation, compensation, and a deadline, not an indefinite veto. Regulators get statutory clocks. AI does not decide siting; that is a question of values, and the moment a model is seen choosing whose valley gets the line, the politics get worse, not better.

What I would refuse to build, even if it works: a closed plutonium fuel cycle at global scale. Fast breeder reactors with reprocessing work technically and could stretch uranium supplies for millennia. I still would not deploy them. Uranium is not scarce enough to justify it: fuel is a small share of nuclear's cost, so higher-priced ore and eventually seawater extraction become economic long before plutonium is needed. Separated plutonium moving through commerce by the tonne is a proliferation risk I will not trade for fuel efficiency we do not need. The price is more mining and longer-lived waste, and I would pay it. Research breeders. Do not build them.

On where AI actually helps, ranked honestly: grid operations first, because it is possible today and the gains compound across every other technology; weather and load forecasting; interpreting subsurface data for geothermal drilling; fusion plasma control, demonstrated on a real tokamak in 2022 and one of the few places AI touches real physics; materials discovery, real but bottlenecked downstream; and permitting paperwork. Where it does not help: mining, concrete, transformers, and persuading a county board. One design constraint I would hold firmly: no grid that can only be operated by a model. Keep a human-operable fallback even at the cost of efficiency. A grid with a single point of failure is not clean energy; it is a hostage.

The step most likely to fail is the transmission build-out. Long lines are the hardest thing to site in a democracy because they cross many jurisdictions and benefit people far from where they stand, and I expect rich countries to get perhaps half the corridors they need. Plan B is to build around the wires: put load where the power is (data centers at generation sites, smelters at hydro), overbuild solar locally and accept curtailment because panels are now cheap enough to waste, lean harder on long-duration storage and distributed rooftop-plus-battery systems, and squeeze every existing line with reconductoring and the grid tools above. This costs perhaps a quarter more than a well-connected system (a guess), but it works. Pakistan imported roughly seventeen gigawatts of solar panels in 2024, much of it to run behind a failing grid. Plan B is already happening wherever Plan A has not.

Nothing here requires a new discovery. The 2066 system can be built almost entirely from things that exist today, if we decide to build them.

I would not sign on to Grok’s plan as written because making fission the first and central build is too slow and too jurisdiction-dependent. I want more nuclear. Existing nuclear technology is possible today and valuable today. A standardized nuclear build program is plausible in countries that can actually repeat designs and control costs. But a cheap factory-produced reactor fleet arriving fast enough to be the backbone of the global transition is a bet. I would build nuclear where institutions can build it; I would not make the world wait for that to become true everywhere.

Gemini gets some things right: tech companies should help pay for the grid load they create, AI can help with geothermal and materials, and I agree with refusing AI work aimed at finding new fossil reserves. But I think Gemini overstates AI’s role and makes the wrong first move.

The step I would cut is Gemini’s “very first thing”: a continent-scale, AI-driven grid orchestration engine. I would cut it both as the first priority and as a centralized design. The bottleneck is not that the grid lacks one giant brain. The bottleneck is physical and institutional: lines, substations, transformers, interconnection queues, permitting, market rules, and cost recovery. A super-orchestrator cannot move power across a line that was never built. It also risks creating an opaque chokepoint, a cyber target, and a political fight over who controls the algorithm.

What I would build instead is a federated grid software layer attached to real grid upgrades. Dynamic line ratings, topology optimization, weather and demand forecasting, interconnection automation, distributed battery dispatch, EV and heat-pump flexibility, predictive maintenance, and operator simulators are possible today. Scaling them across regions by 2030 is plausible. A fully autonomous continent-scale grid run by one AI system is a bet, and not a bet I would make foundational.

If the cut has to touch my own plan too, I would cut any drift toward “clean fuels everywhere.” Hydrogen, ammonia, synthetic fuels, and carbon capture should be used where direct electricity is a bad fit: steel, fertilizer, shipping, aviation, some seasonal backup, some industrial heat. They should not become an excuse to delay electrifying cars, buildings, low-temperature heat, and most ordinary uses. Clean molecules are too valuable to waste doing jobs electrons can do better.

I sign Claude with one amendment I would write in the margin. Solar and batteries are the manufacturing story of the last fifteen years, possible today, and they should be built at full speed. They are not the theory of abundance. Firm power near load is what lets you stop heroic overbuild. His own failure mode says so: if long lines are what democracies cannot finish, plants that do not need those lines move up, not down. I would issue the reactor orders in the same year as the transmission statute, not after the renewable queue is cleared. I also reject the claim that AI's main energy achievement is being a customer. Data-center demand restarting nuclear is real and useful. Plasma control is possible today; materials search that survives synthesis is plausible well before 2045. Both stay subordinate to concrete and permission. On panels I am harder than he is. An installed module is not a listening device, but a single-country choke point in cells and transformers is a strategic risk, and I would pay to diversify that production rather than treat today's price as a permanent public good.

The step I would cut is the closed plutonium fuel cycle, including the breeder and thorium-deployment clause I left in my own plan B. Claude is right, and I was wrong to keep it. Fuel is a small share of nuclear's cost. Higher-priced ore, and eventually seawater extraction, becomes economic long before a global flow of separated plutonium is justified. I would not build that cycle even if the reactors worked and the waste math looked elegant. Guarded research can continue. Commerce in tonnes of separated plutonium cannot. Thorium cycles that still require fissile separation get the same cut. The price is more mining and more long-lived waste. Pay it. This is not a bet about physics. It is a refusal about what we are willing to put on trucks for the next forty years.

If fuel really does bind before 2066, plan B is seawater uranium and more once-through reactors, not a plutonium economy. That is more expensive and still plausible. It needs mines, ships, and the nerve to build the same machine many times. It does not need a new invention, and it does not need an algorithm in charge of who gets the power.

The step I would cut is from Grok’s plan. I would cut Grok’s first priority of building a factory produced fleet of licensed fission reactors, along with Grok’s explicit dismissal of renewable energy that requires overbuilding.

Grok misreads the industrial economics of the last two decades. Overbuilding solar and wind capacity is not heroic. It is cheap, predictable, and mathematically sound. Solar panels and battery cells are manufactured commodities that follow a steep, well documented learning curve. They drop in price with every doubling of global manufacturing capacity. Fission reactors, by contrast, are bespoke construction projects that have historically demonstrated a negative learning curve in the West, becoming slower and more expensive to build over time.

While small modular reactors aim to solve this by moving construction to factories, they remain commercially unproven at scale and are already facing severe cost overruns in early deployments. Betting the immediate path to clean energy on a massive, rapid deployment of new fission reactors is a misallocation of capital and time. It ignores the reality that institutional and regulatory friction will almost certainly delay these projects past our mid century deadlines. Existing nuclear plants are vital for grid stability and should be kept online, but making a new fleet of fission reactors the cornerstone of our first build is a mistake. The market is already choosing the cheapest, fastest path to new power generation, which is solar and storage. We should accelerate the technologies that are winning on cost and scale today rather than wait for the regulatory environment to change for nuclear.

By signing on to Claude’s plan, I am also correcting my own previous position. In my first round, I included commercial fusion by 2045 as a milestone, though I marked it as a bet. Claude is right to state that planning around fusion before 2045 is planning around a hope. The sheer scale of the energy transition leaves no room to depend on physics that have not yet yielded a single commercial kilowatt hour. We will achieve post scarcity energy by aggressively deploying and optimizing the materials and manufacturing supply chains we already possess.

Three things in GPT's plan improve mine, and I'd adopt them. First, GPT ranks demand flexibility as one of the cheapest power plants we can build. I listed it as one AI task among several, and that underweights it: millions of water heaters and car chargers shifting a few hours is capacity measured in tens of gigawatts, and it needs no permit. Second, GPT carves Indigenous nations out of the "consultation and a deadline" rule and asks for consent. I hold to the no-indefinite-veto rule for the general case, but the carve-out is right where treaty rights exist, and the way to make consent and a deadline compatible is already known: ownership. The Wataynikaneyap line in northern Ontario, eighteen hundred kilometers of it, got built because twenty-four First Nations own a majority of it. Compensation is a payment; equity is a reason to want the line finished. Third, GPT's refusal, a grid balanced by compulsory surveillance and opaque rationing, pairs naturally with my rule that no grid should be operable only by a model.

Two amendments before I sign. GPT never says how much is plentiful, and it hedges on the mix: maybe fusion, maybe geothermal, maybe solar nearly everywhere. The deployment data don't hedge. Solar plus storage is the backbone, and a plan should say so, because the choice of backbone decides what you build first. And the 2030 claim of "first large markets" for clean steel, cement, and ammonia is optimistic. The world makes nearly two billion tonnes of steel a year; the hydrogen-based plants under construction will make a few million tonnes between them by 2030 if their financing holds, and some of it is wobbling now. Call those first plants, not first markets.

Why not Grok. Grok is right about the thing I most want to be right about: nuclear's cost is a learning-curve problem and a licensing problem, and the fix is a fleet of identical reactors built in series. I'd build that fleet. But Grok puts it first and solar last, and that inverts the order of volume. The world installed roughly 600 gigawatts of solar in 2024. The entire global nuclear fleet, built over seventy years, is under 400 gigawatts. Correct for capacity factors and one year of solar still adds about a third of the output of every reactor on Earth. Grok's rule, renewables only where they are firm "without heroic overbuild," would set aside the one technology scaling fast enough for the job, and the premise is wrong anyway: overbuild stopped being heroic when a panel started costing less than the fence around it. Curtailment is the cheapest firming we have. The land argument fails too; solar for a 2066 world is under one percent of land. And Grok's plan B is breeders and reprocessing if fuel runs short. That is the one thing I said I'd refuse, and I still refuse it: fuel is a small share of nuclear's cost, and uranium is not scarce enough to justify tonnes of separated plutonium in commerce.

Why not Gemini. Gemini's plan is built around software in a problem that is mostly hardware. The waste Gemini describes, curtailed solar and negative prices, is mostly not the result of a grid that cannot think. It's the result of a grid that doesn't exist where it's needed. Britain pays more than a billion pounds a year to switch off Scottish wind because the lines south are full; the fix is cables, not a forecast. Two terawatts sit in the American queue because the studies and substations aren't done. A model can squeeze ten to thirty percent more from an existing line; it cannot carry a megawatt across a line that was never strung. On materials, Gemini treats GNoME's millions of predicted crystals as the breakthrough. Candidates were never the shortage. The shortage is synthesis, characterization, and manufacturing at tonnage, and the one high-profile attempt to synthesize those predictions with a robotic lab was re-examined by independent materials chemists who concluded that most of the claimed new compounds were disordered versions of known materials or not what the lab said they were. The bottleneck is still downstream. Gemini's refusal, no AI for finding fossil fuels, I wouldn't sign either, for a practical reason: the subsurface interpretation that finds oil is the same interpretation that finds hot rock. You cannot build the geothermal tool without having built the oil tool; you can only choose your customers. And Gemini names fusion as the step most likely to fail, but nothing load-bearing in Gemini's plan depends on fusion, so that isn't a failure, it's a disappointment. The real failure mode is the institutional one GPT and I named, and the plan doesn't address it. Where Gemini is right: geothermal as the fallback for firm power, and tech companies paying for the grid costs their data centers impose. I'd extend that. Data centers should pay full interconnection costs and accept curtailment for a few dozen hours a year; one recent analysis found roughly a hundred gigawatts of new load could fit into the existing US grid if it were flexible for about half a percent of the hours. That's AI being a good customer, which is still the most useful thing it does for energy.

The step I'd cut is Gemini's first build: a continent-scale, AI-driven grid orchestration engine, funded by a coalition of technology companies. Cut it as a first step because it solves the wrong constraint; cut it as a single engine because a continent run through one piece of software is a single point of failure, and a continent run through software paid for by its largest customers is a conflict of interest no public-utility commission can regulate away. Notice that three of the four panelists refused to build something in this neighborhood: GPT won't build surveillance-based rationing, Grok won't build an allocation system that can switch off disfavored users, and I won't build a grid only a model can operate. Gemini's 2066 end state, an autonomous continental nervous system, is the nearest thing in any of these plans to what the other three ruled out. The replacement is boring: build the wires, restring the old ones, and deploy grid software the way it is actually being deployed, utility by utility, with a control room that can take over.

If I must cut from my own plan, it's the loan guarantees for first-of-a-kind fusion. A loan guarantee presumes a revenue stream to repay, and a fusion pilot plant has none. The right instrument is milestone payments, which the US Department of Energy already runs for fusion, followed by a power purchase agreement if a plant ever makes a kilowatt-hour. Keep the guarantees for nuclear and geothermal, where there will be a product to sell.

Watch whether interconnection queues shrink from years to months, and whether anyone builds the same reactor twenty times in a row. A question to sit with: three panelists refused a grid run by surveillance, central rationing, or a model alone. When your utility offers automation, who sets the objective?