SpaceX's $235 AI Theorem: An Audit of Bank of America's Infrastructure Ledger

August 7. The date matters less than the numbers attached to it. Bank of America issued a Buy rating on SpaceX with a $235 price target. The reference price at publication: $125.33. That is an 87.5% premium, and the premium is not justified by rockets. The projections that follow are not revenue estimates. They are audit findings awaiting verification. Negative free cash flow of $43.6 billion in 2026, deepening to $45.4 billion in 2027, and still negative at $37.4 billion in 2028. Cumulative: negative $126.4 billion across three years. No launch provider in commercial history has funded a buildout of this magnitude from operations. The report does not explain who pays for it. The thesis has shifted. The bull case no longer rests on launch cadence or Starlink subscriber growth. It rests on AI infrastructure revenue reaching approximately $24.5 billion in 2026 — 52% of the projected $46.9 billion total. Two counterparties anchor this model: Anthropic, contributing since May, and Google, with compute delivery expected to begin in October. The ledger does not lie, it only waits to be read. This ledger is incomplete. The transition from aerospace contractor to AI infrastructure provider is the most consequential narrative pivot in the sector this decade. Bank of America's report is not alone in this framing; it is simply the most explicit. The revenue trajectory is the argument: $46.9 billion in 2026, $100.7 billion in 2027, $184.8 billion in 2028. The compound annual growth rate implied between 2026 and 2028 is approximately 99%. Let me put that in perspective. Amazon Web Services took roughly a decade to exceed $100 billion in annual revenue. Microsoft Azure followed a similar arc. The projections suggest SpaceX will go from a standing start to $184.8 billion in three years, a rate of revenue accumulation no infrastructure company in modern markets has achieved without serial acquisitions. The report describes capital expenditures as continuously expanding, and the negative free cash flow confirms the description. What remains unspecified is the object of the expenditure. Is the capital going into ground-based data centers competing with CoreWeave? Into satellite-mounted compute in low Earth orbit? Into new launch capacity to place such infrastructure on orbit? The report does not say. It cannot say, because no public technical disclosures exist to support a detailed analysis. This is the central tension of the entire document. The financial model is precise to the decimal. The physical system it models is undefined. A balance sheet without an engineering counterpart is a statement of faith, not a statement of accounts. I have spent 29 years reading ledgers and contracts. That gap between numerical confidence and physical opacity is where value goes to die. The two named customers are instructive. Anthropic is a frontier AI laboratory with a known demand for compute and a reported appetite for unconventional supply arrangements. Google is a vertically integrated cloud provider with its own TPU silicon. Google buying rather than building is either an admission of capacity constraints or a signal about where the market is headed. Both readings are bullish for SpaceX. Counterparty concentration is the first variable. Two tenants underwrite 52% of the projected revenue base. I mapped 47 wallets clustered around insider access to early OpenSea drops in 2021 and identified a $12 million edge concentrated in a group of related addresses. The lesson from that investigation: when a small cluster controls a large share of a system's flow, the system is not diversified; it is levered to the cluster. SpaceX's AI revenue model is levered to the continued willingness of two companies to honor or expand their commitments. Neither commitment has been publicly disclosed with the terms, volume, or pricing necessary to verify the projections. In contract auditing, an undisclosed term is not neutral. It is a deviation from the norm of full disclosure. The growth formula is the second variable. The implied trajectory from $46.9 billion to $184.8 billion in two years requires the AI infrastructure market to incorporate SpaceX as a primary beneficiary in a way no new entrant has ever achieved. My Terra/Luna modeling in early 2022 showed the algorithmic stablecoin's sustainability depended on supply growth at a constant rate. When the growth rate inverted, the system de-pegged within 72 hours. The mathematical issue was identical: the model depended on new capital entering at an increasing absolute rate to sustain promised returns. A 99% CAGR at a $46.9 billion base requires an absolute increment of nearly $138 billion between 2026 and 2028. That increment must come from somewhere in a market that currently has a limited number of counterparties willing to sign long-term compute contracts. The funding gap is the third variable. Negative free cash flow is not objectionable during a construction phase. What is objectionable is the report's silence on the funding sources. Three options exist. Equity issuance would dilute the $235 target price calculation, and the report does not state the assumed share count. Debt financing would add leverage to a business with already massive operational variance. Customer prepayments would require contract structures that are not described. I have audited DeFi protocols where the pool's reserves were the only real asset, and the tokenomics described a fantasy of future value. The same logic applies here. A company bleeding $126.4 billion cannot remain in equilibrium. Either investor capital arrives, or the construction timeline compresses, or the ledger reveals a different outcome. The technical architecture is the fourth variable — and it is unstated. Ground-based compute faces power constraints, land competition, and interconnection costs. Orbital compute faces fundamental physics: free solar irradiance and ambient cooling in space are attractive, but the downlink bandwidth for training clusters is a bottleneck, and satellite-to-satellite laser routing adds milliseconds to every inference. The report provides no chip vendor, no cluster size, no power procurement agreements. In my four-month reverse engineering of EtherDelta in 2018, I documented 14 logical flaws in the order matching engine by reading code, not by reading the founding narrative. In my Curve StableSwap audit, I found arithmetic precision errors that could drain liquidity under volatility. Here, there is no code to read. There is only a revenue curve. The valuation structure is the fifth variable. An 87.5% premium over the reference price, at a time when the market has already priced some AI optionality into SpaceX equity, implies the sell-side views this as the beginning of a repricing, not the end. But the reference price itself is opaque. SpaceX is not publicly listed in a conventional sense; the $125.33 figure likely reflects private market transactions or a related exchange. The imprecision matters. A target price derived from an illiquid reference base has less informational content than one derived from a deep liquid market. I flagged a similar centralization issue in the 2024 ETF custody review: when the security depends on a narrow operational bottleneck, the valuation must discount for the bottleneck's failure probability. Here, the bottleneck is the capital structure. The bulls deserve a fair hearing. Vertical integration is a genuine moat. SpaceX controls launch, satellite manufacturing, and Starlink's global distribution. The reusable launch architecture has inverted the cost curve of access to orbit, and that advantage compounds. If the AI infrastructure deploys compute where terrestrial providers cannot — in orbital platforms or remote power-rich sites — the supply curve is differentiated, not merely competitive. This is the most credible version of the thesis: not another cloud provider, but a new asset class in compute. Google's procurement decision is the strongest evidence. A company with proprietary TPU silicon and substantial hyperscale capacity chose to buy external AI infrastructure. That is either a capacity constraint or a strategic hedge. Both imply the supply-demand imbalance is real enough to make even vertically integrated players seek external compute. And the contract-first model has historical precedent: infrastructure's most successful builds used anchor tenants to underwrite construction. If the reported contracts with Anthropic and Google carry take-or-pay terms, the $24.5 billion revenue forecast is not a hopeful guess; it is an accounting fact. Do not buy the target price. Buy the evidence trail. The ledger will settle this — it will show which customers paid, at what unit price, and with what margin. It will show whether the cash bleed was an investment cycle or a permanent state. The variables to watch: a third named customer, signaling distribution beyond the initial two. Disclosure of contract terms, signaling confidence in unit economics. A capital raise, signaling how the funding gap will be closed. A chip procurement agreement, signaling the technical architecture. When two of those four appear, the thesis becomes testable. Until then, this is a theorem in need of proof. I have modeled algorithmic collapses and traced insider clusters. The pattern repeats. The systems fail where the assumptions are cleanest, and the assumptions are cleanest where the disclosures are thinnest. The final calculation has not yet been entered into the ledger.