Satellites in Sun-synchronous low Earth orbit that follows the day-night boundary—the terminator—could generate solar power continuously, 24 hours a day, unaffected by weather or atmospheric attenuation. As a result, they would need only one-fifth the solar panel area required to produce the same amount of power on Earth’s surface. Because power generation would be continuous, batteries would not be needed either. Another major advantage is that space-based data centres would require neither land for construction nor water for cooling.

On the other hand, their electronic circuits would be exposed to ionizing radiation. Although practical experience in this area remains limited, Google’s calculations suggest that even the most radiation-sensitive component, high-bandwidth memory (HBM), could withstand more than ten years of radiation exposure. Meanwhile, the startup Starcloud has been testing off-the-shelf Nvidia GPUs in space since November 2025. A more challenging issue is maintenance and repair: would it be economically viable to replace failed or obsolete components, or would it make more sense to let satellites to burn up during atmospheric re-entry, sacrificing the opportunity to recycle them?

Although concepts for massive monolithic systems do exist—such was the Ascend feasibility study conducted by Thales Alenia Space in 2023–24—it remains unclear what advantages would justify the additional cost and technical complexity of assembling them in orbit. By contrast, constellations composed of thousands of individual satellites offer far greater scalability: capacity can be expanded simply by adding more satellites, while obsolete or failed ones can be deorbited.

The first plans of the satellites that are probably the closest to practical implementation—named, rather unimaginatively, AI1—were unveiled by SpaceX in June. The 2.15-tonne spacecraft will have a peak power draw of 150 kW, a sustained load of 120 kW, 600 square metres of solar panels for power generation, and 110 square metres of radiators for heat dissipation. Achieving 1 GW of computing capacity would require around 7,000 such satellites.

SpaceX’s plan

The lowest-cost option for placing payloads into low Earth orbit is currently SpaceX’s Falcon Heavy, at approximately $1,500 per kilogram. Through its unique combination of enormous payload capacity and full reusability, Starship promises a step change in launch economics (Falcon Heavy’s second stage is expendable).

Assuming an initial manufacturing cost of $90 million, a service life of 20 flights, refurbishment costs of $5 million per launch, and $2 million in propellant, the total cost per flight comes to $11.5 million. With a payload capacity of 100 tonnes, this translates to just $115 per kilogram. And this is only the beginning: as production scales up and operational experience accumulates, manufacturing and refurbishment costs are expected to fall while vehicle lifetime increases. Assuming a unit cost of $50 million, a lifetime of 100 flights, refurbishment costs of $1 million per launch, and full utilization of the 200-tonne payload capacity of the final (V4) version, the cost drops to just $20 per kilogram—another order of magnitude lower.1

It is difficult to overstate how unique Starship is. It has no true competitor—not only in the United States, but anywhere in the world. While at every other launch provider it will still cost several thousand dollars per kilogram to reach orbit, SpaceX alone will be able to reduce that cost substantially—first by one order of magnitude, and eventually by two.

If a single Starship carries 40 satellites, deploying 1 GW of computing capacity would require 175 launches, at a total cost of approximately $2 billion. Once launch costs fall to this level, they will no longer be the primary cost driver; manufacturing will become the dominant expense. For SpaceX, manufacturing costs are estimated at around $800 per kilogram. Based on the mass of the AI1 satellite, this corresponds to approximately $1.7 million per satellite, or $11.9 billion for a constellation of 7,000 satellites. This means that the cost of building a 1 GW data centre in space—excluding the processors—would be around $14 billion, essentially the same as constructing an equivalent facility on Earth.

It is true that a terrestrial data centre can retain much of its power, cooling, networking, and building infrastructure across multiple generations of hardware, whereas in an orbital system all of this infrastructure must be replaced with each new generation of satellites. However, this is not a decisive difference. According to Epoch AI’s model, the annualized capital expenditure for a 1 GW terrestrial data centre is $7.6 billion. Assuming a five-year service life for the satellites, the equivalent for a space-based system would be $8.8 billion. Moreover, this would be partially offset by the cost of electricity, roughly $600 million per year.

What about the time factor? SpaceX has been increasing its annual launch rate by 30–40 per year, and at that pace it could reach 260–270 launches annually by 2028 and 300 by 2029. To ensure that manufacturing keeps pace, the company plans to open its Gigasat factory near Austin by the end of next year. Its production capacity will far exceed that of SpaceX’s current Redmond facility, which now manufactures around 70 satellites per week. Taken together, it is entirely plausible that, if the Starship program proceeds as planned, SpaceX could be capable of deploying 1 GW of computing capacity in orbit annually by 2028–2029. For comparison, the terrestrial data centre capacity of the largest players—including Google, OpenAI, Amazon, Meta, and Microsoft—is projected to be in the range of 4–6 GW by that time.

This is about more than simply choosing an alternative way to build computing infrastructure. By creating demand for Starship through space-based data centres, SpaceX not only establishes the economic case for its new launch vehicle (much as Starlink did for the Falcon 9) but also gains a unique capability that could evolve into an insurmountable competitive advantage. The integration of satellite manufacturing, launch services, and on-orbit operations (and, in Elon Musk’s long-term vision, eventually even semiconductor manufacturing) within a single company creates a cost advantage that competitors dependent on separately sourced services are unlikely to match.2

In the near term, building data centres on Earth may well remain simpler and more economical. But when the time comes to deploy them in space, there will be only one company in the world that has the capability to do so.3 The key bottleneck is Starship—and that is the lens through which today's launch should be viewed.

Welcome to our world

„The easiest robot to adapt into the world are humanoid robots, because we built the world for us.” – Jensen Huang

Until quite recently, we tended to think of humanoid robots as little more than movie props. We are used to truly useful, high-performance robots being highly specialised machines that bear little resemblance to anything found in nature. But on closer inspection, developing humanoid robots is a very rational approach. After all, the entire world we have built is designed around us, humans. And that means the machines best suited to operating in that environment are those that resemble us and can use the same tools and infrastructure we do.

Bipedal robots first entered industrial operations several years ago. The pioneers were Agility’s robots in Amazon warehouses at the end of 2023, then in 2024 and 2025 several manufacturers began testing their robots in real-world industrial settings: Apptronik at Mercedes, Figure at BMW, Boston Dynamics at Hyundai, and Sanctuary at Magna, while Tesla has long been preparing for similar deployments. In China, UBTECH introduced its robots at factories operated by BYD, NIO, and Zeekr, among others. And this year, after a delay of more than two years, humanoid robots are finally getting to work in Europe as well: Hexagon’s AEON has begun working at BMW’s Leipzig plant, while UBTECH’s Walker has been deployed at Rossmann.

For now, the most common task performed by humanoid robots is handling lightweight items such as small totes or car parts—although the strongest models can already manage loads of 25–30 kilograms. This is a type of work that, even in the most highly automated warehouses, is still performed primarily by humans. In the United States alone, millions of people work in such roles. Employing a worker for these tasks, including the cost of maintaining the necessary infrastructure, typically amounts to around $30 per hour (the worker’s actual wage is significantly lower). Industrial-grade humanoid robots currently cost between $100,000 and $150,000. Assuming a service life of 20,000 working hours and a maintenance cost of a few dollars per hour, this means they are roughly one-third to one-quarter as expensive as human labour. If prices eventually fall into the $20,000–$30,000 range targeted by manufacturers over the long term, that advantage could grow to a factor of eight to ten—provided, of course, that the quality and speed of the robots’ work match those of human workers. For certain tasks, this is already the case today, while for many others, it remains a distant goal.

Factories and warehouses, of course, already make extensive use of robots today, but these machines typically require a carefully structured environment—flat floors and purpose-built infrastructure. A humanoid robot, by contrast, can operate anywhere a human can work: the controlled and safe environment of factories is merely a stepping stone toward a much broader range of applications, including household tasks or caregiving.

What makes a general-purpose robot

„There are really only three hard things for humanoid robots. The real-world intelligence, the hand, and scale manufacturing.” – Elon Musk

This is the vision that makes the development of humanoid robots commercially attractive over the medium term. For them to move beyond inspiring experiments and become genuine products, six factors are key:

• Sensors can already perform virtually every function: robots can see, hear, and—with a few exceptions—even perceive touch.

• The necessary computing power is also available. As with self-driving cars, all critical computations must be performed on-board, since relying on network connectivity would introduce unacceptable latency. For the most part, this challenge has already been solved, even if demonstration videos occasionally still rely on some degree of external processing.

• Batteries have advanced dramatically thanks to the rise of electric mobility. They can be swapped quickly and provide enough energy to power a robot for at least half a shift, if not yet an entire one.

• The cost of actuators (highly precise, fast electric motors1 and the associated transmission systems that deliver remarkable strength relative to their size and weight) is the single most important factor in determining the robot’s overall price—today, these components account for well over 90 percent of the hardware cost. The greatest engineering challenge in building humanoid robots—and at the same time a major competitive differentiator—is the design of the hand. The hand is not merely one component among many; it is the interface through which a robot interacts with the world and ultimately becomes useful. Its importance is so great that some companies have chosen to specialise exclusively in robotic hands. While it is true that many tasks do not require particularly sophisticated dexterity, a truly general-purpose robot—one capable of performing a large share of the tasks humans carry out—will require the finesse and versatility of the human hand.

• Autonomy—the ability of robots to perform tasks by themselves (reacting to their surroundings without human intervention)—is the other defining characteristic that sets general-purpose robots apart from their industrial counterparts. In familiar environments, they are already reasonably capable of handling straightforward tasks involving only a few well-defined steps. Even so, Figure’s robots required more than 60 hours of training to learn the package-sorting task—the key question is how effectively they will be able to generalize knowledge acquired from one task to another. As with large language models, success depends to a significant degree on access to enormous amounts of training data. The challenge, however, is that while text and images already existed in abundance, robotic training data must largely be generated by the robots themselves—which requires either a great deal of time or a great many robots.2 Yet the ability to perform a task—or even a wide variety of tasks—is not enough on its own, robots must be able to operate safely, reliably, and with sufficient efficiency to become an economically attractive alternative to human labour.

• Manufacturing must scale in parallel with all of these developments, followed closely by the build-out of service and maintenance infrastructure. Humanoid robots are still produced largely through manual assembly: last year, only 13,300 units were manufactured worldwide, with 71% of them coming from two Chinese companies, Agibot and Unitree. However, Agibot alone produced 5,000 robots in the first quarter of 2026—the same number it had manufactured during the entire previous year. Among Western companies, Figure began mass production in February, followed by 1X in April. Both aim to reach a monthly production capacity of 1,000 units by the end of the year. Manufacturers are incentivised to deploy as many robots as possible in order to gain more data and experience, reduce production costs through scale, and establish an early market presence, while also having to ensure that they do not create expectations among customers that they ultimately cannot meet. This is a delicate balance, with two distinct strategies: Agibot and Unitree appear focused on maximising production volume, whereas Figure, Apptronik, and others are placing greater emphasis on delivering robots with superior capabilities and higher reliability.

If all of this is met, and autonomous general-purpose robots truly come to cost no more than a used car, they will not only take over boring, repetitive, and dangerous physical labour but will also become our everyday companions: Morgan Stanley projects that by 2050 there could be as many as one billion humanoid robots in operation worldwide.

As with almost every frontier technology, this is a race primarily between the United States and China: among the world’s top ten humanoid robot manufacturers, European companies are scarcely represented. It is telling that, as recently as this time last year, 1X was still a European company—until it announced in July that it would relocate its headquarters from Norway to California.

Diagnosis

Artificial intelligence will reshape the economy fundamentally. We are on the verge of one of the largest, fastest, and most profound technological transformations in history. Change is inevitable and unstoppable, and it will gradually spread across more and more sectors of the economy, increasing productivity. Countries that introduce regulations slowing the spread of machine intelligence will fall behind in the global competition; companies that adopt it half-heartedly will lose market share; and employees who are not proficient in it will be at a disadvantage in the labour market.

This is a decisive race with no way to opt out. Europe has long been aware of its technological decline and loss of competitiveness.1 Although it has always claimed to aim for the technological frontier, in reality, when it sees itself falling behind, it increasingly responds with protectionist measures to buy time for its companies to catch up. As a consequence, not only is it no longer at the forefront of technological innovation, but its citizens are also gaining access to the most advanced technologies later and later than other regions—consider self-driving cars as an example.

When it comes to artificial intelligence, this approach is not sustainable. In the coming decades, automation will spread to an ever-wider range of jobs, eventually affecting almost all white-collar work. An increasing share of office tasks will be carried out by artificial intelligence companies, much like SaaS operates today. In this vast market—amounting to several percent of global GDP—a handful of the most competitive firms will dominate. If none of these companies are European, nothing will offset the lost income, tax revenue, and consumption resulting from disappearing roles; all this revenue will flow outside the EU (not to mention the data). This is no longer merely about a missed opportunity, as in past decades, but about a lasting economic decline measurable in GDP percentage points.2 And if Europe chooses the other path and responds with bans or restrictions, it will deprive itself of the most important technology of our era and will fall permanently and irreversibly behind.

We must do away with illusions. There are three misconceptions that many in Europe cling to:

• Instead of artificial intelligence, we should build on our traditional strengths. Our best companies are specialised in technologies where the potential for breakthroughs is limited. The internal combustion engine is one of the most highly optimised machines in the world, and the automobile one of the most highly developed consumer products, but the innovations driving significant growth in the automotive industry today are in self-driving technology and batteries, two areas where Europe is lagging behind. Europe’s other strength, the pharmaceutical industry, is one of the sectors most exposed to artificial intelligence. The largest research and innovation (R&I) investments in Europe are still directed toward the automotive and pharma industries—in the United States, this was the case twenty years ago. We have reason to be proud of our achievements in these fields—and remain proudly stuck in the 20th century.

• It is sufficient to specialize in certain specific applications of AI. Without its homegrown general-purpose artificial intelligence model, Europe will become entirely dependent on American and Chinese models, paying with its profits, its data, and its sovereignty. This dependence will gradually extend to industry-specific applications in sectors crucial for the Union, such as automotive, financial services, telecommunications, healthcare, transportation, and retail. We do not need to create a universal AI system, but it is naïve to think that any specialized application can compete with those that use ten or a hundred times more computing power, whether for training or for inference.

• If we really wanted to, we could reach the global frontier even with a new company. The performance gap between the leading models of top artificial intelligence companies is fairly consistent. There are no miracles, clearly illustrated by how difficult it is for later entrants such as xAI or the leading Chinese firms to catch up with the OpenAI–Google–Anthropic trio. It is not that new companies will successfully break into the top tier; on the contrary, some of the existing ones will fall out of it.

Europe has only one frontier artificial intelligence company. The continent’s technological sovereignty and competitiveness in the 21st century can only be ensured if it has its own cutting-edge AI firm, over which it exercises strategic control and to which it can guarantee its companies access. There is only one such candidate: France’s Mistral AI. Whether in benchmarks or user evaluations, Mistral’s most advanced models—although somewhat behind the global leaders—are the best in the world outside the United States and China. Clearly, there is no other AI company in Europe that is even comparable. While this is a disadvantage because it does not encourage competition, it is also an opportunity: whereas Anthropic, Google, Meta, OpenAI, and xAI share the same market, Mistral—if supported by regulation—could become dominant in Europe. Moreover, thanks to the widespread use of nuclear energy, France is among the best-positioned countries in the world to meet the energy demands of data centres—a competitive advantage even relative to the United States.

Therapy

For Mistral—and, through it, Europe—to succeed, two things are required:

We must build the world’s largest data centre. Last May, the French national investment bank Bpifrance and the Emirati state investment fund MGX, together with NVIDIA and Mistral, launched a joint venture to build Europe’s largest AI campus near Paris by 2028, backed by an investment of around €8.5 billion. If it began operating today, this would be the largest data centre not only on the continent but in the entire world; however, by the time it reaches its final capacity of 1.4 gigawatts in 2030, it will rank as only medium-sized. This is a good and important step, but for Europe to keep pace with the global frontier, it must set an even more ambitious goal: building a 5-gigawatt data centre by 2030. The capital required to finance it must come from within Europe: France must create the conditions—such as by establishing a joint fund—that enable European countries to invest in the project, thereby giving them a stake in Mistral’s success.3

Regulation should not hinder tech companies, but support them.

• Mistral must be able to play by the same rules as its American competitors—for example, in building and managing large, integrated datasets—otherwise we will see a race in which one runner starts with their legs tied together. Simplification is important, but not sufficient: we must put an end to the practice of tying our companies up in administrative red tape as a precaution against largely non-existent risks. It is untenable that while the United States’ artificial intelligence action plan is titled “Winning the Race,” the European Union’s main ambition is to become the most tightly regulated AI market in the world. This approach comes at a high cost: economic stagnation, a lack of innovation, and ultimately a decline in living standards.

• Authorities must stop harassing innovative technology companies. Excessive, inconsistent regulation and overzealous enforcement by authorities drive dozens of tech companies—and their founders—out of Europe each year, mostly to the United States and to a lesser extent to the Emirates. The creativity of these people no longer enriches Europe; they go on to succeed with their ventures anyway, only no longer within the European economy.

• EU regulation must ensure the single (“one”) market and Mistral’s competitive advantage. Mistral’s greatest potential edge—alongside French nuclear energy—is its access to the EU market. Regulation should not merely guarantee this; it should also encourage European companies—and, in the case of European governments, require them—to choose a European provider.4 This restriction on competition should be offset by giving all European countries the opportunity to share in Mistral’s success through investment in the data centre.

There is no other way. Some of the assumptions above might turn out to be wrong. It may be that artificial intelligence will not have such a transformative impact on the economy, or that Mistral will fall behind its competitors despite all joint efforts. But even in that case, this roadmap is still worthwhile: it will lead to the creation of one of the world’s largest data centres and the strengthening of European innovation.

On the other hand, what happens if these scenarios materialise while we fail to act? In technology, Europe first slipped to second place behind the United States over the past century, then to third place behind China in the past decade. The Gulf region will be the next to overtake us—and not the last. Technological decline will lead to economic decline: in the long run, we cannot provide our citizens with a high standard of living if we do not produce the most advanced, highest-value-added products and services ourselves, but instead have to rely on others for them.

This is Europe’s last chance. The world is changing, unstoppably and inescapably; the only question is where we will end up at the end of the transformation. Those who take the lead will win; those who hesitate will lose. If we want the continent to be a shaper of the 21st century, not just a bystander, we must act. True, this plan is an implicit acknowledgment of France’s technological dominance5—but the alternative is not a fairer and more equal share in the AI revolution, but being left entirely behind. This is Europe’s best—and last—chance.

Cryonics involves people arranging for their bodies to be cooled to extremely low temperatures after death, in the hope that advances in medical science (usually envisioned as a kind of nanotechnology) will one day make it possible to revive them. Most people instinctively dismiss it, largely for two reasons: either because they view eternal life as belonging to the realm of religion rather than reason; or conversely because, on rational grounds, they consider anything involving resurrection to be pseudoscience.2 Still, the question remains: can cryonics work?

Subscribe now

It is based on the assumption that personality is determined by the brain’s physical structure; therefore, if the brain can be preserved in a sufficiently intact state after death, the personality could be preserved as well. To achieve this, before brain tissue is damaged, the blood is replaced with a solution with an extremely low freezing point—similar to those used in organ transplants—and the body is cooled to minus 124 degrees Celsius over a few hours. At this temperature, all chemical processes in the body—and thus all further decomposition—come to a halt: this is the process known as vitrification. (The body—or just the head, since preserving the brain does not require the rest of the body—is ultimately stored at minus 196 degrees Celsius, but this is for practical reasons: it is the boiling point of nitrogen.) In theory, this state can be maintained indefinitely without further damage.

The six unknowns

That is the idea. For it to work in practice as well, we must be able to answer yes to each of the following six questions:3

[1] Can the information the brain contains be preserved in sufficient detail? (Recording)

Current cryonic procedures are capable of preserving brain tissue in a state indistinguishable from that of a living person. Whether this is sufficient, however, we do not know. Even if we believe that we can already record every essential detail, it may turn out that this is far from adequate. Only those individuals whose personalities have been recorded with sufficient thoroughness will be able to be revived—but we do not know where this technological threshold is.

[2] Will the revived person be identical to the one who died? (Identity)

If every essential fragment of information can be stored and perfectly restored, then in theory yes. This process will, of course, involve some degree of error, but we have no idea what level of error would be acceptable. It may be that the person would merely have a few false memories, but it is also possible that their personality would be completely different. It may be that preserving personality would require such a low error rate that it will never be achievable in practice. Moreover, even if the revived person claims to feel identical to their former self, there is no objective way to verify this.

[3] Will revival ever be possible? (Outlook)

Since no further damage occurs during cryostasis—regardless of how long it lasts—from this perspective it makes no difference whether revival takes place in 50 or 250 years. If we give humanity enough time, we may assume that the necessary technology will eventually become available; the real question is rather whether it will be possible to restart those brains that are preserved at today’s level. As preservation technology is expected to advance over time, the later our death occurs, the better.4

[4] Can vitrification be started in time? (Neurodegeneration)

The more time passes after the onset of clinical death (the cessation of breathing and circulation), the greater the damage to brain tissue. Ideally, vitrification should begin as early as possible, but this is not legally permitted as long as there is hope for resuscitation—therefore, those preparing for cryopreservation must have a DNR order in place. Under normal circumstances, the threshold after which resuscitation is only possible with some degree of brain damage is about three to five minutes; however, numerous studies suggest that the damage is actually caused not so much by the cessation of blood circulation as by its resumption. If the restoration of circulation takes place under special conditions (such as cooling the body), a person can remain in clinical death for up to ten minutes without subsequent brain damage.

In reality, however, this does not have much practical significance. Whether the medical team performing the perfusion has three or ten minutes to arrive at the location makes no difference—this is completely unrealistic unless they are already on standby next door. If cryopreservation is to be taken seriously, it is only conceivable if, as death approaches, the patient is admitted to a hospital where the procedure can be carried out.

Although it may seem self-evident, it is still worth emphasizing that even if restoring the brain ever becomes possible, it could only be done in the state it was in at the time of death. This becomes particularly significant with the growing prevalence of neurodegenerative diseases, since someone who wishes to be frozen but suffers from Alzheimer’s may, upon realizing this, be tempted to hasten death. (In the United States, Alzheimer’s disease accounts for 14% of deaths among those over 85, but far more people suffer from the disease: the prevalence in this age group is 30–40%.) Of course, those who want their entire body cryopreserved must also avoid autopsy.

[5] Will the company be capable of safely preserving the body (or head) until reanimation? (Institutions)

From this perspective, it does matter a great deal whether we are talking about 50 or 250 years. Although natural disasters, sabotage, and war should also be taken into account, the main problem is that the stored individuals are legally dead (obviously, since otherwise they could not have been preserved): from a legal standpoint, vitrification is just an eccentric burial. It is possible that a law could be enacted that prohibits freezing (in which case the question arises as to what happens to those already preserved), but the greatest threat is that, for one reason or another, the company goes out of business. However, there are more than five thousand companies in the world that are over two hundred years old, so there is at least some chance that it could survive.

[6] Will there be someone who is willing to carry out the resuscitation and pay for it? (Costs)

The price of cryopreservation usually includes only vitrification and storage, not revival—whose cost we cannot yet even estimate. (The best-known provider, Alcor, does set aside a third to a half of the payments in a fund that serves as a financial safety net for storage and revival, but it is impossible to say today what that will be sufficient for.) Over time, it will presumably become cheaper, and beyond a certain point it may no longer be unaffordable, but even then: why would anyone revive a stranger they have never known, even if that person was their ancestor? It is similarly unlikely that revival will occur as a result of some government measure, since it creates far more problems than the societal benefit it brings.5 As for the role of the state, the most optimistic expectation is that it will not obstruct revival, not that it will assume responsibility for it.

Share

Estimating the probabilities

Since for cryonics to work it must overcome each of these obstacles, the probability of success can be expressed as the product of the probabilities of completing each individual step:

C = R x Y x O x N x I x C

It is helpful to estimate the six factors separately. R, Y, and O address the same question from three different angles: is it possible to reconstruct the personality from the brain at all, what information would need to be preserved for this, and whether current cryonics procedures actually achieve this. The problem lies with the latter two: we do not know exactly what would need to be stored in order to reboot consciousness; cryonics is based on the idea of preserving as much information as possible in the hope that it will be sufficient—but it may be that much of what we store is unnecessary or redundant, while essential elements are left out. Since there are no experiments involving revival, there is no feedback on the success of storage, no compounding knowledge, and thus assessing whether the improvement in the quality of information preservation is actually meaningful for revival is based on theoretical reasoning, not on verified scientific evidence. One reason for scepticism is that we aim to preserve the brain in an “off” state, even though it may have properties that can only be observed while it is functioning. These are scientific and technological questions for which we currently have no answers, so when we assign probabilities to them, we can rely at most on our intuitions. To move forward, let us assume—admittedly arbitrarily—a value of RYO in the range of 0.01–0.1.

In contrast, N, I, and C refer to social factors which we can estimate with somewhat greater confidence. For example, it seems fairly unambiguous that N can be maximized if death occurs in a hospital specially prepared for this; in that case, its value could be as high as 1.6 As regards I, the two oldest cryonics providers, Alcor Life Extension Foundation and the Cryonics Institute have operated for 54 and 50 years, respectively, demonstrating considerable resilience. Based on the Lindy effect, we can expect them to continue to exist for many more decades: the probability that these organizations will still be active in 50 or even 100 years lies roughly between 0.35 and 0.5.7 In the case of C, the ability to bear the costs is likely to increase over time (since the procedure is becoming cheaper both in absolute terms as well as relative to the economy), but the willingness to do so will most probably decline much more sharply over time after death. Personal involvement matters a great deal: C remains at its maximum—perhaps around 0.5 at most, due to regulatory uncertainty—while the deceased’s children are still alive, but it is at least an order of magnitude lower once even the grandchildren are no longer living.

So what is the product of the formula above? The later we die, the higher RYO is likely to be thanks to scientific progress (unless it turns out that cryonics is impossible even in principle), and the less time passes between our death and our revival, the higher I and C will be (while we assume that N can already be 1 today). The optimal scenario is to die as late as possible and be revived as soon as possible afterward, whereas the worst case is to die early and for revival to become possible only much later.

The probability of success—based on the above figures—is 0.02%, that is, 1 in 5,000.8 This also puts a price on cryonics: if resurrection is worth $100,000 to someone, then cryonics is worth $20 to them.9 And even this holds only if everything is geared toward success: if one is willing to be admitted to a hospital, forgo potentially life-extending resuscitation, and give up spending one’s final days at home surrounded by loved ones—to trade certainty for uncertainty. An even more difficult dilemma faces those diagnosed with Alzheimer’s: is there a point when it is worth putting an end to the deterioration of the brain? And those who are religious are effectively excluded from the outset: someone who believes in the soul can hardly imagine that revival would somehow force it back into the body, and those who believe in heaven would not even want that in the first place.

Leave a comment

Not a religion or pseudoscience—but not a technology either

In spite of all this, cryonics is not inherently irrational. Let us not forget: the brains of those born in the 1980s could be preserved at the technological level of the 2070s, using methods far more advanced than those used today. We will understand far more about how the mind works, and by then there may already have been experiments proving that it is possible to revive mammals after freezing.10 We won’t need to rely on a cryopreservation company for a hundred years, only twenty; the financial structure will be clear and predictable, and the legal environment well regulated.

The three major cryonics providers—the two Americans, Alcor and the Cryonics Institute, as well as Germany’s Tomorrow Biostasis—recommend that their clients take out life insurance policies naming them as beneficiaries and thus cover the cost of preservation, which ranges between $30,000 and $230,000, in instalments.11 The main argument is that, since we can’t know when death will come, it makes sense to prepare in advance. But as we’ve seen, if we were to die now, our chances would still be effectively zero—whether or not we have a cryonics contract—much like buying a lottery ticket doesn’t meaningfully improve our odds of winning the jackpot.

For someone seriously considering cryonics, it is much more reasonable to accumulate savings and revisit the question in 30 years: we will know much more about the possibilities, there may be new methods, and there may also be new providers to choose from.12 Not to mention that cryonics is just one of the methods to escape mortality, competing with the others: those who don’t die have no need for resurrection.

Since the mid-19th century (the earliest period for which we have reliable data), life expectancy has nearly doubled: in the Scandinavian countries— where it has long been higher than in most other countries and remains among the highest in Europe—it has risen from 45 years to almost 85.1 The largest rise—roughly from 50 to 70 years—occurred between the 1890s and the 1950s, thanks to three revolutionary changes: the spread of sewer systems, vaccinations, and antibiotics increased life expectancy at birth by up to three to four years per decade. By the middle of the century, infant and maternal mortality had essentially been eliminated, and we had brought most previously fatal infectious diseases under control. (In developed countries today, only 2–3 per thousand people die before their first birthday, and only one in twenty thousand women dies in childbirth.2 In a normal year, contagious diseases account for two to three percent of deaths.)

Meanwhile, maximum lifespan has changed much less: among those born in France between 1820 and 1920, the longest-lived 1% reached the age of 90, whereas someone born today has the same chance of living to 102.3 Thus, gains have come less from people living much longer lives and more from fewer people dying young. This is no longer enough to drive further increase—for that, almost everyone would have to live longer.4

From the middle of the 20th century, after these revolutionary changes, life expectancy gains have slowed significantly, to about one and a half to two years per decade. Over the preceding decades, we learned to defeat external pathogens—bacteria and viruses that make our bodies ill from the outside—but this line of progress has largely been exhausted: we are now confronting chronic and degenerative diseases caused by cellular deterioration and malfunction.5

The problem is that many systems of the human body—from the heart muscle and neurons to the hip and knee joints—begin to fail almost simultaneously after the age of 90. This could be seen as a kind of “design lifespan,” up to which the human body functions (reasonably well) under optimal conditions. To significantly exceed this, we need to change the current approach of trying to eliminate diseases one by one—if we reduce mortality caused by one age-related disease, we merely increase mortality from another. The solution would be to understand and inhibit the mechanism of ageing, which would delay the onset of all age-related diseases simultaneously.

Contra vim mortis non est medicamen in hortis6

The problem is that—although it perfectly fits the definition—ageing itself is not recognized as a disease, which makes it almost impossible to initiate—let alone fund—medical research specifically aimed at treating it.7

Subscribe now

There are, however, workarounds. The most obvious is animal experimentation, but the substances that work in mice prove surprisingly ineffective in humans—who, compared to animals of similar size, already have an exceptionally long lifespan.8 A somewhat better method is to study biomarkers used to estimate epigenetic age. These typically focus on one hallmark of biological ageing, DNA methylation (a chemical modification of DNA related to gene activity)—but this reflects only one of the many aspects of ageing.9

Third, if a drug treats several different diseases associated with ageing (such as diabetes, cardiovascular diseases, cancer, and neurodegenerative disorders), this at least raises the possibility that it acts on a shared underlying cause. Truly convincing evidence, of course, would be to demonstrate that an intervention meaningfully reduces all-cause mortality in healthy people—because that is what we are ultimately interested in, and because what matters is not avoiding death from a particular disease, but avoiding it altogether.

There are some promising candidates.10 Acarbose, used for decades to treat diabetes, by slowing the breakdown of carbohydrates, may indirectly help reduce body fat—particularly visceral fat—along with its many associated beneficial effects. GLP-1 receptor agonists, such as semaglutide, may yield similar results. Rapamycin, originally used to prevent organ transplant rejection, exerts beneficial effects by activating autophagy (the degradation and recycling of defective proteins), but due to its immunosuppressive effects, it is not currently suitable for routine long-term use in healthy individuals. 17α-oestradiol and common anti-inflammatory drugs like aspirin increase lifespan by reducing chronic inflammation—at least in animal experiments. And there are also quite a few substances that once generated great enthusiasm—such as metformin, resveratrol, or fisetin, a purported senolytic (a compound that selectively destroys senescent cells)—but they do not appear to have a meaningful effect on lifespan—not even in mice, let alone humans.

Hachi bun me11

What if we were to find a solution that increases healthspan more reliably and to a greater extent than any of these, without side effects—and that doesn’t even cost anything? There may be one: caloric restriction—a reduced intake of calories not causing malnutrition—has extended lifespan in every animal study to date, including those involving primates.12 Its effects extend far beyond the benefits of reduced body fat, and the drugs listed above appear to mimic precisely those effects.13 It is yet unclear whether continuous restriction or fasting—intermittent or periodic—is more beneficial.14 In any case, beyond the well-known lifestyle recommendations— avoiding smoking, minimising alcohol consumption, maintaining a healthy diet, exercising regularly, getting proper sleep— caloric restriction is currently the most effective tool for extending healthspan, far outperforming any drug that we know of.

Share

The past

The worst global catastrophe of the past two millennia was the mid-14th-century plague pandemic. The disease, originally of Asian origin, first appeared in Constantinople in the summer of 1347, from where it spread over the following years to the Middle East and—via Italy—across Europe. Within a few years the Black Death claimed at least 40 million lives (most of them in 1348). At least a third of the continent’s population perished, mortality rates of 50–60% were not uncommon in large cities. And since the world’s total population at the time was around 400 million, that means the plague wiped out one tenth of humanity—in today’s terms, this would correspond to an almost unimaginable 800 million people. This was the greatest catastrophe in recorded history. But was there any global catastrophe that outright threatened the extinction of humanity?

One nearly happened—and if it had, it could have had an impact comparable to the plague. On October 27, 1962, at the height of the Missile Crisis, a group of American destroyers dropped signalling depth charges on an unidentified Soviet submarine in international waters near Cuba, to force it to surface. The captain of the submarine, having sailed without radio contact for days and believing this meant war had already broken out, ordered the launch of a nuclear-armed torpedo against the American warships. He obtained the political commissar’s approval, but Vasily Arkhipov, the commander of the Soviet flotilla on board, disagreed and persuaded him to surface instead and to request orders from Moscow. Most historians agree that a Soviet torpedo launch would have escalated into nuclear war: according to Arthur M. Schlesinger, a historian who worked in the White House under the Kennedy administration, this was “the most dangerous moment in human history.” (That judgment is, of course, highly subjective, but this incident—as well as the false alarms of November 1979 and September 1983— are widely seen as the moments when we came closest to nuclear war.)

At least how many people have to survive for our species to persist? The number might be surprisingly small. After having set adrift Captain Bligh and his loyal sailors in the ship’s boat in April 1789, the following January Fletcher Christian settled with only 14 other men and 11 women on Pitcairn Island, where he burned and sank their ship, the Bounty. The island proved an ideal refuge: it was practically inaccessible, yet rich in food, fresh water, and arable land. When the sealing ship Topaz arrived in 1808, it found a thriving community of 46 people. The colony created by the Bounty mutineers shows that a group of just over two dozen people can survive long-term without any outside help—but their story also warns what happens when social order breaks down: by 1799, only two of the 15 men who landed in 1790 survived; the rest had fallen victim to murders, typically concluding disputes over women (and one committed suicide).

That may be enough for a community to function, but preserving genetic diversity over the long run requires a larger population. In an isolated population descended from very few common ancestors, hereditary diseases can, in unfavourable circumstances, accumulate within just a few generations. On the Micronesian atoll of Pingelap, only about 20 people survived a typhoon in 1775; one of them, the chief, carried achromatopsia, a rare congenital eye disorder causing complete colour blindness. That single carrier was enough for one tenth of Pingelap’s present-day population of 250 to be colour-blind, and for one third to carry the gene responsible for the condition. A similar pattern appears on “the world’s most remote island,” Tristan da Cunha in the South Atlantic, inhabited only since 1816 and home to roughly 250 people, who all trace their ancestry to just seven women and eight men. Two of the 15 settlers had asthma—five generations later, nearly half of the island’s inhabitants were affected by the disease.

However, genetic evidence shows that Eastern Polynesia was settled by a group of only about 70 people roughly a thousand years ago, which proved large enough to prevent such concentrations from developing. A study that modelled the minimum crew size for multigenerational starships similarly concluded that approximately 70 individuals are needed for long-term survival.2

The future

Could there be a catastrophe that leads to the collapse of human civilisation, or even to the extinction of mankind?

Throughout history, we have become increasingly resilient to natural disasters, and today it is hard to imagine one that would truly shake civilization. The eruption of a supervolcano or the impact of an asteroid a few hundred meters across could cause global effects through the debris injected into the atmosphere by the explosion. It is debated how much cooling such an event would cause and over what geographical extent; in any case, agricultural crops are extremely sensitive to cold—a single frosty night during the growing season can wipe out the annual rice harvest.3

We are also better equipped to deal with epidemics than ever before: we know incomparably more about them than we did even 200 years ago, and we operate modern health systems to cope with them. Our experience with pathogens is also a reason for optimism: the inverse relationship between infectiousness and lethality is well known, as is the fact that major epidemics tend to become milder over time—for these reasons, it is highly unlikely that a naturally emerging infectious disease could be more devastating than the fourteenth-century plague.

While we are better able to protect ourselves against natural disasters, technological progress has brought the emergence of new threats. It is often said that humanity has stockpiled enough nuclear weapons to destroy itself many times over; yet it is a mistake to assume that a total nuclear war would (everywhere) lead to Mad Max–style post-apocalyptic conditions.

In an extreme case, nuclear explosions could indeed claim the lives of hundreds of millions of people (though a figure in the tens of millions is more likely), but they would be concentrated in relatively small areas of land—almost exclusively in the Northern Hemisphere—and neither would the radioactive dust released into the atmosphere spread everywhere. The most significant global effect would be several years of cooling, caused by large amounts of soot reaching the stratosphere (the middle layer of the atmosphere, beginning at an altitude of ten kilometres), which could make agriculture fail and trigger famine.

However, regions that did not suffer direct strikes, where tropical latitudes or the ocean moderate the cooling, and where fruit cultivation, grazing, or fishing can provide sufficient food, could emerge from the cataclysm largely intact, preserving their technology and institutions. A global nuclear war would represent an unprecedented catastrophe and radically transform the face of the world, but when assessing its impact on human civilization and humanity’s long-term prospects, what matters most is how severely development is set back in the regions which survive Armageddon relatively unscathed.4

None of these catastrophes threaten the destruction of humanity, and not even that of civilization. During the medieval plague, at least 35–40% of Europe’s population died (comparable to the number of victims these disasters might cause in the worst imaginable scenario), yet it did not lead to collapse (even if it did reshape social structures). As a matter of fact, it is quite difficult to imagine an event that would result in our extinction. Perhaps the impact of a 10–15 kilometre-wide comet would qualify (this was roughly the size of the asteroid that wiped out all large terrestrial life forms, including the dinosaurs, 66 million years ago), but the probability of such an event is negligible: about one in a million per century; and we are getting close to being able to prevent such a collision.5

Share Letters from Prometheus

One can also imagine a scenario that would not involve the extinction of humanity, but the permanent loss of its freedom of action. If—for example through machine superintelligence—an oppressive totalitarian system emerges and becomes dominant across the globe, it might become nearly impossible to overthrow.6 Such a system would not only be free from temporary external constraints; no external force could even potentially counterbalance it. That also means that if civil liberties were to be curtailed in this empire, there would be nothing to prevent, mitigate, or reverse this.7 The new technologies—such as brain–computer interfaces—which carry the promise of an unprecedented expansion of human freedom may also become the ultimate tools of oppression.

What can be said about the probability of these catastrophes? Supervolcanic eruptions occur on the order of once every hundred thousand years, asteroid impacts by objects a few hundred meters across on million-year timescales, and those involving asteroids several kilometres in size on the order of tens of millions of years. It is true that pandemics decimated humanity far more often, every few hundred years, but our ability to defend ourselves against them has improved dramatically. The situation is different with human-made risks: nuclear war is a danger that has existed for only 80 years, yet among those listed it is arguably the greatest threat to humanity today. This warns us to take the challenges of the future seriously: the emergence of an oppressive world government seems unimaginable today, but may become the greatest threat by the second half of the century.8

The present

Of recent global catastrophes, the coronavirus pandemic naturally comes to mind first—with seven million deaths and a three percent contraction in the global economy, it clearly qualifies as one. There is, however, an ongoing catastrophe that claims far more victims: ageing.

To distinguish deaths caused by ageing from those due to other causes, it is not enough simply to count deaths that occur in, or are typically associated with, old age: any death that is more likely to occur in an older person than in a younger one should be attributed to ageing. For a 30-year-old, the probability of dying from cancer or heart disease within the next year is about 1 in 4,600—at age 60, the chance is 1 in 190, and at age 90, it is 1 in 10. The probability of death caused by an infectious disease is 1 in 19,000 at age 30, and 1 in 75 at age 90.9 If these probabilities remained at the level typical of youth—that is, if we eliminated the effects of ageing—human life expectancy could extend to several hundred years.10

If we do the calculation this way—arguably the only meaningful way, one we seldom adopt because we can do little about ageing and tend to accept it as inevitable11—we arrive at the conclusion that four-fifths of all deaths are caused by ageing.12

The first is replacing a car powered by an internal combustion engine with an electric one. The manufacturing of an average-sized European car results in roughly 7 tons of CO₂-equivalent greenhouse gas emissions. The additional emissions from producing an electric car mainly come from the production of its battery. The exact level of these additional emissions depends largely on how clean the electricity used in the manufacturing process is: producing one kilowatt-hour of capacity results in 60–100 kg of emissions, which corresponds to 5–8 tons for an 80 kWh battery. This is the upfront emissions penalty electric cars start with.1

How quickly can it make up for this through cleaner operation? Again, it depends on how the electricity needed for propulsion is generated. If it comes entirely from carbon-free sources, emissions are (almost) zero; if from gas and coal power plants, they can reach as much as 60 grams per kilometre. For cars powered by internal combustion engines, emissions from the fuel cycle—generated during the fuel’s production, refining, transport, etc.—must be added to tailpipe emissions, increasing it by 30% (that is, if the car’s emissions are 150 g/km, a total of 195 grams must be taken into account). As can be seen, there are quite a few variables, however, in the most favourable case for the electric car (when the extra manufacturing emissions are “only” 5 tons and it operates entirely on green energy), the balance tips after a little more than 25,000 kilometres (and in the least favourable case, at around 60,000 kilometres).2

If we hold two assumptions constant, we can also compare the replacement of a conventional car with an electric one to decisions made in other areas of life aimed at reducing emissions. Assuming it runs 15,000 kilometres per year over a 15-year lifespan, the total carbon-dioxide emissions of the above-mentioned conventional car amount to 51 tons, while that of the electric car is 12–29; that is, by switching from a conventional to an electric car, we save 1.5–2.6 tons of emissions per year.3

Importantly, we barely need to sacrifice anything in return: practically everyone who has driven a modern electric car says that it is no worse than one powered by an internal combustion engine, and aside from trips abroad, we no longer need to worry about range either. It is true that today they are still generally more expensive (though not significantly so in every market segment), but as batteries become cheaper, this difference will gradually disappear in the coming years. It is indeed true that operating an electric car requires access to a private charger, which is difficult to arrange in the city centre—but for those who live in the city centre, owning a car may be less necessary in the first place.

Share

Beyond replacing a conventional car with an electric one, there are three individual (family) decisions that can achieve a significant reduction in emissions.

Air travel is probably the most straightforward to estimate: as a rule of thumb, take-off adds about 100 kg of CO₂ per passenger (in economy class), and each 1,000 km adds about 150 kg more. For a family of four on a round trip, that is eight passenger-legs in total, meaning that a relatively short 800-kilometer flight results in a total of 1.8 tons of emissions.4 If we go by car, this is 0.4–0.5 tons, if by electric car, 150 kilograms at most. That is a difference of 1.3–1.7 tons.

Replacing a gas boiler with a heat pump is a bit more difficult to quantify. Burning 1,200–1,800 cubic metres of natural gas (the typical per-household average consumption in Europe, though much depends on the climate) emits about 2.2–3.4 tonnes of CO₂. If the heat pump runs on fully clean electricity, essentially all of that can be avoided. The crucial question is what share of electricity generation is emission-free (nuclear and renewables), but in general it can be said that a heat pump reduces household carbon-dioxide emissions by one-half to two-thirds (1.1–2.3 tonnes), and if combined with rooftop solar panels, it can even bring emissions down to zero. Although converting existing buildings from gas heating to heat pumps is only very rarely considered, it is clear that when building a new home, heat pumps are the climate-friendly solution.

Finally, compared with the measures mentioned above, reducing meat consumption—especially beef consumption—can bring a small, but still noticeable reduction. Producing the protein found in 100 grams of beef results in 50 kilograms of carbon-dioxide emissions, while for pork and poultry the same figure is only 6–8 kilograms. (Cattle farming also requires far more land: 164 square metres per 100 grams of protein, whereas pork requires only 11 and poultry 7 square metres.) Skipping one hamburger every two weeks saves 250 kilograms of emissions annually (or you can switch from beef to pork or poultry, which still saves 210–220 kilograms).5

With driving, flying, heating, and meat consumption, we have largely exhausted the measures capable of producing substantial reductions in our carbon dioxide emissions. Calculations show that, compared with these, other individual actions yield almost negligible savings.6

One of the most common misconceptions is that locally produced foods come with significantly lower greenhouse gas emissions. In reality, distance matters a lot less than people think: transportation accounts for only about 5% of emissions related to food production, and a much larger share of this comes from road transport than from shipping (ocean freight is incredibly efficient). If our goal is to support local farmers, choosing locally produced food does indeed make a lot of sense—but from a climate protection perspective, it is essentially meaningless. Food packaging does not matter much either: it contributes even less than transportation, only about 4%, to the carbon footprint of food, and this is more than offset by the fact that packaging protects food and extends its shelf life, thereby reducing waste.7

Subscribe now

I think there are three things worth taking away from all this:

• If we want to reduce our personal carbon footprint, it is worth doing the math instead of relying solely on our gut feelings or the last newspaper article we got to read. A hundred years of reading in dim light or washing clothes in cold water will not offset a family weekend in Dubai (literally: lowering the washing temperature saves about 40 kilograms of emissions per year, while a 4,000-kilometer round-trip flight for a family of four results in 5,600 kg of emissions), and consistently shutting down all household appliances instead of leaving them on standby saves about 35 kilograms of carbon dioxide per year—equivalent to the emissions of four Big Macs. There is nothing wrong with overseas travel, nor with household asceticism; just do not think small household savings cancel out a flight.

• Reducing carbon dioxide emissions does not have to make our lives worse. Electric cars are no worse than internal combustion vehicles, just as heat pumps are no worse than gas boilers.

• Finally—and most importantly—while electric cars and heat pumps are essential for making use of clean electricity, it matters at least as much how that electricity is generated. That requires (the combination of) nuclear power and solar energy.

1. we know how to do it, and all the necessary technology is at our disposal;

2. we committed ourselves, and the data shows that we are heading in the right direction (though perhaps not fast enough);

3. human history is full of examples that prove we can adapt very quickly and very efficiently when necessary.

We’ve got all it takes

The central challenge is to curb global warming, chiefly by reducing carbon dioxide emissions. The solution is electrification—the widespread use of already available clean electricity-generation and storage technologies.

Throughout its entire history, humanity has met almost all of its energy needs by burning fuels: wood, coal, oil, natural gas. This is extremely wasteful—exactly how wasteful depends on the method, but overall we utilize only about 35% of each unit of energy produced, while 65% turns into waste heat.1 In the coming decades, this will change: by the middle of the century, wherever possible, fossil fuels will be replaced by clean electric energy.2 Electrification will lead to a fourfold increase in efficiency in road transportation, a threefold increase in heating and cooling, and a two- to two-and-a-half-fold increase in electricity supply and low-temperature (below 200°C) industrial processes (meaning that correspondingly less energy needs to be generated for the same amount to be used). These four areas, which can be totally electrified with currently available technology, together account for about 85% of total primary energy demand—as well as 85% of greenhouse gas emissions from energy production. Of the major sectors, only parts of industry (such as steel, chemical, fertilizer, and cement production) and aviation are not (yet) covered.3 Once this transition will be complete, the only remaining application that still relies on combustion will be rocket propulsion.

A commonly used metric for comparing energy-production technologies is energy return on investment (EROI), which shows how much usable energy an energy source provides compared to the energy needed to extract, process, and deliver it. This also depends on the area of use: for example, it makes a difference whether natural gas is used for electricity generation or heating, because in the former case significant losses must be taken into account, while in the latter, although heat production is more efficient, the gas still needs to be transported to the point of use. To make EROI values truly comparable, the energy that is actually usable at the final point of consumption (for lighting, heating, propulsion, etc.) must be considered—including all energy inputs up to that point—and the calculation must cover the entire life cycle of the system, as well as any supplementary infrastructure needed to ensure security of supply (including energy storage if necessary). Due to these methodological difficulties, the literature presents quite different EROI values; however, there is general agreement that when it comes to electricity production, nuclear energy (alongside hydropower) has the most favourable energy-return performance.4 Critics of electrification often use EROI calculations to argue that the transition will lead to energy deprivation, but the numbers show the exact opposite: substituting fossil-fuel electricity with a combination of two-thirds renewables and one-third nuclear power would not lower, but actually raise EROI. Fossil fuels offer a genuine advantage only in applications where heat generation itself is the primary objective, and even there, they are increasingly replaceable—any analysis suggesting otherwise fails to account for technological advances over the past decade.

Share

Thanks to nuclear energy, we have access to a clean, weather-independent, affordable, reliable, and safe energy source with almost limitless potential.5 We have not yet exploited this, but currently 62 reactors are under construction,6 and small modular reactors are expected to offer more flexible—and hopefully somewhat faster and cheaper—ways to satisfy demand.7 Recent technological changes have made nuclear fusion a focus of growing interest,8 spurring increased private-sector involvement, with some companies even planning to begin production around 2030. This optimism is partly fuelled by fewer bureaucratic obstacles compared to public (and especially international) projects and by shorter prototype development cycles; nevertheless, the technological challenges remain enormous. The problem with fusion is that by the time it actually works, its significance will be much smaller: we need to eliminate fossil-fuel-based electricity production now, not in 25 years.9 For this, fission is already a mature, deployable technology, not a future miracle solution.

For the short-term storage of electricity (ranging from a few hours to a few days), as well as for use in mobile devices, advanced, high-performance batteries are required. Although there are other emerging technologies—such as sodium-ion or solid-state batteries—today’s electric vehicles and utility-scale energy storage systems almost exclusively use lithium-ion batteries. A key factor in the expansion of electromobility—besides the development of a high-performance charging infrastructure—is the increase in battery energy density: for long-distance road transport, 400 Wh/kg, for electric aviation, 650 Wh/kg is required.10 The battery demand for the full electrification of transportation could reach 14 TWh per year by the mid-2040s—nearly thirteen times the current production.11

The transition to sustainable energy requires the increased extraction and refining of raw materials essential for electricity generation and storage—primarily lithium, nickel, and manganese, as well as cobalt, graphite, and copper. Even currently identified reserves are more than sufficient—in the case of lithium, the scarcest of them, enough to cover demand about five times over—and from the late 2030s onward, recycling will play an increasingly important role. The theoretical availability of raw materials does not, of course, eliminate the risk of temporary supply shortages and resulting price fluctuations, nor the geopolitical challenges arising from the fact that reserves are concentrated in the territories of just a few countries—but this is equally true for the fossil fuel market, with the key difference being that different countries are involved (which, of course, matters a lot).12

We’re on the right track

If we want to assess our success in curbing climate change, how can we determine whether we are moving in the right direction? What is the most important metric, what evidence helps settle the debates?

Since global warming is (primarily) caused by carbon dioxide emissions, this is where we need to start: total global emissions in 2024 amounted to 38.6 billion tons.13 According to forecasts by the International Energy Agency, emissions will peak this year or within a few years at most, followed by a decline—initially slow, but gradually accelerating. This is possible because, although humanity’s total energy demand continues to grow, part of this growth is already being met by clean energy. In recent years, the annual increase in clean energy production was only enough to keep pace with the rise in demand, but as it gains momentum, it will increasingly enable the replacement of existing fossil fuel capacities year after year. In the four years between 2020 and 2024, solar power generation rose from 2,100 to 5,200 TWh, and wind power from 3,900 to 6,100 TWh, representing average annual growth rates of 25 and 12 percent, respectively—in other words, the amount of electricity generated by solar power doubles every three years.14

What ultimately matters is the share of total energy consumption that comes from clean energy: this proportion is still only 16 percent today,15 but IEA’s forecasts for 2030, 2040, and 2050 are increasingly optimistic year by year, indicating that the transition is accelerating.16 Overall, an ambitious but realistic goal is to reach 20 percent by 2030 and 40 percent by 2040.

But it’s not simply about a decrease in carbon dioxide emissions: it’s a decrease that occurs while the economy continues to grow. In the fifteen years between 2008 and 2023, the United States reduced its per capita consumption-based carbon dioxide emissions by 22%, and Germany by 29%, while their per capita GDP grew by 22% and 13%, respectively.17 These are not isolated cases: there are at least 30 countries where this decoupling has occurred.18 It is true that this has not yet happened globally—per capita emissions have remained constant at 4.7 tons for 20 years19—but these examples clearly show that it works: economic growth does not necessarily rely on fossil fuels.

Leave a comment

After greenhouse gas emissions, humanity’s second most significant negative impact on the environment is arguably the expansion of agricultural land. This directly leads to a reduction in forested areas and indirectly to a decline in biodiversity. 71% of the Earth’s land surface is fertile (the rest being desert or ice), and 46% of that is used for agriculture. One might assume that no increase in crop yield could ever fully meet the growing food demands of an expanding population, and consequently the area used for agricultural production must inevitably increase as long as the population grows. But this is not the case at all: in fact, the size of agricultural land has remained unchanged for the past 30 years (4.8 billion hectares).20 During that time, the human population has grown by one third—how is this possible?

Largely thanks to fertilizers (and pesticides). Many fear that we are merely replacing one unsustainable practice with another, and that fertilizers may ultimately cause as much or even more harm to nature than farming on a larger area without them. It is indeed true that the amount of agricultural land per capita decreased to two-thirds between 1960 and 1990, while fertilizer use nearly tripled—and without fertilizers, it would be impossible to produce enough food to support humanity.21 However, fertilizer consumption has been stagnant for at least ten years (at 210–215 million tons): it is not yet decreasing, but it seems we may have passed the peak.22 Although fertilizer use continues to rise in poorer countries, wealthier ones have for decades achieved gains in agricultural productivity (food output per unit of land) through increased efficiency.

Whether we look at carbon dioxide emissions, the consumption of fossil fuels, changes in agricultural land use, or fertilizer usage, the data shows that we have either already reached the turning point or are very close to it. As for the pace, there is no reason for complacency, but at least we are already moving in the right direction.

We’ve done it before

We can be optimistic about curbing climate change not only because we already have everything we need to do so and because we are on the right track, but also because this will not be the first similar challenge that humanity has overcome.

The two major environmental problems of the 1980s were acid rain and the thinning of the ozone layer. The former threatened freshwater and soil with contamination, while the latter involved increased ultraviolet radiation, which can cause skin cancer and eye damage. It was clear that these were issues no single country could tackle alone—international cooperation was necessary.

Today, both problems are a thing of the past. Thanks to strict regulations introduced in Europe and the United States, sulphur dioxide emissions—the cause of acid rain—have now dropped to one-fifth of their 1980s levels in these regions (and by half globally).23 In 1989, 43 countries pledged to cut global emissions of ozone-depleting chlorofluorocarbons (CFCs) by half within a decade; by the turn of the millennium, 174 countries committed to a complete ban. The pace of reduction surpassed all expectations: CFC emissions dropped by 30% in a single year and by 85% within ten, and have virtually ceased by today.24 When we commit ourselves, we can achieve results very quickly.

Of course, tackling climate change is more complex than banning freon or filtering the smoke of coal power plants. It requires a radical transformation not just of a small segment of the economy, but of the entire system. However, there have been precedents for such changes as well.

Throughout most of human history, the most important raw material was wood. It was the primary source of energy and the main building material, not to mention that deforestation freed up land for food production. This had clear consequences: in France (where we have reliable historical data), by the end of the 17th century, half of the forests had disappeared, leading to a dependence on imported timber for shipbuilding. Despite the comprehensive regulations introduced in response—the first of their kind in the world—forested areas continued to shrink by another third over the following 150 years.25 It seemed impossible to separate the economy from wood use and that the process was unstoppable.

Then came coal. In France, between 1820 and 1860, the use of wood as fuel dropped to a quarter, and by the turn of the century, it had essentially ceased in all areas where it could be economically replaced.26 (Meanwhile, rye was replaced by potatoes, which could meet the demands of a growing population using less land, so forests no longer had to be cleared for food production either. These days, 31.4% of France’s territory is covered by forest—more than at any time since 1600.)

Today, many continue to believe that the economy is inextricably tied to oil and natural gas. While this was indeed the case twenty years ago, just as coal replaced wood as the primary energy source over a few decades, so are nuclear and solar energy nowadays replacing fossil fuels, confining them by mid-century to specialized domains such as (long-distance) aviation, specific industrial processes, military uses, and spaceflight.27

These examples show that both intention and technology are necessary for change—neither is worth much without the other. Today, we have both. As environmental data scientist Hannah Ritchie points out, we no longer have to choose between creating a good life for people living today and providing opportunities for future generations; the two can go hand in hand.28 It is possible, but it will not happen on its own: it will be the result of the tireless work of people committed to improving the world and willing to act in the long-term interest of humanity—an effort to which the doomers contribute nothing.

Overpopulation doom

The dogma that humans—especially too many humans—are destroying the Earth is as old as civilization itself.1 The modern formulation of this thesis is credited to Thomas Malthus, who, in his 1798 book An Essay on the Principle of Population, warned that population growth inevitably leads to scarcity and, ultimately, famine. We know what has happened since then: the Earth’s population has grown more than eightfold, from 985 million to 8.1 billion, while the proportion of people living in extreme poverty—on less than two of today’s dollars a day—has dropped from 80 percent to 9 percent.2 (In fairness to Malthus, it should be noted that he was largely correct about the periods preceding his time, and it was the Industrial Revolution, just emerging in his day, that changed everything.) But it is, in fact, not Malthus from whom the intellectual roots of today’s doomsayers originate.

The 1960s were one of the most successful—could there be clearer evidence than the Moon landing?—and wealthiest decades in the history of the United States: between 1958 and 1968, GDP per capita rose by 40 per cent, a surge matched only by the explosive economic growth at the turn of the previous century. Ironically, it was precisely during this time of great prosperity that biology professor Paul Ehrlich (a lepidopterist by profession, not to be confused with the German Nobel laureate physician of the same name) wrote The Population Bomb. In this book—ignoring completely the 170 years that had passed in the meantime—Ehrlich took Malthus’s theses to the extreme: “In the 1970s and 1980s hundreds of millions of people will starve to death in spite of any crash programs embarked upon now. At this late date nothing can prevent a substantial increase in the world death rate,” he wrote. In an interview promoting the book, he also said: “in the next fifteen years, the end will come, and by the end I mean an utter breakdown of the capacity of the planet to support humanity.”3 The promotion was highly successful—the book sold over two million copies, making its author wealthy.

By now, it is clear that just as Malthus was wrong before, so was Ehrlich: since then, the population of humanity has doubled again, and in the meantime, the proportion of people exposed to hunger in developing countries has dropped to a third, from 35% to 13%.4 The rapid population growth in the second half of the last century occurred because the drop in the number of children families chose to have lagged about one generation behind the sharp decline in child mortality (in earlier eras, families had on average five to six children, just like in the mid-20th century, but back then, every second child died). The fertility rate—the number of children a woman gives birth to over her lifetime—has by now fallen below 2.0 across Asia and South America as well, down from 5.8 in the mid-century, just as it had earlier in Europe and North America, where this decline began around the turn of the 1900s.5 In Europe, this figure is currently 1.4—meaning that the number of Europeans is halved in two generations. In fact, the only region in the world where the population is still growing dynamically is Africa, and everywhere else, it is economic growth—along with accompanying changes in healthcare, education, and labour markets—that has led to the slowdown in population growth. We are already past the point when the number of children peaked (this occurred sometime between 2017 and 2021, depending on whether we consider children under the age of 5 or under 15), and when they reach old age, the human population will also peak, only to enter a centuries-long decline.6

Ehrlich did not settle for simply outlining the problem; he also proposed radical “solutions.” „We can no longer afford merely to treat the symptoms of the cancer of population growth; the cancer itself must be cut out,” he wrote.7 Among other ideas, he suggested that in wealthy countries substances inducing temporary sterility could be mixed into drinking water (and expressed some frustration that this was probably unfeasible), and proposed that men undergoing vasectomies should receive financial incentives. But this was nothing compared to the fate he envisioned for the inhabitants of poor countries: according to his proposal, in the case of “hopeless” countries—such as India—wealthy nations should have halted all food aid, letting people starve to death.

While he and others like him were lamenting the now “inevitable” catastrophe, others were working to prevent it. Indian agricultural researcher Mankombu Sambasivan Swaminathan began cultivating a new, higher-yielding dwarf wheat variety on a one-hectare experimental plot in 1964. His efforts were crowned with success: in 1968, the very year Ehrlich’s book was published, India’s wheat harvest surged to 140 percent of the previous year’s yield, and a few years later the government declared the nation food-self-sufficient and renounced all further aid. One could hardly imagine a more striking example of the difference between where the Ehrlichian and the Swaminathanian mentalities lead.

Resource doom

Alongside overpopulation, another central trope of the apocalypse industry—treated as a self-evident truth which scarcely requires proof—is the claim that the Earth’s natural resources are running out. The idea is something like this (although it’s hard to pin down, since its proponents rarely articulate it clearly): a certain resource exists for a while, and then suddenly it doesn’t. (Perhaps some take a more nuanced view, but in any case, the possibility of gradual adaptation and substitution has no place in their worldview.)

But the world doesn’t work that way. If demand for a raw material consistently exceeds supply, its price begins to rise. In response, economic actors either increase production or seek alternative, substitute materials (or technologies). Both lead to a decrease in prices—the former through increased supply, the latter through reduced demand. Of course, this adaptation doesn’t always happen quickly: the price of oil, for example, only returned to its previously typical range 13 years after the 1973 embargo, in 1986—but then it remained there steadily for nearly 20 years.8

In 1980, the most prominent doomsday celebrity of his age, Paul Ehrlich made a bet with economist Julian Simon about how the prices of five raw materials he selected (copper, chromium, nickel, tin, and tungsten) would change over the next ten years. Ehrlich bet on prices rising, while Simon wagered they would fall. The inflation-adjusted prices of these five materials—once again: chosen by Ehrlich—dropped to less than half (on average to 42%) between 1980 and 1990: the prices of tin and tungsten fell to one-third, and not a single one of the five increased.9 (Meanwhile, global population and economic growth—the two factors which, in Ehrlich’s reasoning, were supposed to lead to resource depletion—increased by 20% and 35%, respectively.) The outcome speaks for itself, but Ehrlich (unsurprisingly) claimed that Simon had simply been lucky with the time period.

Would the outcome really have been different if another decade had been chosen? In the 1970s, the prices of all five materials increased—but this isn’t particularly surprising, since Ehrlich obviously selected ones whose prices had been rising. In the 1990s, the prices of all five decreased; in the 2000s, all five increased again; and in the 2010s, all but tungsten fell once again. If any conclusion can be drawn from this—or from the oil example mentioned earlier—it’s that a ten-year period is simply too short to make any meaningful statement about the depletion of natural resources.

What about the longer term? Over the five decades from 1970 to 2020, the average price of the five raw materials increased by just 4.5%—which means that, adjusted for inflation, they cost almost exactly the same as half a century ago. Of the five, only chromium saw a significant price increase during this period, but even that doesn’t support the resource depletion narrative when viewed over a longer timespan: in 2020, this shiny metal cost almost exactly the same as it did in 1900. The same holds true if we look at other materials, ones not included in the Simon–Ehrlich bet but crucial to the energy transition: the prices of lithium, manganese, and cobalt are essentially the same (or even lower) than they were in the 1970s; only graphite shows a notable increase. The truth is that, over historical timescales, the price changes of non-energy raw materials show absolutely no consistent trend, despite the fact that their extraction volumes have increased dramatically—for example, between 1970 and 2020, cobalt production increased nearly tenfold, nickel sixfold, and copper three and a half times.10

The Ehrlichians’ prophecies about the depletion of raw material resources should be taken just as seriously as those about overpopulation, or claims like “England will not exist in the year 2000.”11

Growth doom

The other similarly harmful—currently more fashionable—ideology, degrowth, is based on the core assumption that a lifestyle making use of modern technological advancements is inherently environmentally destructive. This is false: measured on a consumption basis, per-capita carbon-dioxide emissions have been declining in wealthy countries for at least fifteen years even as their economies have continued to grow—showing that decoupling emissions from economic growth is not only possible but already underway.12 The wealthier a society is, the less polluting energy sources it uses for cooking and heating: it moves from wood to coal, from coal to natural gas, and from natural gas to electricity. Today, 40% of the world’s population still predominantly uses wood and charcoal—the single most effective means of reducing global air pollution is to help them move up this energy ladder.13

It is economic growth that makes environmentally friendly technologies such as solar panels or electric cars increasingly efficient, affordable, and therefore competitive—a prerequisite for their broader adoption. This allows clean energy to cover part of the increase in humanity’s energy demand even today and to increasingly replace existing fossil-fuel capacities in the coming years. Technological advancement has also led to household appliances, light bulbs, screens, heaters, etc., consuming less and less energy while delivering much better performance. If we want to do good for the environment, we should not cut back the economy—on the contrary, we need economic growth.

Let us also not forget that 9 percent of the world’s population still lives in extreme poverty, on less than two dollars a day.14 It was due to technological progress and the economic growth it enabled that this share has fallen by three-quarters over the past thirty years. To halt growth would, in effect, mean abandoning the poorest.15 (The fact that any nation embracing the ideology of degrowth would be committing economic suicide and would permanently fall behind in the competition among nations is merely the icing on the cake.)

Degrowth is portrayed by its advocates as a kind of utopia, where the relentless pursuit of economic growth is abandoned in favour of other aspects of quality of life, environmental sustainability, and a fairer distribution of resources. While it is undoubtedly unique in the sense that it does not claim to bring about prosperity—on the contrary, it explicitly promises to lead to impoverishment—this does not mean that it is a novel idea or an economic system that has never been tried before. As a matter of fact, there is a country that has been living according to the principles of degrowth for 75 years.

When the Democratic People’s Republic of Korea was established in 1948, its per capita GDP was around $2,460—significantly higher than that of its southern neighbour. Today, that figure stands at $1,570 per person, which corresponds to an annual decrease of 0.6%, meaning the country has followed a moderate degrowth trajectory. During the same period, per capita national income in the South has risen above $40,000, meaning that today, Southerners live roughly 25 times better than Northerners.16 Let us set aside the lack of freedom—which is almost as destructive to the human spirit as poverty itself—and consider only the quality of everyday life: is there anyone who would rather live in North Korea than in South Korea?

Of course, one could argue that Korea is an extreme example, and that degrowth only applies to countries already living in relative prosperity. But even if we move beyond absolute poverty, relative poverty still remains. In 1990, at the time of reunification, the gap between East and West Germany was not 25-fold, but only twofold, and it’s not as if life in the East had been entirely unbearable—yet people still wanted a Mercedes, not a Wartburg. Did any sane person ever suggest that the standard of living in the Federal Republic should be brought closer to that of the Democratic Republic, rather than the other way around? When the Wall fell, in which direction did people cross?

Doomsday hype isn’t just wrong, it is harmful

The main problem with doomsday prophetism is that it is factually unfounded: the conditions said to lead to a coming ‘serious crisis,’ ‘catastrophe,’ or ‘collapse’ (whatever the pundits may mean by these terms) within a few decades either do not exist at all (as in the case of overpopulation or looming resource exhaustion) or do not produce the outcomes claimed (as with climate change).17 And we could stop here—but there are three additional serious issues with this ideology.

• It is not falsifiable. Doomsday predictions are either framed so vaguely that they can’t be tested, or they conveniently schedule the fall of civilization for exactly twenty years in the future. In the words of Nobel laureate Indian economist Amartya Sen: „The new prophets have learned not to attach specific dates to the crises they foresee, and past failures do not seem to have reduced the popular appetite for this creative genre.”18

• It fuels anti-scientific sentiment. By constantly scaring people with nightmarish scenarios that never come true, and wrapping these in the guise of science, the heralds of the apocalypse create a climate in which people no longer listen to real science either.19

• It discourages action. Naturally, not every environmental problem can be addressed with equal effectiveness at the same time, but that is not the goal—the (most pressing) goal is to curb global warming, for which the means is reducing carbon-dioxide emissions, and the solution is electrification. It is entirely achievable, but it will only succeed through serious effort: the result of determination, collective thinking, and a great deal of work—which denial and defeatism not only fail to support but actively work against. We must not give up the chance to address the most important, most urgent problems and to seek and find solutions to them.

Although his name is less well-known outside the Anglo-Saxon world, the toxic ideas of Ehrlich and his followers have seeped into everyday life and pop culture.20 What is truly shocking, however, is that Ehrlich can still be considered a point of reference today, despite the fact that his bestseller is essentially a crude caricature of a theory that had already been disproven a century and a half earlier, and not a single one of his claims has ever come true. (Though it is also true—and in a way more understandable—that many people, on the other hand, remain silent about—or outright deny—following in his footsteps.)

The characters of Ehrlich and Swaminathan may both feel familiar today, and while Swaminathan reminds us of the great achievements that human ingenuity, determination, and willingness to act can bring about, it is chilling to imagine a world in which people like Ehrlich had influence over how affairs are run, and their delusion-driven, inhumane proposals were enacted—instead of tackling the challenges and working to find solutions to them.

Diverse approaches, similar conclusions

The first significant, comprehensive study on the labour market effects of machine intelligence was conducted by researchers at the University of Oxford more than a decade ago, when large language models were not yet in existence.1 Since the emergence of ChatGPT, analyses have projected rather different visions of the future, depending on the temperament and background of their authors.

Goldman Sachs’ March 2023 report on the economic growth impacts of generative artificial intelligence identified 13 out of 39 general workplace activities listed in the O*NET employment database (such as information gathering, communication, coordination) as being exposed to automation by machine intelligence, and assumed that AI is capable of performing tasks rated at level four on O*NET’s seven-point difficulty scale.2 If so, this would mean the automation of roughly a quarter of current workplace activities—allowing employees to spend more time on tasks that machines cannot perform. However, in 7% of jobs, artificial intelligence could fully replace human labour.3 Altogether, over a decade or more, this could result in a 16% increase in productivity—a much greater change than, for example, the spread of personal computers brought about. The investment bank’s analysts also examined what would happen if machines performed at level six instead of level four: in that case, the share of replaced work time would increase only slightly—from 25% to 27%—but the productivity gain would be significantly greater, reaching 33%. (Level four tasks include things like filling out tax returns for a small business or creating a work schedule; level six involves tasks such as organizing the programme for a large conference or gathering data for a complex scientific report.)

McKinsey also based its analysis on the O*NET database but took a different approach: it examined approximately 2,100 workplace activities to determine which of 18 different skills (such as perception, navigation, coordination, reasoning, creativity) are required to what extent for each activity.4 By comparing this with the timeline of when machines are expected to reach the necessary level in each skill, it becomes possible to estimate when machines will be able to replace humans in specific tasks. However, technical feasibility is only the first step: automation also requires integration into existing work environments and processes, it must be economically viable, and it needs to be accepted by users.5 The transition will be fastest in countries that are best prepared and where wages are high—since replacing human labour is most cost-effective there. According to the consulting firm, in the most advanced economies, artificial intelligence could replace up to half of the time currently spent on work activities as early as the end of this decade—and around 85 percent by the mid-2030s—contributing an annual 3.5 to 4 percentage point increase in productivity.6

The GATE model—developed by the research institute Epoch AI—adopts an entirely different approach and can be summarized as follows.7 Investment in the development of artificial intelligence enables an increase in hardware quantity, as well as improvements in both hardware and software efficiency. This leads to a growth in computational capacity, which is used partly to train new models and partly to operate existing ones: the new models enable the automation of an increasing number of tasks, while the existing models perform economically valuable work. This work generates revenue, which partly drives consumption and partly feeds back into AI development, further reinforcing the cycle.

GATE takes into account nearly forty parameters, which can be adjusted via an interactive interface.8 With the default settings, the model predicts automation of 69% of activities by 2030 and 100% by 2034. With “conservative” settings (which involve changing the values of only three parameters), the projected level of automation is 51% in 2030, 81% in 2035, and 100% by 2040.

It is noteworthy how similarly McKinsey (in blue-grey) and Epoch (in green), despite working with completely different methodologies, see the development of automation:

However, these are necessarily simplifying models that do not take into account a range of constraints: never-ending training runs, exponentially increasing electricity demand, potential data scarcity—not to mention raw material supply or regulatory hurdles. It is worth thinking of this as the pace that the most advanced large economies, such as the United States, can at most achieve over the next five years (and which city-states like Dubai, Hong Kong, or Singapore may even surpass), but it is certain that automation cannot continue unabated beyond that. Two factors will slow it down: first, the physical limits of increasing computing capacity (energy and raw materials) will gradually become more and more decisive; second, in order to carry out physical tasks, robots will need to be manufactured.9

The trillion-dollar business

How will automation develop in the coming years, all things considered? Within a few years, the time will come when machine intelligence must move from being a promise of the future to becoming a business of the present, which means that leading artificial intelligence companies will have to finance their infrastructure investments from their own revenues—amounting to hundreds of billions of dollars annually by the end of the decade. What products or services will generate these enormous revenues?10

• The revenues from large language models currently mostly come from consumer subscriptions for personal use: employees pay for them themselves to make their lives easier and their work more productive. These typically cost around twenty dollars per month, which, presuming 50 million subscriptions, would amount to 12 billion dollars in annual revenue.11 Research assistants capable of independent data collection represent a separate category, with a current subscription fee of around 200 dollars. It is not unrealistic to assume that about one-tenth of users will opt for this highest-capability category, which would increase revenues by 12 billion dollars a year.

• At first glance, corporate subscriptions differ only in the sales channel, and while they now represent a negligible share compared to individual ones, their significance is steadily growing. The main conceptual difference is that they make the use of artificial intelligence an expectation placed on the employee. Their price depends on the size of the company; assuming an average price of one hundred dollars, 10 million subscriptions would correspond to another 12 billion dollars in annual revenue.12

• The real transformation will be brought about by autonomous agents capable of independently performing a significant portion of the tasks associated with certain jobs. While they typically won’t be able to fully replace an entire position, they will handle a large part of the related tasks. This means fewer employees will be needed in those roles, with the remaining workers focusing only on tasks the agents cannot perform. If all goes well, for those who remain, this will effectively amount to a promotion: they will (typically) be relieved of the most unpleasant tasks, and their salaries will increase due to improved productivity. The subscription fees for autonomous agents will be benchmarked against the wages of the employees they can replace, and the costs will vary greatly depending on the type of job. In customer service, this could be around one thousand dollars per month, while in fields such as administration, accounting, finance, and law, it could reach two thousand. Replacing half a million jobs at those rates would total roughly eight billion dollars.13 The other major—and much more profitable—area will be software development and research & development, where subscription fees could reach up to ten thousand dollars per month. With 200,000 subscriptions, this amounts to 24 billion dollars in annual revenue.

• Finally, as with all online services, free users will likely have to put up with advertisements sooner or later: based on data from Facebook and Google, advertising revenues could easily reach 20–30 billion dollars within a few years.

Let us see in summary:

It does not seem unrealistic at all to achieve all this by 2027–28.14 We’re talking about replacing seven hundred thousand workers, while the number of employed people in the United States is approximately 160 million—that’s barely four per thousand (though it corresponds to a much higher automation rate, since both individual and corporate subscriptions replace many current work tasks without affecting the number of employees). It’s equally certain that not everyone will succeed: for now, Anthropic, Google, Meta, OpenAI, and xAI each have room to thrive, but once they have to compete for the same limited market, that coexistence will turn into a fight for survival.15

After this, increasingly widespread automation, covering more and more job roles, will start to gain momentum, eventually affecting almost all office work. This doesn’t necessarily mean layoffs, but it does mean that companies that can afford to—those that are well-prepared and not constrained by regulations—will not fill vacant positions, will gradually eliminate entry-level roles, and will not renew contracts with subcontractors. The situation of early-career professionals will change the most, as the availability of traditional office jobs continues to shrink.16 Physical labour will gain value, as replacing these jobs with robots is far more capital-intensive, not to mention the location-dependent nature of such work.

There is considerable variation in the forecasts regarding the extent to which automation will generate economic growth. As we have seen, Goldman Sachs projects a one-time expansion of 16–33% over one to two decades, while McKinsey estimates that productivity could double by 2040. (These translate to an annual growth surplus of 1.5–3.0% and 3.5-4.0%, respectively.) Greater growth than this is difficult to imagine: as automation becomes more widespread within a sector, prices will decrease, limiting the growth of those sectors’ share in total production.17 To put these figures into context: even Goldman Sachs’ more conservative estimate exceeds the productivity growth resulting from digitalization (the spread of computers, the internet, and mobile phones), while the less cautious estimate surpasses the growth experienced during the most dynamic decade in human history, the 1960s. And if McKinsey’s projection proves accurate, it would mark the most transformative shift in leading economies in two hundred years.18

Maintaining control over our future

The societal changes brought about by the emergence of artificial intelligence capable of competing with humans carry fundamental risks. The most immediate challenges include distinguishing machine-generated content from reality, transferring control to autonomous devices, and the adaptation of the labour market to rapid and profound changes. In the longer term, however, the real question is: how much control over our future are we willing to relinquish?19

This is what Stuart Russell calls the WALL·E problem: as we gradually hand over more and more tasks involved in the management of our civilization—for the first time in history, not only the storage of information but also its processing—to machines, we may unknowingly reach a point beyond which this process becomes irreversible, and humanity, voluntarily, gradually and irreversibly relinquishes the ability to shape its own destiny.20

The proposition that artificial intelligence poses an existential threat to humanity takes this argument one (significant) step further.21 The best formulation of this problem to date comes from Swedish philosopher Nick Bostrom.22 The theory of artificial superintelligence rests on the assumption that machine intelligence can engage in recursive self-improvement, enabling it to surpass human cognitive capabilities over time. According to Bostrom, this could grant it six “superpowers”:23

• Being more intelligent than any of its opponents, it can outwit them—devising superior strategies that account for every factor and possibility, and thinking countless moves ahead.

• Through deception, bribery, or blackmail, it can manipulate people into advancing its objectives, thus mobilizing social resources.

• It can seize control of computer networks, hack communication channels and financial systems.

• It is capable of developing new technologies—for instance, surveillance systems, advanced weapons, or autonomous robots.

• Using these capabilities, it can obtain virtually unlimited financial resources, enabling it to purchase services and materials.

• Through continuous recursive self-improvement, it can enhance all of these abilities even further.

It is important to clarify that the concern is not that artificial superintelligence—should it ever emerge—will inevitably turn against humanity, but rather that even a very small chance of such an outcome poses a risk serious enough to warrant attention. According to Bostrom, this threat arises partly from the extreme difficulty of translating human values and objectives into a form that a machine can interpret without the risk of catastrophic error, and partly from the likelihood that a superintelligent system will not only strive to achieve its goals, but also to neutralize efforts aimed at preventing them. In order to be able to execute its objectives, it will seek to ensure its own survival and maximize its capabilities and resources—regardless of the interests of humans.24

In order to give credence to fears concerning the threat posed by superintelligence, we must accept two assumptions: that the creation (or the mere existence) of intelligence surpassing that of humans is at all possible, and that the process leading to it could accelerate suddenly, escaping human control. We know nothing about the first; neither is the second particularly convincing. As we have seen, the performance of large models primarily depends on the amount of computation carried out during their training, which in turn is limited by hardware capabilities. The pace of development of the physical infrastructure is constrained, which—barring magical breakthroughs—rules out sudden leaps. One such magical breakthrough would be a software intelligence explosion.

The advancement of algorithms enables increasingly efficient use of resources: the computational power needed for a language model to reach a given performance level halves roughly every six months. This implies that artificial intelligence gets better even if the hardware remains unchanged. A software intelligence explosion is a hypothetical scenario in which this improvement accelerates from its current pace to many times faster. However, the data shows no sign of this; in fact, the rate of change has remained fairly stable since 2017.25

The hypothesis suggests that when artificial intelligence becomes reliably capable of autonomously executing programming tasks that take several hours—which METR’s research estimates will happen around 2027–28—these agents will be deployed in massive numbers to accelerate the development of machine intelligence.26 The more advanced models thus created will in turn result in the emergence of even better agents, which will also join the research process, and so on, repeatedly. It will become possible to fully automate the entire software development process from start to finish, eliminating the need for human intervention. With humans no longer being the bottleneck, the process will accelerate explosively.27

However, it is far from obvious that multiplying the number of researchers would significantly accelerate progress. The correlation is certainly not linear—doubling the number of programmers hardly doubles the pace of progress—there is no empirical data on the precise nature of this relationship. It is possible that the effect would be barely noticeable, for example because the development of software (beyond a certain threshold) depends less on the number of programmers and more on the abilities of the best of them, or because the pace of algorithmic advancement is limited by computational capacity.28

The crucial point, however, is that any computational capacity used by artificial programmers would have to come from the same pool available to their human colleagues. Whether the programmer is human or artificial, running the same experiment requires the same amount of computational power—except that, in the AI’s case, additional capacity is also needed to run the agent itself (even if that extra cost is small). Consequently, to justify allocating capacity to the machine instead of the human, the AI would need to outperform at least some humans—not necessarily all, but certainly the least capable ones. Thus, artificial programmers will not flood the workforce by the millions as new labour, as proponents of a software intelligence explosion suggest, but will merely replace the lowest-performing programmers at any given time—to a far more limited effect. There is no reason to expect automation in software development to follow a different pattern than in other fields; it will unfold in exactly the same way. Its impact will be significant, but far from explosive.29

Today, both the United States and China regard the advancement of machine intelligence and the strengthening of their relative positions within this field as strategic priorities, which guarantees that progress continues at the fastest possible pace. The earlier large-scale automation arrives, the sooner its economic rewards can be realised: if it indeed yields a four percent growth surplus, as McKinsey forecasts, even a few years’ advantage or delay can translate into a substantial divergence in outcomes.

Exciting years are ahead of us.

What is certain is that expectations have grown dramatically in recent years. While in almost every survey conducted in the 2010s among artificial intelligence researchers, professionals in the field—computer scientists, mathematicians, philosophers, psychologists alike—consistently predicted the emergence of “high-level” or “human-level” machine intelligence around the middle of the century (between 2040 and 2060),3 these timelines have significantly shortened since the arrival of ChatGPT.4 (It is also telling that in all such surveys, only about 5% of respondents were firmly convinced—estimating a probability above 90%—that such intelligence would never be achievable.) Of course, this doesn’t say much about when these changes will actually occur—after all, these are people who spend their days working on artificial intelligence; it would be surprising if they didn’t believe in the success of their work—but what is clear is that there is a prevailing sense that a fundamental breakthrough has taken place in recent years, leading to significant acceleration.

In what follows, I will map out the road ahead and examine how artificial intelligence might impact our lives by the time we reach the end of this road. We will see that progress cannot continue at its current pace indefinitely, and that within a relatively foreseeable timeframe, it will emerge whether machine intelligence based on artificial neural networks will truly reshape the world as fundamentally as the Industrial Revolution once did, or if it will remain an unfulfilled promise. In this post, I will discuss what artificial intelligence is actually capable of today, what promises are made regarding its future capabilities, and how we might get from here to there. In the second part, I will examine how, according to these hopes (or fears), it might affect the economy and the labour market, and finally, I will dedicate a separate chapter to its dangers.

Measures of intelligence

How do we know what artificial intelligence is capable of? How can this be measured? In practice, the most reliable indicator is its ability to handle more and more workplace tasks: according to Nick Bostrom’s definition, we can define “high-level machine intelligence” as that which performs at least as well as a typical human in most professions (setting aside the perhaps interesting, yet somewhat forced issue of what exactly a “typical human” means).5

The classic answer, however, comes from Alan Turing. In his 1950 article Computing Machinery and Intelligence, he posed the question: how can we know whether a machine thinks?6 As a solution, he proposed the following test: a judge asks two remote conversation partners questions on a wide range of topics—science, emotions, everyday matters, personal experiences—one of whom is a human, the other a machine. After five minutes, the judge must decide which is which. If the machine is truly indistinguishable from the human, the judges will identify the machine as human, and vice versa, purely by chance—i.e., in about half of the cases. Turing argues that since humans are undeniably intelligent, if a machine cannot be distinguished from a human, then it too must be considered intelligent.

There are generally two main criticisms of the Turing test.7 One is that we cannot truly know whether the machine actually thinks or merely pretends to think. But why should machines be held to stricter standards than humans? How do we know that other people really think and not just pretend that they do? „[I]nstead of arguing continually over this point,” Turing writes, “it is usual to have the polite convention that everyone thinks.”8 The other criticism is that the test is not sufficiently precisely defined. This objection overlooks the fact that, while there is nothing preventing it from being operationalised, the Turing test is much more of a thought experiment than an actual exam.9 Rather than getting hung up on the details of its implementation, one should instead conduct one’s own modified Turing test—assign a task to an artificial intelligence, and ask: did it complete the task just as well as I would have done it myself?

At the same time, the demand for more quantifiable tools of comparison is understandable. One such tool is MMLU, a multiple-choice test comprising around 14,000 questions across 57 different domains, ranging from mathematics through history and ethics to psychology.10 One advantage of MMLU is that it allows for good comparison with human capabilities: human performance at the 95th percentile is 89.8%—meaning only one in twenty people can outperform this score. Another benefit is that it has been in use for a relatively long time, making it easy to track progress over time. This is also its limitation: cutting-edge large language models now perform so well on it (well above 85%) that the test can no longer differentiate between them.11

This flaw is addressed by tools like Chatbot Arena, which enables direct comparison under real-world conditions: users choose the better of two anonymous chatbot responses, and these votes form the basis of a ranking.12 The Turing test, MMLU, and Chatbot Arena are all tools for evaluating the ability to communicate in natural language. And for good reason, since language is a generally suitable tool for describing the world, and it is also the natural medium for interaction between humans and computers. But we’re increasingly moving beyond this—we are no longer just interested in how well artificial intelligence performs as a response-generating machine, but rather in how effectively it can take over tasks from us.

This has likely been the most cited, analysed, and shared chart in artificial intelligence literature in 2025:13

The graph by METR research institute shows how long it takes a professional programmer to complete the longest programming task that machine intelligence can perform autonomously with 50% reliability: while GPT-2, released in early 2019, could only manage tasks lasting two seconds, Claude 3.7 Sonnet—released almost exactly six years later—can be entrusted with tasks lasting up to an hour. The trend is striking: the duration of tasks that can be handled doubles every seven months (which equals a 3.3-fold increase per year).

Several caveats are in order. First of all, programming is a field where artificial intelligence performs particularly well, since its success criteria are clearly defined and easy to verify. Therefore, it would be an overstatement to assume that progress will be equally rapid in all other domains. Secondly, the 50% task completion rate is merely indicative, and can hardly be interpreted as really useful work that someone could rely on. A third question is how applicable this result is to real-world tasks—those that are less precisely defined, assume implicit contextual knowledge, require the real-time coordination of multiple parallel workflows and collaboration with other actors, and are carried out in a changing environment.

As concerns the expected success rate, increasing it to 80% significantly (by about one-fifth) reduces the task time horizon, but the doubling rate remains seven months, and artificial intelligence already surpasses human accuracy in tasks shorter than 15 minutes. It is also true that it performs much worse in less structured tasks, but the growth rate remains unchanged here as well. And finally, it should be added that it operates much more cheaply: executing 80% of the successfully completed tasks cost less than 10% of what it would have cost if performed by a human.14

Perhaps the most balanced way to view the study is as suggesting that by early 2028, the tasks that can be performed by artificial intelligence will, in some areas, be comparable to what a junior employee can now accomplish in a workday15—provided that the pace of progress remains consistent with that of recent years. Will it remain so?

The end of scaling

The progress in artificial intelligence performance over recent years has been made possible by the exponential growth in the size of artificial neural networks, the data used to train them, and above all, the amount of computation performed during training. As early as in a 2020 study, OpenAI raised the possibility that there may be a point beyond which the capabilities of transformer-type language models no longer improve, even with further scaling; at that time, this threshold was estimated to be around 10¹² parameters, 10¹² tokens, and 10²⁴ operations.16 We have long since surpassed these figures, yet there is still no sign of such a breaking point: it appears that as resources increase, so improves performance, and no clear limit is in sight.17 However, scaling has inherent limitations. Four factors come into play: computation, data, energy, and capital.

The amount of computation performed during the training of the most advanced models is on the order of 10 septillion (10²⁵), and increases four- to fivefold a year.18 This is made possible by the growing performance of supercomputers, supplemented by the increasing duration of training.19 The growth in supercomputer performance results roughly equally from increases in the number of chips and in their individual performance.20 There is enormous demand for these chips, and their production is a reliably profitable business (their most significant manufacturer, Nvidia, has increased its market value tenfold in recent years and has become one of the three most valuable companies in the world), which leads to an expanding supply. In recent years, the duration of training runs for leading models has increased by 40% annually, and if this trend continues, it could reach nine months by late 2027 and a year by the end of 2028.21 This is likely to be the limit beyond which further increases in duration, while not technologically impossible, will no longer be practically feasible.22 After this point, one of the contributing factors to the growth in training compute will be eliminated, which—unless compensated by the increase in hardware performance—will lead to a slowdown in the rate of growth.23

When it comes to data, the bottleneck is text: the amount of publicly available, high-quality, human-generated textual data suitable for training is 300 trillion (3 × 10¹⁴) tokens.24 If the current growth rate continues, there could be a language model by 2027 that fully utilizes this amount. After that, increasing the size of models may still improve performance for a while even without increasing the volume of data, but a real solution can only come from new data sources. Instant messaging applications owned by Meta, such as Messenger and WhatsApp, generate about 165 trillion tokens of text annually, but it’s questionable whether their quality makes them suitable for training. Emails are likely to be more useful, generating a total of around 800 trillion tokens per year, one-third of which is created at Google. In any case, since we’re dealing with exponential growth, even all these non-public sources combined can only delay the exhaustion of human-generated text by no more than a year or two. Text is critical because—unlike images, videos, scientific data, or experiences gained in virtual worlds—human text is only produced by humans. Although OpenAI was generating 100 billion words per day already in early 2024, training language models on synthetic (machine-generated) text has, according to experience, not improved but rather degraded their performance, leading to outputs that are less and less similar to human language, more homogeneous, and increasingly meaningless.25

As concerns electricity, the demand of the most powerful supercomputers doubles every year.26 Starting from the current record holder xAI Colossus’s demand of 300 megawatts, this means that by the end of 2026 we could reach 1 gigawatt, the typical output of a modern nuclear reactor.27

It is worth examining both the cost of training a model and the cost of building the necessary infrastructure—both are informative, though in different ways. The cost of training is useful for understanding the economic processes behind the development and deployment of a model, while the capital investment reflects the scale of funding required to develop competitive artificial intelligence. The cost of training is primarily determined by hardware depreciation: in the case of leading models, this is on the order of hundreds of millions of dollars, and it increases 2.4-fold annually, in proportion to the growth in hardware prices and the time spent on training.28 To get a complete picture, one must also account for failed experiments and development paths that lead to dead ends; those who can avoid these—even by drawing inspiration from others’ work—can save a lot. It is also true that creating models with a given performance level becomes cheaper over time, as processor price-performance ratios improve and algorithms evolve.29

The price of the supercomputers used for development doubles every year: while Summit, the world’s most expensive system at the time, cost 200 million dollars in 2018 (adjusted for today’s prices), Colossus, completed in the spring of 2025, cost 7 billion. And to that must be added the cost of building the power supply infrastructure, which comes to ten million dollars per megawatt (i.e. nearly three billion in the case of Colossus). If this trend continues, the cost of building the world’s leading supercomputer could exceed 100 billion dollars within three years.30 This entry threshold is out of reach for independent institutes, universities, and most governments, placing the future of artificial intelligence in the hands of the largest technology companies.

All things considered, a rather precise timeline is taking shape: there is nothing to suggest that the development of recent years could not continue until about the end of 2028. In 2029, prolonged training times, data scarcity, energy demands, and exponentially rising costs will almost simultaneously begin to impose limits on further growth. By around 2030 at the latest, machine intelligence will either fulfil its promises and truly reshape the world, or—at least with the current paradigm based on scaling artificial neural networks—it won’t at all: its capabilities will plateau, and it will have to generate revenues proportionate to the investments required for its creation and operation with the abilities it will have attained by then. Of course, an earlier slowdown cannot be completely ruled out, but those who have been gleefully declaring the AI bubble burst (or at least deflate) for years have been continuously proven wrong by reality. Nor should we expect a significant acceleration: the pace of expansion is primarily determined by the time needed to build physical infrastructure, and in that area, we are already moving full steam ahead.31

The framework of human existence has been defined by six determining factors:1

• work carried out throughout the entire day, except for a very narrow segment of the population, can barely secure subsistence;

• one-third of all children die before the age of five;

• human lifespan is at most 80, in exceptional cases 90 years;

• intelligent life exists in only one place in the entire universe: on Earth;

• consciousness is confined within the body, which fundamentally limits its freedom and makes it vulnerable to physical impacts;

• man is the most intelligent of all living beings, and therefore in control of the planet and of his own fate.

Of these six, the first two are already a thing of the past. We take it for granted today, but this is actually quite a recent change: in 1850, 96% of Western Europe’s population lived on less than ten dollars a day (today it’s 3%), and child mortality was still around 25% in the 1890s (today it’s 0.4%).2 This is a revolutionary, sweeping, paradigm-shifting transformation that has sidelined these limitations as if they had never existed. Two hundred years ago, no one thought this was possible.

When it comes to the long-term future of humanity, the question is how enduring the other four limiting factors will prove to be over the next 100 or 1,000 years — the future of certain technologies only matters insofar as it helps guide us in answering this question.

The four limits

When we instinctively say that a certain technology will never come true, it’s worth considering: are we really judging its feasibility, or are we simply sceptical about the time frame? Is the fact that creating a technology is very difficult enough reason to assume that it won’t happen in 100, 1,000, or 10,000 years?

In reality, what we most often mean is that we lack the necessary knowledge or that the material resources aren’t available. However, these are temporary limitations: as long as scientific progress and global economic growth continue, we have every reason to hope that these barriers can eventually be overcome. Just as a journey to the Moon seemed fantastical when Jules Verne’s novel was published in 1865 — and, if we adjust the cost of the Apollo program accordingly, would have amounted to one and a half times the GDP of the United States at that time3 — so too are we today unable to build interstellar probes capable of reaching even a few percent of the speed of light, with projected costs in the trillions of dollars that lie beyond comprehension.4 But if the global economy grows at the same pace as it has over the past 100 years, doubling three more times by the end of the century, then this goal will be no more unattainable than the building of the International Space Station was in the 1990s.5

But of course, there are still constraints that may prevent the development of a useful technology.

The first and most obvious such limit is if it is impossible according to the laws of nature. Nowadays, every high school student knows that building a perpetual motion machine is impossible, but the fact that there is a theoretical obstacle to it has only been known since the mid-19th century — until then, those who failed might reasonably have thought that it was merely the method of implementation that was flawed. Similarly, it is conceivable that, although we are currently unaware of any fundamental obstacles, it may turn out that, for example, mind uploading (transferring human consciousness into a computer) is impossible.

Another possible such reason is the decision to abandon a given technology. During the development of the atomic bomb, Edward Teller raised the concern that the atomic explosion might create temperatures never before experienced on Earth and conditions similar to those found in the cores of stars, potentially triggering a self-sustaining chain reaction in the surrounding air and igniting the atmosphere.6 Calculations ordered by Oppenheimer showed that this would not happen, as throughout the entire process the energy loss through radiation would always be several times greater than the energy production. However, one can imagine a situation where such calculations are not so clear-cut and indicate that a scientific experiment, expected to yield extremely high returns, carries a significant, say 1% chance of leading to the end of the world. In such a case, would we be willing to abandon it? And returning to the historical example: would the United States have been willing to give up the atomic bomb under such circumstances? And what if, in the meantime, news arrived that the Nazis were also close to developing it? (The later history of mutual deterrence based on nuclear weapons actually shows that as long as both parties act rationally and assume the same of the other, these risks can be managed quite well.) Even today, there are technologies whose development we have deliberately halted: for instance, human reproductive cloning (the creation of an entire individual, not just specific cells or tissues) is banned worldwide — although, in this case, the potential benefits are doubtful in the first place.

It is also conceivable that certain technologies may never be developed because their practical utility lies on a timescale far beyond the human scope. For example, it is questionable how appealing the prospect is that, at least in theory, the entire galaxy could be colonised within a few million years using von Neumann probes (self-replicating spacecraft). To what extent could star systems on the far side of the Milky Way — with which even a single message exchange would take 200,000 years — be considered part of human civilization?

Finally, one can also imagine a situation where technological progress comes to a halt. A 1984-like total dictatorship, if it achieves global dominance, could abandon innovation due to the lack of external competitors, or even restrict it in order to maintain its stability. And if its existence is threatened by neither external nor internal factors, overthrowing it could become nearly impossible. It is also possible that it is precisely technology that ultimately blocks the path to further change — for example, if humanity chooses to continue to exist not in physical reality, but in a computer-generated simulation.

What matters is not only whether every technology that is possible and desirable eventually comes into existence, but also the order in which this happens: some technologies, once dominant, can slow down, hinder, or (temporarily) block the spread of alternatives that might prove more beneficial in the long run (think of fossil fuels versus electricity).7

Beyond material abundance

The medieval peasant’s utopia was Cockaigne, a land where there is no deprivation and feasting lasts all day long. A place where fences are made of sausage, fish jump out of the water already fried, roasted pigs roam the streets, and cheese rains from the heavens. There is no illness, every day is Sunday, music plays constantly, there is plenty of time for merrymaking and rest, and one can glimpse under women’s skirts. In the developed world today, we more or less live in this utopia: food can simply be taken from the fridge or ordered to your door, life expectancy is double that of the Middle Ages, work is much less exhausting, and our lives are largely easy and pleasant (and if someone wants to look under skirts, there are plenty of opportunities for that too).

In 1930, John Maynard Keynes wrote that within the next hundred years, productivity would increase to such an extent that the inhabitants of the most developed countries would need to work only 15 hours a week to satisfy all their material needs.8 As far as productivity is concerned, he was right – in his homeland, the United Kingdom, per capita production today is four and a half times higher than it was a hundred years ago – but the amount of time spent on work has decreased only much more modestly, by a quarter (and even that reduction occurred almost entirely between 1950 and 1980).9 Thus, it seems that the surplus generated by productivity growth is being used to a much greater extent to increase production (and consumption) rather than to reduce working time. Yet, even if the British worked only 15 hours today, as Keynes predicted, they would still live more than twice as well as they did a hundred years ago.

We spend a lot of time thinking about how the future could go wrong (which is understandable — after all, to avoid pitfalls, we must first identify them), but what if everything turns out for the best, and technological progress over the next hundred years improves our lives just as much, or even more, than it has over the past centuries? What if artificial intelligence truly works, and takes over every task we’re willing to entrust to it, creating unprecedented abundance? What if we go beyond today’s post-scarcity utopia (a realized Cockaigne, where every good exists in abundance) and move closer to a post-work utopia (Keynes’s vision, where we hardly need to work — or not at all — for these goods)? What would we spend our income on if it far exceeded our needs (and those of our children)? In a world where we can achieve anything without effort and every wish can come true, would anything we do still have a purpose, meaning, or stakes? What would we find interesting in our activities, and what joy would we find in them?10

We can predict with reasonable precision what we would spend it on: largely the same things as today.11

• Above all, on even better products and services: luxury items (real estate, vehicles, jewellery, works of art, vacations, etc.), healthcare (not only medical treatments, but also those aimed at improving physical or mental capabilities), education, and entertainment.

• Secondly, to improve our relative social status, partly for the recognition that comes with it, partly for the security it provides. This is not about achieving a specific material standard, but rather about reaching a certain position in the hierarchy relative to others. Under such circumstances, the extent of spending is not limited by the value of the products and services themselves, but by the amount of resources others are willing and able to devote to the same purposes.

• And finally, for social purposes, to make a change in the world.

But material prosperity is only one aspect. Indeed, the fact that we live in greater wealth, safety, and comfort than ever before is largely due to technology, but the real question is how technological progress expands the boundaries of human existence — whether, in time, it will dismantle the third, fourth, and fifth frontiers (meanwhile hopefully preserving the sixth); and ultimately, how it increases human freedom — this, if anything, is its final purpose.

After the death of the world-renowned mathematician John von Neumann, his friend and colleague Stanisław Ulam wrote a memoir for the journal of the American Mathematical Society. In it, he recalled a conversation they had about the ever-accelerating progress of technology, which, according to Neumann, foreshadowed the approach of a turning point “beyond which human affairs, as we know them, could not continue.”1

The essence of Neumann's observation is that the more advanced the technology available to humanity becomes, the faster it enables further development. Therefore, as we move forward in time, we can say not only that progress continues, but also that it continues at an increasingly rapid pace.2 If this is true, then the time will eventually come when the rate of technological advancement becomes so unimaginably fast, and its impact so widespread and profound, that it will fundamentally change humanity’s role, possibilities, and limitations in the world — this is the technological singularity.

Historical experience supports this observation: even if not consistently across all regions, overall, technology has become increasingly advanced from generation to generation, century to century. As a result, each new generation could start from a slightly higher technological level than its predecessors. Until quite recently, this process was so slow that it was imperceptible within an individual’s lifetime — until everything suddenly changed in the past two hundred years. However, for progress to accelerate indefinitely, this alone is not enough. Most of the time, taking the first steps is easier, continued advancement becomes increasingly difficult, requiring ever more knowledge and resources. The singularity can only occur if the pace at which technological capabilities accumulate consistently exceeds the pace at which difficulties increase.

Think exponential

Few have done more to popularise the idea of technological singularity than Ray Kurzweil. In his 2005 book The Singularity Is Near, published exactly twenty years ago today, he took examples from the fields of computing, artificial intelligence, nanotechnology, and biotechnology to support the claim that the singularity is not merely a distant vision, but a reality that could arrive within our lifetimes — by 2045.3

In spite of his successful business career, Kurzweil is primarily known for his bold long-term technological predictions that are concrete enough to be held accountable later. In late 2012, the community of LessWrong examined 172 of his claims made in 1999, which were supposed to come true within ten years.4 According to the volunteer reviewers, 42% of the claims turned out to be true, 46% false, and only the remaining 12% were too vague to assess. Considering that these statements weren’t ambiguous at all, but rather concrete (specific enough to even serve as the basis for business decisions),5 this is quite a remarkable achievement. In 2020, the same online community reviewed the predictions Kurzweil made in 1999 that were meant to come true within twenty years (i.e., by 2019): this time, only 24% of the 105 claims were found to be accurate, while 67% were deemed false.6 This marks a significant decline compared to the ten-year assessment and is not a particularly strong result. However, the analysis also revealed something else: many of the predictions for 2009 that had not come true by 2012 had, in fact, materialized by 2020. In other words, even when Kurzweil was wrong, he generally pointed in the right direction — only he was overly optimistic about the pace of change. How did he know? What is his secret?

He tracks exponential growth. In order to see how the technologies of the future will grow out of those already surrounding us today, we need to shift our way of thinking. On a linear scale, certain technological innovations seem to emerge suddenly, while on a logarithmic scale, they are perfectly predictable. Take the example of the internet: in the chart on the left, we see that the number of computers connected to the internet began to grow explosively around 1993–94, seemingly out of nowhere, increasing nearly a hundredfold in just a few years. The second chart shows the same data, but with the number of computers plotted on a logarithmic scale, which is better suited to demonstrating exponential growth. This shows that the growth rate was predictable and consistent, increasing at nearly the same pace year after year (and if there was any surprise, it was a slight slowdown in 1986, not the steady growth in the mid-1990s). A similar pattern can be seen with many other technologies that Kurzweil discusses.

This also implies the limitations of his forecasts. No matter how good he is at predicting the direction of development, few trends maintain the same intensity for twenty years, so the further ahead he looks, the more inaccurate his predictions become. Secondly, although an authority in the field of computer science, he is not exceptionally skilled at estimating the social acceptance and impact of innovation — and is notoriously reluctant to admit when his predictions go wrong.7

We’ll be the Borg

Alright, but what happens by 2045 — what is it that brings about the singularity? Put most generally: the exponential growth of the performance-to-price ratio of computers drives progress in all other fields.8 According to Kurzweil’s 2005 prediction, by the mid-2040s, one thousand dollars’ worth of computation will reach 10²⁶ operations per second, so the machine computation capacity created per year will be one billion times more than the combined biological computation capacity of humanity.9 One can actually find an independent estimate that supports this: back in 2011, two researchers calculated the machine-based information processing capacity of humanity in 1986, 1993, 2000, and 2007. The growth was accelerating: while it was only 42-fold between 1986 and 1993, it was already 81-fold between 1993 and 2000, and 130-fold between 2000 and 2007.10 If we project this trend into the coming decades, we do in fact arrive at a scenario where by the mid-2040s the annually produced artificial computational capacity reaches 10³⁵ operations per second — if it made sense to extrapolate a trend observed between 1986 and 2007 into 2025–2045.11

Reality, however, cannot keep up with this: in 2023, the computing performance per thousand dollars was supposed to have been around 10¹⁶ operations per second, but even according to Kurzweil's own data collection, it fell short by two orders of magnitude.12 By 2045, perhaps the 10¹⁹ magnitude might be reached at most, but 10¹⁵–10¹⁶ is much more likely – it's no wonder that the sequel published last summer, The Singularity is Nearer, although it continues to focus on the exponential growth of the performance-to-price ratio, refrains from making quantified predictions.

It must be emphasized that, according to Kurzweil, it is not the increase in computational capacity itself, but rather the technological advancements enabled by it in other fields that result in the singularity. As for its tangible content, two well-defined, successive steps lead to technological singularity:

1. Artificial intelligence will reach human level by 2029, as evidenced by passing the Turing test. (This is not just some vague reference to the test: Kurzweil precisely defined the detailed rules as early as 2002.)13

2. A brain-computer interface with a few million to a few tens of millions of channels will enable merging with artificial intelligence by 2045, thereby turning artificial intelligence into an extension of consciousness.14

Regardless of whether we believe this would be beneficial, it is no doubt a rather concrete scenario, which would fundamentally alter human existence, thus fulfilling Neumann’s criterion. We could say that we do not want to become Borg, but it’s undeniable that being Borg is fundamentally different from being human.15 Moreover, while these two steps are challenging, neither can be dismissed as definitively impossible.

Time for the Roaring Thirties

But is it really true that technological development is accelerating? If we rely on our impressions, there is reason to be sceptical. Humanity's arguably greatest achievement so far, the Moon landing, has not been surpassed in more than half a century (and has only been repeated a few times). The speed of the fastest mode of transportation has remained virtually unchanged for 70 years, since the advent of jet-powered passenger aircraft. And if we exclude digital technologies, the world looks more or less the same as it did in the 1970s: by the end of that decade, three-quarters of American households had a colour TV and a washing machine, half had air conditioning, a dishwasher, and a dryer. Of the conveniences of modern life, only the microwave oven, spread in the 1980s, was missing.16

And what do the numbers show? As Nick Bostrom points out, a good measure of the overall pace of technological development is not the growth of a few selected leading sectors, but productivity as a whole, which encompasses all technological innovations (understood not only as machines and devices, but also procedures, processes, and institutions), weighted by their economic significance.17 In the world’s leading economy, the United States, productivity grew by 112 percent between 1950 (since when we have reliable data) and 1984, but only by 73 percent between 1985 and 2019 (a similarly 35-year period). The same trend appears when we compare the first and second halves of the last 30 years: between 1990 and 2004, growth was 36 percent, but only 18 percent between 2005 and 2019. No matter which two consecutive periods of equal length we compare, we consistently see that the pace of growth was higher in the first than in the second — if we were truly approaching the singularity, this trend would have to be the other way around.

The change in the productivity rate shows that the pace of growth slowed from its peak in the 1960s to 1.2–1.3 percent annually by the 1980s, followed by a temporary strengthening lasting a decade and a half, after which it again declined to that level. This acceleration, which began in the second half of the 1990s, was due to digital technologies: the personal computer, the internet (and the online commerce built on it), and the mobile phone. But why is it that these undeniably transformative technologies had only a relatively moderate and time-limited impact on productivity? Robert Gordon wrote the following about this in 2012: “Attention in the past decade has focused not on labor-saving innovation, but rather on a succession of entertainment and communication devices that do the same things as we could do before, but now in smaller and more convenient packages. … These innovations were enthusiastically adopted, but they provided new opportunities for consumption on the job and in leisure hours rather than a continuation of the historical tradition of replacing human labor with machines.”18 This will hardly bring about the singularity.

Among proponents of the singularity hypothesis, a dominant (though not exclusive) idea is that it will be the creation of artificial intelligence reaching or surpassing human intellectual capabilities what triggers accelerated development.19 Even if the recursive self-improvement envisioned by Irving Good, Vernor Vinge, and their modern followers does not materialize, it is still reasonable to expect that the widespread application of machine intelligence could lead to significant productivity gains—since, unlike the “innovations” of the past two decades, this is a technology that is indeed capable of replacing human labour.20

According to a relatively conservative forecast by Goldman Sachs, artificial intelligence could increase productivity by a total of 16 to 33 percent over the course of one or two decades—already much more than the 9 percent boost attributed to the spread of digital technologies.21 McKinsey estimates that this growth could be considerably greater, potentially reaching 3.5 to 4 percent annually, which would amount to a doubling over two decades.22 Robin Hanson goes even further, suggesting that we could be on the brink of a transformation as profound as the agricultural revolution 5,000 years ago or the industrial revolution around 1800: not just an increase in productivity, but our whole economy entering a new, steeper growth trajectory.23 Even if the singularity does not occur, we may yet be approaching a period of expansion unmatched since humanity’s most dynamic decade to date, the 1960s.

Regardless of the direction and pace of technological development in the 21st century, one thing is certain: those who turn singularity into a religion are bound to be disappointed — it will bring them neither judgment day, nor the afterlife, nor salvation from their current existence. On the other hand, it is also untenable for someone to accept the theoretical possibility of the technological singularity occurring at some point, while simultaneously rejecting from the outset any contemplation of technologies such as mind uploading (the transfer of human consciousness to a computer) or superhuman machine intelligence — the very definition of the singularity, according to von Neumann, is that it fundamentally transforms human existence.

The onset of growth is usually attributed to the Industrial Revolution: between 1810 and 2020, the per capita national income in the world’s leading economy (which was the United Kingdom until the turn of the century, and the United States thereafter) increased seventeenfold, representing an average growth of 14% per decade. This means that each generation, on average, experienced one and a half times greater material prosperity than their parents — for the first time in history.

It certainly is true that growth changed pace in the early 19th century, but it had in fact begun earlier: in Great Britain, material wealth per capita had already been increasing by an average of 4–5% per decade during the preceding 150 years, while the Netherlands had achieved an even higher growth rate of 6% per decade as early as the 16th century (see the Netherlands in orange, Britain in burgundy, and the United States in blue in the chart). The most important reason was trade with India and the Americas, made possible by a series of technological innovations.

How volta do mar set the world in motion

When a sailing ship heads south from Europe along the coast of Africa, there comes a point beyond which returning the same way — along the coast — is no longer possible due to ocean currents. Instead, the ship must steer westward, out into the open ocean, from where the prevailing winds will carry it back to Europe in a vast arc — the farther south it goes, the wider the detour. Without this technique, known as volta do mar, developed and perfected by Portuguese sailors in the 15th century, return from the southern coast of Africa is not possible. It was volta do mar that led first to the Portuguese colonization of Madeira (1419), the Azores (1432), and Cape Verde (1462), and ultimately to the circumnavigation of Africa and the discovery of the sea route to India (1497–98).

Besides the excellent sailors, this was above all to the credit of the Portuguese prince, Henry the Navigator (1394–1460), who provided the financial background for concentrating the required knowledge and experimental spirit in one place. This led to a revolution in navigation and cartography, as well as to the development of a new, more efficient type of ship, the caravel, capable of sailing against the wind. Together, these turned oceanic navigation from a suicidal venture into a scientifically grounded and refined technology. (The first center of global trade, however, became not Lisbon, but rather Antwerp — and later Amsterdam — which, thanks to its advanced financial system, was better suited to assume this role and was at that time already taking over the position of the commercial hub of the Low Countries from Bruges.)

With the emergence of global maritime trade made possible by open-sea navigation in the 16th century, the era of sustained, centuries-long economic growth began. However, this growth was still too slow to bring about noticeable change within a single human lifetime: in Great Britain, per capita national income increased by an average of 4% per decade between 1650 and 1800 — roughly one-third the pace of today's growth. Had this rate of growth persisted, everyday life today would resemble that of the 1860s. That this is not the case is due in no small part to a handful of Englishmen (and one Scot).

When machines gained power

In 1768, Richard Arkwright invited John Kay to Preston to use the modest fortune he had earned from his periwig business to perfect the invention Kay, a clockmaker, had been working on for years. The result was the spinning frame, which was capable of performing the work of dozens of people much more cheaply and quickly. (Kay received no share of the credit: Arkwright patented the machine solely in his own name, and although the patent was eventually annulled in 1785, by that time he had already reaped all its financial benefits.)

The spinning machine was too large to be powered by human strength, so after some experimentation with horse-powered operation, Arkwright built a new facility in Cromford in 1771, where the machinery was driven by a waterwheel. The idea proved successful: by the following year, 200 workers were employed, working around the clock in two 12-hour shifts — more than the population of the nearby villages — so Arkwright had houses (as well as shops, pubs, a church, and a school) built for them. This mode of production was radically new: the necessary machines, tools, and raw materials were provided by the industrialist, the work required no special skills, and the concentration of capital made continuous, large-scale production economically viable — far beyond what was possible in the domestic system. The basic unit of the modern economy, the factory, was born.

In 1766, Birmingham businessman Matthew Boulton was seeking a reliable power source for his new silverware factory – the stream running beside it did not provide sufficient water flow during the summer – so he contacted James Watt, an instrument maker at the University of Glasgow. Watt had recently completed a prototype of his new steam engine and was looking for an investor to help perfect it. (He initially partnered with John Roebuck, and in 1769 they jointly secured the patent in a one-third to two-thirds ownership split, however, Roebuck was unable to pay his debts and sold his share to Boulton.) Boulton eventually convinced Watt to move to Birmingham, and in 1775 they founded their joint company, the Boulton & Watt steam engine manufacturing firm. Boulton managed to have the patent extended until 1800; over the following fifteen years, the company sold around 450 steam engines. In the 1780s, Watt modified his invention to be capable of producing rotary motion, which allowed it to directly power machinery (previously it had only been suitable for pumping water). This innovative application of steam engines made it possible for factories no longer to be located strictly near water sources, but instead where other resources – such as raw materials or labor – were available.

The mechanized factory and the steam engine turned Great Britain into the "workshop of the world," where goods could be produced so efficiently that they could compete with locally made products anywhere in the world. The way these new technologies emerged, and how successfully the Brits exploited their potential, is a prime example that new technical innovations alone do not lead to prosperity, but only when combined with the right social environment. What stands out at first in both the stories of Arkwright and Kay, as well as Boulton and Watt, is the strong protection of intellectual property and the role of patents (and, of course, the differing fates of Kay and Watt). Secondly, while everyone knows the Industrial Revolution began in England, it is far less well known how little the imperial capital, London, had to do with it: it was, in fact, something in the air of rural England that allowed these two stories to unfold independently, essentially at the same time and within barely a hundred kilometers of each other.

The Birmingham lunaticks

Just as Henry the Navigator played a key role in the emergence of global trade through the enabling of open-sea navigation, so too did Matthew Boulton and his contemporaries change the fate of humanity by initiating industrialization. Henry's merit lay not so much in his personal achievements as in his support for innovation, which led to the gathering of the finest knowledge of his time. What his court in Sagres was to 15th-century Portugal, the Lunar Society was to late 18th-century England.

Boulton and Erasmus Darwin (the grandfather of Charles Darwin) had been regularly dining together for years when, in 1765, William Small, former tutor of Thomas Jefferson, returned from America and moved to Birmingham. It was at this point that the dining circle expanded into a club, initially calling themselves the “Birmingham Philosophers,” and soon counted James Watt among its members. After Small's death in 1775, to prevent the group from disbanding, Boulton made his own home the club’s permanent headquarters and introduced a more regular schedule: from then on, meetings were held monthly, always on the Sunday closest to the full moon (hence the name), so that members could return home more safely in the absence of street lighting. In the 1780s, Joseph Priestley — known for the discovery of oxygen — became the leading figure of the society (the meetings were even moved to Mondays to accommodate his pastoral duties), but by the 1790s, the historical moment that had given rise to the society had passed.

For two decades, the Lunar Society attracted the finest minds to Birmingham from hundreds of miles around (or, seen from another angle, the city’s unique spirit of innovation provided fertile ground for the society). The key to its success lay in the combination of scientific and technical interests among its members; however, it wasn’t the meetings themselves that had the greatest significance, but rather what emerged from the members’ collaborative work between those gatherings.

We live (far) better than history’s richest

When Richard Arkwright died, he left behind a fortune of half a million pounds – the equivalent of £60 million ($80 million) today. Despite this fabulous wealth, Arkwright’s life was, in almost every essential aspect, far worse than that of any of us today.

Few things in life are more horrifying to imagine than losing a child. Yet in the late 18th century, this was a part of everyday life: child mortality (the proportion of children who died before the age of five) was around 30%, and pregnancy was not only a source of joy but also a fatal danger. When Richard Arkwright junior was born in 1755, life expectancy in Great Britain — largely due to childhood and maternal deaths — was just 39 years.

It would be a mistake to think that the wealthy were exceptions. Matthew Boulton adored his children and wrote: “Nothing could in the least palliate this long, this cold, this very distant separation from my dearest wife and children but the certain knowledge that I am preparing for their ease, happiness and prosperity, and when that is the prise, I know no hardships that I would not encounter with, to obtain it.” Yet even he could do nothing to prevent having to bury three of his daughters as infants. James Watt had it even harder: of his seven children, only James survived him — three died in early childhood, three between the ages of 15 and 30. Neither Arkwright’s, Boulton’s, nor Watt’s first wife lived to see their 40th birthday. If we include all the Arkwrights, Boultons, and Watts — fathers, wives, children, a total of 25 people — we get exactly the same average lifespan of 39 years as in the general population. (It is true, however, that the wealthy had better prospects after age 40: Richard Arkwright Jr. lived to 88, James Watt to 83, and Watt’s second wife to 94.)

But it wasn’t just health. In the last fifteen years of his life, Richard Arkwright lived in a fairly large but rather plain house on the hill opposite his Cromford mill; due to his early death, he never moved into the mansion he had built for his family. Matthew Boulton opted not to build, but to buy instead, and thus could spend the second half of his life in the charming Soho House, which — thanks to the host’s efforts — became one of Birmingham’s most elegant and fashionable buildings by the turn of the century.

The building stood out not only for its exterior but also for its state-of-the-art engineering solutions. Central heating, hot running water, and a flush toilet were exceptional luxuries in that era: not only was every room warm, but there was no longer any need to carry water for washing, cooking, and bathing — or coal and firewood for heating — into each room separately, nor to haul out wastewater and ash afterward. Lighting, however, was still provided by candles and oil lamps (along with the accompanying smoke), just as it had been for millennia before — although Boulton and Watt’s steam engine factory was among the first buildings in the world to receive gas lighting in 1798, the technology only began to spread after Boulton’s death (and it would take another 30 years to spread to middle-class homes).

With rare exceptions like Boulton, two hundred years ago even the wealthiest people lived far more uncomfortably than the average person today, and their lives were not significantly different from those of a Roman patrician two thousand years earlier. Boulton and his contemporaries, no matter how wealthy they were, did not experience the modern luxury of being able to travel across the country in a few hours (the journey from Birmingham to London, 200 kilometers away, took a whole day), let alone the marvel of crossing the entire world in mere hours. There was no such thing as a vacation in Paris or Rome; it was customary among the British upper class to go on a European “grand tour” upon reaching adulthood, which typically lasted from six months to a year. These trips were the only opportunity to see the most important classical works of art or to acquire foreign books — and for the average person, this was not only inaccessible but downright unimaginable.

It is trite to say that, as European or American citizens, we belong to the luckiest one-eighth of the Earth's current population — the one billion living in the best conditions. However, we seldom consider that this also means that our fate places us among the top 1–2% of the roughly 110 billion people who have ever lived on Earth. The fact that losing a child is now a rare tragedy rather than an everyday reality, that our days are more than a mere struggle for survival from sunrise to sunset, and that we have the chance to discover the world and can freely access its beauty — all of this is the result of the technological change and the ensuing economic growth of the past 250 years, set in motion by people like Richard Arkwright, Matthew Boulton, and others of their kind.