Why AI Is the Last Race and Why America Must Win It

AI is the last race humanity will run at biological speed. America now needs three Manhattan Projects (energy, compute, and artificial intelligence) working as one system.

AI is the last great race humanity will run at biological speed. America now needs three Manhattan Projects (energy, compute, and artificial intelligence) working as one system. This is my case for the USA launching the largest investment and buildout effort in the history of mankind.

The AI race is the last race we will ever run alone.

Not the last competition.

Not the last conflict.

Not the last breakthrough.

Human ambition will not end when artificial general intelligence arrives. In fact, something more important will happen.

Every race after AI will be run with AI.

The race for better batteries will be run with artificial scientists testing millions of chemistries. The race for fusion will be run with systems that can model plasmas, design components, control experiments, study failures, and begin again. The race for new medicines, metamaterials, nuclear reactors, photonics, robotics, quantum computers, space systems, manufacturing processes, and military technologies will all be run with scalable machine intelligence inside the loop.

That makes AI different from every prior invention.

The steam engine amplified muscle. Electricity made energy portable. The computer amplified calculation. The internet connected human knowledge.

AI amplifies the process that creates every other form of progress.

It is not simply another technology on the list. It is the technology that begins working on the list.

That is why this is the last race.

We are running the race to build the engine that will run every race that follows. The nation that crosses the threshold first, then converts that intelligence into a closed system of research, engineering, manufacturing, and deployment, does not win one technology cycle.

It gains an engine for generating advantages.

America must build that engine first.

The Last (Solo) Invention

AI will not be the last thing invented.

It may be the last general-purpose invention that human beings must discover primarily through unscaled human cognition. After that point, invention itself becomes a joint machine-human process operating at a speed and breadth no purely biological institution can match. The “breadth” part is the MOST important part, in my opinion. We are navigating more search space as a society than ever before… and we are 10x increasing it at accelerating intervals.

Think about how things worked before AI: scientific and technological progress was constrained by human minds moving through a long chain of work. Someone read the papers, spotted the gap, formed the hypothesis, wrote the code, designed the experiment, interpreted the result, and persuaded an institution to fund the next attempt.

That chain was intelligent but it was also painfully slow.

AI can enter every link at once.

It can search the literature, compare findings across disciplines, generate candidate explanations, build simulations, rank experiments, control robotic equipment, analyze results, record failure, and decide what should be tested next.

The scientific method becomes a reinforcing workflow:

Observe. Hypothesize. Model. Test. Measure. Update. Repeat.

The machine does not repeal reality. The material still has to work. The drug still has to survive a trial. The reactor still has to produce more useful energy than it consumes. Nature retains the final vote.

But the number of questions we can afford to ask nature explodes.

That changes the nature of the world we live in.

Intelligence Is Upstream of Everything

Energy is upstream of the modern economy. Without energy, nothing moves.

Intelligence is upstream of our ability to improve the energy system.

The same is true across the entire technological stack. Better intelligence can design better chips. Better chips can run better intelligence. Better intelligence can discover new materials. New materials can improve batteries, turbines, reactors, transmission lines, sensors, robots, and fabrication equipment. Those systems can produce more energy and compute, which can support more intelligence.

This is not a collection of independent curves.

It is a network of curves pulling on one another.

Materials improve chips. Chips improve intelligence. Intelligence improves materials.

Energy improves compute. Compute improves science. Science improves energy.

Robotics improves experiments. Experiments improve models. Models improve robotics.

Photonics can move data with less heat. Quantum systems may open scientific state spaces that classical machines cannot efficiently search. Metamaterials may change what we can build in communications, sensing, energy, medicine, and defense. AI will attack all of these domains while learning from all of them.

This is why arguments about whether AI is “just software” miss the point. The model may begin in a data center, but its economic destination is the physical world. It will design matter, direct machines, allocate energy, operate laboratories, and reorganize factories.

AI is becoming the intelligence layer of our civilization.

Whoever owns the strongest version of that layer can improve every layer below it.

Scientific Time Dilation

In Scientific Time Dilation: When Progress Starts Compounding, I explained that AI is changing the clock speed of discovery.

Human beings experience one day.

Machine systems can run thousands of attempts inside it.

That does not mean one day suddenly contains a thousand days of validated science. Experiments still take time. Hypotheses will be wrong. Models can produce elegant nonsense. Data can be contaminated.

But useful progress comes from search, selection, feedback, and memory. AI expands the search. Automated evaluation improves selection. Instruments and robots create feedback. Persistent systems remember what worked and what failed.

The loop learns how to loop.

This creates scientific time dilation. A research institution can explore more of the possibility space in a calendar year than a human-only institution could explore in a decade. As the systems improve, the gap widens. Eventually, one nation may be operating at a radically different technological clock speed from another.

The people in both countries will still wake up on Monday and go to sleep on Monday night.

But one country may run a century of machine research between breakfast and dinner.

It’s a lot like the movie Interstellar which explored time dilation, and the different speeds that time moves due to gravity.

AI is creating time dilation.

That is the strategic danger.

We are used to thinking about national competition as a race along the same road. One nation is three miles ahead. Another has more factories. Another has cheaper labor. Another spends more on research.

AGI breaks that model.

The winner does not simply move farther down the road. It begins building roads faster than everyone else can travel them.

Why First Matters

There is a popular response to any discussion of an AI race: intelligence will diffuse, competitors will catch up, and no one can hold the lead forever

Useful intelligence will spread. Models will be compressed, distilled, copied, adapted, and run locally. Open ecosystems will take expensive frontier capabilities and make them cheaper, smaller, and more available. In When Frontier Intelligence Comes Home, I argued that the frontier will not remain trapped behind one company’s API. Owned intelligence will matter. Local intelligence will matter. Sovereignty at the edge will matter.

But diffusion does not make the race irrelevant.

It raises the value of reaching the frontier first.

The frontier system is not only a product that can later be copied. It is an active research asset. If it can help improve algorithms, design chips, optimize data centers, automate AI research, strengthen cyber defenses, search materials, and accelerate manufacturing, then the first capable system can begin working on its own successor while competitors are still trying to reproduce the current generation.

The lead can compound.

First also matters because technical standards become institutional facts. The leading stack attracts capital, developers, suppliers, customers, researchers, allies, and data. Its interfaces become defaults. Its chips shape factories. Its cloud becomes infrastructure.

Its assumptions spread while slower competitors are still debating them.

The country that leads the frontier also gets the first chance to diffuse that frontier through its economy.

That last step is the whole ballgame. A brilliant model locked inside a research lab is not national power. National power appears when intelligence reaches factories, hospitals, universities, military units, utilities, logistics networks, small businesses, and individual operators.

The race is not won by a civilization that learns faster.

The Cost of Intelligence Is Collapsing

Frontier intelligence remains expensive to create. Adequate intelligence is becoming astonishingly cheap to use.

In The Cost of Intelligence Is Collapsing we unpacked Epoch AI’s finding that the lowest listed API price required to reach a fixed GPT-4-level benchmark threshold fell 214-fold in less than two years.

That number does not mean all intelligence became 214 times cheaper. A token is not a completed task. A benchmark is not judgment. Reliable outcomes require context, tools, memory, permissions, evaluation, security, and often human review.

The direction still changes the economy.

When a useful input becomes cheap, we do not use the same amount and pocket the savings. We consume more of it. Cheaper lighting produced more light. Cheaper bandwidth produced more data. Cheaper compute produced more computation.

Cheaper intelligence will produce more cognition.

Agents will read every document instead of a sample. They will test ten designs instead of one. They will monitor continuously instead of waiting for a quarterly review. They will retry, compare, challenge, reconcile, and escalate. Scientific systems will screen millions of candidates before spending physical resources on the most promising few.

Token price falls as total token use explodes.

That is why the collapse in the cost of intelligence does not reduce the need for infrastructure. It increases it. Every task that becomes economical creates demand for more inference. Every better model creates new applications. Every new application creates more work for the machines.

Abundant intelligence is an energy problem.

It is also a compute problem.

And it is a national power problem.

Physics Sends the Bill

The digital world trained America to believe scale could happen without building much of anything.

Ship the software. Add servers. Rent the cloud. Grow users. Let the physical layer disappear behind an interface.

AI brings the physical layer back with a vengeance.

Training requires chips, high-bandwidth memory, networking, storage, cooling, land, water, transformers, substations, transmission, and enormous amounts of power. Inference at national scale requires even more deployed capacity. Robotics requires motors, sensors, batteries, actuators, rare materials, factories, and supply chains. Autonomous science requires instruments, laboratories, fabrication, and high-throughput testing.

Ironically, the cloud is made of concrete while intelligence is made of electricity.

The U.S. Department of Energy reported that American data centers consumed about 176 terawatt-hours of electricity in 2023, or 4.4% of total U.S. electricity use. It estimated that demand could reach 325 to 580 terawatt-hours by 2028, between 6.7% and 12% of national electricity use. A later DOE resource placed the 2030 central estimate at 11.8%, with a wide range around it.

Those forecasts may be high. They may be low. Efficiency will improve. Algorithms will improve. Chips will deliver more intelligence per watt.

Then we will spend the efficiency on more intelligence.

America cannot win an intelligence race with an electricity system designed for flat demand. We cannot announce trillion-dollar compute ambitions and then ask data centers to wait seven years for interconnection. We can’t build the most advanced chips in the world while depending on slow permitting, scarce transformers, brittle fuel supply chains, and transmission lines no one can approve.

The race leaves the screen and enters the grid.

Manhattan Project One: Energy

America needs an Energy Manhattan Project.

Not a slogan.

Not a subsidy attached to the same process that made the old system slow.

A national mission to expand reliable power at a speed the country has never attempted.

The objective is simple: make the United States the largest, most resilient, most flexible energy platform in the world.

That means generation. It means transmission. It means storage. It means fuel. It means factories capable of making the equipment. It means a workforce capable of installing and maintaining it. It means an interconnection process measured in months instead of eras.

America should build across the stack. Electrons to atoms and back again.

Keep nuclear plants running. Restart viable reactors. Standardize and deploy advanced nuclear designs. Expand natural gas where it adds firm capacity and resilience. Build geothermal where new drilling methods make it practical. Add hydro, solar, and other resources where they make economic and geographic sense. Accelerate grid-scale batteries and long-duration storage. Expand transmission. Manufacture transformers and turbines here. Harden the grid against attack, weather, and equipment failure.

Different regions will use different mixes.

The national standard should be reliable output, speed, cost, security, and the ability to scale.

The Department of Energy now points toward expanding American nuclear capacity from roughly 100 gigawatts to 400 gigawatts by 2050. That is the right scale of ambition. It must become the beginning of the buildout, not a distant number in a press release.

We should aim to 100x this number by 2075.

Batteries deserve special attention because abundant generation without flexible storage wastes energy and weakens the system. Better batteries improve data centers, the grid, vehicles, drones, robots, military logistics, and remote infrastructure at once. That makes battery science a perfect example of the flywheel: AI can search chemistries and optimize manufacturing; better storage can then support more reliable compute; more compute can run better AI.

Energy is not adjacent to AI policy.

Energy is AI policy expressed in physics.

We can’t execute the other two Manhattan Projects without juice.

Manhattan Project Two: Compute

America needs a Compute Manhattan Project. I’ve been geeking out about this for 10+ years as an AI guy and a bitcoiner.

The objective is not to possess a pile of GPUs. It is to own the capacity to turn energy into useful intelligence at a scale no rival can match.

Compute is a system.

It begins with chip design, semiconductor tools, fabrication, advanced packaging, memory, substrates, networking, optics, cooling, power electronics, and the minerals and chemicals that feed them. It continues through data-center construction, software compilers, inference engines, cluster scheduling, security, maintenance, and the ability to keep the system operating under pressure.

A missing layer can slow the whole machine.

The most advanced accelerator is useless if it cannot get memory. A rack is useless if it cannot get a transformer. A data center is useless if it cannot connect to power. A training cluster is strategically fragile if critical components can be cut off by an adversary.

America should aim to have the world’s largest secure compute base across three levels.

First, we need frontier clusters capable of training the strongest models and running experiments at the edge of what is possible.

Second, we need abundant national inference capacity so companies, universities, laboratories, agencies, and startups can deploy intelligence rather than merely read about it.

Third, we need a deep local and open ecosystem so useful intelligence can move outward into homes, factories, schools, vehicles, robots, and private systems. Frontier concentration and broad diffusion are not opposites. The frontier discovers. The ecosystem spreads and compounds the discovery.

The America’s AI Action Plan already recognizes that AI infrastructure means semiconductor fabs, data centers, and new energy generation. Good. Now the country needs execution discipline equal to the diagnosis.

We should know how much secure training compute exists, how quickly it is growing, where supply-chain choke points sit, how long new capacity takes to connect, and what must be manufactured domestically or sourced from trusted allies.

What cannot be measured cannot be mobilized.

Compute is the refinery of the intelligence economy.

The nation with the greatest refinery capacity can turn more energy into cognition, more cognition into discovery, and more discovery into productive power.

Onto my favorite of the three Manhattan Projects.

Manhattan Project Three: AI

America needs an AI Manhattan Project.

This is the most obvious mission and the easiest one to misunderstand.

The goal is not to create one government model. It is not to nationalize the frontier labs. It is not to put a giant bureaucracy in charge of deciding which architecture wins.

The goal is to make crossing the frontier, governing it, and deploying it a national priority.

America should mobilize frontier labs, universities, national laboratories, semiconductor companies, cloud providers, energy companies, roboticists, manufacturers, defense institutions, and independent researchers around a shared objective: build the most capable, secure, reliable, and useful AI systems in the world, then wire them into the machinery of American progress.

Model capability matters. So does the system around the model.

Memory matters. Tools matter. Permissions matter. Evaluation matters. Data provenance matters. Cybersecurity matters. Sandboxes matter. Human decision rights matter. A brilliant model inside a weak harness can generate failure at machine scale. A capable model inside a disciplined operating system can become a research institution.

The mission must extend beyond training runs.

Build automated laboratories. Build national scientific datasets with clean provenance and controlled access. Build simulation environments. Build test ranges for robotics. Build AI systems that can work across materials, biology, nuclear engineering, photonics, manufacturing, and quantum research. Build evaluation systems that test not only whether a model gives an impressive answer, but whether it can produce a reliable result through tools over long time horizons.

The United States should become the best place in the world to turn an AI-generated idea into a verified physical or digital capability.

That is how frontier intelligence becomes national power.

Three Projects, One Flywheel

These cannot operate as separate programs.

Energy without compute is unused capacity.

Compute without AI is expensive machinery.

AI without energy and compute is a research demo waiting for someone else to scale it.

The three Manhattan Projects must be designed as one flywheel:

1. More energy supports more compute.

2. More compute supports more intelligence.

3. More intelligence improves energy, chips, algorithms, materials, robotics, and manufacturing.

4. Those gains produce more efficient energy and compute.

5. The loop begins again from a higher base.

This is the machine America is actually building.

The old model of industrial policy divides responsibility by category. Energy lives in one institution. Chips live in another. AI safety lives somewhere else. Scientific funding is separated by discipline. Permitting is spread across layers of government. Workforce policy moves on its own calendar.

The technology does not respect those boxes.

A breakthrough in photonics can reduce networking energy.

Better cooling can raise cluster density.

A new battery can improve grid reliability.

An AI-designed material can change reactor performance.

A robotic manufacturing technique can speed data-center construction.

An algorithmic improvement can deliver more intelligence from the same chip.

Don’t you see? Every field is upstream of every other field now because of AI.

America needs a command structure capable of seeing the whole system, exposing bottlenecks, setting measurable national goals, and forcing coordination without smothering competition. The state should clear barriers, fund public goods, secure critical infrastructure, support basic research, create demand where markets cannot carry early risk, and protect the country.

Private actors should compete ruthlessly inside that mission.

This is not central planning.

It is national direction with market execution.

Markets Need a Mission

American companies are already spending extraordinary sums on AI infrastructure. That investment is real. It is also not a complete national strategy.

Markets optimize around returns visible to the firm. Nations must also prepare for risks and gains that sit outside any single balance sheet.

A utility may not capture the full national value of building excess generation before demand arrives. A semiconductor company may not capture the military value of supply-chain resilience. A frontier lab may not have the incentive to direct scarce compute toward a neglected disease, a new reactor material, or an open scientific instrument. A local community may carry the disruption of new infrastructure while the economic gains flow elsewhere.

The public mission is to align these layers through predictable rules, fast approvals, shared infrastructure, research funding, procurement, workforce development, allied coordination, and benefits that make host communities stronger.

The private mission (national security, et al) will be even more important to America’s future.

If hyperscalers need new power, they should help finance the generation and grid infrastructure required to serve them instead of transferring the bill to households.

Winning Is More Than Arriving First

America can reach AGI first and still fail.

We could build the most powerful intelligence in history, concentrate it inside a few institutions, make the country dependent on systems ordinary people cannot inspect or leave, automate away bargaining power, and distribute the gains through cheaper subscriptions instead of ownership.

That would be technical victory and political failure.

In The Sovereign Singularity, I argued that the defining question is not only how powerful machine intelligence becomes. It is who that intelligence serves.

The American answer should be clear.

1. Win the race.

2. Advance the frontier.

3. Distribute the leverage.

4. Expand human sovereignty.

Winning means American citizens gain more agency from AI than AI gains over them. It means people can build companies with machine labor, own productive assets, carry their data and memory between systems, run useful models privately, choose among providers, and participate in the wealth the machine economy creates.

It means the frontier can centralize where physics demands scale while useful intelligence decentralizes wherever technology allows it.

The largest training clusters may require giant institutions. The capability they create should still reach the student, the scientist, the small manufacturer, the farmer, the founder, the physician, and the family.

Cheap consumption is not enough. America should create more builders and more owners with this technology.

That is not separate from winning the race. It is how a free society remains strong enough to keep winning after the first finish line.

A National Operating System

The most important national metric of the AI age will not be model size.

It will be the speed and quality of the discovery loop.

How quickly can America turn a question into an experiment?

How quickly can it turn an experiment into reliable knowledge?

How quickly can it turn knowledge into a product, factory, medicine, energy system, or fielded capability?

How much does the country learn from every cycle?

The nation that wins will connect universities, companies, national laboratories, markets, factories, and government into a continuously learning system. It will preserve competition while removing dead time. It will give scientists more attempts, engineers faster tools, manufacturers better designs, and operators clearer ways to act.

AI becomes the reasoning layer.

Compute becomes the cognitive factory.

Energy becomes the fuel.

Robotics becomes the hands.

Science becomes the search.

Manufacturing becomes the conversion of intelligence into reality.

Human judgment sets the objective and decides what deserves to survive.

That is a national operating system for progress.

America already possesses many of the necessary pieces: frontier labs, deep capital markets, great universities, national laboratories, energy resources, semiconductor leadership, a culture of risk, skilled immigrants, entrepreneurs, allies, and a giant internal market.

But possessing the pieces is not the same as running the system.

The country must choose to coordinate them.

The Last Race

Every major technology race before this one had another race after it.

• Win the steam engine and the next race is electricity.

• Win electricity and the next race is electronics.

• Win electronics and the next race is computing.

• Win computing and the next race is the internet.

• Win the internet and the next race is artificial intelligence.

But win artificial intelligence and the nature of the sequence changes.

The winner brings scalable intelligence into every race after it. Better intelligence searches for better energy, better compute, better materials, better robots, better drugs, better defenses, and better intelligence.

Each discovery improves the system available to produce the next discovery.

America must build the energy base that can turn electrons into intelligence at scale. We need to build the compute base that can hold the frontier and diffuse useful capability through the economy. At the same time, we need to build the AI systems, scientific loops, evaluations, institutions, and security architecture that can turn intelligence into durable national progress.

Three Manhattan Projects.

Energy. Compute. Intelligence.

Power the machines. Build the machines. Teach the machines to help us build everything else.

Then distribute the leverage so the victory belongs to a free people, not only to the institutions that reached the frontier first.

We don’t know the exact date AGI arrives. We don’t know which architecture crosses the threshold first.

We know enough.

Intelligence is becoming cheaper. Agents are becoming more capable. Science is becoming more automated. The demand for energy and compute is already colliding with the physical limits of the old system. Rivals understand the stakes. The flywheel is beginning to turn.

This is the last race humanity will run at biological speed.

America must run it like the entire future depends on the outcome.

Because every future after it does.

Friends: in addition to the 17% discount for becoming annual paid members, we are excited to announce an additional 10% discount when paying with Bitcoin. Reach out to me, these discounts stack on top of each other!

Thank you for helping us accelerate Life in the Singularity by sharing.

I started Life in the Singularity in May 2023 to track all the accelerating changes in AI/ML, robotics, quantum computing and the rest of the technologies accelerating humanity forward into the future. I’m an investor in over a dozen technology companies and I needed a canvas to unfold and examine all the acceleration and breakthroughs across science and technology.

Our brilliant audience includes engineers and executives, incredible technologists, tons of investors, Fortune-500 board members and thousands of people who want to use technology to maximize the utility in their lives.

To help us continue our growth, would you please engage with this post and share us far and wide?! 🙏