I’m a fan of people trying things, even if they seem silly. Dismissing risky ideas misses the point of research.

But thoughtful criticism can direct effort to more promising fields. To that end, I’m going to try to make my criticism as constructive as possible, with concrete reasons for my pessimism and closely related research areas which are promising (and stand to benefit even from failures in the field I’m criticizing).

Not every section will live up to that standard; some arguments are built on vibes alone. I might be wrong, so I still want to see people work on these problems.

Preliminaries

Exhaustible resources more expensive long-term

The theory of exhaustible resources suggests that “the real price of an exhaustible resource should grow at a rate equal to the real interest rate.” The data are consistent with this hypothesis^[1].

Innovations in search and extraction can dramatically lower prices in the short term, but in the long term, we converge on a correct way to do things. Once extraction is solved, prices return to their slow upward climb.

Idiot index

Raw materials are just a fraction of the final cost of a good. You need to take out loans to buy the materials. You need land and trucks. You need to buy equipment. You need people to design, build, and test.

(And robots do not solve labor costs. Robots are made of raw materials and equipment and loans. Most importantly, like humans, robots have an opportunity cost, other uses of the robot bid up the price. Robots will have a non-zero “wage”.)

It’s usually straightforward to estimate input costs, but these other costs (land, equipment, wages, etc.) are tricky. I’ll try to get around this with the idiot index: the ratio of the final cost of a good to the cost of the raw materials. An example for metals:

Source

We’ll use the idiot index as a heuristic to think about the hard-to-estimate costs of producing a good. If we know the raw materials cost and an idiot index for similar products, we can guesstimate the final cost. Simpler processes have a lower idiot index while complex processes a higher one.

But don’t take this too seriously, the index I use for different areas is just an educated guess. If It’s Worth Doing, It’s Worth Doing With Made-Up Statistics.

BOTEC’s are inherently optimistic

Armchair estimation of costs is inherently optimistic. Reality has a surprising amount of detail, you often miss major cost contributors until you actually start doing the thing.

This is a feature, I’ll be employing an optimistic cost benefit analysis for the technologies I’m criticizing. I want to make the best possible case before I rule out an idea. But don’t be fooled, even if something looks slightly better on paper, complicated plans often fail in practice.

Preliminaries aside, let’s start critiquing.

e-Fuels

Producing fuel from CO2 capture and then burning it again doesn’t make sense:

… e-fuels and sustainable aviation fuel are silly. Capturing carbon *and *making fuel is probably more expensive than just capturing carbon. Making e-fuels requires capturing carbon plus making green hydrogen *plus *using renewable energy *plus *building reactors plus new fuel pipelines. Capture is simpler and has the same effect of making transport carbon neutral.

Even if it did, it’s hard for e-fuels to beat fossil fuels on price. In this section^[2] I made some pretty favorable assumptions for green hydrogen and it wasn’t a slam dunk compared to fossil fuels. And DAC CO2 at $100/tonne is more expensive per mole carbon than fossil fuels.

Even assuming $1/kg hydrogen and $100/tonne CO2, natural gas comes out to $1.03/kg, roughly 6x higher than today’s prices. Further in the future, exhaustible resource economics will bite and the fossil price will pass the e-fuel price, but that takes a while.

So methane prices are about as low as they’ll ever be^[3]. There will be price fluctuations, and better tech will help, but eventually we run out of breakthroughs. The economics of exhaustible resources demand that prices rise over the long term.

Instead of fuel, this tech would be valuable for making green ammonia, doing carbon capture, or making synthetic food. See my series on solving climate change for more.

E-fuels aren’t going to save us, but most methane users can switch to alternatives. One application can’t go without it: rockets.

Launch costs below $100/kg to LEO

So e-fuels won’t save us money and methane prices won’t be lower than today. Now we have a lower bound on fuel costs for chemical rockets (which predominantly use methane and liquid oxygen).

Fuel is just one component of rocket costs, there’s engines and staff among other things. It’s hard to estimate prices for these, particularly as technology improves and manufacturing gets automated. That’s where the idiot index comes in.

Treating fuel as the “raw material” that goes into rockets, how many times larger should the final product (launch) be?

Today, fuel for a Starship launch is $1-2 million or 1% of overall costs; an idiot index of 100. SpaceX’s stated goal is to bring total launch cost down to $10 million, an idiot index of 10.

Looking at the idiot index for metals, they sit around 5. For passenger air travel, fuel costs are 15-25% of overall costs, an idiot index of 4-6. Launching and reusing rockets is more complicated than either of these, so an idiot index of 10 is generous.

Now to fuel prices, starting with methane. Looking at historical Henry Hub natural gas price let’s assume $3/MCF. With 19.26 kg of natural gas per MCF, that’s $0.158/kg.

For liquid oxygen, I get conflicting estimates. Wikipedia has a lovely (but outdated) page on Prices of chemical elements claiming that oxygen is $0.154/kg. Let’s be generous and round that down to $0.1/kg, lower than most estimates^[4].

Using propellant masses here, we can estimate fuel costs for a Starship launch.

Booster fuel cost:

($0.158/kg * 700 tonnes methane + $0.1/kg * 2700 tonnes LOX )/ 100 tonnes payload = $3.81/kg

Starship fuel cost:

($0.158/kg * 330 tonnes methane + $0.1/kg * 1170 tonnes LOX )/ 100 tonnes payload = $1.69/kg

Totaling $5.5/kg in fuel costs. Applying a 10x idiot index, the overall cost is $55/kg. This is the most optimistic case I can make for chemical launch costs.

The fuel price will be higher than calculated here. For instance, natural gas consumers pay over $10/MCF because of markups and delivery infrastructure. Rockets need high-purity liquid methane, which is probably why Starship fuel costs are 2-4x higher than my estimate. I doubt SpaceX^[5] is leaving money on the table, this is the best price they can get.

SpaceX Falcon Heavy costs closer to $2000/kg despite booster reuse, lots of flight experience, and a payload of 50 tonnes. Costs are higher than the theoretical minimum because this is literally rocket science!

$100/kg is probably the lowest cost we can hope for^[6]. It would still be a revolution in launch, but if we want to go further, we need to start thinking about space tethers and sourcing lunar material.

Asteroid mining

Don’t take my word for it, Casey Handmer says There are no known commodity resources in space that could be sold on Earth. The moon in particular is made of the same crust as Earth, but with fewer geological processes to form veins of ore. There’s little reason to fly there and back.

Perhaps there are precious metals in the asteroid belt? Generally, Earth has more concentrated pockets of metal due to ore-forming processes, but asteroids do have higher concentrations of platinum group metals.

Source

One problem: by the time someone grabs an asteroid, extracts the precious metals, and lands on Earth, the demand for these metals may have collapsed. Consider that “[a]bout 32% of the total Pt, 85% of the total Pd, and 90% of the total Rh were consumed by the automotive catalyst industry”. If we switch to electric cars during the decade they’re mining, the venture may be in the red^[7].

The broader problem is that asteroid mining might cost more than the price of the materials being collected. Challenges include:

• A large delta V to most asteroids, with the “closest” requiring about 4.5 km/s of delta V from LEO. Compare to 5.7 km/s for the Moon.
• Large financing costs since you need to buy expensive equipment and wait a long time for it to make the round trip. If it doesn’t break.
• Poor solar resources and little fissile material in asteroids, making energy a big challenge.

So maybe mining asteroids for precious metals isn’t the best idea. What about using asteroids for structural material in space? Well, the Moon is 30x larger than the entire asteroid belt, faster to get to, more amenable to space tethers, and awash in energy resources^[8]. So asteroids aren’t great for this.

There’s little reason to go to the asteroid belt^[9]. Establishing the capability to fly to the belt and fling rocks at Earth is also not great. I think people focused on asteroid mining should mine the moon instead.

The broader lesson is that there’s no reason to bring anything back to Earth from space. Leave the mass up there where it’s most valuable.

Space data centers

Let me be clear about one thing: space data centers are entirely feasible from a technological perspective. Starlink already has solar, compute, cooling, interconnect, communications, etc. Nothing stops you from putting computers in space. It’s also entirely possible to make a profit doing computing in space.

But space data centers are not cost-competitive with terrestrial data centers.

People are comparing space data centers on paper to the terrestrial data centers of today. That’s not a fair comparison. You need to look at where progress in terrestrial data centers is going and correct for the optimism inherent to an on-paper assessment.

My argument examines the different parts of a datacenter (chips, interconnect, comms, cooling, energy, etc.) and shows that in space the cost per unit of performance is worse at every step. If every step is more expensive in space, then terrestrial data centers must be cheaper.

Even if you don’t buy my argument, I have some useful things to say about the design of such datacenters.

Chips

This point is relatively uncontested: computer hardware has a lower price-performance in space.

Chip performance is worse in space because:

• Radiation causes bit flips and damages chips.
• Space debris can destroy your data center^[10].
• Maintenance is more infrequent, meaning more downtime and out-of-date hardware relative to terrestrial.

Chip costs are higher in space because:

• Chips need to be redesigned for launch and the space environment, including radiation shielding.
• Maintenance is more expensive.
• You need to pay launch costs^[11] to get the hardware into space.
• Additional costs of stationkeeping^[12].

The size of these effects varies widely. For example, chips are very small, the cost to launch them is minuscule. BUT it’s a non-zero additional cost you pay to put chips in space. This necessarily makes it (slightly) more expensive to do computing in space relative to Earth.

These factors lower the price-performance of computing in space relative to Earth. They apply equally well to other parts of the server such as interconnects, racks, wiring, and so on.

While servers and networking are 73% of data center costs, they’re not the only cost. If other areas become much cheaper in space, then the cost of putting chips in space might be worthwhile.

Cooling

Chips need to be near each other to keep interconnect latency low. These chips take energy from solar panels and produce heat. That heat needs to be dissipated into a radiator.

Rather than look at terrestrial data center cooling costs and speculate about the design of future satellite radiators, consider a simple test:

Imagine we design a radiator for space. We make two copies: one goes to space and the other stays on the ground. Which rejects more heat?

This is a very rough test; but it works for our purposes. If the space radiator has better performance on Earth, we can safely claim that space datacenters have higher cost-per-unit-cooling.

Why? One, because launching the radiator into space adds additional cost and lead time. Two, because a radiator designed for space makes the terrestrial performance look worse. We don’t use space radiators on Earth because we have better options.

In other words, this is an optimistic case for space radiators, enabling us to rule them out if they perform worse in space.

For both space and Earth, radiators enjoy radiative heat transfer, governed by the Stefan-Boltzmann Law:

Where ε is the emissivity of the radiator material, σ is a constant, T is the temperature of the radiator, and T_env is the temperature of the environment.

Earth also enjoys convective heat transfer governed by Newton’s law of cooling:

Where h is the convective heat transfer coefficient and A is the area of the radiator.

Some assumptions:

1. In both cases, the radiators are shielded from the sun’s radiation by the solar panels. The space radiator is also perfectly shielded from the Earth’s radiation. In practice the latter is not feasible because the Earth takes up a large part of the satellite’s horizon, but this is an optimistic case for the satellite.
2. Environmental temperature on Earth is 293 K, environmental temperature in space is 2.7 K (CMB)
3. The radiator is 1 m² of anodized aluminum with an emissivity of 0.9. Emissivity can’t be greater than 1, so there isn’t much room for better materials here.
4. Choosing the heat transfer coefficient is tricky. After looking at some examples, 10 W/m2-K seems reasonable^[13]. Adding a fan would 10x this number but that messes up the cost comparison.

Now we can look at the heat radiated as we change the temperature of the radiator (different from the temperature of the chips).

Code here.

Terrestrial chips operate around ~75°C. At this temperature, the radiator on earth rejects more heat than the one in space, and the gap grows wider at higher temperatures^[14]. Adding a fan to the Earth radiator (h=100) extends its advantage down to around 25°C.

For a wide range of radiator temperatures, the space radiator works better on Earth than in space. Given the additional costs of sending the radiator to space, cooling is cheaper on Earth than in space.

I’m glossing over details here, like the different temperatures for the chips and the working fluid; optimizations around chip temperature, leakage current, and compressor energy consumption. It’s plausible that we invent chips that make space radiators look better (see discussion of reversible computers). Then again, it’s plausible we invent chips that make the Earth radiator look better.

For now, it’s safe to say that cooling costs will be higher in space than on Earth.

The Earth’s atmosphere is a giant radiator. It emits 174 petawatts into space, seven orders of magnitude more than all AI capacity today^[15]. We’re nowhere close to saturating Earth’s compute capacity. This is why space AI is only interesting if demand outpaces Earth’s capacity to provide it.

See also:

Is It Really Impossible To Cool A Datacenter In Space? - YouTube

A humorous but information dense thread from Andrew McCalip.

Energy

Energy is ~7% of overall costs, so even if energy was free in space, costs would be at most 7% lower^[16]. Given that hardware costs 10x more and should increase in cost in space, that already looks like a bad trade.

There are three challenges for solar in space. The first was covered in the hardware section; hardware generally has a lower cost per unit performance in space including radiation, debris, maintenance costs, and launch costs.

The second is heat. Panels must face the sun and cool radiatively, leading to high equilibrium temperatures. Panels on the ISS reach 70°C in full sunlight, which hurts efficiency. This means using efficient and expensive panels (since any light not converted to electricity becomes heat, higher efficiency lowers equilibrium temperature).

The third challenge is stiffness. Panels need to be stowed in the rocket in such a way that they don’t get shaken apart by the ~40 Hz vibration during launch. Avoiding this requires stiffness which adds mass and cost. Once deployed, there are limits to how thin the panels can get. Varying gravitational forces can induce thin panels to flop and moving parts on the satellite can drive vibrational modes.

Now for a calculation:

1. The panel is in dawn-dusk SSO with constant illumination at 1362 W/m2.
2. A 5 year lifespan before the computer hardware has to be swapped out (with the upfront cost of swapping out the chips and the risk of panel failure, it will likely be more economical to replace the panels at the same time).
3. The GaAs panels are 30% efficient at 70°C
4. A panel mass per meter squared of 3 kg/m2. This is slightly heavier than the panels on the ISS, but accounts for additional structural mass required in the fairing and during deployment.

Result: $16.76/MWh before accounting for the idiot index. What index should we use to account for panel costs, design, assembly, and testing? Surely a factor of 3 is generous? The resulting $50/MWh is comparable to terrestrial solar today, but solar costs will continue to fall (see section on solar costs).

Space solar will continue to improve. More efficient and thinner cells may be achievable. Instead of areal density, let’s look at specific power. This page has a large list of different space panels and their specific power. It seems like 200 W/kg is an optimistic goal. A 5-year lifetime gives us $34/MWh with a 3x idiot index. Once again, probably not better than terrestrial solar.

It’s possible that new designs will improve the specific power enough to outpace advances in terrestrial solar, but once again, cutting down the small fraction of energy costs while damaging the hardware does not look promising.

Communications

Satellites are good for long distance communications. Light travels faster in a vacuum than in optical fiber, and satellites enjoy long lines of sight. So if you’re trying to communicate with someone far away, satellite internet works well.

But for AI inference, it’s easier to park compute close to users on Earth than in space. This is particularly true for tasks that demand extremely low latency. Robots, drones, and autonomous vehicles will need compute onboard, excluding space AI from these lucrative use cases^[17].

It’s a closer race for generative AI running in distant data centers. But space AI still struggles to compete. While Starlink latency is comparable to cellular data and only slightly slower than fiber internet, it has lower bandwidth and higher cost for most users. These facts are unlikely to change, Starlink is not planning to beat traditional internet on cost or bandwidth.

AI in particular adds new latency challenges. Satellites in LEO stay visible by your antenna for ~5 minutes. They spend some amount of time flying over oceans and sparsely populated areas. To avoid low utilization, queries need to be dynamically routed to different satellites. This problem is solvable, but the routing means that responses take longer to reach users relative to terrestrial.

So once again, the cost performance is worse than terrestrial counterparts. This is particularly true if we start parking data centers in cities and running more fiber.

Permitting

A common argument for space compute is that the permitting challenges are easier. I have yet to see an argument for why permitting is somehow easier when building launch pads, firing off hundreds of rockets, and filling up limited orbital slots. Rocket launches have permitting too! Indeed, there is more bureaucracy in launch because you have to deal with local ordinances, aviation restrictions, communications law, and international law^[18]. As we learn more about launch pollution, restrictions may tighten further.

I’ll edit in an additional response here if I see a good argument for why permitting will be easier for rockets than laying down solar panels.

A better way

This criticism lays the groundwork for a better approach to space compute.

For example, people are talking about increasing the chip temperature in space to decrease the size of the radiator. This is silly because increasing chip temperature increases things like leakage current, thus raising energy cost per FLOP and counteracting the benefits of a smaller radiator^[19].

Go the other way! As we saw in the cooling section, space radiators look better at low temperatures. Launch costs are proportional to mass, and the mass of the hardware and radiators is small relative to the panels. You can lower the energy cost per FLOP if you cool the chips or do reversible computing. The panels can shrink while the radiator grows more slowly^[20], lowering mass in net.

Reversible computers create an advantage to space over Earth. Space is cold, it’s easier to keep your chips cold in space than on Earth. That’s why reversible computing featured prominently in my post on the end of semiconductors.

What should space compute be used for? Onboard processing of satellite imagery, military intelligence, telecommunications, and stock markets in space are all far more profitable opportunities than AI inference. See that linked post for other ideas.

If you insist on AI inference in space, the economics push for large numbers of chips linked together with lots of fiber (perhaps spread out on the back of the solar panels). A single massive datacenter in orbit with satellite internet gathering requests from across the globe makes the most sense^[21]. The latency of this compute means it’s best for the low speed, low cost niche suited to autonomous AI workers.

So … space AI?

To be clear, it’s entirely possible to compute in space. It’s also entirely possible to make a profit computing in space. But it probably won’t be cost competitive with terrestrial datacenters. Every part of the tech stack has a lower performance in space relative to terrestrial.

You could fight this conclusion by stacking tethers and lunar materials and reversible compute, but then we’re not making a fair comparison to terrestrial compute. Terrestrial compute has a lot of opportunities to improve^[22]. By the time we build space infrastructure, we might be synthesizing solar from dirt or using a laptop AI to do inference over a hypercompressed local internet.

Space doesn’t need to be profitable

My fellow space enthusiasts, your dreams don’t have to be profitable to be worthwhile. For most endeavors, AI will teach us that anyways.

Going to space provides little economic benefit for those who are content to stay on Earth^[23]. It’s not clear that even the promise of a Dyson sphere could coax our economy into space^[24].

Instead, we will go to space out of desire, not necessity. Because we enjoy such widespread abundance on Earth that anyone can reach the heavens.

Further reading on space data centers, with an emphasis on good technical arguments and actually doing math.

Skeptics:

Economics of Orbital vs Terrestrial Data Centers

Do Orbital Data Centers Make Sense? - by Andrew Cote

The Truth About SpaceX’s “Orbital Datacenters” - YouTube

Notes on Space GPUs - by Dwarkesh Patel

Will we really put data centers in space?

Casey Handmer is more neutral here, though he thinks space AI would be 2x more expensive than terrestrial:

His Orbital inference Spreadsheet.

Direct Current Data Centers – Casey Handmer’s blog

Space AI: I guess we’re doing Moon factories now – Casey Handmer’s blog

SpaceX’s AI Data Centres Might Actually Be A Good Idea. Here’s Why - YouTube

Optimists:

Space Intelligence

Can We Build AI in Space?

Why Space-Based AI Data Centers Are Inevitable: 3 Levels of Analysis

Ramjets and demand for Mach 3+ flight

After my post on rocketplanes, I looked at hypersonic flight times between major cities. Mach 3 can get you between most major cities in under 3 hours.

Mach 3 is achievable with existing tech. The SR-71 blackbird used a turbofan with afterburner for an official top speed of Mach 3.3 (the real number is likely higher). In this regime, turbofans are more efficient than advanced concepts like ramjets.

So Mach 3 is fast, fuel-efficient, and uses existing tech. Is it worth the headache of going Mach 5 if you only shave an hour or so off your flight time? It’s not even clear the time savings will materialize if a Mach 5 plane needs more pre-flight checks or has to stop more frequently to refuel.

I’m still excited about companies like AstroMechanica, it’s great to see people try new things here. But after the dust settles, the future of passenger air travel looks to be in Mach 3, improving traditional flight, and flying cars.

Passenger rail

Trains are great for moving stuff over land, but moving people? Nope.

As Eli Dourado says, trains are not an abundance technology. If you want to go long distances quickly, an airplane can get you there faster and cheaper. And without the headache of building massive amounts of track with low utilization.

What about public transit in dense urban areas? Well, Ed Glaeser says:

Forty years of transportation economics at Harvard can be boiled down to four words. Bus good, train bad.

Simply maintaining roads, optimizing bus stops, and using congestion pricing to maximize throughput will produce a cost effective transit system. Ride-sharing autonomous vehicles and busses are about to make the case for roads even stronger.

Fusion

Fusion generates energy by colliding two atomic nuclei and releasing energy in the form of neutrons, electromagnetic radiation, and charged particles. The radiation can be converted to energy in two ways. One, charged particles can generate energy via direct energy conversion. Or two, X-rays, gamma-rays, and neutrons can be absorbed by radiation shielding, producing heat^[25].

We can probably rule out the second method. Fission reactors already do this using simpler, de-risked technology. They’ve been unable to compete with other forms of energy generation^[26]. Using fusion to produce even-more-expensive neutrons is not promising.

Direct energy conversion is interesting. Avoiding steam turbines is a big reason why solar is cheap. Maybe fusion could be cheap too.

I’m going to focus on D-D fusion for our optimistic analysis. While few companies are pursuing this (it’s harder than D-T fusion), it looks better for producing charged particles and the cheaper fuel lowers the on-paper cost. Assumptions:

1. The reactor performs Deuterium-Deuterium fusion without fuel reprocessing. When including side reactions with Tritium, this produces 8.35 MeV from 5 Deuterium atoms. That is 1.67 MeV of charged particles per Deuterium atom. Much of this energy comes from side-reactions with Tritium. This is the largest amount of energy a D-D reactor could produce from charged particles alone.
2. Charged particles can be converted into electricity at a 50% efficiency (probably higher than achievable at scale). Each deuterium atom produces 0.835 MeV.
3. Deuterium costs $4K/kg. This is highly uncertain, I’m basing it off of an out-of-date Wikipedia page. See the footnote^[27] for more.
4. An idiot index of 100. Also tricky to estimate, but the idiot index for fission plants is 10 and fusion promises to use much fancier equipment and more technical staff.
5. We ignore the value of energy produced via neutrons from the D-T side reaction. Traditional fission plants have not been able to generate energy cost effectively from such sources despite decades of effort.
6. We ignore any recirculating power required to maintain the fusion plasma.

The result is electricity at $36/MWh, with a wide range of uncertainty^[28]. That’s pretty good, comparable to an optimistic estimate of solar costs today (see next section). But fusion needs decades to mature, solar will be cheaper then.

I’ll admit, this estimate is not very rigorous, but it’s the clearest argument I can make for why I’m not excited about fusion energy. Brian Potter came to a similar conclusion:

Even if we could produce a power-producing reactor, fusion will never be anywhere near as cheap as simpler technology like the combined-cycle gas turbine, much less future technologies like next-generation solar panels or advanced geothermal. By the time a reactor is ready, if it ever is, no one will even want it.

(See also: Fusion power experience rates are overestimated)

What about using fusion power for rockets? Unfortunately, the energy density of fusion reactions is only ~2-4x higher than fission reactions, even without the higher induced mass for a fusion reactor. For space propulsion, fission fragment rockets or antimatter are more promising.

Are there better uses for fusion research? Well, proton-boron inertial confinement fusion should enjoy much lower fuel costs, I’d like to see more work here. The economics depends critically on the cost of particle accelerators and Bremsstrahlung losses.

The ability to control plasmas has applications elsewhere: MHD aerobraking might lower launch costs and wakefield plasma acceleration might give us cheap particle accelerators for research, lithography, and cancer treatment.

So while I’m skeptical that fusion will change the energy landscape, I’m happy to see people working on it.

Non-solar energy sources

Solar is very cheap; falling battery costs are fixing the intermittency issue.

All existing energy technologies are more expensive. There’s little prospect for them to catch up with solar today, much less the solar a decade from now. No proposed technology beats solar on paper, and in practice the costs will be higher.

To make a fair comparison to other proposed energy sources or space compute, we need an optimistic BOTEC for solar. Assume solar panels cost $0.4/W and batteries cost $50/kWh (China is already achieving these numbers). Assume we need 5 kW of panels and 17 kWh of battery capacity to achieve 1 kW of supply with 97% uptime.

For a 25 year design lifetime, we get $13/MWh. The idiot index for a future solar farm should be low, as the industry is moving towards containerized batteries and simple racking systems. The 2020 cost breakdown below isn’t perfect, but suggests an idiot index of 3-4. So at today’s prices the optimistic estimate would be $39-$52/MWh.

These costs will continue to fall as solar scales to terawatts. CATL claims its sodium ion batteries cost $19/kWh at the cell level, and aims to reach $19/kWh at the pack level in the future. I see many other opportunities to make solar cheaper using robotic installation, pre-assembly, thinner panels, multi-junction cells, metasurfaces, recycling, and incorporating solar energy into solar manufacture.

$20/MWh is achievable without resorting to science fiction, no other technology can say the same.

Further reading

How to Make Off Grid Data Centers Affordable - Austin Vernon's Blog

Simple Solutions Power Solar's Advance - Austin Vernon's Blog

Expanding the Universal Marginal Energy Source - Austin Vernon's Blog

Future of Energy Reading List – Casey Handmer's blog

Quantum computing

Consider what Scott Aaronson, preeminent quantum computing researcher, has to say here:

The trouble for the optimistic vision is that the applications, where quantum algorithms outperform classical ones, have stubbornly remained pretty specialized. In fact, the two biggest ones remain the two that we knew about in the 1990s:

1. simulation of quantum physics and chemistry themselves, and
2. breaking existing public-key encryption.

#2 is nearly obsolete. The Signal protocol already uses cryptography that’s expected to be secure to quantum computers. Ethereum and other cryptocurrencies are in the process of switching as well. The world simply needs to update their software^[29].

Once old cryptography is broken, old internet and personal data will be revealed. This might cause a few diplomatic incidents, but it’s not a big deal. Though breaking modern cryptography seems like a robustly bad thing to accelerate^[30]!

Now on to simulating chemistry on a quantum computer. This would be interesting, but it’s not clear how useful it will be. We already have good methods of simulating chemistry on classical computers. Quantum computers will be far more expensive and limited in scale for an unknown improvement in accuracy.

I’m also skeptical of unconventional computing more broadly, though I’m holding out hope for reversible computing in some applications.

Gene therapies in humans

Gene therapy is hard to do, hard to reverse, prone to error, and immunogenic. Most traits are virtually omni-genetic, meaning that you have to change lots of genes to get the desired effect. It is often hard to even identify the relevant genes or causality, much less target and change them. There have been many expensive clinical trial failures here.

Contrast with the wonders of small molecule drugs, which avoid many of these issues. Worse, consider all the modalities that might substitute for gene therapies: RNAi, RNA-targeted drugs, vaccines, antibodies, CAR-T, in-vivo CAR-T, stem cell therapies, and 3D printed organs. Each has an advantage over gene therapies in some aspect, and many are likely to be cheaper.

That said, monogenic disorders like Sickle cell, Tay-Sachs, Hemophilia, Duchenne Muscular Dystrophy, and Huntington’s are good directions for human gene therapy.

(Of course, genetic modification in plants and animals dodges many of my concerns and seems promising)

Brain-Computer Interfaces for healthy adults

Brain-computer interfaces are great for addressing disabilities. But it’s not clear that healthy humans can use BCI’s to output far more information than they already do.

Across a wide range of domains, we seem to be bottlenecked at 10 bits/s information output. Combine typing, speech-to-text, and eye tracking and you’re probably near the limit.

Things to work on instead:

• Ultrasound neuromodulation for adjusting mental states
• Increasing information throughput into the brain via VR, ultrasound tactile stimulation, digital smells and so on.
• Subvocal recognition
• EMG wristbands
• Highly interactive voice models

“Nanotech”

Philosophy spawned many disciplines and gets none of the credit. Once a field becomes a successful scientific endeavor, it gets called something other than philosophy.

Nanotechnology has the same problem. Chemistry, semiconductor manufacturing, and biotechnology are nanotechnology in their own right. We can build complex molecules, pattern arbitrary materials with nanoscale precision, and direct nanobots to modify our bodies. But we refuse to call this nanotechnology because it already exists.

These fields will continue to change the world. My skepticism lies with the nanotechnology that remains.

Early visions of nanotechnology had tiny robots that could carry out precise tasks. In harsh conditions and high temperatures, such robots are infeasible; their fine structures would degrade quickly, limiting their value. If you’re willing to accept standard conditions, proteins are a remarkable nanotechnology that can perform almost any desired physical transformation. No need to invent a new field when biotech can do it all.

For stuff we have today, it makes more sense to build it directly than to try to build it with nanotechnology^[31]. Say you want a car. You could feed the bots some metal, and they could assemble a car, but isn’t that just a factory?

So maybe the nanobots go out into the environment and gather metal. They need to be self-repairing and have some sort of search function and we need some way of beaming energy to them to break chemical bonds. And they need to transport all this metal to a concentrated location and assemble it into a car.

Instead of tackling dozens of near-impossible technical problems, you can just build a car the old-fashioned way: processing many atoms of ore simultaneously, arranging many atoms of metal all at once.

Perhaps there are some things we simply can’t make with current techniques? Not really, as discussed in the next section, we’re already approaching the limits of what’s possible in materials science. For computers, one of the few applications that requires nanoscale precision, lithography will take us to the limit of what’s possible. Nanomachines need not apply.

It’s in lithography that the last holdout of nanotechnology, atomically precise manufacturing, shows the most promise. The throughput and equipment costs are too high to replace industrial chemistry, but APM could be used for producing lithography masks, fixing defects, and metrology.

“New materials”

Structural materials

I watched a talk by a materials scientist where someone asked him “what material is the real-life equivalent of Vibranium?” The audience member expected him to choose something exotic like graphene or high-entropy alloys. But nope, he chose steel.

Steel is incredible; it’s extremely tough and possesses a strength-to-weight ratio that rivals aluminum in practice. It’s also cheap and can be made with a variety of properties and specifications.

Concrete is also an incredible material. You can ship some dust to a construction site and mix it with water to get instant stone with high compressive strength, consistent properties, and low cost.

And then there’s wood, a material synthesized from air and sunlight with high strength-to-weight, low cost, and integrity against fire (unlike plastics). It’s easy to cut to shape and you can improve its properties substantially with chemical modification.

For bulk structural materials, I don’t think we’re going to beat these three on cost per unit of performance any time soon. They’re local maxima that take advantage of available resources.

For alloys, we can only paint with the periodic table, and we have to make these out of Earth-abundant elements while obeying constraints on what elements alloy with each other. Much of this space has been explored, resulting in superalloys with incredible properties that make things like air travel possible. We’re pretty close to the ceiling here, though high-entropy alloys deserve more attention.

High strength-to-weight materials

When researching space tethers, I realized that there are very few options for high specific strength fibers. You pretty much have to make it out of a carbon allotrope or some sort of glass. For this reason, my excitement over discovering new tether materials waned^[32].

Electrical materials

Pure elements are generally more conductive, and copper is the ~highest conductivity element in practice. It’s also cheap and recyclable (aluminum works well too). It’s unlikely we’ll replace our power lines and wires with something else.

That said, superconductors are interesting as a new electronic component, particularly for reversible computers^[33]. I’m skeptical of analog computing, but if it takes off memristors will play a central role.

Optical materials

Glass is pretty great for optical fiber. Earth is made of the same stuff and the fibers can beam information accurately over ~100 km. The main contender here is ZBLAN fiber, also interesting because it might need to be manufactured in space.

For optical computing, the main challenge is making materials with consistent optical properties at scale. Fortunately, many familiar materials in semiconductor devices have good properties as waveguides. For the moment, putting exotic new materials on a chip runs counter to progress.

That said, optical metamaterials are pretty cool. The opportunity to replace precision-engineered lenses with something that benefits from the enchippening is exciting.

Catalysts

Many of the most important chemical processes^[34] are the product of decades of research, simulation, and testing. A big advance in major petrochemical processes is unlikely; we mastered organic chemistry in the late 20th century.

Some things we can make with ethylene.

There are still some opportunities here. I’m particularly interested in nanocrystals, modular peptides, and nanochannel glass materials as catalyst platforms. The shale revolution made methane a viable feedstock, raising the status of methane activation and methanol-to-olefins processes. Cheap renewable energy might make new electrochemical processes viable. Recycling atoms from crop residues, plastics, or waste is also interesting.

So … new materials?

Hopefully this hand-wavy tour gave you a sense of the frontier. The same issues are repeated across chemistry and materials science, namely that we’ve found cheap and effective solutions in most applications.

The frontier of materials science looks more like “try lots of existing materials and processing conditions for your specific application” rather than discovering some new composition on a computer.

Economics of home robots

Robots in the home have to walk a knife edge. Labor-saving innovation is valuable if human time is valuable, but if robots are very competent they make human labor cheap. So you need a weird situation where human labor is super scarce but also replaceable with robots but only in the home.

Outside of work, people spend 4 hours a day watching TV and 7 hours a day sleeping^[35], limiting the time savings from home robots. Historically, labor saving in the home has come from specialized items like the dishwasher. Robot vacuums were supposed to be the next step here, but the market has been lackluster.

The economics of robots become much better in an environment designed to use them all the time. Instead of folding laundry once a week, run the robot 24/7 in a laundromat. Everyone can get rid of their washer and dryer.

One of the biggest opportunities is food prep. Robotic ghost kitchens can churn out healthy, high-quality meals at scale. It would be far more expensive to have a dedicated robot and kitchen to feed you a few times a day.

(As an aside, there are reasons to be skeptical of the humanoid form in general. Robots will be adapted to their task, we’ll see an explosion of different designs.)

Conclusion

Consider this post via negativa for technology. Skepticism about one idea is optimism for something else.

Instead of fusion, our society will be rebuilt around solar. Instead of trains, hypersonic planes and flying cars will pull the world closer. We’ll build cozy cities afforded by self driving cars and hyperlogistics. And even if we never see AI in space, the AI of today is enough to automate anything.

1. ^{^}

   One thing that complicates this story is the shift in demand due to Herbert Simon style innovation. As prices rise, there’s consistent pressure to shift away from the resource, lowering demand and slowing price growth.

2. ^{^}

   That post also contains some other skepticism on green ammonia, biomass recycling, and point source CO2 re-use that may be of interest.

3. ^{^}

   This was written before the war in Iran. When I talk about natural gas prices, assume I’m referring to the long term price level, not the recent spike.

4. ^{^}

   These prices are already pretty low, it’s hard to find any form of matter significantly below $0.1/kg besides things like dirt, aggregates, and water.

5. ^{^}

   Elon invented the idiot index to criticize people who were buying products for much more than their true cost. This is something they think about a lot!

6. ^{^}

   Laser launch is another promising option, but unlikely to break the $100/kg barrier. Matterbeam is generally optimistic in his analyses, yet estimated $100/kg for laser-based launch. I am also skeptical of alternatives to chemical rockets.

7. ^{^}

   There’s also a risk that we figure out nuclear transmutation or start mining polymetallic nodules during this time.

8. ^{^}

   Solar, fissile material, microwave beams from Earth.

9. ^{^}

   See also this study on getting propellant from the asteroid belt: Assessing the economics of asteroid-derived water for propellant. They estimate a price of $3000/kg, worse than what Starship will likely achieve and much worse if using space tethers and atmospheric scoops.

10. ^{^}

    Satellites are also far more vulnerable during a hot war. Consider that a single exoatmospheric nuke can virtually wipe out LEO:

11. ^{^}

    Launch has an additional time delay. And you might need several launches to get enough hardware to inference large models. This time delay increases depreciation costs.

12. ^{^}

    For example, the satellites need to adjust positions so they can communicate with each other and the ground. They also need to raise/lower orbit regularly.

13. ^{^}

    A vertical plate in still air is closer to 6, but the induced convection from wind increases this. See equations 6 and 7 for example.

14. ^{^}

    Notice that the space radiator has an advantage at lower temperature, more on that later.

15. ^{^}

    Far more radiative capacity is available if we’re willing to use the deep ocean as a cold sink or allow Earth’s temperature to increase.

16. ^{^}

    It’s not clear how this will change over time. As chips get more efficient and solar gets cheaper, the energy cost share should fall. However, as hardware improves, the cost per token should fall as well, muddling the story.

17. ^{^}

    I also think local AI devices and interaction models will be big opportunities that require nearby compute.

18. ^{^}

    Orbital slots are limited economic land that deserve an appropriate allocation mechanism.

19. ^{^}

    The higher temp allows you shrink the radiators in theory, but leakage increases exponentially with temperature while radiation only increases with the fourth power of temperature so I’d guess the net effect is to increase the radiator mass?

20. ^{^}

    Less energy per FLOP means less heat to radiate away. So even though the lower temperature demands a larger radiator (or a heat pump) the lower energy demand counters this.

21. ^{^}

    For interconnect, RF over fiber is pretty interesting. Lightweight, cheap, less energy demand, and almost as fast as optical.

22. ^{^}

    For example, I’m excited about underground thermal energy storage, adsorption refrigerators, climate simulation for optimal siting, and using the deep ocean as a cold sink.

23. ^{^}

    With some notable exceptions covered here.

24. ^{^}

    A Dyson swarm kind of breaks the concept of profit. To be feasible at all, you need to bootstrap from a small initial investment. Once built, it’s not clear whether terrestrial beings will want so much high-latency compute.

    But assuming that being uploaded is in demand, what would the owners of the Dyson sphere possibly want from Earth in exchange for transcendence? We’ll sooner build this as a public good than as a profitable enterprise.

25. ^{^}

    Alternatively, radiation from the reactor can be used to drive a fission reaction.

26. ^{^}

    And I doubt a slightly different regulatory burden for fusion will change that.

27. ^{^}

    Deuterium has only low-volume industrial use today but if deuterium found use in fusion reactors, it may rise or fall in price. It is unclear if the current production process for deuterium can be improved. Could viable fusion suddenly make deuterium cheap? In the short term, prices would go up with higher demand. Long term, prices could fall. A casual look at other raw materials before and after they become economically valuable hasn’t revealed huge falls in price. If anything, prices rise and remain stable.

28. ^{^}

    Just guessing, the deuterium price could be off by half an OOM (3.2x) and the idiot index another half an OOM, so overall a 10x range.

29. ^{^}

    This is also why quantum cryptography and quantum key distribution aren’t very interesting. We just need to change our classical software to have good cryptography, no need for quantum mechanics to save us!

    There are a few cryptographic tricks you can’t do without quantum computers, so these might eventually find niche applications.

30. ^{^}

    Certified randomness, quantum summoning, and device independent-quantum randomness are neat though and might find some applications in cryptography.

31. ^{^}

    Like asteroid mining and quantum computing, nanotechnology is also dangerous, which makes it even less exciting. I don’t want to be turned into grey goo.

32. ^{^}

    And because building more tethers out of a weaker material can substitute for one strong tether.

33. ^{^}

    Though to be clear, superconducting electronics mostly just uses aluminum right now.

34. ^{^}

    Such as Haber-Bosch, steam reforming, electrolysis, cracking, catalytic reforming, and Fischer–Tropsch.

35. ^{^}

    Which is better addressed with sleep need reduction.