RSS preview of Exponential View

Rss preview of Blog of Exponential View

📈 Data to start your week

2026-03-17 02:30:29

Hi all,

Here’s your Monday round-up of data driving conversations this week in less than 250 words.

Let’s go!

Bricks and bytes ↑ Data center spending overtook office spending in December 2025.
Silicon gravity ↑ Global semiconductor sales hit $82.5 billion in January 2026, up 46.1% year-on-year.
Doctor’s orders ↑ Over 80% of US physicians now use AI at work.

🔮 Exponential View #565: Autoresearch; the solar supercycle; an agentic nation; ChatGPT Olympian, seeing fraud & moving asteroids++

2026-03-15 16:12:15

I signed up because you are just getting better and better. – Amir, a paying member

Subscribe now

Radical optimism

In the week when oil crossed $100 a barrel and triggered the biggest oil shock in history, we published a radically optimistic analysis of the near future of energy. The world right now may feel like a place where the short run is gorging on the long run – and yet! there are forces that no political decision can eclipse.

We argue that solar power could unlock a cascade of civilizational problems that cheap electricity can solve. Solar energy enjoys learning curves, cheaper with scale, while fossil fuels, for all their remarkable properties, face depletion curves.

At three cents per kilowatt-hour, desalination will no longer be a luxury and water scarcity will stop being “a law of nature”. At one cent, carbon capture will approach economic viability and synthetic aviation fuel will close on fossil jet fuel.

This is the solar supercycle, a self-reinforcing loop where every cost reduction opens a new market and every new market funds the next cost reduction. To show its potential, we built a model grounded in fifty years of Wright’s Law data that maps when each threshold arrives and which markets open up. The model allows you put put in your own assumptions: bearish, bullish or something in between. Paying members have had the model since Thursday. Today it opens to everyone: solar.exponentialview.co 🌞

An agentic nation

Chinese local governments are competing to normalise AI agent adoption. Around 1,000 people lined up outside Tencent’s headquarters to get OpenClaw installed last week. Paid installers appeared on Chinese consumer platforms almost immediately, charging for setup even though the software is free. Six major platforms pushed out hosted or one-click versions within days. Users from China make up nearly 40% of the 200,000 publicly visible OpenClaw agents.

I recommend reading ’s essay on this:

From DeepSeek in 2025 to OpenClaw now, Chinese media have been hammering one narrative non-stop: learn this AI tool, get a high-paying job.

Government agencies are offering multimillion-yuan subsidies to startups using OpenClaw. Alibaba’s DingTalk gave out unlimited API calls through March 31; developers could deploy for as little as $1.4. A whopping 83% of survey respondents in China said AI products and services were more beneficial than harmful. In the United States, only 39% agree.

That gap will have a cost. Sequoia partner Alfred Lin estimates that the top 5-10% of builders using AI coding tools are now 3-5x more productive than a year ago because they orchestrate fleets of agents to code for them. Through my own use of OpenClaw, I’ve found it to be self-compounding, meaning the more I use it, the better it gets, the more productive I am.

Autoresearch

One of the most interesting things that happened in AI this week was seeing ’s autoresearch run 700 experiments over two days. In that time, it found around twenty improvements that cut the time to train a language model to GPT-2 level of capability from scratch, from 2.02 hours to 1.80 hours (some 11% faster).

Each grey dot is one experiment the agent tried and discarded. Each green dot is an improvement it decided to keep. The green line is the running best score — and it only moves when something genuinely works. Lower is better on the y-axis, so that staircase descending toward the bottom right is the machine getting smarter, one kept idea at a time.

Autoresearch started on a small set of engineering problems, but similarly to tools like Claude Code, it can be applied to any problem you can define, measure and test cheaply:

🔮 The lantern and the flame

2026-03-14 01:32:47

AI is not a tool I pick up and put down. It’s become completely ambient, embedded in every process I run at work, daily.

A couple of weeks ago, in the first AI Vistas conversation, I sat down with , , and to discuss exactly this: do we use our tools, or do they use us? That conversation pushed me to lay out how my own thinking process has evolved.

There is a useful distinction here, drawn by researchers Shaw and Nave, between cognitive offloading and cognitive surrender. I used to know dozens of phone numbers by heart. Now I know two, and I don’t miss the rest. That’s cognitive offloading: a strategic delegation that costs nothing. Cognitive surrender is something different; an uncritical abdication of reasoning itself. And there is something about AI, about its allure and potency, that could make surrender far more widespread.

Thinking is my livelihood. If I stop thinking new things, I’m not doing my job. This week, I want to share how I navigate this.

Watch on YouTube or listen on Apple Podcasts or Spotify

What I outsource

Roughly 100 million tokens a day flow through systems my team and I have built and the most immediate change has been to my attention. Herbert Simon observed fifty years ago, that a wealth of information creates a poverty of attention. He was right, and the problem has only compounded since. I want to avoid missing an important signal, without drowning in everything else. So I built synthetic personas modelled on thinkers I value, Vinod Khosla for venture patterns, John Paulson for macro risk, Clayton Christensen for disruption logic, each scanning hundreds of items a week through their own intellectual lenses.

I've made an even bigger change to how I stress-test my reasoning. I’ll start writing and before I’ve finished the paragraph, an argument engine trained on 100,000 words of my writing might flag a structural weakness: I’m asserting where I should be evidencing. As a complement, House Views codify what we already believe at Exponential View, from learning curves to how Anthropic’s strategy differs from OpenAI’s, so a new argument faces challenge rather than confirmation. The friction is useful because it catches what I might miss.

On a bad day when I’m not doing my best writing, I will reach for a tool I built called The Stylometer, a Claude skill trained on 60,000 words of my prose, flags where sentences have gone slack and where the rhythm has drifted from my own voice. Synthetic editors interrogate the frame of the argument. The benchmark is always my own past best, not whatever I happened to produce that afternoon.

Subscribe now

What I protect

What I’ve described above is the artificial scaffolding. It works because I deliberately and fiercely protect my actual thinking and what makes it mine.

The first thing I safeguard is the space where ideas arrive before they’re shaped. For me that’s a walk, a long shower or a blank piece of A4 in landscape mode with a fountain pen. These are the conditions under which something genuinely new can appear. I let the process be non-linear, messy and iterative, because straightening it out too early kills what’s interesting.

These moments are profoundly personal. They depend on a world model that took decades to build, assembled from every conversation, every argument I’ve lost, every book that changed how I saw something. That specific arrangement of experience and association belongs to no one else and can’t be replicated. I safeguard that lived interiority carefully, because the things worth saying tend to come from it.

And as pointed out in the AI Vistas discussion

Once you figure out where your generative constituent of competence lives, that’s the thing you protect from offloading.

This is the best arrangement I’ve found for the work I need to do right now. Ten uninterrupted years of thinking would likely yield something different and perhaps better in many ways. And all of this will evolve in the coming years, so keep your minds open.

Azeem

Subscribe now

Missing the inevitable

For decades, almost everyone with a forecasting model got solar wrong. The IEA’s 2010 models predicted the levelized cost of solar electricity would fall 2.6% per year through 2020. The actual decline was closer to 17%, nearly 7x faster. Policy papers written around 2010 got their timelines wrong by a factor of ten. In 2001, the only institution that predicted anything close to the correct trajectory was Greenpeace in collaboration with the European Photovoltaic Industry Association – and even they undershot. What forecasters got wrong is that they modelled a price. They should have modelled a process.

Solar photovoltaic panels operate on Wright’s Law: for every doubling of cumulative production, costs fall by a fixed percentage. This is a result of “learning” that happens when you make a lot of something – workers get faster, designs are refined, error rates are lower, waste is reduced and more steps get automated. The more you’ve made, the cheaper the next unit becomes. That learning rate for solar has held at 23.7% per doubling over the last 48 years,2 confirmed across multiple independent studies.

Fossil fuels cannot do this. They are extracted, not manufactured. Their costs are set by geology, control, demand and supply signals. Production experience helps; new technologies are, of course, more efficient, but they fight the overwhelming gravity of the depletion curve. Of the roughly $16 trillion invested in upstream oil extraction since 1900, half has been invested since 2010. Despite that, oil output has only increased 13%. Depletion curves give you less from more.

Over the past century, coal costs roughly the same in real terms as in the 1880s. Oil has doubled, natural gas has doubled. The only certainty about fossil fuel prices is volatility. All the while, solar’s price curve moves in one direction, down. For the 85% of solar panels made in China, manufacturing capacity now stands at 1,045 GW, almost double the actual production in 2024 of 587 GW.

The relative prices of electricity generated by gas in the US since 2013, of crude oil, the cost per kWh of utility-scale solar (as tracked by BNEF) and the price per kWh storage of Lithium-ion batteries. Two of those curves are not like the others.

Subscribe now

The supercycle

The process of learning is a flywheel. Again, more production means lower cost. Lower costs increase appeal to customers, which drives more demand. And more demand drives production. At some point, a product tracking Wright’s Law will see its price fall enough that customers who never thought of buying it start to. A new market opens and the flywheel accelerates.

If that sounds familiar, it’s because you’ve lived it. This is the exact flywheel that built the computer industry over the past fifty years. In 1971, Intel’s first microprocessor had 2,300 transistors. Today’s chips hold 100 billion because a learning curve cut the cost per computation by a factor of 10 billion. That curve did not produce a single product. It produced an industry: mainframes gave way to PCs, PCs to smartphones, smartphones to cloud computing, cloud computing to AI. Each market was unimaginable at the price point of the previous one. Each one funded the next doubling.

Solar photovoltaic is a manufactured technology on a learning curve. It is the energy equivalent of the microprocessor. In fact, early solar borrowed heavily from semiconductor manufacturing technology and materials, including silicon purification know-how and wafer-cutting equipment3. And it has a similar flywheel.

The first solar panels were expensive, $1000 per watt. Only one customer could afford them – NASA. In 1958, Vanguard 1 became the first satellite powered by photovoltaic cells, six small panels generating less than a watt.

The technology was so exotic and expensive that, when the James Bond film The Man with the Golden Gun came out in 1974, its plot revolved around a stolen solar energy device, a technology worth killing for.

But the flywheel was turning. By the late 1970s, panels had fallen to around $76 per watt, cheap enough for Casio to put them on calculators and for off-grid telecom towers in remote locations. This expanded the market, cumulative production doubled and doubled again, and prices came down. By the 1990s, they were affordable enough for rural electrification programs across the developing world.

By 2010, $2 per watt. By 2015, below $0.60, the point at which solar photovoltaic-based systems could reasonably compete on a price basis with fossil fuels for new grid generation. Today: $0.07 per watt. Each threshold opened a market that had not existed at the previous price and each market funded the next collapse.

You can see that the story of solar is not just a story of an energy transition – it is what we call the solar supercycle, the self-reinforcing loop in which every cost reduction opens a new market and every new market funds the next cost reduction, which opens new markets.

The flywheel has been turning for fifty years. Now at a global weighted-average of as low as four cents per kilowatt hour and $0.01-0.02 in optimal locations, solar already cost 56% less than fossil fuel and nuclear alternatives in 2023. It has displaced coal and gas for new power generation in most of the world.

Solar is now large enough to power everything else, and the question we are asking is, what market opens next?

Subscribe now

Market creates market

Solar now generates 9.5% of global electricity, from 1% a decade ago. If the market were fixed, if electricity demand stayed where it is, there would be only three-and-a-bit doublings left before solar hit 100%. But the market is not fixed. Every time the price falls far enough, the ceiling moves and a use that was previously uneconomical becomes viable.

Electricity is only 17% of primary energy today and solar is 1.4% of that. But roughly two-thirds of fossil energy is wasted as heat – electrons do the same work with a fraction of the input. Solar’s real share of useful energy is closer to 4%. From 4% to 100% is five doublings and each one will unlock new markets that fund the next.

Grid demand itself is growing. As the global south grows richer, the demand for refrigeration, air conditioning and other mod cons will accelerate. Residential cooling electricity demand alone is projected to grow from 1,220 TWh per year in 2020 to 1,940 TWh in 2050, and that assumes up to four billion people still live without air conditioning. Data centers will add more to the demand as the economy moves to a compute substrate. It’s plausible that global electricity demand will double over the next two decades, given these rather predictable trends. Yet another turn of Wright’s Law.

Ground transport was electrification’s first big market beyond the grid. At eighteen cents per kilowatt-hour, the going rate across much of the world, it costs less to own and run an electric vehicle than a petrol one, as long as oil sits above $40 a barrel (for most of the decade it’s not been below $40). Electrifying all road transport would consume roughly 10,000 TWh – a third of global grid energy – creating a demand for more than 12 billion solar panels. And panels and batteries are symbiotic: batteries can soak up electricity when it is plentiful and fill gaps when it isn’t. Grid-scale virtual batteries made up of thousands of ordinary cars are becoming viable.

As the price keeps falling, harder markets open. There is a breakeven for each where clean electrons compete with dirty joules, even accounting for the cost of new equipment. The threshold for each is the point at which solar economics are compelling enough to drive adoption at full scale globally, including in markets with subsidized fossil fuels4.

Test the model here

Notes on the breakdown under footnotes: 5 6 7 8

Two billion people lack access to safely managed drinking water. We’ve had the core technology to fix a lot of this for decades in reverse osmosis desalination. The electricity bill is a major hurdle. Running a modern plant to turn seawater into fresh water takes 3 to 3.5 kilowatt-hours of power for every cubic meter. At twelve cents a kilowatt-hour, that alone is 35 to 40 cents before you’ve built the plant, hired anyone, or laid a single pipe. Total production costs run to $1 to $2.50 per cubic meter, at or above what American households pay for tap water. So desalination stayed a rich-country luxury. At three-cent solar, the energy cost falls to a dime per cubic metre. When the price of electricity goes even lower, energy will stop being the main constraint altogether.

Agricultural water is a different challenge. Farmers pay $0.01-0.05 per cubic meter and seawater desalination has a capex floor that cheap electricity alone cannot close. But agricultural water stress is rarely an ocean salinity problem – it’s usually a brackish groundwater problem from depleted aquifers and saline rivers where salinity is 10-30x lower than seawater. For these, capacitive deionisation and electrodialysis require a fraction of what seawater reverse osmosis demands and are on their own early-stage learning curves. At one-cent electricity, the energy cost for brackish treatment falls below $1 per thousand liters. The capex constraints that make seawater desalination uncompetitive for agriculture do not apply to the same degree here – the agricultural water ceiling is not a law of physics.

Below two cents per kilowatt-hour, the hardest sectors to decarbonize stop being quite as hard.

Green hydrogen, made by splitting water with electricity, currently costs $3 to $6 per kilogram – electricity is the majority of the bill. At two-cent solar, production falls to $1.50 to $3.00 per kilogram, undercutting the fossil-derived grey hydrogen that dominates today. We may never use hydrogen to move a car or a truck, but at that price, we will have cheap green steel, synthetic ammonia for fertilizer and industrial heat. These are the exact sectors skeptics say renewables can never reach. The machines that split the water – electrolyzers – are on a learning curve and their costs are falling 12-20% with every doubling of installed capacity, a rate comparable to solar panels two decades ago.

Direct air capture could cross a threshold, too. It costs $600 to $1,000 to pull a tonne of CO₂ from the atmosphere today. For current deployments, solid sorbent systems like Climeworks, energy is roughly 30-40% of cost and much of it is heat, not electricity; the rest is capital and sorbent replacement. But the next generation of DAC is electrochemical: Verdox’s electroswing adsorption uses electrons directly to capture CO₂, with no thermal regeneration step and ocean-based approaches like Equatic use electrolysis to move seawater chemistry at scale. For these, electricity dominates operating cost and they sit at the steep part of their own learning curve, roughly where solar panels were in 2005. At one-cent electricity, electrochemical DAC approaches sub-$100 per tonne in techno-economic models, provided that capex on membranes and electrodes follows learning curves down, approaching the price where a genuine carbon removal market becomes possible.

Aviation cannot electrify without ultra-cheap solar. Jet fuel carries around 50x more energy per kilogram than a lithium-ion battery and no plausible battery chemistry on the horizon closes that gap. Long-haul flights will burn liquid hydrocarbons for decades; the only real question is which hydrocarbons. Synthetic kerosene, made from water, captured CO₂ and renewable electricity, is chemically identical to fossil jet fuel but, net, emits no carbon. Today it costs 10x more than conventional aviation fuel. Electricity is a major component of that cost, but so is the capital cost of the electrolyzers that split water into hydrogen and the Fischer-Tropsch reactors that assemble it into fuel. Both are on early-stage learning curves: electrolyzers have been halving in cost roughly every five to six years. Cheap, abundant electricity is what makes running them continuously economical. The two cost drivers are not independent: as solar prices fall, it becomes rational to build larger electrolyzer fleets optimised for continuous operation rather than intermittent use, thereby reducing capex per unit of fuel. At two cents per kilowatt-hour, the combined effect pulls e-kerosene toward roughly twice the cost of fossil jet fuel at current carbon prices. A carbon price in the range already imposed by the EU, around $80 per tonne and rising, may close the remaining gap.

PFAS, the “forever chemicals,” are detectable in 99% of people tested. They are linked to cancer, liver damage, and reduced fertility. Remediation would cost $120-175 billion in the US alone, while global annual societal costs run at tens of trillions of dollars. We know how to break the carbon-fluorine bond – the strongest in organic chemistry, which is the reason these molecules persist for millennia. Electrochemical treatment has demonstrated up to 99% destruction in lab settings. Scaling it is partly an engineering challenge, but it is fundamentally an energy problem. At one cent per kilowatt-hour, remediation becomes economically tractable.

Even further out, the outlook is like science fiction. Every material object is atoms in a configuration. A landfill and a warehouse contain the same elements – carbon, silicon, aluminum, phosphorus – just differently arranged. Rearranging them is chemistry, not physics, and chemistry is an energy cost. Today that cost is prohibitive. Converting waste streams into plasma, sorting atoms by mass and reassembling them into useful materials is technically possible but economically absurd. But so was desalination in 1990 and direct air capture in 2010. At a hundredth of a cent per kilowatt-hour, you are in the territory of the Star Trek replicator.

Source: *Star Trek, The Next Generation,* screenshot

The chart below shows our mid-case model for the trajectory of the solar supercycle – and you can access the full model and simulator here.

The model operates using the learning rates and Wright’s exponent (b) that encapsulates the cost declines as a power law of cumulative production.

The cost (C) at a cumulative capacity (Q) is described by,

where C0 and Q0 are values from 2024, which are used to anchor our model. The exponent b contains the learning rate (b) via

We have used lambda = 0.24 for the base case. Both the levelised cost of electricity (LCOE, C0 = $0.043/kWh) and the module price (C0 = $0.085/W) follow this same curve with their respective 2024 anchors.

The LCOE is bounded below by a cost floor (cfloor), ensuring costs never fall below a theoretical minimum:

The cumulative capacity grows with each year’s accumulating installations. Baseline solar deployment, driven primarily by grid electricity, follows a logistic S-curve with a ceiling (K), described in our model as baseline deployment ceiling, and steepness (the growth rate, k):

The growth rate sets the initial annual percentage increase in installations (~20% in the base case) and decreases as we approach the ceiling. The value, c, follows from requiring the curve to pass through the known 2024 value of 600 GW/yr, and t_sat is the year at which overall industry deployment reaches its inflection point when annual additions stop accelerating and begin to plateau.

An effective LCOE is calculated and used to evaluate market unlocks. This effective LCOE accounts for the fact that real-world deployment requires grid integration and storage. Hence, the effective LCOE adds a grid integration penalty (storage costs, starting at $0.005/kWh in 2024) to the LCOE; this penalty declines at 8% per year as battery prices fall.

As the effective LCOE crosses a market’s threshold, that market generates additional demand beyond the baseline capacity buildout. We compute the total solar capacity needed to serve each market at full scale, then ramp that demand linearly9 over fifteen years. Once a market is fully ramped, its capacity is considered built and no further additions are needed. This additional deployment feeds back into cumulative capacity, accelerating the Wright’s Law cost decline.

Market scale is expressed in terawatt-hours of annual electricity demand at full penetration. This is the quantity used to compute the solar capacity needed for each market’s demand ramp.

Test the model here

Fifty years, no floor

Forecasters have predicted a cost floor for solar since 1990. They were wrong every time and it’s backed by academic evidence. A 2022 paper in Joule by Way, Ives, Mealy and Farmer at Oxford showed that solar has followed Wright’s Law reliably across decades, costs dropping as a power law of cumulative production. A full shift to renewables in their fast-transition scenario, achievable within 25 years, saves roughly $12 trillion compared with no transition, without counting climate co-benefits.

Casey Handmer, a Caltech physicist turned entrepreneur whose company Terraform Industries is building machines that turn sunlight and air into synthetic fuel, evidences that in recent years the solar learning curve has accelerated, not slowed down.

A number of commentators have consistently predicted that solar cost improvements will level out ‘this year’ since about 1990, and yet if anything deployment, cost decline, and the learning rate have only improved.

Production is doubling roughly every 18 months. Casey expects another halving of cost within three to five years; no miracles are required.

Every technology not on a learning curve – nuclear, hydropower, biopower – has seen flat or rising costs over the same period. Learning curves reward mass manufacturing: you stamp out millions of identical solar panels and each doubling of cumulative production drives another predictable percentage drop in cost. Nuclear plants and dams are bespoke construction projects, each shaped by its site, regulations and politics. Even China, which has driven nuclear costs to roughly $2 per watt by standardizing designs and building in series, has achieved a one-time step change, not a compounding curve. All the while, the solar curve is not slowing down. New tandem solar cells – stacking two light-capturing layers instead of one – have already exceeded the theoretical efficiency limit that governed the technology previously.

For most of human history, energy scarcity was a physical problem. It isn’t anymore. The photons arrive free and the question was always whether we could convert them cheaply enough to matter. Solar panels are, of course, only part of the equation. They create energy, but they don’t store it, transmit it, or do the work to create useful outputs. Batteries and electrolysers are on their own learning curves. And China, with over 50,000km of HVDC transmission lines, is proving that transmitting electricity from sunny areas to cloudy ones is economically feasible.

What remains is solving the administrative constraints, which, unlike much of physics, can be changed. Over 900 gigawatts of solar projects sit in US interconnection waiting lists, with average wait times exceeding four years. Interconnection costs have risen fourfold since the early 2000s, from $25 per kilowatt to $138-167 per kilowatt today.

The flywheel turns regardless, every panel manufactured today makes the next one cheaper. Every new market unlocked increases demand and accelerates the next doubling. The solar super flywheel has been turning for fifty years. In 2024, it installed more capacity in a single year than existed on earth in 2010. What do you do with electricity that costs almost nothing?

China understood this before anyone else. Beijing is investing $580 billion in grid infrastructure over the next five years, with nearly half earmarked for ultra-high-voltage direct current transmission lines that can move solar electricity across 3,000 kilometers with minimal loss. Pakistan understood it too. In 2024, Pakistan became the third-largest importer of solar panels on Earth, $1.4 billion worth from China in the first half alone. By 2025, solar was Pakistan’s largest electricity source. Half the households in parts of rural Punjab and Sindh have solarized. In India, rooftop solar is reaching villages that grid infrastructure never did. And across the Global South, hundreds of millions of people are getting their first reliable electricity from panels that cost seven cents a watt.

Subscribe now

Testing the bears

A few objections recur even from the well-informed, educated participants in the space. They all make the same error against solar as they treat a dynamic system as static.

Intermittency and storage

Intermittency is the most intuitive objection, so we’ll start with it.

Battery storage costs have fallen 97% since 1991. They follow the same curve, with lithium-ion cells having a learning rate of roughly 18%, consistent with Wright’s Law. BloombergNEF’s 2025 survey finds stationary storage reaching $70 per kilowatt-hour, low enough to make daily solar cycling economically rational. High-voltage direct current transmission enables continental-scale resource sharing with only about 3.5% losses per 1,000 km. South Australia already showed that operation with 93% inverter-based renewable generation is possible, supported by synchronous condensers and grid-forming technology. The state installed the world’s first grid-scale battery, then added grid-forming inverters. It has not lost a single hour of electricity to load-shedding since 2018 and Australia didn’t stop there. EnergyConnect linked three state grids in 2025. Marinus Link will run undersea cables from Tasmania to Victoria, and Sun Cable is building a 20-gigawatt solar farm in the Northern Territory, connected to Singapore by an undersea cable.

The multi-day storage problem – the intermittency objection that seems hardest – is closing faster than most expect. Iron-air batteries leverage the chemistry of rust; they store energy for 100 hours at roughly one-tenth the cost of lithium-ion batteries. Form Energy’s 30 GWh project in Minnesota is a wind farm that can supply continuous power through a week of calm.

The hardest version of the intermittency challenge is not daily but seasonal: winter in northern Europe or the northeastern United States means shorter days, lower solar angles and higher heating demand arriving simultaneously. A solar panel in Germany generates roughly 8.7x less power in December than in May; 100-hour storage closes the daily gap but not the seasonal one.

The honest answer is that solar, alone, does not solve winter power at high latitudes – and the model does not claim it does. Solar’s supercycle is a claim about trajectory and cost, not about geography. Where the sun is abundant, the tropics and subtropics, where three billion of the next two billion energy consumers live, solar is already unambiguously dominant. At higher latitudes, solar’s falling cost pairs with offshore wind (which peaks in winter), continental HVDC grids and long-duration storage.

You can’t learn that far

Some will argue that you can’t learn that far. After all, the internal combustion engine reached its limits at some point, went through a radical period of learning, then slowed down, and cars started to get more expensive. And of course, that did happen. But that was because the internal combustion engine ultimately hit a wall. Walls, as I write in my upcoming book, are problems that are limited by the laws of physics; they are practically unsolvable. The combustion engine approached the limits of the Carnot cycle, which is the absolute maximum efficiency you can achieve in a heat engine. The rest of the costs of that car and running that engine are a function of marketing dynamics, market structure, market power and, of course, the fact that the fuel that powered them was essentially a commodity oil that has, over a hundred years, been on a highly volatile, but generally rising cost curve. With solar photovoltaics, the path to continued learning remains strong for several more doublings. The physics alone show that this is possible.

For decades, solar was limited by the so-called Shockley-Queisser Limit, a calculation that describes how photons that are too energetic and not energetic enough are not used efficiently, either being transformed into heat or passing straight through the silicon unaffected. The upper limit as a result of this behavior was an efficiency of around 33.7%, derived for a single solar cell. But that efficiency value has since been broken.

By printing perovskite on top of the silicon, you essentially create a device that captures different parts of the solar spectrum in each layer – the perovskite absorbs higher-energy light while letting lower-energy light pass through to the silicon beneath. This opens up more of the spectrum for efficient conversion into electrical energy. This increased the theoretical efficiency limit to ~45%. Continuing to add layers could increase this further. Assuming a hypothetical scenario with infinite layers, you might think that 100% of the spectrum could be captured. Unfortunately, that’s not the case, but the theoretical limit is around 68%, significantly higher than the highest efficiencies today. That limit is set by thermodynamics itself. Any object at a finite temperature – including a solar panel – unavoidably emits thermal radiation. That emission cannot be eliminated, and it’s what prevents even a perfect infinite-junction cell from reaching 100% efficiency.

Panels are cheap, power isn’t

US energy prices may appear to be defying our exponential gravity. John Arnold, the legendary energy trader turned infrastructure reformer, frames this friction as panels are cheap; power isn’t.

North American solar power purchase agreement prices climbed toward $62 per MWh by early 2026, a nearly 50% rise from their 2020 lows, even as the price of the silicon modules themselves halved again. A solar panel now accounts for barely 20-25% of a utility-scale project’s total cost. The remaining 75% is a cocktail of structural friction: interest rates that doubled the cost of capital, grid connection queues that have ballooned to over 2,300 GW and a tightening labor market for specialized engineers.

This is the strongest version of the bear case. The bear is not arguing that modules won’t keep getting cheaper but that it doesn’t matter, because the non-module costs dominate and don’t follow the same curve.

But they may well do. Each line item in that 75% is on its own cost curve – earlier-stage than modules, but the same shape. Robots from companies like Cosmic Robotics install 60 modules per hour at project sites, a labor learning curve. Inverters and batteries ride their own Wright’s Law trajectories, with lithium-ion cells at an 18% learning rate. Automated permitting is compressing soft costs. Building-integrated photovoltaics eliminate racking, land acquisition and mounting entirely and the building becomes the balance of the system. And interest rates are cyclical, not structural; they are the weather, not the climate.

Market structures also impose high costs. In a merit‑order power market, generators are stacked from lowest to highest marginal cost: solar and wind at the bottom, then nuclear and coal, and gas peakers at the top. The price is set by whichever unit is needed to meet demand at the margin, usually gas, and that marginal gas price is then paid to all generators.

Solar’s falling cost and near‑zero marginal cost push more cheap supply into the left of the merit order, lowering prices when the sun is strong. But when demand extends beyond renewables, the market “lands” on gas, and that gas price anchors the wholesale. For households, this means bills stay high even as the solar cost curve comes down, because the retail price is effectively held aloft by where gas sits at the top of the merit order. This isn’t a law of physics. As the energy mix changes, the merit-order pricing will look increasingly antiquated and, ultimately, intractable.

Land use

Critics of hyper-scale solar deployment often point to a curious physical trade-off, albedo. In the simplest terms, albedo (measured from 0 to 1) is the Earth’s “reflectivity index.” Darker surfaces absorb more of the sun’s energy; lighter surfaces, such as Saharan sands or Arctic ice, reflect it back into space. By blanketing these regions in dark, photon-hungry silicon, are we inadvertently turning up the planetary thermostat?

In the most extreme example, Li et al. considered the impact of installing solar panels in the Sahara. You’d likely shift the local climate, with more rain and vegetation as darker panels absorb heat and slow the wind, setting off a reinforcing feedback loop.

However, we must zoom out to see the full picture. To focus solely on local heat while ignoring displaced carbon is to fall into a thermodynamic trap; it is like measuring a computer’s value by its thermal output while ignoring its computational work. Research by Gregory Nemet provided some valuable context. While solar farms do reduce local albedo, the carbon-offsetting power of the clean energy they generate is roughly 30x greater than the warming effect of their darker color. In fact, a 2024 analysis by Wei et al. found that the “albedo debt” of a solar farm – the heat it adds by being dark – is typically paid back by the CO₂ it avoids in less than a single year. We are trading a minor, local warming effect for a massive, compounding global cooling advantage. Meanwhile, energy and extractive sectors are projected to expand into an additional 1.26 million square kilometers of intact natural land by 2050, more than 3x the urban expansion footprint.

We use more than 30 million hectares of land to grow biofuels. That’s about the size of Poland; it powers just 4% of land transport. If we were to cover the same area in solar panels, we would meet the global electricity demand today. And land where solar is placed can be used for other things. Rooftop potential alone – the space that already exists on buildings – could generate over 60% of current global electricity consumption.10 Agrivoltaics, floating solar on reservoirs, and dual-use installation along roads and canals expand the supply of suitable land further. Fossil fuel land use is cumulative and destructive; solar is permanent infrastructure that can improve the land it sits on.

Subscribe now

Back to reality

The Strait of Hormuz is closed. A literal and metaphorical chokepoint. But the learning curve can plot a course, a long one with may obstacles to be negotiated, to eliminate that and many other chokepoints. China realized this a decade ago and as the first electrostate is insulated from the current energy shock.

With solar now below the cost of fossil fuels almost everywhere on earth, other nations will follow China’s lead. In the United States, solar is winning support even from MAGA insiders who see it as essential to powering AI and competing with China, a former DOGE aide is championing solar on socials and Energy Secretary Chris Wright, who once called it a “parasite,” now concedes its commercial role. Solar energy won’t be the sole contributor to our energy system in 30 years' time; other sources, mostly carbon-free, will also contribute. But solar will be the dominant player.

The Strait of Hormuz will reopen. It always does. But sometime in the not-so-distant future, it might close again, because that is what chokepoints do. And each time, it will matter a little less. Because while politics is powerful, the laws of physics will always win. The learning curve will keep on turning. Solar has no strait. It has a learning curve and the learning curve does not close.

Thanks to Ramez Naam and Jan Rosenow for reading an earlier version of this analysis.

Learning rate calculated with an OLS regression from historical global weighted-average module prices, using data from Nagy et al. (2013) and the PCDB, and BNEF from 1976 through 2024.

Based on data used to compute our learning rate for solar modules, using data from 1976 to 2024.

Around 2009, solar got big enough to build its own dedicated polysilicon refineries and manufacturing lines. Wacker Chemie in Germany and GCL-Poly in China built plants exclusively for solar-grade silicon. It became its own industrial stack..

Geography compresses the timeline because the flywheel turns fastest where sunlight is cheapest and fossil alternatives most expensive – those early deployments pull the global average down for everyone else. Policy can accelerate or interrupt that sequence: carbon pricing pulls markets forward; trade barriers and subsidy reversals push them back.

Grid electricity already below threshold at anchor (2024); of the IEA NZE 2050’s ~80,000 TWh total electricity projection (WEO 2024), roughly ~50,000 TWh serves grid displacement, transport and building electrification.

~40% of 150 EJ global industrial heat is electrifiable via arc furnaces and resistance heating at ~$25/MWh (IEA Tracking Industry; McKinsey 2024).

Green H₂ TAM is IRENA 1.5°C full-transition scenario, not the current market (≈$130B/yr). SAF physical demand: 350B L/yr × 20 kWh/L.

DAC requires ~8.5× reduction in the net cost of carbon removal from the current ~$850/tonne – likely needs technology breakthrough beyond learning-curve scaling (DOE Carbon Negative Shot). PFAS $250B is an upper-bound annual estimate; 90B m³/yr × 50 kWh/m³.

A more realistic approach would model each market as its own logistic S-curve, but that would add three parameters per market (peak capacity additions, growth rate and time to saturation), substantially increasing complexity. The linear ramp spreads demand evenly across a 15-year window – a sufficient first approximation for capturing the flywheel’s cumulative-capacity dynamics under Wright’s Law.

The authors calculate the rooftop solar photovoltaic potential at around 19,500 TWh, while the world consumption of electricity in 2024 sat at ~31,000 TWh.

📈 Data to start your week

2026-03-09 23:53:08

Hi all,

Here’s your Monday round-up of data driving conversations this week in less than 250 words.

Let’s go!

History’s biggest oil shock ↑ Closing the Strait of Hormuz knocked some 20 million barrels a day off global supply, about the same as the next five biggest oil shocks combined.
Claude’s takeover ↑ In Feb 2025, ChatGPT held 90% of US business AI subscriptions. A year later, Claude commands nearly 70%.
AI bug-hunting ↑ Anthropic’s Claude Opus 4.6 found 14 high-severity bugs in Firefox in two weeks in January, almost 20% of all high-severity bugs patched in Firefox in 2025.
AI hiring divide ↑ Companies that hired fewer AI specialists than is typical for their industry delivered, on average, 19% lower equity returns than their peers.

Exponential View Modify