Grok Bot is growing fast—but Elon Musk says its costs are ‘unsustainably high’ compared to revenue

Grok Bot reached 800,000 paid weekly users, double September’s figure, but Elon Musk said its serving costs were “unsustainably high relative to revenue” as users increase their usage. SpaceX’s Super Intelligence division reported Q3 revenue of $7.56 billion, up from $813 million a year earlier, and $431 million in operating profit versus a $1.7 billion loss; however, SpaceX spent $16.4 billion on AI capex, 84% of total Q3 capex. Musk said SpaceX aims to split compute capacity evenly between its own products and leasing to other companies, though Grok demand could change that allocation.

Analysis

The key economic tension is not user growth versus no growth; it is whether agent usage can be monetized faster than it consumes scarce compute. If Grok Bot takes capacity that could be rented to hyperscalers at rising dollars per watt, each additional task has an opportunity cost even before accounting for infrastructure investment. A 50-50 allocation target is therefore not a stable operating plan: strong agent demand could displace contracted external revenue, while prioritizing rentals could constrain product quality and adoption. Reported division operating profit does not resolve this question; verify product-level inference cost, realized revenue per user, utilization, and the economics and duration of external leases.

Near term, this is a sentiment risk for SPCX, not enough on its own to establish that AI demand or the capex thesis is broken. Over 1–3 months, watch for pricing or usage limits, revised compute allocation, and evidence that lease renewals remain attractive. Over 6–18 months, inference-efficiency gains could restore attractive agent economics; failure to improve would make usage growth a capacity and cash-investment burden. META faces the same inference-cost test, but any efficiency target should be treated as an execution claim until demonstrated. ANT and GOOG may benefit from access to leased capacity, though dependence on a supplier allocating more compute to its own products is a counter-risk. The contrarian point: high demand is not automatically bullish for the compute owner if its best use of capacity is shifting from paying tenants to under-monetized internal agents.

Key Decisions for Investors

• SPCX: avoid chasing growth headlines; treat as a watch/hedge candidate pending disclosure of agent monetization and compute allocation. Reassess positively if revenue per active user rises or inference cost per task falls while external lease renewals remain firm; thesis weakens if usage limits/pricing changes appear alongside reduced lease capacity or weaker economics.
• META: monitor, rather than trade the stated efficiency target in isolation. A catalyst would be evidence of lower inference cost per unit of usage without deterioration in product engagement; failure to show progress over coming quarters would reinforce the margin-risk interpretation.
• ANT and GOOG: monitor lease renewal terms and capacity availability. A shift of SPCX compute toward Grok could create supply or continuity risk for tenants; do not infer a direct earnings impact without contract duration, sourcing alternatives, and pricing data.
• For the next 1–3 months, seek product-level disclosure: paid-user monetization, compute cost per task, capex-to-capacity utilization, and the share of capacity committed to external customers. These are the missing measures needed to distinguish healthy investment from costly usage growth.