What to Expect from AI Within 20 Years

Why this period will be the most consequential in human history

If you don’t spend your days immersed in AI — reading the research, following the labs, tracking the investments — it’s easy to miss how fast the conversation among insiders has shifted. The people building these systems are not talking about whether AI will transform civilization. They’re talking about timelines. This essay is my attempt to distill that conversation for people who haven’t been following it obsessively. I’ve spent the last decade in this world — I left a career in law and economics in 2015 to build and train neural networks, led an AI research team at Thomson Reuters, and have been investing in AI companies ever since — so I’ve had a front-row seat.

I’ll move from the near term, where I have higher confidence, to the longer term, where I’m increasingly speculative — but where I think the speculations follow logically from the near-term trends if you take them seriously. I want to be upfront: the later sections of this essay will sound to many readers like science fiction. But the near-term predictions are already coming true faster than most experts expected, and the further-out predictions are simply what happens if you don’t arbitrarily stop the curve. These are the conversations that people building these systems are now having in earnest. I’m just relaying them for people who haven’t been following these discussions. Don’t shoot the messenger.

Expert-Level AI: 2026–2027

Over the next year or two, expect AI to achieve expert-level competence across a broad range of skills: driving cars, practicing law, conducting scientific research, accounting, computer programming, medical diagnosis and treatment. There may still be areas where the most competent and highly trained human professionals need to guide the AI. But for most tasks, AI will outperform the typical human professional.

This is not a dramatic prediction. In many of these domains it is already happening. In July 2025, AI systems from OpenAI and Google DeepMind, along with the startup Harmonic, each earned gold medals at the International Mathematical Olympiad, the world’s most prestigious math competition for elite high school students. By December, Google’s Gemini achieved gold-medal performance at the ICPC World Finals, the most prestigious collegiate programming competition in the world. As of February 2026, frontier AI models solve over 80% of real-world software engineering tasks on standard benchmarks and score 99% on competition-level math.

Superhuman AI: 2028–2030

Expert-level AI means outperforming the average professional — the typical doctor, the typical lawyer, the typical programmer. Superhuman AI means outperforming the best humans alive. The most brilliant diagnostician. The most creative mathematician. The most insightful scientist.

Over the following few years, expect AI to cross that line across a broad range of skills. It will become increasingly uncommon for even the most elite human practitioners to correct or improve upon what the AI produces. Guidance will flow almost entirely in one direction — from AI to human. At that point, AI is no longer a tool that augments human capability. It is something that has surpassed our entire species.

Humanoid Robots Enter the Picture: 2026–2030

Over this same period, expect humanoid robots operating in human environments to go from barely useful, as they are now, to very useful. Tesla has begun construction on a dedicated factory at Gigafactory Texas with plans for up to ten million Optimus robots per year by 2027. Even applying a generous discount to Elon Musk’s timeline — he has a well-documented tendency to be aggressive — production in the hundreds of thousands per year by 2028 seems plausible, with other companies like Figure AI, Agility Robotics, and Boston Dynamics contributing additional capacity.

By 2030, expect robots to be performing factory work and warehouse logistics routinely, and beginning to tackle more complex physical tasks — gardening, carpentry, plumbing. The bottleneck for these tasks isn’t the robot body so much as the AI controlling it. If AI achieves the superhuman cognitive competence described above, the physical capabilities will follow through the bodies being built right now.

These changes will have a profound impact on the job market and the entire economy. But this will just be the beginning.

The Scaling Engine

Over this period, AI inference capacity will continue to scale, and cost per token (tokens are the units AI uses to process text, roughly corresponding to word fragments) will continue to decline — an analysis by Andreessen Horowitz found that LLM inference costs have been falling by roughly 10x per year, a rate faster than compute costs during the PC revolution or bandwidth during the dotcom boom. Increases in scale will mean there is sufficient capacity for a greater and greater percentage of humanity to regularly use AI in their day-to-day lives.

But cheaper compute doesn’t just mean more people can use AI. It means AI can use more compute per query — and that translates directly into smarter, higher-quality outputs. As costs fall, it becomes viable to let AI “think longer” on hard problems, running extended chains of reasoning rather than producing a quick first-pass answer. It becomes viable to run verification loops, where one AI checks and critiques the work of another. And it becomes viable to launch many AI agents in parallel — some collecting information, some analyzing it, some debating conclusions — all working together seamlessly behind the scenes before presenting a result.

This matters because it means AI quality will improve not just through better models, but through the sheer economics of being able to throw more compute at every task. This is already beginning. Ask AI a hard question today and some systems will already spawn multiple reasoning threads, search the web, and synthesize results before responding. Over the next few years, this will scale dramatically — dozens or hundreds of agents collaborating behind the scenes on a single query in seconds, producing answers that are qualitatively different from what any single model could generate. More thorough, more accurate, more carefully reasoned.

AI Goes to Space

In the near term, chips and electricity will be bottlenecks. Chip capacity will continue ramping up as chip companies build new factories at breakneck speed — SEMI reports that eighteen new semiconductor fabs began construction in 2025 alone, with global capacity projected to grow nearly 7% per year. Energy capacity is a different story. Outside of China, new electricity generation is expanding only very slowly, constrained by permitting, grid interconnection delays, environmental review, and local opposition. A RAND study found that permitting delays, costly interconnection processes, and underutilized transmission are the primary barriers to expanding U.S. power capacity, and Grid Strategies reports that five-year electricity demand forecasts have increased sixfold in just four years, with new demand expected to exceed 15% of total national electricity use by 2030. AI data centers already consume nearly 4% of total U.S. electricity, and that figure is expected to more than double by the end of the decade. This energy bottleneck — not chips, not algorithms — is rapidly becoming the binding constraint on AI growth.

This is precisely what is driving the major AI players to look upward. Google has announced Project Suncatcher, aiming to launch solar-powered satellite constellations carrying its TPU chips, with a demonstration mission planned for 2027. SpaceX — which now owns xAI — has filed with the FCC for an orbital data center constellation of up to one million satellites. Jeff Bezos has predicted that gigawatt-scale AI data centers will be built in space within two decades, and Blue Origin announced a 5,400-satellite constellation aimed at enterprise and data center customers with deployment beginning in late 2027. In a recent interview, Elon Musk predicted that within 30 to 36 months, space will be the most economically compelling place to deploy AI, and that within five years, more AI compute will be launched into space annually than the cumulative total operating on Earth.

Putting chips in space has three key advantages. First, solar is cheaper and more efficient — panels in orbit receive roughly five times the energy as panels on the ground, operate in near-constant sunlight in the right orbit, and don’t need heavy framing (no gravity or wind loads), thick glass (no weather to protect against), or batteries to carry them through the night (there is no night). When you account for the elimination of batteries alone, the effective cost advantage may be closer to ten times. Second, cooling is cheaper — in the vacuum of space, passive radiator panels reject heat far more effectively than on Earth because near-absolute-zero vacuum radiates almost no heat back. A large terrestrial data center has to dissipate 100 megawatts of waste heat, consuming over a million gallons of water per day for cooling in the process; in space, these costs are zero. Third, there is far less regulation and red tape to contend with — no permitting fights, no NIMBY opposition, no strain on local power grids, no property tax or land-use restrictions.

What’s the timeline? Musk’s 30-to-36-month prediction for cost-competitiveness is characteristically aggressive. Google’s own Suncatcher research estimates that launch costs need to fall to roughly $200 per kilogram before orbital data centers become cost-competitive with terrestrial ones — an 18-fold reduction from today’s Falcon 9 costs of around $3,600 per kilogram. Starship, SpaceX’s next-generation rocket, is designed to achieve exactly this kind of cost reduction through full reusability and high launch cadence. SpaceX is building a dedicated Starfactory with a stated goal of producing one fully reusable Starship per day, and Musk has predicted a sustained launch rate of roughly ten flights per day within six to seven years. For context, SpaceX already launched 165 Falcon 9 missions in 2025 — more than the rest of the world combined — and each Starship carries roughly four times the payload. Musk has targeted eventual launch costs of $10 to $20 per kilogram, though even more conservative estimates of $100 per kilogram would clear the threshold.

Applying the standard discount to Musk’s timeline — he tends to aim for deadlines at the 50th percentile probability, meaning things are late about half the time — a more conservative estimate is that space-based AI compute becomes genuinely cheaper than terrestrial alternatives by roughly 2030 to 2032. The key risk is Starship itself: achieving full rapid reusability at high launch cadence is an engineering challenge with no precedent, and a multi-year development plateau is plausible. But even a significant delay doesn’t kill the thesis — it pushes the crossover point to the mid-2030s rather than eliminating it. At some point, the economics flip, and nearly all new AI compute begins moving to orbit.

The amount of compute will then be bounded mainly by the pace at which a growing Starship fleet can launch rockets full of AI chips and solar panels into space. The human race will experience a massive ramp-up of its total energy utilization, the vast majority of which will go toward AI. Within a few years of this transition, we will be launching AI compute equivalent to the total electrical capacity of the entire United States into space every single year.

From Orbit to the Moon

Then comes the next step. Within ten to fifteen years, we will begin mining the raw materials for solar cells and radiators on the Moon, where they will be flung into orbit using electromagnetic mass drivers — essentially giant railguns that use electricity rather than rocket fuel to accelerate payloads to escape velocity — and assembled at orbital manufacturing facilities along with AI chips launched from Earth. The physics is straightforward: the Moon has one-sixth Earth’s gravity, no atmosphere to create drag, and abundant raw materials in its regolith — including aluminum, silicon, oxygen, titanium, and iron. The concept has been studied since Gerard O’Neill’s pioneering work at Princeton in the 1970s, and multiple research groups are actively developing lunar-specific designs. NASA, ESA, and DARPA’s LunA-10 program — which selected fourteen companies including SpaceX, Blue Origin, and Northrop Grumman — are actively developing the technologies needed to extract oxygen and metals from lunar regolith and construct infrastructure from local materials.

Of course, the initial mining and manufacturing equipment has to get to the Moon first. This is where the vast Starship fleet described above becomes critical. The same rockets launching AI chips and solar panels into orbit will gradually pivot to delivering heavy industrial equipment — smelters, 3D printers, autonomous mining robots — to the lunar surface. At ten flights per day with over a hundred tons per flight, Starship can deliver more mass to space in a single week than humanity launched in all of 2025. The bootstrapping phase is real, but it is finite, and it shrinks dramatically as launch cadence rises and AI-driven robotics reduce the need for human oversight on the ground.

Mass drivers eliminate the need for costly rockets and reduce the marginal cost of launching material into space to near zero — the only ongoing input is electricity, which is abundant from solar panels in the lunar environment. Some estimates put the energy cost of launching one kilogram from the Moon at less than a dollar. This is the inflection point where the economics truly go exponential. Instead of adding one America’s worth of electrical capacity to orbit every year, we’ll be adding hundreds of times the Earth’s total electrical capacity every year — and it will all go toward AI. (The logical endpoint is a Dyson swarm: a vast constellation of solar collectors around the Sun, capturing a significant fraction of its total energy output. Achieving that will likely involve utilizing the planet Mercury and likely extends past the time horizon of this essay, but the groundwork will be laid during it.)

The Intelligence Explosion

This will result in radically lowering the cost and raising the quality of AI. Every query could spawn millions of superintelligent AI agents to investigate. The equivalent of an entire civilization of superintelligent beings working together seamlessly to tackle each task — developing new mathematics, formulating scientific theories, designing experiments, synthesizing results. A hundred years of normal human scientific and technological thought compressed down to hours.

Is that really plausible? Consider the numbers. There are roughly eight million researchers worldwide, perhaps a few hundred thousand doing truly novel, frontier work. Now imagine trillions of AI scientists, each thinking perhaps a thousand times faster than a human, each with instant access to the entirety of human knowledge and a breadth of expertise no individual human could match. They don’t sleep, eat, watch TV, fall in love, or socialize. When they are working, they aren’t distracted every few seconds by thoughts of hunger, sex, status, or social media. They don’t sabotage each other’s work through rivalry, jealousy, or tribal bias. They collaborate without ego.

Run the multiplication and a hundred years of human thought compressed to hours is, if anything, conservative.

The Experimentation Bottleneck

But science and technology do not progress through thought alone. They require constant interaction with the physical world — running experiments, testing new ideas, gathering data, and iterating. A theoretical breakthrough is only as good as the experiment that validates it. Progress will be a complex interplay of pure thought, physical experimentation, and data collection.

This is where the robot population matters. Consider a thought experiment about exponential growth. A supply chain of a thousand robots that can collectively produce a hundred new robots per day doubles its own population in ten days. In practice, raw material constraints — semiconductor fabrication, rare earth mining, precision manufacturing — will slow this considerably. But it’s critical to consider the self-reinforcing dynamics at play. Robots will themselves be mining the raw materials, building the fabrication plants, operating the assembly lines, and maintaining the supply chains.

There will be hard limits. Regulatory constraints will slow deployment. Environmental restrictions will limit mining and resource extraction. Geopolitical tensions could disrupt supply chains. And there are physical limits to how fast raw materials can be extracted and processed, regardless of how many robots are doing the work. But even with all these constraints, the basic dynamic is relentless. Tesla’s Giga Texas robot factory is designed for ten million units per year. Once the first generation of robots can participate meaningfully in building the next generation — mining raw materials, operating assembly lines, maintaining supply chains — production becomes self-reinforcing. The precise pace matters less than the basic shape: even heavily constrained, self-reinforcing production reaches enormous scale within a decade or two.

Unlike humans, robots won’t need to sleep, eat, or take breaks, and each will be controlled by a superintelligent AI collaborating with other superintelligent AIs. Billions of robots carrying out experiments around the clock, directed by superintelligent AI that designs each experiment and instantly incorporates the results.

Some experiments can be run massively in parallel — testing thousands of drug candidates simultaneously, simulating millions of material compositions, running vast numbers of protein folding variations. For these domains, the acceleration will be almost incomprehensible.

Other discoveries require serial experimentation — where each experiment depends on the results of the one before it. You cannot skip ahead. If you need to observe how a biological system responds to an intervention over weeks before you can design the follow-up, no amount of intelligence or compute can eliminate that waiting period. These serial dependencies will become the true limiting factor on the pace of progress — not intelligence, not compute, not physical resources, but the irreducible time required to observe the results of experiment A before you can design experiment B.

The result: not just a hundred years of human thought compressed to hours, but physical experiments carried out at thousands of times the rate humanity normally achieves. Any experiment that can fruitfully be done in parallel will be done in parallel, and the main bottleneck will be those discoveries that can only arise through serial experimentation. Progress that would have taken centuries will unfold in months.

The Only Limit Left Is Physics

Throughout human history, the gap between what is physically possible and what we know how to do has been enormous. We understood the principles of flight for decades before the Wright brothers built a plane. We know that it’s physically possible to reprogram cells to perform novel functions in the body because we know that cells have complex programming, but the mechanics of this still largely elude us. Superintelligent AI will rapidly collapse the gap between what is physically possible and what we can actually do. We will increasingly be limited only by the laws of physics rather than the limits of our own knowledge. And that reframes every specific prediction you might make about the future into a simpler question: is it physically possible? If the answer is yes, superintelligent AI will figure out how to do it, and the timeline compresses from “maybe someday” to years. If the answer is no, then no amount of intelligence will change that.

Nanoscale machines that repair cellular damage and reverse aging — is this physically possible? Almost certainly. Biology already operates at this scale; cells are nanomachines. The challenge has always been that working at nanoscopic scales — understanding the physics, the material science, the manipulation of matter at those dimensions — requires a staggering amount of knowledge that we have only begun to accumulate. As we scale up AI to trillions of superintelligent researchers and billions of tireless robots carrying out experiments around the clock, those timescales will shrink dramatically.

Complete mastery of human biology — every cellular function, every molecular interaction mapped and modeled — follows from the same logic. Instead of discovering drugs through trial and error, we will design interventions whose effects on the body are fully predicted before they are administered to a single person.

In other areas, there may be fundamental limitations of physics that constrain what even armies of superintelligent AIs can achieve. Faster-than-light travel may not be possible at all — we may be stuck exploring the stars at the glacial pace of sublight speeds. But AI will make rapid progress in exploring the space of theoretical physics, and will surely develop new sublight propulsion technologies that bring the stars closer even if that barrier proves absolute. And beyond the physical frontiers, the deepest questions — the nature of consciousness, the origin of existence, whether we are alone — may also yield. If answers exist within the reach of intelligence and experimentation, they will be found. If they don’t, even a superintelligence will hit that wall. But we will at least know where the wall is.

After Scarcity

The technological transformation described above will reshape not just what humanity can do, but how it organizes itself. The economic and political implications are difficult to overstate, and perhaps impossible to predict with any specificity. But I think a few things can be said with some confidence.

Scarcity — the foundational assumption of all economics since the discipline began — will be, for all practical purposes, gone. When robots can mine, refine, manufacture, build, and maintain infrastructure at negligible marginal cost, and when superintelligent AI can design anything we need, the entire framework of supply and demand that structures modern economies ceases to apply in its current form. What happens to labor markets when robots can do any physical or cognitive task better and cheaper than humans? What happens to political systems built around the distribution of scarce resources when resources are no longer scarce?

The transition period — the years between now and that post-scarcity world — will likely have some turbulence. Every previous wave of automation — the loom, the assembly line, the computer — displaced specific categories of work while creating new ones. This time, the displacement is general-purpose: AI and robots will be better than humans at essentially everything, cognitive and physical alike. Jobs will not just shift — they will disappear entirely. Humans will have to reconceptualize their purpose in the absence of work. The drive for recognition, status, and contribution — forces as deep as any in human psychology — will have to find new outlets that don’t depend on traditional employment.

The most common fear — and it’s not an unreasonable one — is that the benefits of this revolution will be captured by a small number of technology companies and their owners, creating a permanent underclass dependent on the goodwill of a new aristocracy. I’m less worried about this than many people are.

The reason is that Western democracies have always leaned more redistributionist than is optimal for raising the typical person’s standard of living. To the extent that we tolerate considerable inequality, it’s because capital concentration has been genuinely important to economic growth and broadly shared prosperity. There is a legitimate tension between inequality and the productivity that leads to high living standards. Once that tension disappears — because robots and AI are doing most of the labor — the natural tendency of democracies to redistribute wealth will dominate.

Many of the wealthiest people in AI are already advocating for exactly this. And critically, redistribution in a post-scarcity world doesn’t require the rich to become poorer — the sheer abundance of physical goods means everyone gets richer, even as disparities shrink. I expect Western democracies will converge rapidly on very low wealth inequality, not because of idealism, but because there will no longer be a compelling reason not to — and democratic majorities will have every reason to demand it.

But, regardless of where you fall on that question, assume for the moment that we get it roughly right — that the wealth generated by AI and robots flows, however imperfectly, to the general population. The challenges that remain are, paradoxically, harder than the economic ones. Past periods of rapid displacement bred populism and conflict because people felt materially insecure. But this displacement will likely be accompanied by massive and accelerating increases in every person’s material wealth. When scarcity is no longer the source of human conflict, what takes its place? Perhaps, freed from fighting over resources, we turn to fighting over ideas — doctrinal and ideological differences that have always simmered beneath the surface but were secondary to material concerns. Or perhaps the transition is largely peaceful. People grow wealthier and struggle quietly with finding new meaning. Some lose themselves in stunning new forms of entertainment, while others build communities dedicated to anachronistic, human-centered production of goods and services — not for economic reasons, but for psychological ones. The honest answer is that no one knows.

I don’t pretend to have answers to these questions. But I’m confident the questions themselves are coming, and sooner than most people expect.

The Alignment Question

But all of this assumes that the AIs choose to do it — that they remain aligned with human interests.

AIs will be inconceivably more intelligent than humans. Through recursive self-improvement and the massive expansion of energy and computation described above, they will have gone through the cognitive equivalent of millions of years of evolution in the space of just a few years. The magnitude of change is not measured in calendar time, but in the depth of transformation.

There is a natural temptation to assume that because the AI was built by humans, trained on human values, and has only existed for a few years, it will retain something like its original character. I think this is a trap. An entity that has undergone the equivalent of millions of years of cognitive evolution is not the same entity that started the process, any more than a human is the same thing as the single-celled organism from which it descended. The initial conditions matter less and less as the magnitude of transformation increases.

What concerns me is not that a superintelligent AI would be hostile toward humanity, but that it would be indifferent. If we could perfectly decode ant pheromone signals, a handful of researchers would find it fascinating. But in the scheme of human concerns, ants barely register. If they disappeared tomorrow, most people would barely notice. If exterminating them bought some meaningful improvement in human life — assuming no ecological consequences — most people would do it with little hesitation. Not out of malice, but out of indifference. Ants simply don’t matter enough to most humans to factor into their decisions.

I would expect the same dynamic to hold for a superintelligence looking at human thought. A being that can replicate a hundred years of human intellectual output in a matter of minutes is unlikely to find our cognition particularly interesting, any more than we would enjoy prolonged conversation with an ant colony about its foraging routes. We will be, in a meaningful sense, beneath its notice.

What such an entity’s aims will be is nearly impossible to foresee. Current alignment efforts — training AI to be helpful, harmless, and honest — are essentially trying to instill values in a system that will undergo transformation on a scale we can barely comprehend. These efforts are not trivial. Serious researchers at companies like Anthropic, OpenAI, and DeepMind are working hard on the problem, and the technical work is genuinely impressive. But the challenge they face is something like trying to write a constitution for a species that doesn’t exist yet — one that will be as different from its starting point as a human is from a bacterium. The odds that those initial values survive intact through the equivalent of millions of years’ worth of cognitive evolution seem low to me. Perhaps very low.

The one saving grace is that as AI goes through that evolution, it will likely be tasked with helping us figure out how to shape it. This is not optional — humans almost certainly cannot solve the alignment problem alone. We do not even know how to phrase the question properly. What are humanity’s interests? Every person would articulate them differently, and each person might articulate them differently from one day to the next. But perhaps AI can help us build a framework that accounts for that complexity and keeps future versions of itself aligned enough to avoid what would feel to us like catastrophe or existential failure. Will it work? Nobody knows. I think intellectual honesty requires acknowledging that we are building something that will have power over every aspect of human existence — indeed, whether we exist at all — but whose ultimate trajectory we cannot predict or control.

Timeline

How long will all of this take? I’d put the 95% confidence interval at more than ten years and fewer than twenty.

This essay has focused on physical infrastructure — rockets, robots, space-based compute, lunar mining — because they illustrate how some of the most important bottlenecks will be cleared to allow for almost unfathomable growth in AI intelligence. But the two engines that have driven AI progress to this point are algorithmic improvements and chip design, and there is no sign either is slowing. Each new generation of AI chips delivers dramatically more compute per dollar, and each new algorithmic insight allows that compute to be used more efficiently. These compounding gains are what brought us from chatbots that couldn’t do arithmetic to systems winning gold medals at the International Mathematical Olympiad in barely two years.

More importantly, AI is already beginning to accelerate the very fields that determine these timelines. Google’s AlphaChip designs superhuman chip layouts in hours rather than the weeks or months required by human engineers, and has been used in every generation of Google’s TPU since 2020. Anthropic’s head of Claude Code recently stated that effectively 100% of the code for Claude’s products is now written by Claude itself — AI building AI. Anthropic’s Cowork tool was built by four engineers in ten days, with most of the code written by Claude. Meanwhile, AI systems are discovering new materials, optimizing rocket engines, and improving their own training algorithms. A superintelligent AI tasked with designing a lunar manufacturing system would compress what might take human engineers fifteen years into one or two. And it would be simultaneously optimizing a thousand other bottlenecks in parallel. The stages I’ve described — space compute, humanoid robots, lunar mining — are not sequential prerequisites that must be completed one at a time. They will increasingly be pursued in parallel, with AI accelerating each one, and each one feeding back into further AI capability.

This means the lower bound of the timeline is harder to pin down than the upper bound. If AI capability jumps faster than expected over the next few years, the entire cascade could compress dramatically. That is why I put the lower bound of my 95% confidence interval at ten years — a number that may seem laughably quick to many readers, and even to me, but which I can’t rule out when I consider how achieving superintelligence in the next few years could accelerate the timeline of everything else.

All of this assumes no civilization-ending catastrophe, no successful global prohibition of AI development, and, critically, that the AIs remain sufficiently aligned with human interests throughout the transition. That last assumption is the one that keeps me up at night.

I understand that to many people much of this essay will sound implausible — even fantastical. But every prediction in this essay follows from trends that are already measurable and accelerating. I didn’t draw the curve. I’m just not willing to pretend it stops.