Who Counts the Capex

Correction, 15 October 2026: This piece said AMD was putting "up to $5 billion of its own cash into a customer", Anthropic, "with no warrants attached". AMD's release says it "has committed to make a strategic equity investment of up to $5 billion in Anthropic in the future", and its quarterly report records the commitment as "subject to certain contingencies". Nothing found supports "its own cash" or "no warrants attached". The details are in In Exchange for the Guarantees, dated 12 October 2026.

Interest disclosure: Tessera PE issues T-SpaceX, a tokenized loan participation right providing economic exposure linked to SpaceX, and T-OpenAI, providing economic exposure linked to OpenAI. Following the SpaceX listing, the investment underlying T-SpaceX is being divested. Redemption opens only once a Redemption Start Date is announced, and as of 20 September 2026 none has been. This piece therefore discusses a listed company in which Tessera's founder, and entities he controls, hold exposure that is currently being unwound. Read it as the view of an interested party, and discount it accordingly.

Two disclosures, two days apart, both about spending enormous amounts of money on semiconductors.

On 4 August, SpaceX reported second-quarter capital expenditure of $18.37 billion against a consensus estimate near $13.22 billion and quarterly revenue of $7.81 billion. Of that capex, $15.83 billion went to AI infrastructure. The next session the stock closed at $108.27, down 13.6%, its lowest close since listing.

On 6 August, SpaceX and Tesla confirmed Terafab, a vertically integrated semiconductor campus in Grimes County, Texas: a first phase of approximately $16.8 billion, more than 100 million square feet planned, and total investment across phases that SpaceX has indicated in filings could reach $119 billion.

The second disclosure deserves care, because it is easy to overstate. Terafab was not a surprise; it appeared in the May IPO filing and had been trailed publicly in March. What was new on 6 August was the site, the first-phase figure, and the scale. The interesting part is not whether it was hidden (it was not) but how it was characterized: per reporting on the filing, as a general framework carrying no binding commitments, with both parties free to walk away. It was announced on the company website, and it went unmentioned on the earnings call held two days earlier.

The stock rose that day, and rose again the next.

What Cannot Be Attributed

The strongest-looking evidence here needs disarming before anyone leans on it. The price move cannot be attributed, by this masthead or by anyone else. 6 August was also the day roughly 911.5 million shares came out of lock-up, close to $100 billion of insider paper, a far larger event than a site announcement and one covered here last Monday in Eligible Was Not Sold. The supply did not arrive. Analyst revisions landed in both directions in the same window. Short covering was reported.

And the whole move round-tripped. SPCX closed at $138.74 on Monday 10 August, above its $135 offer price for the first time since 15 July, reached $141.29 on 12 August, and closed at $140.56 on Friday 14 August, roughly 30% above the 5 August low. Another 320 million shares become eligible this week.

A signal that reverses completely inside five sessions was not a verdict on capital allocation. No argument below rests on it, and if this piece reads as a claim that the market punished capex and rewarded the announcement, it has been written badly. The argument is about the reporting regime, and it holds whether or not the tape does anything at all.

The Asymmetry That Does Not Depend on the Tape

Incurred capital expenditure, once reported, is a permanent fact about a company. It sits in the cash flow statement, depreciates through the income statement for years, gets reviewed, and gets re-examined every ninety days against whatever it was supposed to produce. It cannot be withdrawn.

Announced, phased, non-binding capital expenditure creates none of that. No obligation attaches, no period is committed, and nothing is measured against it, because there is nothing yet to measure. If it never happens, no line item ever records that it did not.

The dollar amounts in these two disclosures are comparable. Their contractibility is not, and only one of them starts a clock. That asymmetry is structural: it lives in the accounting rules rather than in investor sentiment, and it would be there on a week when nobody traded at all.

Concentration, Then Circularity

The AI buildout is not a set of independent balance sheets, but it interlocks in two different ways, and they are worth separating because they get conflated constantly and one is far more ordinary than the other.

The first is customer concentration. Per SpaceX's own filings, Google has agreed to pay $920 million a month from October 2026 through June 2029 for access to approximately 110,000 Nvidia GPUs and related components. Anthropic agreed in late May to pay approximately $1.25 billion a month for the full capacity of the Colossus 1 data center near Memphis, which xAI, now part of SpaceX, originally built for internal use. Reuters put the two together at roughly $26 billion a year. SpaceX's audited 2025 revenue, per the S-1, was $18.67 billion.

That is leasing rather than circular financing, and leasing is how every cloud business on earth works. What makes it worth noting is the concentration and the tenor: two counterparties, both contracts terminable by either side on 90 days' notice after 31 December 2026. A run rate of that size resting on two cancellable agreements is a contract rather than an annuity, and the distinction matters when it is being capitalized into a valuation.

The second interlock is genuinely circular, and it runs in both directions at once. AMD issued warrants to OpenAI in October 2025 and to Meta in February 2026, each for up to 160 million AMD shares at a cent apiece, roughly a tenth of the company each, vesting against deployment and share-price milestones. There, customers hold paper in their supplier. In July 2026, AMD went the other way with Anthropic: up to $5 billion of its own cash into a customer, tied to compute-capacity milestones, with no warrants attached.

Follow the sequence and revenue and share price stop being independent variables. The same silicon is one firm's capital expenditure, another's rented capacity, and the milestone against which a supplier's warrants vest. The arrangements are disclosed, and the accounting for them is not uniform across the parties, which is precisely why the aggregate exposure is hard to see from any single income statement.

To that, add a form of spending that bypasses the capex line entirely: SpaceX completed a $60 billion all-stock acquisition of Anysphere on Friday. Stock is not cash, and an all-stock deal never touches capital expenditure.

Who Is Paying for It

Hyperscaler capital expenditure for 2026 is now forecast at roughly $700 billion, up from about $387 billion last year, on Moody's estimates for the six largest US operators. Increasingly it is not funded from operating cash flow: FactSet's July analysis found incremental annual debt rising from about 9% of capital spending in fiscal 2024 to roughly 32% for the twelve months to mid-2026.

That is the number worth watching most closely, because it is the one that converts a valuation argument into a financing argument if revenue timing slips. Equity funding of research is a shareholder's problem. Debt funding of research is a wider one.

It should not be overstated. The same FactSet work notes total debt to EBITDA remains around 1x or below for four of the five hyperscalers it covers, and lower still net of cash. This is a change in the funding mix rather than a balance sheet in distress. What has changed is that the marginal dollar of research now arrives with a maturity date attached.

The Older Problem Underneath

Strip out the specifics and this is a financing-structure question that predates all of it. Basic research is the equity tranche of a capital structure: highest variance, longest duration, no reliable claim on near-term cash flow. Commercialization is the mezzanine and senior tranche: shorter duration, contractible, financeable. The arrangements that funded fundamental research at scale in the past respected that distinction. A regulated monopoly's protected cash flow carried it in one era; in others, a state balance sheet carried the science while a private operator commercialized the output: ITRI and TSMC in Taiwan, JNR's research program and JR Central's operating business in Japan.

Those examples pick winners, knowingly. For every ITRI there is a fifth-generation computer project or a Minitel, and the failure rate of state-carried science is exactly what an equity tranche should be expected to produce. That is the point rather than an objection to it: the tranche is supposed to fail often, which is why it needs a holder who can absorb the failures without being asked to explain them quarterly.

What is happening now differs in structure, not merely in scale. The equity tranche of a science program is being held by public shareholders and re-underwritten every ninety days. Nobody compelled that; buyers volunteered, at a price. But a quarterly reporting cycle is a poor instrument for pricing a research bet whose payoff distribution runs to decades.

The dispersion in analyst targets is what that mismatch looks like when it reaches a screen. On effectively identical financials: Piper Sandler cut to $140 with a Neutral rating; Wells Fargo trimmed to $215; Argus upgraded to Buy at $160; JPMorgan raised to $240; Bank of America reiterated $235; Morgan Stanley has stood at $300; Morningstar's fair value estimate is $62. The published range across the full analyst base runs from $62 to $800 against an average near $236.

Wide dispersion on a newly listed, hard-to-model company is normal, and Amazon ran wide for years. But a thirteen-fold spread goes beyond a disagreement about next year's margin; it amounts to an admission that nobody has a shared method for pricing the thing at all.

The Incentive That Follows

Here is the conclusion, and it requires nobody to have been punished by anything. If committing to spending creates a permanent, auditable, quarterly-marked obligation, and announcing spending creates none, then the cheapest place to hold an uncertain long-duration program is in the announced column. Not because the market rewards it (that, as above, cannot be demonstrated) but because the announced column carries no reporting cost, no measurement, and no downside if the program changes shape. Optionality is free there. Everywhere else it is expensive.

That is the regime the frontier labs are lining up to enter. Anthropic filed confidentially on 1 June and has been in a quiet period since. OpenAI has also filed confidentially and has reportedly pushed its listing to 2027. Databricks, able to raise privately without accepting a public mark, has stayed out. The option to remain private has become a valuation input in its own right.

None of this is a prediction, and none of it rests on explaining a two-day price move, which nobody can do from the outside. It rests on something duller and more durable: a reporting regime in which the spending a company commits to in writing is marked against it every ninety days, while the spending it merely announces is not marked at all, arriving at the exact moment the largest research programs in the economy are deciding how much of themselves to put in writing.

Sources: SpaceX S-1 and subsequent filings (SEC), including FY2025 consolidated revenue of $18.674bn; SpaceX Q2 2026 results, 4 August 2026, with consensus per FactSet via CNBC; SpaceX and Tesla Terafab announcement, 6 August 2026, with scope and IPO-filing characterization per TechCrunch and Electrek; lock-up expiry of 911.5m shares, 6 August 2026; Google and Anthropic compute agreements per SpaceX filings and reporting of late May and early June 2026, with the combined annual figure per Reuters; AMD announcements of October 2025, February 2026, and 22 July 2026; Anysphere acquisition completed 14 August 2026; Moody's hyperscaler capex estimates, June 2026; FactSet Insight on hyperscaler external financing, 23 July 2026; analyst ratings and price targets as published 4–12 August 2026, and consensus range data as of 14 August 2026. Share prices as stated. Figures are as of the dates given.

This is market commentary, not investment advice. Nothing here is a view on any Tessera product or on whether to hold exposure to any company mentioned.

T-Tokens are tokenized loan participation rights, not equity. High risk. DYOR. Not financial advice. Not available in the US or other restricted territories. https://tessera.pe/terms