Taking Half of Anthropic and OpenAI's Revenue Would Be Worth $27.84 a Share to SpaceX

Update, August 19, 2026: the Model layer vertical described here is no longer a what-if. It is now part of the SPCX model, which moves the published fair values to $39.25 bear, $187.12 base, $467.81 bull and $882.19 Elon — the last of those on a 60% terminal share of the pool. The fair-value tables below were run before that revision and price the vertical as an overlay on the old $172.13 base; read them as the sensitivity study that produced the committed assumptions, not as the model's current output.

The argument goes like this. The frontier model layer is a duopoly: Anthropic and OpenAI booked a combined $18.3 billion in the June quarter and, on the estimates circulating, hold something like 70% of an annualised market near $145–150 billion. Grok is an order of magnitude behind. But models are not databases — there is no schema to migrate, no data to move, and the integration is an API key and a prompt. If Grok 4.7 or 5 matches the frontier on quality while beating it on speed and price, then the thing keeping developers and enterprises on Claude and ChatGPT is habit, and habit can move in a quarter. SpaceX's AI segment could therefore take a large share of an $18 billion quarterly pool faster than anyone is modelling.

Every step of that is plausible. We modelled it anyway, because the interesting question is not whether it could happen but what it would be worth if it did, and the answer is smaller than the argument implies — for a reason that has nothing to do with Grok.

What SpaceX actually reports

Start with the disclosure, because the pool the argument is about does not appear in it.

The AI segment is $2,561M, up 247% year over year, and $2,194M of it is AI solutions and infrastructure — capacity sold by the megawatt, driven by Cloud Services Agreements against $14.1 billion of contracted sales. Advertising, the X platform's contribution, is $367M and shrinking. Grok is not broken out. There is no line in SpaceX's reporting that corresponds to the model-layer market the argument is about, which is the first thing to know before sizing a share of it.

Our own model says the same thing in a different way. The Terrestrial AI vertical carries "Grok / xAI" as a stream, but its driver is capacity: 1,800 MW energised at $2.03M per MW per quarter, 70% utilised, priced down 1.5% a quarter. That is a landlord's model, not a model-vendor's. It prices the building, not the tenant.

The prize, against the company

Now the pool, on booked quarterly revenue rather than annualised estimates:

Neither company files. Both figures are press-reported and neither is auditable, which is a caveat that has to travel with every number below.

Against SpaceX's $7,814M, the two of them together are 2.34x the entire company and 7.1x its AI segment. That is what makes the argument attractive. A pool more than twice your size, in a business you are already in, with switching costs measured in hours.

So we put it in the model.

What it is worth

We added a Model layer vertical to the SPCX model — a pure revenue line ramping from Grok's reported run rate to a target share of the pool over twenty quarters, at software economics gliding to a 50% EBITDA margin and 8% capex intensity. To avoid crediting the same dollar twice, the starting run rate is reclassified out of Terrestrial AI, so total June-quarter revenue is unchanged; only its character changes. Fair values are the model re-projected, not estimated.

The first table assumes the pool never grows again — that Anthropic and OpenAI are frozen at $18.3B a quarter forever, and SpaceX takes a share of that fixed pool by 2031 Q2:

Taking half of Anthropic and OpenAI's current revenue — a commercial event without much precedent — moves the base case from $172.13 to $199.97. $27.84 a share, about 16%. Taking a quarter of it is worth $14.64.

For scale: our Developer Platform vertical, the entire Cursor acquisition at $60 billion, is worth $8.40 a share in the same case. So the total destruction of the model-layer duopoly's current revenue base, half of it landing at SpaceX, is worth about three and a bit Cursors.

That is the finding, and it is not what the argument predicts.

Why the number is small

Because $9.15 billion a quarter is not large inside this company's own projections. The base case already carries a terminal quarter near $56 billion of revenue, most of it Starlink subscriptions and megawatts. A model-layer line arriving at $9B a quarter in 2031 is roughly a sixth of that, at similar margins, discounted five years. The valuation is dominated by the terminal multiple — 20x exit EV/revenue in the base case, which our own model notes is the largest and least derivable input in it. Nothing about who wins the coding-agent market moves that slider.

Change the assumption that actually matters and the answer changes completely:

Read the middle two rows together. Taking 10% of a pool growing at 10% a quarter is worth more than taking 25% of a pool growing at 5% — and both are worth more than taking half of a pool that stops. The share shift is the small variable. Whether the model layer keeps compounding is the large one, and SpaceX's exposure to that is mostly through the megawatts it already sells to everyone, including to the incumbents it would be displacing.

Ten quarters of the three of them

So model the pool rather than the share of it. Three players, ten quarters, three cases — and these are cases for the model layer, not the Bear/Base/Bull columns of the SPCX model above, which are a different set of assumptions about a different company.

Each case fixes two things: how fast the combined pool grows, and where xAI's share of it lands by 2028 Q4. None of the three has the pool shrinking; the argument being tested is about who wins a growing market, and a case that kills the market answers a different question. Anthropic and OpenAI split whatever is left in today's proportion, 63/37, so any share xAI takes comes out of both of them evenly. The June quarter is the actual; everything below it is arithmetic.

Bear — the pool crawls at 3% a quarter. Token deflation eats most of the volume growth, roughly 13% a year survives, and Grok grinds to 6% of it.

Base — 6% a quarter. Volume outruns price cuts by a fair margin, the pool roughly 1.8x's in two and a half years, and Grok reaches 12%.

Bull — the compute thesis. The pool grows 10% a quarter, Musk energises megawatts faster than anyone can match, and Grok takes 30% of the market by the end of 2028.

What the bull case actually says

Read the three endpoints side by side and the first conclusion is not about Grok at all:

Anthropic and OpenAI grow in every case. In the case where Grok takes 30% of the market, they grow 87%. That is the whole answer to "taking a big chunk out of the other two": nothing is taken out of anyone. Once the pool compounds — and it compounds in all three, because a model layer that stops growing is a different argument than this one — share is a ratio, not a transfer. Anthropic ends the bull case at $21.7B a quarter, nearly double where it started, having lost two thirds of its share of a duopoly that no longer exists.

The spread between the columns says the same thing from the other side. Going from the bear to the bull adds $6.6B a quarter to Anthropic and $3.8B to OpenAI — more, in absolute dollars, than Grok's entire bear-case business. The incumbents are far more exposed to how fast the market grows than to how much of it Musk takes. So is SpaceX: the difference between Grok at $1.5B and Grok at $14.7B is overwhelmingly the pool growth rate, not the share ramp, which is exactly the result the fair-value table found before any of this was projected.

Which leaves the argument needing something it has not yet supplied. For "Grok takes a big chunk out of the other two" to be a claim about dollars rather than percentages, the pool has to stall while Grok scales — growth near zero and share near a third at the same time. That combination is not in the table because it is close to self-contradictory: the thing that would stall the pool is price collapse, and price collapse is the mechanism by which Grok is supposed to be winning the share.

Does the compute hypothesis survive its own arithmetic?

The bull case rests on one claim: that Musk can scale compute faster than Anthropic's and OpenAI's suppliers can, and that serving capacity converts to share. Our model has a number for the first half of that, so it can be checked rather than asserted.

The Terrestrial AI vertical energises 1,800 MW today and adds 400 MW a quarter, compounding at 10%. Ten quarters of that is:

Compute grows 4.5x. The bull case needs revenue to grow 26.7x. The gap is a factor of six, and there are only three places it can come from: xAI taking a far larger slice of SpaceX's own capacity than it has today, revenue per megawatt rising, or capacity being bought from somebody else. The second is the one the thesis implicitly assumes and the one the timeline contradicts — this is a business whose competitive weapon is cutting the price of a token, and our own model already assumes lease prices fall 1.5% a quarter for exactly the reason that everyone is building at once.

There is a harder version of the same check. At 8,175 MW, 64.6% leased to third parties, at $1.75M per MW per quarter after ten quarters of price drift, the entire ground-compute leasing line bills about $9.2B a quarter in 2028 Q4 — every megawatt SpaceX owns, sold to every tenant including the hyperscalers and, on the timeline, to the incumbents Grok is supposed to be displacing. The bull case has Grok alone at $14.67B — more than half again the entire compute business that is supposed to make it possible.

That is not a refutation of the compute thesis. It is a statement of what the thesis costs: for Grok to reach 30% of a market growing at 10% a quarter, SpaceX has to build materially faster than 400 MW a quarter compounding at 10%, hand a dominant share of it to its own model rather than to paying tenants, and do it while the price of what the model sells falls. Each of those is defensible on its own. Together they are a different company than the one in the model, and the place to argue for them is the capacity slider, not the share slider.

The problem underneath all of it

There is a reason to doubt the growth assumption, and it arrived on the timeline the same week: average AI token prices have fallen from roughly $2.10 to $1.00 per million tokens in two months.

Revenue is price times volume. A 52% price decline over two months compounds to roughly −67% a quarter. To grow revenue even 5% a quarter through deflation like that, token volume has to grow 220% a quarter — and sustaining that for twenty quarters implies a volume increase of about ten billion times. It will not happen. Either the deflation slows sharply, or model-layer revenue growth is far below the rates in the tables above, and possibly negative. That is a warning about our bear case, not only about the bull one. The bear assumes the pool still compounds at 3% a quarter — 13% a year, and $25.3B a quarter by the end of 2028 — because a case in which the market shrinks is answering a different question than the one this piece was asked. If token prices keep halving, the 3% floor is the optimistic assumption in the set, and every column of every table above is too high.

That reframes the whole argument. The case for Grok taking share rests on being cheaper and faster. Cheaper and faster is exactly the mechanism destroying the revenue pool being fought over. Our own model already says so in a place nobody looks — the Developer Platform's ARPU rationale reads "up 1% a quarter as agents run longer jobs. The load-bearing number, and inference price cuts work against it."

The company positioned to survive that is not the one with the best model. It is the one that owns the megawatts, bills for them by capacity rather than by token, and sells the editor the tokens are consumed in. Which is, as it happens, a fair description of what SpaceX bought this month.

The case against this piece

The pool numbers are press reports about private companies, and Anthropic's +143% sequential quarter is a figure no auditor has signed. The ARR estimates behind the "$145–150 billion market" are estimates, and we are dividing by them.

The Model layer vertical is our construction, not a disclosure. SpaceX does not report Grok, so the reclassification of $0.55B a quarter out of Terrestrial AI is an assumption about a number the company has never published; if Grok's actual revenue is materially different, the ramp starts in the wrong place.

The ten-quarter projection is arithmetic on assumptions, not a forecast. The pool growth rates — 3%, 6% and 10% a quarter — and the share endpoints of 6%, 12% and 30% are chosen to bracket the argument, not derived from anything. All three grow the pool, which loads the deck in a specific direction: it guarantees the incumbents grow in every column, which is the finding. A fourth case with the pool flat or falling would show real dollar losses at Anthropic and OpenAI, and the section above on token prices is the argument that such a case deserves to exist. It is not in the tables because the claim being tested is about winning a growing market. The share ramp is also a straight line, which no market has ever followed, and holding the Anthropic/OpenAI split fixed at 63/37 assumes the two are equally exposed to Grok, when the June quarter — +143% against +18% — is evidence they are not.

And the reason the answer comes out small is partly structural: a twenty-quarter DCF with a 20x exit multiple will always be dominated by the terminal assumption, so any new revenue line of this size looks modest. That is a fair criticism of the method rather than a defence of it. The right response is that the same method priced Cursor at $8.40 and nobody objected, and that a reader who thinks 20x is wrong should move that input first — it is worth more than everything in this article.

What to watch

1. Whether SpaceX ever breaks out Grok. Today the model layer is invisible inside a capacity line. A separate disclosure is the moment this stops being a modelling exercise.
2. Q3, the first quarter carrying Cursor — six weeks of it — and which segment it lands in. Internal inference consumption is not AI-segment revenue, exactly as internal Starlink launches are not Space revenue.
3. Token prices. If the $2.10-to-$1.00 path continues, the model-layer revenue pool shrinks in dollars while growing in tokens, and every share-capture argument written this year is measuring the wrong quantity.
4. Whether Anthropic's +143% quarter repeats. One quarter is not a rate. If it does repeat, the bottom row of the second table is the one to reread.

SpaceX segment figures are from its Q2 2026 release as stored on this site: revenue $7,814M, Connectivity $4,291M with $1,656M of operating income, AI $2,561M with a $(1,257)M operating loss, Space $962M with a $(542)M operating loss, consolidated operating loss $(143)M. Anthropic's $11.6B and OpenAI's $6.7B June quarters, OpenAI's $5.8B March quarter and the ~$145–150B market-layer estimate are press-reported figures about private companies; neither company files with the SEC and none of these numbers is auditable. The ten-quarter three-player projection is our own construction on the same press-reported base: the pool grows at 3%, 6% or 10% a quarter, xAI's share ramps linearly to 6%, 12% or 30% by 2028 Q4, and the residual splits 63/37 between Anthropic and OpenAI throughout. The capacity check uses the Terrestrial AI driver as published — 1,800 MW energised, 400 MW a quarter of adds compounding at 10%, third-party utilisation gliding from 55% toward 67%, price drifting −1.5% a quarter — which gives 8,175 MW and about $9.2B of quarterly leasing revenue in 2028 Q4. That utilisation is net of the fifteen points of the fleet the August 19 revision reassigns to Grok's own inference; the pre-revision figures were 70% toward 82% and $11.4B. The $2.10-to-$1.00 token price move is a single third-party claim and is used here as an order of magnitude, not a measurement. Fair values are our SPCX model at its August 18, 2026 revision, re-projected with an added Model layer vertical whose starting run rate is reclassified out of Terrestrial AI so that reported June-quarter revenue is unchanged; the published columns reproduce the model exactly. The price basis is the August 18, 2026 close of $146.23 across 13.176 billion shares, and it will not be the number you see today.