What Expires

Across June, the market delivered a range of verdicts on AI infrastructure spending that cannot be reconciled by spending scale alone. Nvidia reached $5.1 trillion ^[1]. Amazon recovered from a 10% dip to record highs ^[2]. Meta fell 6% ^[3]. SpaceX erased roughly $1 trillion from its June 16 peak ^[4]. And KKR launched a $10 billion AI data-center platform designed to sit outside the public-market storm entirely ^[5]. What separated reward from punishment was whether the revenue that spending generates has an expiration date, and whether the capital behind it sits in public markets where every reassessment is violent and visible. The largest combined capital raise in history, SpaceX's $85.7 billion IPO and $25 billion in bonds, produced roughly $1 trillion in value destruction from peak ^[6][4]. The stock peaked near $3 trillion, then fell to approximately $2 trillion. This is not the market doubting that AI infrastructure is real. SpaceX has $26 billion a year in contracted compute revenue: Google pays $920 million a month for 110,000 Nvidia GPUs, Anthropic pays $1.25 billion a month for the Colossus 1 data center in Memphis ^[7]. The revenue is booked, delivered, and generating cash. The market is questioning whether it lasts. Google's own language says it does not. The deal is, in Google's words, "short-term" "bridge capacity" to meet surging demand for Gemini Enterprise while Google builds its own Tensor Processing Units ^[8]. The contract runs October 2026 through June 2029 and includes a termination clause if SpaceX fails to deliver the full GPU commitment by September. SpaceX's filing confirms the temporary framing, stating the deals "monetize unused compute capacity" with reallocation rights retained ^[7]. The capacity SpaceX is selling is, by its own description, spare capacity it can take back. That is not a cloud business. It is a rental with an exit. The same business earns a different verdict when the contract is durable. Hut 8 signed a $9.8 billion, 15-year AI data-center lease on a "take-or-pay, triple-net basis with no termination for convenience" with a high-investment-grade tenant. Hut 8 shares rose after the announcement ^[9]. Same business: renting data-center capacity for AI workloads. Opposite market verdict. The difference is the contract. Fifteen years with no exit clause versus bridge capacity with a termination clause. The market reads contract terms and prices revenue durability accordingly. That is the first axis of the discrimination. The second is capital structure. On June 11, KKR launched Helix Digital Infrastructure with $10 billion-plus in committed capital, structured as a "permanent operating company funded by private institutional capital" and explicitly not dependent on public debt and equity markets ^[5]. Led by former AWS CEO Adam Selipsky, with founding investors including Kuwait Investment Authority, Nvidia, and Vistra, Helix is building AI data-center infrastructure, the same business SpaceX's compute operation is in, but funded privately and structured as permanent, insulated from the public-market repricing that destroyed SpaceX's equity. SpaceX is the only company in this set exposed on both axes: its compute revenue is explicitly temporary, and its capital sits in public equity and debt where every reassessment of that temporariness translates into a price swing. The hyperscalers calibrate the scale of punishment. Amazon announced $200 billion in 2026 capex, initially dropped 10%, then recovered to record highs after AWS revenue grew 28% year-over-year ^[2]. Alphabet raised 2026 capex guidance to $180-190 billion; its stock rose 10%+ after results showed a Google Cloud backlog exceeding $460 billion ^[10][11]. Meta set 2026 capex at $125-145 billion and fell 6% ^[3], moderate punishment for a company with $55 billion in Q1 ad revenue growing 33% year-over-year. Meta's CFO noted the company could "choose to bring it online more slowly or reduce our spending in future years" ^[3]. The market punished Meta 6% for a vague scaling plan. It punished SpaceX 30%+ for revenue with an explicit expiration date. Nvidia sits at the far end of the durability spectrum: $5.1 trillion in market cap, $81.6 billion in Q1 FY2027 revenue, 85% year-over-year growth, with Jensen Huang projecting $1 trillion-plus in visibility for Blackwell and Rubin through end of 2027 ^[1]. The inference demand driving Nvidia is permanent: models must run indefinitely. The bridge-capacity demand driving SpaceX's compute revenue is temporary for a concrete reason. Hyperscalers are pouring $650-725 billion into their own build-outs in 2026, with Nvidia projecting $1 trillion-plus in data-center capex by 2027 ^[1][12]. Alphabet's CFO said spending would rise "significantly further" in 2027 ^[11]. That self-build will close the 12 GW capacity deficit, a gap Jefferies attributes not to demand but to supply-chain constraints on engineering labor, cooling equipment, and power infrastructure ^[13]. When hyperscalers finish building, the bridge-capacity demand funding SpaceX's $26 billion a year narrows toward zero. SpaceX's path from temporary to permanent compute exists on paper. One million AI1 satellites with Nvidia GB300 chips, solar-powered, linked by laser into a decentralized orbital grid, with first compute demonstrations targeted for late 2027 ^[14]. That vision requires a 60x increase in launch cadence, from 170 launches in 2025 to 10,000 annually. The FAA has told SpaceX it needs to see "a lot more reliability" ^[15]. The Starship Super Heavy booster failed its controlled splashdown on May 28, prompting an FAA mishap investigation before any return to flight ^[16]. Without reusable Super Heavy, launch costs stay too high for orbital data centers to be economical. SpaceX's permanent compute revenue is optionality on a moonshot, and the market is pricing it as such. Post-IPO, analysts are comparing SpaceX's compute valuation to established cloud providers and finding it overpriced ^[17]. The comparison is not SpaceX-versus-aerospace. No aerospace rival has a compute play. It is SpaceX-versus-hyperscalers, and on that axis, temporary bridge-capacity revenue trades at a discount to permanent cloud revenue. Morningstar valued SpaceX at approximately $780 billion against the company's $1.75-1.8 trillion target, a 57% discount ^[7]. The market has learned to ask a question that did not exist when AI capex was simply a lot of money going into servers: does the revenue this spending generates expire? For the companies whose answer is no, capital is abundant. For the one whose answer is yes, and whose capital sits in public markets where that answer is repriced daily, capital became the most expensive thing in the sector.