The Download: AI’s trillion-dollar gamble and OpenAI’s biology data bid

Hyperscalers are projected to spend nearly $1.1 trillion on AI data centers through 2027, and analysis by University of Pennsylvania professor Jessica Wachter suggests AI firms will need extraordinary productivity gains merely to break even by 2030. The article highlights mounting uncertainty over whether AI monetization can justify the scale of infrastructure investment, alongside intensifying regulatory, antitrust, cybersecurity, and AI-safety debates. OpenAI Foundation is also funding the creation of high-quality biological datasets, including potentially valuable data from failed biotech companies, to advance medical AI.

Analysis

The relevant AI equity risk is shifting from chip availability to return-on-invested-capital durability. NVDA can sustain earnings momentum while hyperscalers are capacity-constrained, but its multiple becomes vulnerable once incremental GPU orders are judged against utilization, inference pricing, and enterprise conversion rather than training demand. META is more exposed to this transition than AAPL: its spend can be rationalized through engagement and ad-ranking lift, but investors will demand measurable advertising yield or lower cost-per-impression within the next two reporting cycles.

Over the next 1-3 months, capex commentary and depreciation schedules matter more than model announcements. A widening gap between capex growth and disclosed AI monetization would compress platform multiples first, while leaving the semiconductor supply chain relatively insulated until 2027 purchase commitments are revised. The second-order beneficiary is AAPL if enterprises and consumers increasingly prefer on-device or tightly integrated AI features that avoid recurring cloud inference costs; this is a distribution and margin-defense advantage, not necessarily a near-term revenue acceleration.

Consensus appears too focused on whether AI produces transformative revenue and insufficiently focused on a more mundane outcome: productivity gains may accrue to customers and advertisers rather than model owners. That outcome supports broad software and digital-ad margins but weakens the case for permanently elevated infrastructure returns. For MRNA, scientific-data initiatives are strategically interesting but not investable absent trial-linked evidence that AI improves target discovery, trial enrollment, or development timelines; the near-term valuation driver remains pipeline execution rather than dataset access.

The bear thesis is falsified if META reports sustained AI-driven ad conversion improvement that exceeds incremental depreciation and operating expense, or if NVDA shows continued order visibility accompanied by rising inference demand rather than merely customer prebuild. Conversely, a capex-guide increase without monetization KPIs, or evidence of lower GPU utilization/pricing, should be treated as an early de-rating signal.

Key Decisions for Investors

• Maintain a tactical long NVDA / short META pair over the next 1-2 earnings cycles only if NVDA retains backlog or supply-constraint visibility while META does not quantify AI-related advertising ROI. Target relative upside of 10-15%; exit if META delivers two consecutive quarters of material conversion or pricing acceleration attributable to AI, or NVDA signals a digestion cycle.
• Do not add directional META exposure ahead of the next capex update without a disclosed monetization metric. Use any post-results rally driven solely by model capability claims, rather than advertising yield or expense discipline, to initiate a 3-6 month underweight versus GOOG or the communication-services ETF XLC.
• Accumulate AAPL on weakness for a 6-18 month structural position, with the thesis tied to private/on-device AI preserving ecosystem economics as cloud inference costs rise. Falsify on evidence that AI features require substantial recurring cloud subsidy, or that hardware upgrade rates fail to improve after feature deployment.
• Keep MRNA as a watch item rather than an AI-data trade. Reassess only after management provides independently verifiable evidence that computational tools improve clinical timelines, probability of success, or R&D expense efficiency; absent that, pipeline readouts remain the appropriate catalyst framework.