The SpaceX Bull Case, Taken Literally

A response to a Hyperchange interview with Steven Mark Ryan,¹ recorded shortly after the SpaceX IPO. Quotes are lightly cleaned from the auto-transcript. Companion piece to SpaceX: Amazing Company, Terrible Investment.

Three weeks ago I argued that SpaceX at $1.77T is a great company priced as a religion, and put fair value at $329–591B. Fairness demands engaging the other side, so let's take the strongest bull narrative I've found: a long conversation between Hyperchange and Steven Mark Ryan, a man who spent ten years all-in on Tesla, skipped every private SpaceX round he was offered, and is now moving his savings into the stock at the IPO price. He is sincere, he has clearly thought about this a lot, and he has a valuation model he describes as "the biggest monstrosity I've ever created," running out to 2046.

So let's do something neither of them does in the video: write his claims down, put numbers on them, and run the same discounted-cash-flow machinery from my previous post. Not a strawman version. His numbers, priced charitably. Then I'll respond to the individual claims.

What He Actually Claims

Ryan's foundational thesis is that "demand for intelligence is near infinite," and that SpaceX can bring AI compute online faster, cheaper, and more coherently at scale than anyone else. Google and Anthropic leasing SpaceX compute is his proof: "why would they pay SpaceX when they could do it themselves?" From there the claims stack up. Extracting every number he commits to on air:

Two things about the model itself are worth knowing before we price it. First, he declines to share the actual spreadsheet or its outputs: "I'm not going to give the numbers away here cuz I want people who are listening to do your own homework." Second, the model was built with Grok Build, and its red-team is also Grok: he asks it to critique his assumptions, and when it produced a terrestrial compute estimate he found too low, he told it "bro, you're lowballing this," and it raised the number. The independent auditor of the model is an LLM that gets yelled at until it agrees. Keep that quality-control loop in mind; it is going to come back when we get to benchmarks.

What the Bull Case Is Worth If It All Happens

Now the exercise the video never does. Take his claims at face value, quantify them at their midpoints, and discount them exactly the way my previous post discounted the S-1: 8.2% WACC and 3% terminal growth (both Damodaran's figures), horizon extended to 2040 since his claims all live "in the next decade or two."

Read that again. Grant him three hundred million Starlink subscribers, a terrestrial compute business eighty times its current size, fifteen uninterrupted years of 25–34% compound growth, and the result lands below the price the stock trades at today. Push every input to his upper bound, 400M subscribers and $300B of compute, and you get $1,990B: about 12% above the IPO price, earned by assuming the single most successful corporate execution in the history of capitalism goes perfectly for fifteen years.

If that result feels impossible, here is the cross-check that makes it make sense. At an 8.2% discount rate and 3% terminal growth, the arithmetic of the price is fixed: $1.77T requires roughly $290B of EBITDA in 2040. For scale, Apple and Microsoft each produce about $130–150B today; the IPO price is a claim that SpaceX will be generating two Apples' worth of EBITDA within fifteen years. And Ryan's own headline summary in the interview is that Starlink and compute "cumulatively could be doing hundreds of billions if not half a trillion dollars or more in revenue in the future." Take his top figure, $500B of combined revenue at a blended 55% margin, and you get about $275B of EBITDA: his own maximal claim, discounted at market rates, is the IPO price, almost to the dollar. The market did not fail to hear the bull case. It transcribed it.

The two honest sensitivities run in opposite directions. Upward: my $30/month terminal ARPU and the $250B compute figure are choices, and more generous readings (a $60/month blend, or crediting "half a trillion" to compute alone) land in the $2.1–2.5T range. The most aggressive configuration I can construct, 400M subscribers at today's ARPU held flat plus $500B of compute, reaches $4.0T; but flat ARPU contradicts his own claim that prices fall and that the growth comes from low-bandwidth cell users, and it would make Starlink alone roughly a fifth of the entire global telecom services market. Downward, and larger: everything in this section treats EBITDA as free cash flow. No taxes, no capex, no replacing a GPU fleet that depreciates in five years. A real DCF would cut every number shown, his and mine, by a third or more. The overstatement runs in the bull case's favor throughout, and it still barely clears the ticker.

There's an equivalent way to say this that I find more damning. Solve for the discount rate instead: if his midpoint scenario happens with 100% certainty, buying at $1.77T earns you about 7.9% a year. That is an index fund's return, obtained by taking single-company risk on the most execution-dependent story in the market. Every deviation from the script and you underperform; the only way to genuinely win is the stuff that isn't in the model. Which is exactly why the second half of the interview exists: the orbital "zero or two," the moon mass driver, Grok discovering new physics. The bull case, taken literally, requires the parts that cannot be taken literally.

Claim by Claim

Direct-to-cell: the phone is the problem

His biggest Starlink claim is that direct-to-cell "blasts past" home broadband because each satellite cell can serve "an order of magnitude, maybe one and a half" more customers. This gets the physics backwards. The hard direction of a satellite phone link isn't down, it's up: your phone has to close a link to a spacecraft 550km away moving at 7.5km/s, using a radio designed to reach a cell tower a few kilometers away, with milliwatts of transmit power and an antenna the size of a fingernail. The satellite can compensate on its side with an enormous phased array; your phone cannot, because you are holding it. That link budget closes for messages, voice, and modest data. It does not close for anything resembling broadband, and no amount of Starship cadence changes what a handset can radiate.

And per-user bandwidth isn't even the binding constraint; area capacity is. A terrestrial network reuses its spectrum every few hundred meters; tens of thousands of small cells serve a single city. A satellite beam covers thousands of square kilometers with one cell's worth of shared spectrum. Direct-to-cell is a genuinely great coverage product: oceans, deserts, dead zones, roaming, the international-travel use case he correctly identifies as painful. It is a terrible capacity product, and "hundreds of millions of cell subscribers" is a capacity claim. My previous post left direct-to-cell out of the model as unquantifiable upside, in line with Quilty, who tracks Starlink professionally and doesn't model it either. It stays out.

Grok and Cursor: I used it

His evidence for the AI product is Composer's benchmark chart: roughly 10x better cost-performance, "nothing even close," attributed to a couple of months of training on SpaceX compute. Two problems.

First, the empirical one. I have actually used Composer, which Ryan, by his own cheerful admission, has not (he is "too much of a mid" to code). It is decent. For many tasks it is genuinely usable, and it is fast and cheap. It is also not remotely as competitive with Claude as its published benchmarks indicate. When the gap between a vendor's benchmark chart and the daily experience of using the product is this large and this consistent, I conclude the benchmarks are cooked. My previous post already noted that benchmark contamination is trivially easy and that Gemini got caught doing exactly this; nothing about the incentive landscape has improved since. Note also the epistemic loop in play: Ryan's SpaceX valuation model is audited by Grok, whose capabilities are attested by benchmarks, which are produced by the company he is valuing. Every layer of verification terminates in the same party.

Second, the logical one. If two months of leased compute genuinely produced a 10x cost-performance leap, that recipe is not a moat, because training compute is fungible and Google, Anthropic, and OpenAI have more of it than xAI has. A quantum leap that anyone with a cluster can replicate next quarter is not a differentiator; it's a price cut. To his credit, Ryan hedges here: his thesis "makes no meaningful difference" if Grok never wins, because the real business is renting out the compute itself. Fine. But renting compute is a commodity business with a brutal capex treadmill, and my previous post already walked through what Colossus 1's actual economics look like. "He who controls the AI compute controls the world" is a great line about a business whose product is, definitionally, an interchangeable commodity sold to the actual model builders.

Star Mind: the sun does not send invoices to anyone

The orbital section rests on two sentences: "your power is free once you're in space" and "the cooling is free." Both are wrong in the same instructive way.

Sunlight is free everywhere. It is free in orbit and it is free in the Nevada desert, in the sense that the sun bills neither location. What costs money are the solar panels, and orbit requires the same panels plus a rocket. What orbit actually buys you is capacity factor: a panel in a dawn-dusk orbit produces about 5x what the same panel produces on good ground, there is no night, and so you delete the battery, which is the real structural advantage; I worked through the numbers in GPUs in Space are dumb, right? and the storage swing is worth about $4,000/kW. That's a real edge. "Free" it is not, and the difference matters, because on the ground you can choose between grid power, gas turbines, solar-plus-storage, or a nuclear PPA, and in orbit you have exactly one option and a launch bill. Relocating to a place with fewer options is not an obvious win; it is a bet that the one remaining option's cost curve (launch $/kg) falls faster than the ground alternatives' (battery $/kWh). That race is the entire question, and ground has the head start and the larger industrial base.

The cooling claim is worse. "Space is essentially the coldest place anywhere" is true and irrelevant: vacuum is also the best insulator there is. Nothing convects. Every watt your chips dissipate must leave by radiation, at roughly 245 W/m² near room temperature, which means enormous radiators whose mass is, per my own post, the single binding uncertainty in whether orbital compute ever beats the ground at all. The thing he lists as free is the thing that decides the outcome. And even where the physics does work out, my post's conclusion stands: the launch cadence ceiling (Musk's 100 GW/yr needs ~80 launches a day, 180x the entire 2025 global cadence) means orbit's share of compute climbs over decades. It doesn't switch on. "Add a zero or two" within a model horizon is not analysis; it's the S-1's $22.7T enterprise AI TAM again, arrived at independently by the same method of multiplying enthusiasm by ten until the number feels big enough.

Terafab: bullish, just not for SpaceX shareholders

Terafab is the one under-discussed item where I partly agree with him: it's real, it's enormous, and it matters. We just disagree about who captures the value. Musk wants sovereign control of leading-edge chip supply for Tesla, SpaceX, and xAI, and given how many times he has been burned by suppliers, strategically that is coherent. But read the position honestly: being forced to build your own fab because TSMC won't prioritize you is a cost of your situation, not a moat. Leading-edge fabrication is arguably the hardest industrial undertaking on Earth, TSMC is the most capital-efficient near-monopoly in the world, and buying EUV machines is the easy part; yield is the whole game, and the expertise lives at TSMC, Samsung, and Intel. Terafab is accordingly leaning on Intel as a core partner to make the thing work at all.

So the investable consequences of Terafab are: ASML sells more EUV machines, to a buyer with infinite urgency and Musk's balance sheet behind it, and Intel's process expertise turns from a stranded asset into something a trillion-dollar project depends on. I am long both, in part for exactly this reason. What Terafab is not is a bull thesis for SpaceX or Tesla shareholders: it is a decade-plus of nine-figure-per-week capex at negative margins, chasing the best manufacturer on Earth, in pursuit of a strategic rather than financial return. "Produce more chips than the rest of the world combined" is a goal in the same register as the rest of the interview's numbers.

The rest of it

The moon mass driver, moon-rock heat shields for return payloads, asteroid retrieval, antimatter propulsion to other galaxies, Grok's "ultimate underlying mission to discover new physics," Mars as a new model of government: I'm not going to argue with any of this, because none of it is an argument. It's worldbuilding. Some of it may even happen; none of it has a date, a cost, or a revenue line, which are the three things a valuation is made of. The host says it best, entirely without irony: "they don't have a cell in their spreadsheet for antimatter propulsion." Correct. That is not a defect of spreadsheets.

The Tesla merger is the one item in the back half that is neither physics-constrained nor fictional: a vote probably does happen eventually, and Ryan's 85–90% on that is plausible. But a merger reallocates value between two tickers; it doesn't create any. Whether one overvalued company absorbs another at a negotiated ratio is a question about whose shareholders get diluted, not about what the combined machine is worth. It changes nothing in this post.

The Story, Told by Its Best Believer

My previous post argued that at $1.77T you're paying for $329–591B of modelable business plus a story. This interview is the story, told fluently and sincerely by the people who believe it most, and that's precisely what makes it useful: it shows you the structure of the belief. Every claim that can be priced, priced generously, lands at or below the ticker. The case for buying is therefore carried entirely by the residue that can't be priced, and the interview knows this, which is why it keeps reaching for it. At one point the host says the financials "are an insane delay and lag on the actual value of the technology." That sentence is the thesis of my previous post, stated as a bull point. When the price has detached from the financials, one of two things is true: the market is early, or the price is wrong. Both sides agree on the diagnosis and disagree only on which patient is sick.

The dynamic is the Tesla PsyOp again, and this time you can watch it operate on a single person. Ryan is not a rube; he's intelligent, he does more homework than most professionals, and he was right about Tesla for a decade. He was offered SpaceX in private rounds for years and declined, waiting for the IPO, which means he waited patiently for the single highest price at which the company has ever been available and is now buying at it, at roughly 99% portfolio concentration in one man's companies, on a model whose auditor is that man's chatbot. The 30% retail allocation from my previous post's mechanics section isn't an abstraction; it has a face, and the face is doing a podcast explaining why dissent from Musk requires a track record and everyone else should "shut the [expletive] up."

Same conclusion as last time, sharpened by the exercise: if everything the most bullish informed person on the internet believes comes true, on schedule, without a single stumble, for fifteen years, the stock returns about 8% a year from here. That's the ceiling of the literal case. Everything above it is antimatter.

# Same machinery as the SpaceX IPO post: 8.2% WACC and 3% terminal growth, both
# Aswath Damodaran's figures. Horizon extended to 2040 (15 years) because Ryan's
# claims all live "in the next decade or two"; his own model runs to 2046.
WACC, TERM_G, YEARS = 0.082, 0.03, 15  # 2026-2040

def pv(ebitda_path):
    pv_explicit = sum(e / (1 + WACC) ** t for t, e in zip(range(1, YEARS + 1), ebitda_path))
    tv = ebitda_path[-1] * (1 + TERM_G) / (WACC - TERM_G)
    return pv_explicit + tv / (1 + WACC) ** YEARS

# --- Starlink, Ryan case: 300M subscribers by 2040 ---
# Midpoint of his "100-400M feasible". Blended ARPU declines from today's ~$92/mo
# to $30/mo as the cell-service mix he predicts takes over (his own claim that
# ARPU trends lower). 60% margin is Damodaran's long-run connectivity target.
subs0, subs_end = 10.3e6, 300e6          # 25.2% CAGR for 15 straight years
arpu0, arpu_end = 11.4e9 / 10.3e6, 360.0  # $/year
g = (subs_end / subs0) ** (1 / YEARS) - 1
revs = [subs0 * (1 + g) ** t * (arpu0 + (arpu_end - arpu0) * t / YEARS) / 1e9
        for t in range(1, YEARS + 1)]
starlink = pv([r * 0.60 for r in revs])
print(f"Starlink: 2040 revenue ${revs[-1]:.0f}B, fair value ${starlink:.0f}B")

# --- Terrestrial compute, Ryan case: "multi-hundred billion a year" ---
# He declines to give his number on air; $250B by 2040 is a fair midpoint reading.
# 50% EBITDA margin, and treating EBITDA as cash flow for a business that must
# replace its chip fleet every ~5 years is itself a large gift.
c0, c_end = 3.2, 250.0                    # 33.7% CAGR for 15 straight years
gc = (c_end / c0) ** (1 / YEARS) - 1
compute = pv([c0 * (1 + gc) ** t * 0.50 for t in range(1, YEARS + 1)])
print(f"Compute: fair value ${compute:.0f}B")

launch = 81.0  # unchanged from the prior post; Starship is his enabler, not a revenue line

total = starlink + compute + launch
print(f"Total, granting his midpoint claims: ${total:.0f}B vs IPO $1,770B")

# Upper bound: 400M subscribers, $300B compute -> $1,990B, ~12% above the IPO price.
# Orbital compute ("add a zero or two" = $2.5T-$25T revenue) is deliberately absent;
# see the Star Mind section and the gpus-in-space post for why.

# Cross-check: what does the $1,770B price itself demand?
df, tmult = (1 + WACC) ** YEARS, (1 + TERM_G) / (WACC - TERM_G)
print(f"2040 EBITDA required to justify $1,770B (TV-only): ${1770 * df / tmult:.0f}B")
# ~$290B: two Apples' worth of EBITDA by 2040. Ryan's own "half a trillion cumulative
# revenue" at a 55% blended margin gives ~$275B, i.e. his maximal claim IS the price.

# Sensitivity: the upward knobs (terminal ARPU, compute revenue), all still pre-tax,
# pre-capex, EBITDA-as-cash-flow. $60/mo ARPU -> total $2,121B. Compute $500B alone
# -> total $2,533B. Flat $92/mo ARPU at 400M subs + $500B compute -> $4,033B, but
# flat ARPU contradicts his own pricing claim and equals ~20% of global telecom revenue.

# Implied return: the discount rate at which his midpoint scenario equals $1,770B.
lo, hi = 0.031, 0.20
for _ in range(60):
    mid = (lo + hi) / 2
    def pvw(path, w=mid):
        pe = sum(e / (1 + w) ** t for t, e in zip(range(1, YEARS + 1), path))
        return pe + path[-1] * (1 + TERM_G) / (w - TERM_G) / (1 + w) ** YEARS
    t = pvw([r * 0.60 for r in revs]) + pvw([c0 * (1 + gc) ** t_ * 0.50 for t_ in range(1, YEARS + 1)]) + launch
    lo, hi = (mid, hi) if t > 1770 else (lo, mid)
print(f"Buy at $1,770B, his scenario plays out with certainty: {(lo + hi) / 2:.1%}/yr")  # ~7.9%