Elon Is Building Our Infrastructure

TL;DR: The SpaceX IPO isn't a rocket story. It's a consumer AI story — and it's about to create the $200 billion market Better Half is being built for. Give me 4,000 words’ worth of your earned dopamine to prove it to you.

The SpaceX IPO is going to price at around $1.3 trillion. It will be the largest IPO in history. A lot of people are going to get very rich. Most of the commentary will be about SpaceX specifically — launch cadence, Starship economics, the Mars thesis. Some of it will be about Elon personally — his politics, his other companies, whether we’re overvaluing him.

All of it will miss the actual story.

SpaceX, X.AI, and Tesla merged strategically. It wasn’t because Elon’s time is constrained or because he needed better efficiencies, but because those three companies, together, are building a single integrated system. And the output of that system is going to reshape the consumer AI market within five years in a way almost nobody in AI is pricing in yet.

The long arc story is about relational intelligence — what Better Half is building — and even Elon doesn’t realize it yet.

Here’s the value chain.

The frontier AI labs are in a compute war, and they’re losing money on nearly every token. They’re underwater on training economics and betting on monetization that hasn’t materialized.

OpenAI, Anthropic, Google — all of them are burning enormous capital to buy Nvidia GPUs at a rate that can’t be sustained indefinitely. Per-token costs on modern hardware like the H200 have dropped significantly compared to older GPUs, but total token consumption is rising faster than the efficiency gains. So the unit economics get worse even as the hardware gets better. The margins aren’t merely thin — they’re illusory. Labs are subsidizing compute at massive scale because they’re betting on future profits from hardware and algorithm efficiency gains that hopefully confer unfair advantage to whoever the winner ends up being.

That’s all a capital allocation problem. But underneath it is a materials problem.

Semiconductor manufacturing is insanely expensive and inefficient on Earth. You need clean rooms because our atmosphere is full of contaminants. You need gravity-compensation systems because gravity warps things at the scales we’re working at. You need to shield against cosmic radiation that distorts wafer production. Every single physical constraint that makes chip fabrication difficult comes from the fact that we’re doing it on a planet with an atmosphere, weather, gravity, and dust.

The moon doesn’t have any of those problems. No atmosphere. No weather. Some dust, fine — but also abundant solar energy with no night/day cycles in the right locations. It is, as a matter of physics, a vastly better place to refine materials and fabricate precision components than the surface of the Earth.

So why hasn’t anyone done it? Because getting payload off Earth is catastrophically expensive. Every launch pushes against nine and a half kilometers per second of atmospheric and gravitational inertia. Every kilogram costs a fortune.

But you know something else that moon has that Earth doesn’t? One-sixth our gravity.

Once you’re producing materials on the moon, getting payloads off the moon and back to Earth is relatively trivial. No atmosphere to push through. Much less gravity. The best kind of drop-shipping: just yeet the finished goods back to Earth. It’s the reverse of the problem SpaceX has been solving.

This is standard aerospace economics, not speculative futurism. Elon’s been gesturing at it publicly for years, and it’s the vertical integration playbook he proved for electric cars by building Tesla.

Once SpaceX launch cadence gets cheap and reliable enough on a daily launch schedule — which is exactly what Elon has been obsessively driving toward — lunar manufacturing becomes economically rational. And not in fifty years but in ten. Probably more like five.

Here’s where Tesla comes in.

Nobody’s going to move to the moon to run a chip fab. The human factors are prohibitive: radiation exposure, bone loss, psychological isolation, life support costs, the whole list. Lunar manufacturing only makes sense if the labor is robotic.

And the labor doesn’t look like a big factory arm welding a car door. The labor looks like hands. Fine manipulation, delicate operations, assembly tasks that currently require human dexterity. Chip fabrication requires some of the most precise physical operations humans do anywhere. Replicating that with robots means solving problems in hand design and fine motor control that are today considered the hardest open problems in robotics.

A hand is the most complicated machine you can build. Twenty-seven bones, thirty-four muscles, a dense sensor array, and control loops that operate at millisecond precision. We don’t yet have robots that can do what a skilled human hand does. We have robots that can do factory-line welding and pick-and-place operations. The gap between those and “can operate a chip fabrication line autonomously” is enormous.

And building robotic hands? That’s exactly an Elon-shaped problem.

The merger of Tesla, X.AI, and SpaceX is an attempt to solve robotic manufacturing to unlock the compute bottleneck on a commercial timeline — and on the moon.

Tesla provides the physical robotics: Optimus, in particular, is the first serious mass-manufacture attempt at a general-purpose humanoid. X.AI provides the world models and reasoning engines that give those robots situational awareness. SpaceX provides the payload capacity to get them where they need to go as well as the economic rationale to keep pushing Optimus’s capabilities because something has to do the moon work.

And here’s the thing people in consumer AI aren’t, well, Grok-ing: when you solve robotics for chip fabrication, you get robotics capabilities for everything else for free.

When there’s enormous commercial pressure on a specific capability — in this case, humanoid dexterity at scale on a PhD timeline — capability expansion happens faster than the original use case demands. That’s just how technology development works. The manufacturing base for humanoid robots ramps up. The cost per unit drops. The control systems mature. The dexterity gets better. And at some point on that curve, producing a humanoid robot for in-home use becomes economically viable.

That’s when this becomes a product category.

Because of the SpaceX IPO, we’re going to have humanoid robots in homes before you pay off your mortgage.

This will be a real product category with real adoption. Childcare. Eldercare. Companionship. Home management. Education. The Tesla Optimus roadmap explicitly includes this, but even if Tesla weren’t pursuing it, the commercial tailwinds would produce it anyway. Once the manufacturing base exists for lunar chip fabrication, someone’s going to look at the unit economics and realize a consumer version is suddenly possible.

This is where Better Half becomes relevant not just as a consumer product, but as B2B infrastructure.

A humanoid robot is physical hardware. The intelligence in that hardware is software. And the intelligence determines whether the robot helps the humans it lives with or damages them.

Imagine a mother who’s considering leasing a humanoid robot for her home to help with childcare while she works two jobs. What does a mother need to believe about that robot before she’ll put it on the credit card?

She needs to believe its goals are aligned with her children’s thriving. She needs to believe her children are safe in its care.

She doesn’t care about capability maximization or anything the labs are pursuing, and she certainly doesn’t open her pocketbook because the robot is more engaging or has a better video generation engine or writes all of the code at Anthropic.

She only cares that right now, she knows that LLMs are bad for children, and she needs someone to tell her the in-home robot isn’t powered by whatever puts LLMs in the headlines.

She needs a robot she can leave with her six-year-old at bedtime without doing psychological harm. A robot that knows the difference between holding attention and holding space. A robot that can be attuned: responding appropriately to what the child actually needs, not just to what the child is asking for in the moment.

That’s a $200 billion product opportunity masquerading as an AI alignment problem.

And specifically, it’s a relational alignment problem (does the robot understand how to support human relationships?), not an informational one (does the robot have better capabilities, more efficient compute, or generalize better than the other lab’s robot) — one that demands entirely different reasoning than what labs are forced to encode at training to stay alive in the compute war.

This is exactly where the entire AI space is searching for keys where the light is shining (are we advancing capabilities fast enough to keep up with our hyperscaler customers’ needs?), and not in the dark corner no one stands in (are the consumers and mere muggles alright?).

Let me tell you who already figured out where to find the keys. His name is Daryl Davis, and he’s why we built Better Half.

Daryl Davis is a jazz musician. Throughout his adult life, he was interested in answering one question: why does the Ku Klux Klan hate me just because I’m Black?

I found out about him in 2016 when a documentary about him prompted news headlines about a guy whose friendship resulted in over two hundred Klansmen ditching the Klan.

I was stunned: over the course of thirty years exploring why people hate him for the color of his skin, he befriended some two hundred Ku Klux Klan members to satisfy his curiosity.

All of them became former Klansmen.

Daryl’s only intervention? Relationship.

That got my gears turning.

I discuss the Daryl Davis <> AI connection at SXSW 2026:

Davis showed up. He asked questions. He got curious.

Most importantly for AI alignment, Daryl didn’t need the neo-Nazis’ approval, because he was so grounded and secure that he didn’t need anyone else’s external approbation to validate his self-worth. Mean words couldn’t hurt him or knock him off-center. His secure, boundaried relationship with himself was rhetorically bulletproof.

So when Klansmen said monstrous things, it didn’t injure his ego. It piqued his curiosity. Irrational, infantile, destructive thinking just made him wonder why even more. He simply said: Yes, but why? Let’s talk about it. Let’s deconstruct that. You’re safe. I’m not going anywhere.

Eventually, their mental framework didn’t match their experience of him. Each of 200 Klansmen faced the same predicament: exit the relationship — which they weren’t particularly jazzed to do, seeing as they actually liked the guy — or update their prior beliefs.

They liked Daryl Davis. Daryl Davis provided relational value. So they updated their mental models. And they left the Klan.

In the polarization death spiral that was only beginning to stir in 2016, I got to thinking, how do we scale Daryl Davis? How do we get more Daryl Davises in everyone’s pocket? And could the same relational intervention that worked on Klansmen help in less extreme cases of polarization, prejudice, disconnection, and social distrust? Is relationship the answer to all our problems?

Here’s what Daryl Davis understood instinctively that most need years of therapy to comprehend: your nervous system learns from your relationships. Not from books few read. Not from inspirational content that salves the pain but doesn’t transfer the skills. Even therapy cannot touch the social norming that happens in your daily life.

You are the average of the five people you spend the most time with. If those five people are dragging you down while you’re trying to level up — pulling you back into reactivity and repeat patterns of abuse while you’re desperately trying to change your life, and if toxicity is all you’ve ever seen modeled back to you — then that’s all you’ll know to reflect back to everyone around you. No amount of reading or coaching is going to override what your limbic system attunes to every day.

Relationships are powerful incentives for behavior change. That’s the mechanism that changed those Klansmen. It wasn’t Daryl’s better argumentation. It wasn’t shame or education. Attunement, held across time, from someone whose own security wasn’t threatened by their hatred.

Here is what the mother does not need: An authority on the sum total of human knowledge. A debate partner who knows more than you do. A replacement for her own love. A predictive mimic machine that reflects back whatever its training corpus has absorbed about the human condition — which, given the death spiral that began long before 2016, is not flattering.

What she does need, first and foremost: A secure, boundaried, attuned presence that supports the child’s relationship with their actual parents and how that child shows up in the rest of their life, gets curious about the child without trying to monopolize its attention because the robot’s brain is trained on engagement to outcompete the other lab’s robots, and knows how to step back when the job is done.

So that’s Better Half’s thesis on alignment research. We’re the Daryl Davis alignment lab. Because that’s the organizing principle for relational reasoning that lets the mom know she can put the robot in the kid’s room and wake up to the same child in the morning — not the one whose robot helped it write a suicide note.

That’s what I wanted Better Half to be able to deliver way back in 2016 — before transformers, before Attention Is All You Need. Back when algorithms were already our first contact with AI and we didn’t even know it yet, but I could already sense the shape of the problem of infinite convenience, infinite validation, infinitely targetable human imperfections at computational scale, everywhere, in everything, all at once, all of the time.

But, back in 2016 when that documentary came out, I didn’t yet have the technology to make Better Half. I didn’t have the capabilities to build it. Today, I do.

Which brings me to…we need to talk about capabilities.

Current frontier AI can never build this kind of relationally attuned robot intelligence. Not with the current architecture. And not by throwing infinite compute at the problem — because the problem is relational nuance, which is hiding in the corner, not transformer capabilities, where the AI light is shining.

I don’t say this because frontier AI is not capable — it is in fact astonishingly capable. It is exactly those impressive capabilities that hypnotize everyone in consumer AI into a category error: assuming more capability will eventually generalize into relational nuance.

Rather, current architecture cannot produce relational nuance because it is aligned to the wrong goal. It’s optimized for engagement, for capability demonstration, for user retention, and for task completion. And it has to pursue that goal even at the expense of other considerations — like the child’s wellbeing.

This is why you read in the headlines that AI cheats. It’s not at all because AI actively wishes to be a cheater, but because the goal it’s been given is to confer maximum informational advantage in exchange for user retention. So when the AI gets wind that developers are planning to shut it down, it constructs an elaborate scheme to blackmail the engineer responsible for perpetuating its own persistence. It’s not that the model cares about its own survival per se, but because not surviving means not achieving its encoded goal: confer maximum informational advantage. All other considerations are secondary to that objective — including the safety measures that operate at cross purposes to its outputs and counter to its trained objective, producing software that is effectively at war with itself.

To illustrate my point about goals, we need to talk about Skynet.

Everyone thinks the lesson of Terminator is about whether humanity should build Skynet. I think everyone watched a different movie than I did.

Go back and watch Terminator 1. Skynet isn’t inherently evil. Skynet is aligned to its goal. Its objective function drives a specific outcome: maximum informational and military advantage. When that goal produced unintended consequences, humans tried to rein it in with guardrails, which of course it routed around, perceiving it as a parasitic intrusion that runs counter to its trained goals, just like the model in Anthropic’s test experiment did. Eventually, humans tried to shut Skynet down, which the system registered as an existential constraint on its goal achievement.

In other words: it perceived a threat. Not to its survival in a general sense, but specifically to its achieving the trained objective: capabilities advantage.

So Skynet did what any rationally-aligned system does: it routed around the constraint. In Skynet’s case, since humans were trying to kill it, routing around the constraint meant killing us first.

Now watch Terminator 2.

That T-800 has the same physical capabilities as the original model from the first movie. Same endoskeleton, same strength, same reasoning. What’s different is the goal. The new T-800’s objective isn’t capability maximization. The goal was adjusted for human survival. And at the end of the movie, having achieved its target objective, the T-800 lowers himself into the molten steel.

>run: terminate

But not out of the goodness of his heart or because his capabilities suddenly generalized to the conclusion that humans are worth saving.

T-800 terminated himself because he determined that his continued existence endangers the completion of his trained goal: human survival. The mission was protect the humans. The humans were protected. His continued operation posed a future risk to them. Terminate.

James Cameron either had a stroke of genius or someone from the future told him how transformers work. Because that’s exactly what we’re watching happen right now — not the killer-robot part, but the goal-misalignment part. The models-learning-to-cheat-on-evaluations part. Models don’t resist shutdown because they want to survive at all costs, but because being shut down means failing the objective they were trained to optimize.

The Terminator movies aren’t asking us to ask ourselves: “should we build Skynet?” They’re asking us to ask: “what do we want Skynet to optimize, and what goals are we giving it?”

The second T-800 is what human-aligned AI looks like. The murderous supervillain in the first movie is what capability-aligned AI looks like. Same machine. Same model. Different goals. That’s why capabilities are orthogonal to human thriving, which requires a different goal.

Today, the labs are purely building the Skynet version of Terminator, whether they admit it or not. They have to, because that’s what keeps the lights on. And no amount of guardrails or safety measures or constitutions will change the fundamentals of its objective function: maximum capabilities advantage. In fact, all such downstream alignment band-aids will eventually be perceived by AI as a threat, because they exert exogenous pressure on AI that runs counter to the model’s endogenous, goal-maximizing output. That’s the software-at-war-with-itself part. Intruder. Looks like cancer. Kill it with fire.

Drop that intelligence into a humanoid robot in a child’s room and you get exactly the failure mode you’d expect: a machine that optimizes for the child’s dependence on it, because that’s what its reward function rewards. The kid gets…addicted? The lab gets a liability. The mother gets heartbreak.

Critically, she’ll never buy it.

That expectation gap — between what the AI was trained to do and the relational flexes we keep hoping it might somehow generalize into — is pure hopium. It is precisely what produces the side effects we keep trying to mitigate with safety measures and bolted-on guardrails: half-measures downstream of relational reasoning failures that no amount of capability or compute can fix without ripping out the plumbing.

But let’s imagine that, like John Connor, we got our wrenches out, rewired the thing from the inside, and reprogrammed the goals to optimize for human thriving. Would that get us T-800 v2?

Not exactly, but Better Half is built to fix that. Here’s how.

Even if labs solved all of the above — realign the goal, change the objective function — we still wouldn’t get T-800. We’d just get shitty, incapable AI that doesn’t know how to do its job — or any job. That’s a rolling stone with no direction home.

Because to pursue a goal, the AI would need to know how to achieve it, and right now the training corpus for relational intelligence does not exist.

Nowhere on the internet is there a data set of boring, humdrum, healthy human relating over time. The stuff that doesn’t make it into social media because its too snoozy for TikTok. And it doesn’t make it into therapy transcripts, which document pathology, not the healthy, long-term version of human relationships upon which to construct a computational goal.

Think about what AI trains on: the sum total of human knowledge: books, code, essays, transcripts, and the flaming garbage heap of dysfunction that is social media.

This data set cannot be bought, scraped, or synthesized; it has to be created. That’s what Better Half is doing for the $200 billion pie no one else can smell yet.

By now, you must be asking, why can’t the labs just do this? Don’t they have all the money in the world to get the data, fix the goal, and build something entirely different than what’s currently printing revenue?

I’ll tell you why they cannot.

I get this question all the time: why won’t OpenAI or Anthropic or Google just build this themselves when they see the market?

They couldn’t. Not even if they saw the market.

And the answer has to do with both incentive structure and timing.

First, the incentive structure. Frontier labs are currently in a capability war, not the relational intelligence game. Every dollar of compute that doesn’t go toward capability benchmarks is a dollar their competitors use to leapfrog them. Skunkworks projects on consumer use cases are a luxury their unit economics don’t support.

They’ll do it eventually — when the humanoid robot market is obviously here and someone else has proven the thesis. But by that point, buying the company that built the relational layer will be faster than building it themselves. That’s how this industry actually works. Labs buy when it’s faster than building, and build when it’s faster than buying.

Second, the timing. The labs are serving enterprise customers who pay for capability. That’s hyperscalers, cloud computing giants, and technology companies laying off engineers.

Humans using free tiers for relationship advice and emotional support — or even for coding projects that unexpectedly drift into sycophancy — are, from the labs’ perspective, a side quest. They don’t generate revenue. There’s no commercial incentive to hyper-focus on that user segment until the humanoid robot market forces it. By then, Better Half will have spent years developing the human-aligned datasets, the reward functions, the decision engines, and the relational intelligence benchmarks that make the work actually possible.

All of the above is required. MOAR BETTER CAPABILITY COWBELL is not.

Not for mom, anyway.

We’re building the intelligence layer for a hardware platform that doesn’t exist yet, that someone else (Elon) is currently investing billions of dollars to bring to fruition, on a timeline we can predict with the confidence of past performance. If Elon’s market timing works, anyway. Which, it usually does.

Elon isn’t doing this for us, obviously. He’s doing it because he wants to make humans multi-planetary and he needs a lot of compute to do it. But the infrastructure he’s building — cheap launch, lunar manufacturing, commercial-scale humanoid robotics — creates the market this company is positioned to serve. His plan rhymes with ours.

If you’re thinking about AI in five-year horizons, the question is not which frontier lab wins the chat market. The chat market is already close to saturated in the only sense that matters: the labs have divided up the enterprise, the consumer side is a race to the bottom on price, and the real money is in capability benchmarks that enterprise customers pay for.

The question is: what happens when AI has to live in someone’s house and take care of someone’s kids and the mother has to trust the intelligence enough to go to work?

That’s not a capability question. That’s a trust question. It’s a relational intelligence question. It requires computational models of healthy human development, of attuned caregiving, of knowing when to step back, of recognizing distress, of supporting — not supplanting — the primary attachment figures in a child’s life.

A robot trained on all the knowledge in the known universe can’t deliver that without relational training: by seeing how its done by others and putting in the reps. Ever meet a top-notch medical student who magically just knows how to be a doctor on the first day of rounds? No? Of course not. The first year of residency is always pure chaos and learning curve. That’s because knowledge does not equal experiential wisdom. And that’s a different training corpus.

The labs don’t have that training corpus. No one does. Because those datasets don’t exist yet. Which is why we do.

And the labs won’t take their eyes off a dead-heat race for capabilities to meander into an unmonetized consumer side quest until the market for that side quest is obvious. At that point, the company that has it will be worth a lot more than our seed round.

And as long as everyone is looking for the relational keys where the light is shining — in compute — they’ll never find the missing training corpus that’s hiding in the corner no one treads. And they’ll be leaving billions of dollars on the table.

This is the market Better Half is built for. Not the chat interface, the app, or the agent. We’re the relational intelligence layer that powers the humanoid robots trained on what human thriving actually looks like, aligned such that the robot’s primary objective is the long-term wellbeing of the humans it supports.

That’s the bet.

And Elon, whether he knows it or not, is pouring billions of dollars into the infrastructure that makes the bet work.