AI Doesn't Plateau — It Compounds. Which Wall Falls Next?
AI is the first technology that improves the machine that makes technology — so it doesn't plateau, it compounds. The only question left is which wall falls next: compute, power, or the physical world.
Executive summary
The most important thing to understand about artificial intelligence is not what it can do this year. It is that we may be living through the opening of a compressed century — a stretch of a decade or two in which the scientific and economic progress that would otherwise take a hundred years arrives instead in a handful. That is not a slogan; it is the median reading of serious long-run growth models and the stated working assumption of the people running the largest AI labs. If it holds even partway, the 2030s and 2040s will feel less like faster versions of today and more like a different world stacked on top of it.
Where this is heading turns on a single mechanism that no previous technology had: AI is the first invention that improves the machine that makes inventions. The steam engine could not design a better steam engine; a transistor could not run the research program that produced the next transistor. An AI system can now do a growing share of the work that builds the next AI system — and that is why the honest question is no longer whether progress plateaus but which physical constraint gives way next. Three walls stand between here and the compressed century, and they fall in a knowable order: compute (the raw computation available), already the fastest-falling and now being handed to AI itself to push; power (the electricity to run it), the binding wall of the late 2020s, dated to fall through a nuclear and fusion build-out landing between 2027 and the early 2030s; and the physical world (robots that can act in it), the last and highest wall, which is cracking right now as the recipe that solved language gets pointed at atoms.
One promise upfront, because it is the whole point of this show: everything past the present moment is a projection, not a prophecy. The knowability horizon in AI is brutally short — the field turns over in weeks, and credible experts disagree by years about the same milestone. What follows is grounded in forecasters' curves, lab leaders' roadmaps, and measured trends, each named so you can weigh it yourself. Treat it as the best available map of a fast-moving front, drawn in pencil.
Why this is the question that matters
Keeping pace with AI by reading each week's model release is a losing game — it tells you where the front line sits today and nothing about where it is going. The more useful vantage point is the one economists take when they zoom all the way out. Fit a curve to the whole arc of human material progress and you do not get a straight line; you get a curve that keeps bending upward, with each civilisation-scale leap arriving faster than the last. Understanding this moment in the AI revolution means understanding that curve — what bends it, what stalls it, and why AI is the first technology with a plausible claim to bending it from the inside. That is the thread this report follows: not what AI is, but what topples first to get us where it is heading.
Innovation is a curve that bends — and AI is the first technology that bends it from within
Zoom out far enough and human progress has a shape. For most of our species' existence, a single civilisation-defining advance — agriculture, bronze, the written word — took millennia to arrive and millennia more to spread. The Agricultural Revolution compressed that clock to centuries. The Scientific and Industrial Revolutions compressed it again, to decades. The computer compressed it once more, to years. Economist David Roodman, in a long-run model built for Open Philanthropy, fit a single equation to the entire trajectory of gross world product and found it is best described not as steady exponential growth but as super-exponential — a curve whose growth rate itself accelerates, echoing a 1960 result by Heinz von Foerster. Extrapolated naively, Roodman's best-fit model implies a growth "explosion" with a median date around 2047. He is explicit that the naive extrapolation overshoots — real growth will not literally run to infinity — but the underlying finding is the point: history's engine has been acceleration, and it has stalled only when a binding constraint held.
That word — binding — is the one to hold onto. A binding bottleneck, or forcing function, is the single constraint that actually caps a system's speed: relieve any other and nothing happens, but relieve that one and everything downstream lurches forward. Real change rarely comes from a smooth rising tide. It comes in step-changes, when a binding constraint falls and unlocks a cascade that was waiting behind it.
What makes AI a candidate for the same civilisational weight as agriculture or industry — but with a multiplier potentially orders of magnitude larger — is recursive self-improvement: an AI system helping to design, train, and improve the next AI system, so that capability feeds back into the process that produces capability. Plainly: the tool is starting to build the tool. This is not yet a runaway loop, and serious people doubt it will be soon. But it is no longer hypothetical — Anthropic disclosed that as of May 2026, its own model Claude authored more than 80% of the code merged into the company's production codebase, up from low single digits in early 2025. The machine is already inside the workshop.
So the useful question is not whether the curve keeps bending. It is which wall gives way next to let it.
Compute is the wall that is already falling — and it has just been handed to AI
Of the three walls, compute — the sheer quantity of computation you can throw at training and running a model — is the one already falling fastest, and it is best understood as accelerated, not unlocked. It was always going to grow; AI has simply put it on a tear. The research group Epoch AI finds that the computation used to train frontier models has grown about 5× per year since 2020, doubling roughly every five months. Underneath that, algorithms get more efficient: the same capability now takes about 3× less compute each year, so that every nine months, better methods contribute the equivalent of a doubling of the hardware budget. Stack the two and, if the trend holds, models trained around 2030 could use on the order of a thousand times more effective compute than today's — the kind of leap that has, each prior time, brought a qualitatively new capability rather than just a better version of the old one.
The cleanest way to see this wall falling is not to count chips but to measure what the models can actually do unsupervised. The evaluation group METR tracks the length of task an AI agent can complete on its own — where "length" means how long the job takes a human expert. That horizon has doubled roughly every seven months for six years, climbing from about four seconds of equivalent work in 2019 to more than sixteen hours by 2026. And the pace is not steady — since 2023 the doubling has compressed to about every four to five months. A capability measured in seconds a few years ago is now measured in a workday.
Here is where compute stops being a story about hardware and becomes the master accelerant behind all three walls. The labs are now pointing AI at the job of AI research itself. OpenAI has stated a public target of an AI "research intern" by September 2026 and a fully autonomous AI researcher — able to run its own research projects — by March 2028, arguing plainly that "AI doing AI research will become the determining factor of the pace of progress within the next few years." That is the recursive-improvement loop moving from essay to roadmap. Where to look for the signal: at one year, whether task horizons cross a full autonomous workweek; at five years, whether a materially larger share of frontier-lab research is machine-generated; at ten years, whether the doubling curve bends further up rather than flattening. This is the episode's one genuine wildcard — we went deep on it in our episode on AI improving itself, number 26 a few weeks back, and it is the hinge on which every horizon below turns.
Power is the binding wall of the late 2020s — and it is dated to fall
For a brief window the constraint on AI was chips. That window has closed; the binding wall now is electricity. The heads of the largest labs have converged on it independently — the GPU crunch eased, and power became the thing standing between ambition and reality. The numbers behind the shift are stark. The International Energy Agency projects electricity demand from data centres roughly doubling from about 485 terawatt-hours in 2025 to around 950 by 2030 — with the AI-specific slice tripling. And supply is not keeping up: Goldman Sachs flags a US power shortfall of about 9.3 gigawatts in 2026, widening to 45 by 2028. By mid-2026, an estimated 75 US data-centre projects worth roughly $130 billion had stalled, held up less by money or chips than by transformers and grid connections with lead times running to four years.
A wall this visible attracts capital, and that is precisely why it is dated to fall rather than permanent. The response is a nuclear build-out of a scale unseen in a generation. The four largest hyperscalers have signed more than a dozen deals for close to ten gigawatts of nuclear capacity. Microsoft is paying to restart the undamaged reactor at Three Mile Island — 835 megawatts under a twenty-year agreement, with a federal grid waiver cleared in June 2026 and full power targeted for 2027. Google has signed the first US corporate deal for a fleet of small modular reactors — factory-built reactors of 50 to 300 megawatts rather than the 1,000-plus of a conventional plant — targeting 500 megawatts from 2030. Meta has contracted for up to 6.6 gigawatts, including next-generation designs from Oklo and TerraPower.
Behind fission sits the longer-shot, higher-ceiling bet: fusion — fusing light atoms to release energy, the process that powers the sun, long "thirty years away" and now attached to real purchase orders. Helion signed the world's first fusion power-purchase agreement, committing 50 megawatts to Microsoft by 2028, and has broken ground in Washington. Commonwealth Fusion Systems, whose SPARC demonstration is more than half-built, has a 200-megawatt agreement with Google for a commercial plant in Virginia in the early 2030s and applied for grid interconnection in April 2026. AI data centres turn out to be fusion's ideal first customer — enormous, steady demand from buyers who will pay a premium for reliable power. Where to look: at one year, whether stalled projects reconnect and the first restarted reactors energise; at five years, whether SMRs ship on schedule and Helion delivers real megawatts; at ten years, whether fusion crosses from demonstration to grid. Unlike compute, this wall falls on the timeline of concrete, not code — which is why it, not the chip, is the pacing constraint of the decade.
The physical world is the last wall — and the recipe that cracked language is now pointed at atoms
The highest wall is the one between thinking and doing. An AI that writes, reasons, and codes still lives entirely in the world of information; to change the physical world it has needed a human pair of hands. Embodied AI — intelligence that perceives and acts through a physical body, a robot — is the wall that, once breached, unlocks a categorically different trajectory, because it lets the compounding curve reach into manufacturing, construction, logistics, and the lab bench itself.
For decades robotics was gated on the body: actuators, balance, battery. That is no longer where the constraint sits. The step-change of the mid-2020s is that the foundation-model recipe — pre-train a large model on vast data, then adapt it — is now being applied to physical action. Nvidia's GR00T, released in a new version at the start of 2026, is a general foundation model a robot-maker can start from instead of training motion from scratch; Google DeepMind's Gemini Robotics and the startup Physical Intelligence's models point the same way. The old obstacle — where do you get enough real-world training data — is being routed around with video pre-training and simulation, with Nvidia reporting that synthetic motion data lifted task-success rates by around 40%. As one 2026 industry survey put it bluntly: hardware is no longer the primary bottleneck; intelligence is. Tesla says it has more than a thousand of its Optimus humanoids working in its own factories, targeting fifty thousand — deployed partly to do the work and partly to gather the data that trains the next generation.
Where to look, because robotics is the wall whose fall is easiest to overclaim: not at a slick demo video but at cost per useful hour and at generalisation — one model doing many unrehearsed tasks in a messy environment, not one rehearsed task on a stage. On the horizon spine: now, supervised pilots on factory floors; at one year, the first fleets earning their keep on narrow work; at five years, general-purpose robots plausibly crossing into commercial usefulness if the intelligence curve holds; at ten years, the wall between AI's mind and the physical world genuinely thin. That last step is the one that would let the compressed century touch atoms and not just bits.
Ride the ladder out and you reach a compressed century — just not everywhere at once
Stack the three falling walls and the thing they unlock is the prize the whole trajectory points at: a dramatic acceleration of science itself. Dario Amodei, Anthropic's chief executive, calls it the "compressed 21st century" — the claim that a large population of capable AI systems could raise the rate of biological discovery tenfold, delivering a century's worth of progress in five to ten years after truly powerful AI arrives. The leading edge of that is already visible, and it earns its place with a Nobel: DeepMind's AlphaFold, which predicted the three-dimensional structure of essentially every known protein, won the 2024 Nobel Prize in Chemistry — the first Nobel for an AI-enabled breakthrough — and its successor now models how molecules interact. AI-designed drugs have reached human trials; Eli Lilly signed a deal worth up to $2.75 billion for compounds from one AI-native discovery firm in March 2026.
But the honest trajectory has a brake built in, and naming it is the difference between a forecast and a fantasy. AI accelerates the parts of science that are search and design — problems where the answer already exists in a space of possibilities and the task is to find it — far faster than the parts that require building and validating something in the physical world. DeepMind's GNoME predicted around 380,000 stable new materials; as of the reporting, outside labs had physically confirmed a few hundred. No AI-discovered drug has yet been approved, because designing a molecule is fast and proving it works in a human body is not. This is the texture of the whole compressed century, and it is worth internalising: the acceleration is genuine but uneven. The fields whose bottleneck is search — protein folding, materials on paper, molecule design, mathematics — leap first, because computation dissolves an informational constraint almost on contact. The fields whose bottleneck is a physical loop — a reactor that must be built, a robot that must be trained, a trial that must run its course in real bodies over real months — move on the timeline of matter, and matter does not care how fast the model thinks.
This is also why quantum computing, so often filed next to AI as the other looming revolution, belongs on a later shelf. The progress is real — Google's Willow chip crossed a genuine threshold, becoming the first where adding qubits makes the error-corrected memory better rather than worse. But a useful, fault-tolerant quantum computer is a different order of problem: estimates put the requirement for cryptographically relevant work near twenty million physical qubits against roughly a hundred deployed today, and Google's own public goal for a useful error-corrected machine sits at 2029. Quantum is a wall that has barely been scratched, not one about to fall — a genuine unlock, but one whose horizon is the 2030s at the earliest, and which matters to AI's trajectory far less than the boring miracle of more classical compute, cheaper power, and robots that work.
The knowability horizon is short — here is what would change the conclusion
Everything above is a projection, and the discipline of this show is to say so plainly rather than in a footnote. AI's "notability horizon" — how far ahead the field is even legible — is measured in weeks, and the experts disagree by years. Anthropic's Jack Clark puts the odds of meaningful recursive self-improvement around 30% by 2027 and 60% by 2028; the AI critic Gary Marcus calls a 2027 timeline implausible and doubts even 2030. A survey of forecasters put only a 20% median probability that the next few years compress six years of progress into two. The right posture is neither the true-believer's certainty nor the sceptic's dismissal, but calibration: a set of grounded projections with the hinges named.
So name the hinges — the ways this trajectory jams. If power stays stuck — if the reactors slip and the transformers never arrive — compute growth throttles regardless of how good the chips get, and the whole ladder slows from its second rung. If the robot-data problem proves deeper than video and simulation can solve, the physical wall holds and the compressed century stays trapped in software. And if algorithmic progress quietly stalls — if the recursive loop turns out to hit diminishing returns before it becomes self-sustaining — the super-exponential curve relaxes back into a merely-fast exponential, extraordinary but not civilisational. Each of these is a live possibility, and each is exactly what to watch. The signal that the optimistic reading is winning is not a single breakthrough; it is all three walls continuing to fall roughly on schedule — task horizons lengthening, reactors energising, robots generalising — at the same time.
Bottom line
AI is the first technology with a credible claim to bending history's growth curve from the inside, because it is the first that improves the process that produces it. Whether that claim pays off is not a question about the technology in the abstract; it is a question about three physical walls and the order in which they fall. Compute is already giving way and is now being handed to AI itself to push. Power is the binding wall of the late 2020s, and the capital rushing at it — fission now, fusion behind it — dates its fall to the turn of the decade. The physical world is the last and highest wall, and the recipe that cracked language is, for the first time, being pointed at atoms. Ride that ladder to the top and the payoff is a compressed century of scientific progress, arriving first in the informational sciences and only later in the physical ones. None of it is certain, and the knowability horizon is short. But the direction is legible, the bottlenecks are dated, and the honest thing to say is not that we know what happens — only that, for the first time, we can see which wall has to fall next.
Sources
- Modeling the Human Trajectory — David Roodman's super-exponential fit to all of human economic history; the long-run curve behind the compounding frame.
- This Can't Go On — Holden Karnofsky's accessible companion argument that today's growth rates are historically anomalous and unsustainable in either direction.
- Can AI scaling continue through 2030? — Epoch AI on the ~5×/year training-compute trend and the constraints on sustaining it.
- Algorithmic progress spurs more compute, not less — the ~3×/year efficiency gain and why it accelerates, not replaces, the compute build-out.
- A new Moore's Law for AI agents — METR's task-horizon doubling curve, the cleanest single measure of the compute wall falling.
- OpenAI targets a full autonomous AI researcher by early 2028 — the recursive-self-improvement roadmap in the labs' own words.
- Energy demand from AI — the IEA's data-centre electricity projections (485→~950 TWh by 2030); the power wall, quantified.
- Every nuclear-powered data-centre deal (2026) — the hyperscaler fission build-out, deal by deal, that dates the power wall's fall.
- Fusion energy in 2026: SPARC, Helion and the PPAs — the first real fusion purchase agreements and their timelines.
- Robot foundation models — the 2026 landscape — GR00T, Gemini Robotics and π0, and the shift of the robotics bottleneck from body to mind.
- Machines of Loving Grace — Dario Amodei's "compressed 21st century" thesis; the clearest statement of what the fallen walls unlock.
- Making quantum error correction work — Google's Willow result and why fault-tolerant quantum remains a later-decade problem, not a near-term AI lever.
Transcript
Alex: Here's the sentence that reframes everything about AI. Every technology before it could build a better product. This is the first one that can build a better version of itself.
Sam: So the honest question stopped being "will it slow down." It became "which physical wall falls next" — and it turns out they fall in an order you can actually name.
Alex: Welcome back to Dan's AI Intel — the show that tries to make honest sense of the fastest, strangest shift most of us will ever live through.
Sam: I'm Sam, here with Alex, and today we're doing something a little different. We are not going to tell you what AI can do this week.
Alex: Right, because that's a losing game. If you try to keep up by reading each new model release, you learn where the front line sits today and nothing about where it's actually going.
Sam: The whole show exists for the other thing — taking the one question that actually matters and digging past the hype and the fear to what's really going on, and what it means. Not just the tech, but the economy, the politics, the race between the labs.
Alex: And today's question is the biggest one of all. Where is this whole thing heading? Not "is AI impressive" — where does the curve go?
Sam: So here's the map for the next forty minutes or so. The trigger is a quiet claim that the people running the largest labs now treat as their working assumption — that we might be living through the opening of a compressed century.
Alex: Which sounds like a slogan. It isn't. It's the median reading of serious long-run growth models, and we'll show you the actual curve it comes from.
Sam: But the reason I couldn't stop turning it over is what it lets us ask. If progress doesn't plateau — if it compounds — then the whole question flips. It stops being whether the machine slows down, and becomes which wall gives way to let it speed up.
Alex: And there are exactly three walls. Compute, power, and the physical world. We're going to walk each one, in the order they fall, and by the end you'll be able to see the whole ladder.
Sam: There's a genuinely surprising turn in each — including one technology everyone files right next to AI that actually belongs on a much later shelf. We'll get to why.
Alex: One promise before we start, because it's the whole point of this show. Everything past this present moment is a projection, not a prophecy — and we'll name every source so you can weigh it for yourself.
Sam: Quick thing before we dig in — if you find this useful, follow the show in whatever app you're in. It's free, and it's genuinely the biggest thing that helps a small independent show like this one grow.
Alex: Okay. Let's start with the shape of the whole thing — because you can't judge where AI is going without it.
Sam: So when you say "the shape," you mean literally the shape of a graph?
Alex: Literally. Zoom out far enough — past AI, past the internet, all the way back — and human progress has a shape. For most of our species' existence, a single civilisation-defining advance took an eternity to arrive, and an eternity more to spread.
Sam: Give me the scale. How long are we talking?
Alex: Agriculture, bronze, the written word — those were spread across millennia. Then the Agricultural Revolution compressed that clock down to centuries. The Scientific and Industrial Revolutions compressed it again, to decades. The computer compressed it once more, to years.
Sam: So each big leap arrives faster than the one before it. That's not a straight line — that's a curve bending upward.
Alex: That's exactly the finding, and there's serious work behind it. An economist named David Roodman built a long-run model for Open Philanthropy — he fit a single equation to the entire trajectory of gross world product. All of human economic history, in one curve.
Sam: And what came out?
Alex: The best fit isn't steady exponential growth. It's super-exponential — a curve whose growth rate itself keeps accelerating. And it echoes a result a physicist named Heinz von Foerster published all the way back in 1960.
Sam: Okay, "super-exponential" is one of those words that either means everything or nothing. Give me the intuition.
Alex: Think of a normal exponential as a snowball rolling downhill at a steady pace, getting bigger as it goes. Super-exponential is the same snowball — except the hill itself keeps getting steeper the further it rolls. The growth feeds the growth.
Sam: So where does Roodman's curve actually point?
Alex: Extrapolated naively, his best-fit model implies a growth explosion with a median date around 2047. Now — he's very explicit that the naive version overshoots. Real growth does not literally run to infinity.
Sam: I was about to say. Nothing runs to infinity.
Alex: Right, and that honesty is the point. The exact date is not the takeaway. The takeaway is that history's engine has been acceleration — and it has only ever stalled when a binding constraint held it back.
Sam: You've said that word twice now — "binding." Why does it carry so much weight?
Alex: Because it's the whole mechanism of how real change happens. A binding bottleneck — some people call it a forcing function — is the single constraint that actually caps how fast a system can go.
Sam: Meaning if I fix any other problem, nothing moves.
Alex: Nothing moves. But relieve that one — the binding one — and everything downstream that was piled up behind it suddenly lurches forward. Change rarely comes from a smooth rising tide. It comes in step-changes, when one wall falls and unlocks a cascade that was waiting behind it.
Sam: Okay, that genuinely reframes how I picture progress. It's not gradual. It's wall, then flood, then the next wall.
Alex: And that's the lens for the entire episode. If innovation moves when a binding wall falls, then the useful question about AI becomes — what are AI's own binding walls, and in what order do they come down?
Sam: But hold on. That's true of every technology. What makes AI special enough to bend that giant civilisational curve? Steam didn't do that. Electricity didn't do it on this scale.
Alex: This is the crux, and it's that very first sentence. What makes AI a candidate for the same weight as agriculture or industry — but with a multiplier potentially orders of magnitude larger — is something called recursive self-improvement.
Sam: Translate that.
Alex: An AI system helping to design, train, and improve the next AI system. So capability feeds back into the very process that produces capability. Plainly: the tool is starting to build the tool.
Sam: And that's the part that gives me a little vertigo. Every prior invention sat outside the workshop. The steam engine could not design a better steam engine.
Alex: Exactly right. A transistor couldn't run the research program that produced the next transistor. But an AI can now do a growing share of the work that builds the next AI. And before anyone files that under science fiction — it's already partly real.
Sam: How real? Give me a number, not a vibe.
Alex: Anthropic disclosed that, as of May 2026, its own model — Claude — authored more than eighty percent of the code merged into the company's production codebase. Up from low single digits in early 2025.
Sam: Wait — eighty percent? In about a year, from basically nothing?
Alex: In about a year. Now, I want to be careful here — that is not yet a runaway loop, and serious people doubt it will become one soon. But it is no longer hypothetical. The machine is already inside the workshop.
Sam: Okay. So that's the frame. History is a curve that bends when a wall falls, and AI is the first thing that might bend it from the inside. So — which wall is falling first?
Alex: The first wall is compute — the sheer quantity of computation you can throw at training and running a model. And the key thing to hold about this one: it's best understood as accelerated, not unlocked.
Sam: What's the difference? Those sound like the same word.
Alex: Big difference, and it runs through the whole episode. "Unlocked" means it becomes possible at all — a door that was fully shut. "Accelerated" means it was always coming, but something suddenly put it on a tear. Compute was always going to grow. AI just stepped on the gas.
Sam: How hard on the gas?
Alex: The research group Epoch AI finds the computation used to train frontier models has grown about five times per year since 2020. That's a doubling roughly every five months.
Sam: Five months. For comparison — the old Moore's Law doubling was every two years, right?
Alex: Every two years, and everyone treated that as breakneck. This is far faster. And there's a second engine running underneath it. The algorithms themselves keep getting more efficient — the same capability takes about three times less compute each year.
Sam: So it's a double whammy. More hardware, and the software squeezing more out of every chip.
Alex: Stack them together and, if the trend holds, models trained around 2030 could use on the order of a thousand times more effective compute than today's. And every prior time we've had a leap of that size, it hasn't just produced a better version of the old thing. It's produced a qualitatively new capability.
Sam: Here's my problem with "a thousand times more compute," though. That's an input. I don't care how many chips are humming — I care what the thing can actually do. Is there a way to measure the output?
Alex: And that's the question that actually matters, because there's a clean way to measure it. Forget counting chips. An evaluation group called METR tracks the length of task an AI agent can complete on its own, unsupervised.
Sam: Length meaning time?
Alex: Length meaning — how long would this job take a human expert. So they're really asking: what's the biggest chunk of real work the AI can finish, start to finish, without a person stepping in.
Sam: And that number's been climbing.
Alex: That horizon has doubled roughly every seven months, for six years straight. It's gone from about four seconds of equivalent expert work back in 2019, to more than sixteen hours by 2026.
Sam: Four seconds to sixteen hours. So a few years ago it could handle a task you'd blink at — and now it's a full workday of expert effort.
Alex: A full workday. And here's the part that got me — the pace isn't even steady. Since 2023, the doubling has compressed to about every four to five months. So the curve isn't just rising, it's bending upward. That's the super-exponential shape showing up in the real data.
Sam: Okay, that's the most convincing chart in this whole story for me — because it's not a promise about the future. It already happened.
Alex: It already happened. And this is where compute stops being a story about hardware and becomes the master key for all three walls.
Sam: Go on.
Alex: The labs are now pointing AI at the job of AI research itself. OpenAI has stated a public target — an AI "research intern" by September 2026, and a fully autonomous AI researcher, able to run its own research projects, by March 2028.
Sam: So they're not just using AI to write code faster. They're aiming it at the discovery process itself.
Alex: And they say it about as plainly as you can — that, in their words, AI doing AI research will become the determining factor of the pace of progress within the next few years. That's recursive self-improvement moving from an essay to a roadmap with dates on it.
Sam: That's the vertigo again. If that loop actually closes, compute isn't just falling — it's falling and then reaching back to shove every other wall over.
Alex: Right — reaching back to push the other two over faster than they'd fall on their own. This is the episode's one genuine wildcard, and it's the hinge everything else swings on. We went deep on it — AI improving AI — in our episode number twenty-six, a little while back. Worth a listen after this one.
Sam: But even if you're skeptical about the loop, compute itself is clearly the wall that's already coming down. So what's the wall standing right in front of us — right now?
Alex: Electricity. For a brief window the constraint on AI was chips — you simply could not get enough of them. That window has closed. The binding wall now is power.
Sam: And this one honestly surprised me. We spend all our time talking about chips and models, and the thing that's actually stuck is — the electrical grid?
Alex: The grid. And the heads of the largest labs converged on this independently. The chip crunch eased, and power became the thing standing between ambition and reality. The numbers are stark.
Sam: Hit me.
Alex: The International Energy Agency projects electricity demand from data centres roughly doubling — from about four hundred and eighty-five terawatt-hours in 2025, to around nine hundred and fifty by 2030. And the AI-specific slice of that triples.
Sam: A terawatt-hour is one of those units that just washes right over me. Give me a handle on it.
Alex: Fair. Picture the whole data-centre demand as the electricity use of a large industrialised country — and we're saying that entire national-scale appetite doubles in five years. That's the order of it.
Sam: And the grid can just — add that?
Alex: That's the problem. It can't, not fast enough. Goldman Sachs flags a US power shortfall of about nine point three gigawatts in 2026, widening to forty-five gigawatts by 2028. And by the middle of 2026, an estimated seventy-five US data-centre projects — worth roughly a hundred and thirty billion dollars — had stalled.
Sam: Stalled on what? Money? Chips?
Alex: Neither. Held up by transformers and grid connections, with lead times running to four years. So you've got the chips, you've got the capital, and you are sitting there waiting on a piece of steel-and-copper grid equipment.
Sam: There's something almost funny about that. The most futuristic industry on earth, bottlenecked by a four-year wait for a transformer.
Alex: It's the whole texture of this wall. It's physical. Which — and this is the important move — is exactly why it's dated to fall, rather than permanent.
Sam: Unpack that for me. Why does "physical and stuck" mean temporary rather than permanent?
Alex: Because a wall this visible, this expensive, attracts an enormous amount of capital aimed straight at it. And when that much money targets one constraint, it doesn't stay a wall for long. The response here is a nuclear build-out at a scale we haven't seen in a generation.
Sam: Nuclear. Old-school fission — splitting atoms.
Alex: Fission first. The four largest hyperscalers have signed more than a dozen deals for close to ten gigawatts of nuclear capacity. Microsoft is paying to restart the undamaged reactor at Three Mile Island — eight hundred and thirty-five megawatts, a twenty-year agreement, with a federal grid waiver cleared in June 2026 and full power targeted for 2027.
Sam: Three Mile Island. That name carries a lot of baggage — it is not exactly a reassuring brand to bring back. And they're restarting it to run AI?
Alex: The undamaged reactor next to it, yes. And then there's a newer flavour of this. Google signed the first US corporate deal for a fleet of small modular reactors.
Sam: Small modular — meaning what, physically?
Alex: Factory-built reactors, fifty to three hundred megawatts each, instead of the thousand-plus of a conventional plant. The whole idea is you build them on an assembly line rather than as one-off mega-projects. Google's targeting five hundred megawatts from 2030. And Meta's gone even bigger — contracted for up to six point six gigawatts, including next-generation designs from companies like Oklo and TerraPower.
Sam: Okay — that's fission, which is proven, we know it works. But I feel like you're about to say the word "fusion," and my skeptic alarm is already going off, because fusion has been thirty years away for my entire life.
Alex: Your alarm is perfectly calibrated, and I'm glad you said it, because it's exactly the right amount of skepticism. Fusion — fusing light atoms to release energy, the process that powers the sun — has been "thirty years away" forever. What's genuinely different now is that it's attached to real purchase orders.
Sam: Purchase orders. Somebody's actually signing a contract to buy fusion power.
Alex: Helion signed the world's first fusion power-purchase agreement — committing fifty megawatts to Microsoft by 2028 — and they've broken ground in Washington. And Commonwealth Fusion Systems, whose demonstration machine, called SPARC, is already more than half-built, has a two-hundred-megawatt agreement with Google for a commercial plant in Virginia in the early 2030s.
Sam: So why now? What makes fusion suddenly bankable when it wasn't for fifty years?
Alex: This is the elegant part. AI data centres turn out to be fusion's ideal first customer. Enormous, steady, around-the-clock demand, from buyers who will happily pay a premium for reliable power. The technology found its perfect first market at exactly the moment it needed one.
Sam: So where do I look to know whether this wall is really coming down, versus another decade of promises?
Alex: Three checkpoints. At one year — do the stalled projects reconnect, do the first restarted reactors energise. At five years — do the small modular reactors actually ship on schedule, does Helion deliver real megawatts. At ten years — does fusion cross from demonstration to the grid.
Sam: And the honest headline is that this one moves on the timeline of concrete, not code.
Alex: Concrete, not code — and that gap is the whole decade. Compute moves at the speed of software; power moves at the speed of pouring concrete and pulling permits, and that's precisely why it, not the chip, is the pacing constraint right now.
Sam: Alright. Compute's falling, power is the one we're actively climbing right now. What's the last wall — and why is it the hardest?
Alex: The highest wall is the one between thinking and doing. An AI that writes and reasons and codes still lives entirely in the world of information. To actually change the physical world, it has needed a human pair of hands.
Sam: Right — it can write you a brilliant essay about building a house, but it cannot pick up a hammer.
Alex: Exactly. And the name for closing that gap is embodied AI — intelligence that perceives and acts through a physical body. A robot. And this is the wall that, once it's breached, unlocks a categorically different trajectory.
Sam: Why categorically different, though? We've had robots for decades. Car factories are full of them.
Alex: We've had robots that do one bolted-down, pre-programmed motion, forever. What we have never had is a robot that can act flexibly in a messy, unrehearsed world. Breach that, and the compounding curve reaches into manufacturing, construction, logistics — and the science lab bench itself.
Sam: So what was actually stopping it? If robotics has been around for decades and it's still not there, that sounds like a genuinely hard problem, not a wall about to fall.
Alex: Here's the twist, and it's the single most important idea in this whole section. For decades, robotics was gated on the body — the actuators, the balance, the battery. That is no longer where the constraint sits.
Sam: The bottleneck moved.
Alex: The bottleneck moved. From the body, to the mind. And that changes everything about the timeline.
Sam: Okay, walk me through the move. Why is the mind suddenly the easier half?
Alex: Because of one specific recipe that's been working spectacularly everywhere else. The foundation-model recipe: pre-train a large model on a vast amount of data, then adapt it to your particular task. It's the recipe that gave us the language models. And it is now being pointed at physical action.
Sam: So instead of programming a robot move by move, you — pre-train it on what, exactly? You can't download the physical world.
Alex: That was precisely the old objection. Where on earth do you get enough real-world training data. And the answer is, they're routing around it. Nvidia released a general foundation model for robots called GR00T — a new version at the very start of 2026 — that a robot-maker can start from, instead of training motion from scratch. Google DeepMind has Gemini Robotics; a startup called Physical Intelligence has its own models pointing the same way.
Sam: And the data problem itself?
Alex: Video pre-training and simulation. You train the robot partly in a simulated world, and on video, not only on a real machine moving in real time. Nvidia reported that synthetic motion data — data generated inside a simulation — lifted task-success rates by around forty percent.
Sam: Forty percent, just from practising in a simulator. That's the training-montage version of robotics.
Alex: It's basically a flight simulator for hands — you get thousands of repetitions the real world could never give you. And here's the line from a 2026 industry survey that captures the whole shift: hardware is no longer the primary bottleneck. Intelligence is.
Sam: And is anyone actually deploying these, or is it still all lab demos?
Alex: Tesla says it has more than a thousand of its Optimus humanoids working in its own factories, targeting fifty thousand. And there's a clever double purpose built in — they're deployed partly to do the work, and partly to gather the data that trains the next generation.
Sam: Oh, that's sneaky. The robots working the floor are also the data-collection fleet for the robots that replace them.
Alex: The fleet teaches its own successor. Now — I have to put a big flashing caution light on this one, because robotics is the wall it is easiest to overclaim.
Sam: Good, because a slick humanoid video is basically engineered to fool me.
Alex: And you should assume it's edited to. So don't watch the demo. Watch two things instead. One — cost per useful hour. Is it actually cheaper than the alternative. And two — generalisation. One model doing many unrehearsed tasks in a messy environment, not one rehearsed task on a stage.
Sam: So on the horizon spine — where does that put us?
Alex: Now: supervised pilots on factory floors. At one year: the first fleets earning their keep on narrow work. At five years: general-purpose robots plausibly crossing into commercial usefulness, if the intelligence curve holds. At ten years: the wall between AI's mind and the physical world genuinely thin.
Sam: And that last step is the enormous one, because that's the moment all of this stops being about bits and starts being about atoms.
Alex: That's the moment the compressed century gets to touch the real world. Which is exactly where we should head next — what actually falls out the far end of this ladder.
Sam: So stack all three walls — compute falling, power climbing, the physical world cracking. What do they unlock together? What's the prize at the top?
Alex: A dramatic acceleration of science itself. And it has a name — Dario Amodei, Anthropic's chief executive, calls it the "compressed twenty-first century."
Sam: Which means what, concretely? Because "acceleration of science" is the kind of phrase that could mean almost anything.
Alex: He's specific about it. The claim is that a large population of capable AI systems could raise the rate of biological discovery tenfold — delivering a century's worth of progress in five to ten years, after truly powerful AI arrives.
Sam: A century of biology in five to ten years. That is an enormous claim. Is there anything real underneath it, or is that just a CEO dreaming out loud?
Alex: There's a Nobel Prize underneath the leading edge of it. DeepMind's AlphaFold predicted the three-dimensional structure of essentially every known protein — a problem biologists had been stuck on for fifty years — and it won the 2024 Nobel Prize in Chemistry. The first Nobel for an AI-enabled breakthrough.
Sam: So this isn't speculative at that edge. It's already been honoured with the highest prize the field has.
Alex: Already banked. And its successor now models how molecules interact, not just their shapes. AI-designed drugs have reached human trials. Eli Lilly signed a deal worth up to two point seven five billion dollars for compounds from one AI-native discovery firm, back in March 2026.
Sam: Okay, so I'm ready to be swept away here — but you've got that tone. The one that means there's a "but" coming.
Alex: There is a very important "but," and naming it is the whole difference between a forecast and a fantasy. The compressed century has a brake built into it. AI accelerates the parts of science that are search and design far faster than the parts that require building and validating something in the physical world.
Sam: Define those two halves for me, because that distinction feels like it's carrying a lot of weight.
Alex: It's carrying all of it. Search and design: the answer already exists somewhere in a huge space of possibilities, and the task is just to find it. Which protein folds this way. Which material has these properties. Computation eats that kind of problem alive, because it dissolves an informational constraint almost on contact.
Sam: And the other half?
Alex: Building and validating in the real world. And here's the number that makes it vivid. DeepMind's system called GNoME predicted around three hundred and eighty thousand stable new materials. As of the reporting, outside labs had physically confirmed a few hundred of them.
Sam: Wait — three hundred and eighty thousand predicted, and a few hundred confirmed? That's — that's a rounding error. That's a fraction of a percent.
Alex: About zero point two percent. And that gap is not a failure of the AI. It is the shape of the near future. The AI can propose a scientific revolution far faster than the physical world can confirm it.
Sam: Because someone still has to actually go into a lab, make the material, and test it.
Alex: Make it, test it — and that takes real-world time that no model can compress. Same story with drugs. No AI-discovered drug has been approved yet, because designing a molecule is fast, and proving it works in a human body, over months, in real trials, is not.
Sam: So the honest picture isn't "everything accelerates." It's "the informational stuff leaps, and the physical stuff waits."
Alex: That is exactly the texture of the whole compressed century, and it's worth really internalising. Protein folding, materials on paper, molecule design, mathematics — the fields whose bottleneck is search — those leap first. The fields gated on a physical loop — a reactor that must be built, a robot that must be trained, a trial that has to run its course in real bodies over real months — those move on the timeline of matter. And matter does not care how fast the model thinks.
Sam: That line's going to stick with me. Matter doesn't care how fast the model thinks.
Alex: And it's the key to not getting fooled in either direction — not the breathless hype, not the flat dismissal.
Sam: Okay, I have to ask about the one you teased at the top. What's the technology everyone files right next to AI that you're saying belongs somewhere else entirely?
Alex: Quantum computing. It gets shelved right beside AI as the other looming revolution. And the progress is genuinely real — Google's Willow chip crossed a real threshold. It became the first one where adding more qubits makes the error-corrected memory better, rather than worse.
Sam: And normally, right here, I'd expect you to say "so quantum's about to blow up too." But you're shaking your head.
Alex: I'm shaking my head, because a useful, fault-tolerant quantum computer is a completely different order of problem. The estimates put the requirement for cryptographically relevant work near twenty million physical qubits. We have roughly a hundred deployed today.
Sam: A hundred, against a target of twenty million. That's not a gap, that's a canyon.
Alex: It's a canyon. Google's own public goal for a useful, error-corrected machine sits at 2029, and broader usefulness is a 2030s story at the earliest. So quantum is a wall that's barely been scratched — a genuine unlock someday, but its horizon is far out.
Sam: So if I'm ranking what actually matters for AI's trajectory in the near term, quantum is way down the list.
Alex: Way down. It matters far less than the boring miracle of more classical compute, cheaper power, and robots that actually work. Those unglamorous three walls are the whole story. Quantum is a later chapter.
Sam: Okay. So we've climbed the whole ladder — the three walls, the payoff at the top, the honest unevenness of it. But you keep saying that word — "projection." So let's be adults about it. How could this whole thesis be wrong?
Alex: This is my favourite part, honestly, because it's the discipline of the show. Everything we've said is a projection, and we say so plainly — not buried in a footnote at the end.
Sam: How short is the horizon, really? When you say we can't see far ahead, quantify it for me.
Alex: The field's own legibility horizon — how far ahead it's even readable — is measured in weeks. And the credible experts disagree by years about the very same milestone. Let me give you the actual spread, because it's wide.
Sam: Please, because I want to know who to believe.
Alex: Anthropic's Jack Clark puts the odds of meaningful recursive self-improvement — that loop we kept coming back to — at around thirty percent by 2027, and sixty percent by 2028. Meanwhile the AI critic Gary Marcus calls a 2027 timeline implausible, and doubts even 2030.
Sam: Those aren't small disagreements. One serious person says likely by 2028, another says maybe not even by 2030.
Alex: And a survey of forecasters landed on only a twenty percent median probability that the next few years compress six years of progress into two. So the right posture is neither the true-believer's certainty, nor the skeptic's dismissal. It's calibration — a set of grounded projections, with the hinges named out loud.
Sam: So name them. What are the specific ways this jams?
Alex: Three hinges. One: if power stays stuck — the reactors slip, the transformers never arrive — then compute growth throttles no matter how good the chips get. The whole ladder slows from its second rung.
Sam: Because it doesn't matter how brilliant the model is if you can't plug it in.
Alex: Exactly. Two: if the robot-data problem turns out deeper than video and simulation can solve, the physical wall holds, and the compressed century stays trapped in software — brilliant on paper, unable to touch the world.
Sam: And the third?
Alex: The third is the quiet one, and maybe the most important. If algorithmic progress just stalls — if that recursive loop hits diminishing returns before it becomes self-sustaining — then the super-exponential curve relaxes back into a merely-fast exponential. Still extraordinary. But not civilisational.
Sam: So how do I, sitting at home, tell which way it's actually breaking? Is there a single tell?
Alex: There's no single breakthrough that settles it. The signal that the optimistic reading is winning is all three walls continuing to fall roughly on schedule, at the same time. Task horizons lengthening, reactors energising, robots generalising — together. If you see all three moving, the ladder is real. If one of them jams, the whole timeline stretches.
Sam: Alright. Bring it home for me. If someone's been half-listening on a run and they get exactly one thing, what is it?
Alex: One thing: AI is the first technology with a credible claim to bending history's growth curve from the inside — because it is the first that improves the process that produces it. Everything else follows from that one sentence.
Sam: And the "which wall falls next" framing — give me the order, one more time, tight.
Alex: Compute is already giving way, and it's now being handed to AI itself to push. Power is the binding wall of the late 2020s, and the capital rushing at it — fission now, fusion right behind — dates its fall to the turn of the decade. The physical world is the last and highest wall, and for the first time the recipe that cracked language is being pointed at atoms.
Sam: And at the top of the ladder — the compressed century. But arriving unevenly.
Alex: Arriving first in the informational sciences, and only later in the physical ones. None of it is certain, the knowability horizon is short — but the direction is legible, the bottlenecks are dated, and for the first time we can actually say which wall has to fall next.
Sam: And honestly, that's the thing I'm walking away with. I came in asking "will AI slow down," and I'm leaving with a much better question — which of three specific, nameable, datable walls comes down next. That's a map, not a mood.
Alex: A map drawn in pencil. Which is the most honest kind there is.
Sam: So that's the picture — where this is all heading, and why. I hope you came away seeing a little more clearly that this isn't magic, and it isn't doom. It's three physical walls, an order they fall in, and a set of hinges worth watching.
Alex: It's a genuinely complex, fast-moving picture, with a brutally short horizon on what's actually knowable — and honestly, that is exactly what makes it worth following this closely. There has never been a more interesting front to watch.
Sam: One honest note on how this show gets made, because it matters. It's AI-generated. AI moves too fast to keep up with, so Dan built a custom stack of AI tools to research, analyse, verify and illustrate the questions worth understanding — mostly to learn them himself, and he shares what he finds. AI-assisted, fact-checked, and always worth a second look.
Alex: And before we go, one genuinely useful thing you can do. Follow the show. Whatever app you're listening in right now, there's a follow or a plus button — it's one tap, it's free, and it does two things.
Sam: You'll get every new episode the moment it lands. And honestly, for a small independent show like this one, a follow is the single biggest lever there is for helping it reach other people who are trying to make sense of all this too.
Alex: So if the last forty minutes were worth your time — go ahead and hit follow. That's the whole ask, and it genuinely matters.
Sam: One last thing before we wrap. This show runs on what you tell us. If there's a claim in here you'd push back on, or a thread you want us to pull harder on next time — a wall we got the order wrong on, a technology we underrated — write to us. The address is podcast at connective shift dot com.
Alex: We read every single message, and it genuinely shapes what we dig into next. Want to help decide which question we take apart next time? That's how you do it.
Sam: Thanks so much for listening. We'll see you on the next one.
Alex: Take care.