SpaceX Built AI Data Centers So Fast, Some Ran Without Backup Power for Months. Now It Faces a $920 Million Deadline.

SpaceX is racing to deliver roughly 110,000 Nvidia GPUs by September 30 under a Google compute contract worth $920 million per month at full capacity, but data-center power, cooling, engineering, and turbine-supply issues have impaired reliability. AI-related capex reached $15.8 billion in Q2 and $23.6 billion in the first half, driving a $3.7 billion AI-division loss despite AI revenue more than tripling to over $2.5 billion and compute capacity rising from 0.4 GW to 1.4 GW year over year. The company has delayed expansion projects, is pursuing in-house turbine-component production, and has replaced data-center executives; failure to improve uptime could lead Google to reduce payments, accept fewer GPUs, or terminate after a one-month grace period.

Analysis

The principal economic exposure is not Alphabet’s but the operator’s: an availability shortfall converts a nominal ~$11 billion annualized customer commitment into immediate revenue leakage while fixed depreciation, power reservations, and GPU financing continue. A 5-10% capacity or uptime shortfall would imply roughly $0.6-$1.1 billion of annualized revenue at risk before considering service credits and the reputational cost to subsequent hyperscale customers. The market should also distinguish installed power from monetizable compute; redundant generation, cooling and network commissioning—not GPU delivery—are now the binding constraints.

GOOG has substantial negotiating leverage because alternative capacity can be sourced, even if with delay, whereas the GPU fleet remains economically valuable to the operator if this customer reduces its take-rate. NVDA faces a timing risk to revenue recognition only at the margin: constrained capacity deployment can defer incremental accelerator orders, but deployed GPUs can be redirected to other buyers in a supply-constrained market. The more meaningful second-order beneficiary is BE and other behind-the-meter power providers, as reliability remediation favors modular, redundant generation over greenfield utility interconnection schedules; however, the equity upside depends on disclosed bookings and margins rather than headline deployment claims.

Consensus may over-penalize the failed-build narrative while underestimating the cost of fixing it. Emergency power, mobile cooling, redundant electrical architecture and accelerated turbine sourcing can preserve revenue, but likely reduce AI-infrastructure EBITDA margins for the next 2-4 quarters and raise capital intensity. A credible operating-data release showing sustained uptime, commissioned redundant capacity, and no customer-volume reduction would reverse the near-term bearish thesis; conversely, any extension, repricing, or reduced minimum commitment would make the revenue risk structural over 6-18 months.

Key Decisions for Investors

• Do not initiate a directional position in SPCX until ticker/entity, contract terms, and independently reported capacity are verified; treat this as an event-driven watch item, not an investable recommendation on the supplied record.
• Long BE only on confirmation of incremental contracted backlog or a named deployment conversion, preferably via a 3-6 month position with a 10-15% downside stop. Risk/reward is favorable only if new AI-related orders are additive rather than merely shifted from existing pipeline; falsifier is backlog stagnation or gross-margin dilution on accelerated projects.
• Maintain GOOG as the relative winner versus the compute provider over the next 1-3 months: its contractual optionality limits direct earnings exposure while its ability to secure capacity is strategically valuable. This is a low-beta relative-value expression, not a standalone catalyst for GOOG; exit if management confirms full contracted capacity is delivered without concessions.
• Avoid short NVDA on this development. Use any broad AI-infrastructure de-rating to add selectively only after checking supplier commentary for order deferrals; the thesis is falsified if multiple hyperscalers cite power/cooling constraints as causing GPU shipment cancellations rather than deployment delays.