‘Almost unlimited’: Execs says AI demand remains strong even as enterprises move to ‘valuemaxxing’

Chip stocks’ volatility is being questioned as a potential sign of AI demand slowing, but multiple AI executives said compute demand remains “extraordinary.” Quotes from Nebius, Cerebras, Rebellions, and Lumentum point to ongoing supply constraints—e.g., Lumentum is sold out for the next five years and CFO-led “rationalization” is shifting spend toward ROI rather than abandoning demand. While Meta and xAI selling/renting excess AI capacity raised overcapacity questions, executives characterized those cases as unique and maintained that hyperscalers are still short on compute and data center inputs.

Analysis

The market is overfitting to the idea that any monetization of spare compute equals demand destruction. More likely, it is a utilization story: as enterprises move from experimentation to ROI screens, spend shifts away from indiscriminate token burn toward workloads with clear payback, which preserves aggregate demand but redistributes economics to the scarce bottlenecks. That favors infrastructure owners with real capacity constraints—NVDA on accelerators, NBIS on delivered cloud capacity, and LITE on connectivity—while leaving legacy CPU-centric vendors like INTC with less pricing power and weaker mix leverage.

The second-order effect is that frontier model pricing may compress faster than compute demand. If open-source or task-specific models handle routine tasks, marginal inference migrates to cheaper stacks, but the total number of deployed workloads can still rise; that keeps the capex machine turning while lowering the narrative premium on "largest model wins." Meta-like capacity leasing is therefore more of a balance-sheet optimization signal than a broad overcapacity tell, unless it starts showing up across multiple hyperscalers and independent neoclouds.

Near term, the key falsifier is not sentiment but utilization: if next quarter’s commentary shows easing lead times, lower booked backlog, or slower data-center power deployment, the scarcity trade will de-rate quickly. Over 6-18 months, energy availability is the binding constraint, so the winners are likely the names that can convert constrained watts into billable inference fastest. The consensus is still underestimating how much of the AI capex pie gets captured by plumbing rather than applications.

Key Decisions for Investors

• Go long NBIS on pullbacks over the next 1-3 weeks; thesis is scarcity-driven capacity monetization with asymmetric upside if utilization stays constrained. Falsify if management signals backlog normalization or lower contracted load in the next update.
• Initiate a pair trade: long NVDA / short INTC for 1-3 months. NVDA remains the direct toll collector on AI intensity, while INTC has limited leverage to the bottlenecked parts of the buildout. Cover the short if INTC shows meaningful AI design-win traction or if NVDA order commentary weakens.
• Use NVDA call spreads instead of outright equity into the next earnings/capex cycle. The risk/reward is best if AI infrastructure spend stays firm but volatility remains high; the trade should be cut if hyperscaler capex guides down or lead times shorten.
• Do not short META on the capacity-leasing headline alone. Treat it as an internal utilization optimization, not a structural demand rollover, unless multiple hyperscalers start monetizing excess capacity over the next 1-2 quarters.