NVDA Earnings Preview: AI Capex Duration in Focus
Market consensus sees NVIDIA's quarterly earnings with data-center revenue still growing 60-70% YoY, with focus shifting from "will it beat" to "where does 2026 capex top out." Blackwell ramp-up and sovereign AI orders provide a 12-18 month moat, but watch for the medium-term structural risk of inference workloads migrating to ASICs.
- Blackwell platform supply constrained; 2026 H1 remains seller's market
- Sovereign AI orders (Middle East, Europe) chart clear next-wave growth path
- Inference compute market losing ground to ASICs; training moat holds steady
- 33x PE at historical median; beat triggers 8-12% Davis double play
Neutral-to-bullish: valuations aren't cheap, but demand curve hasn't turned yet
Short-term: technically bullish on a post-earnings surge over the next quarter. Medium-term: monitor three key metrics — anchor client capex growth, in-house chip penetration, and sovereign AI order fulfillment timing over 6-12 months.
- Data center Q4 consensus $35-36B; hyperscaler capex guidance upgrade backs demand
- Blackwell production on track; early supply-demand tightness to persist through 2026Q2
- Meta, Google ASIC adoption accelerating as inference share rises, pressuring GPU ASP
- US risks tightening chip export restrictions on China; H20, other near-flagship products' fate uncertain
Rather than keep debating GPU oversupply, the real concern: when large customers' depreciation cycles push H100 generation retirement pressure to H2 2026, existing compute absorption and new orders will create a structural mismatch.
Core Thesis Shift
NVIDIA's market focus has clearly pivoted entering 2026: from "can it keep breaking records" to "where does this AI capex cycle actually peak." This quarter's earnings aren't about data center revenue — consensus already sits at $35-36B, up 60-70% YoY, largely priced in — but rather what management says about next-quarter guidance and full-year 2026 order visibility.
Demand: Three Drivers
1. Hyperscaler capex still ramping
Microsoft / Meta / Google / Amazon combined capex crossed $350B in 2025, with 2026 guidance raised to ~$420B. Sixty percent going to AI infrastructure. Blackwell shipments are the biggest beneficiary of that incremental spend.
2. Sovereign AI orders
Middle East (Saudi HUMAIN, UAE G42) + Europe (France Mistral, Germany Telekom Cloud) have contributed ~$18B in backlog, with deliveries scheduled across 2026-2027.
3. Enterprise penetration
Traditional industries (finance, healthcare) moving from PoC to scaled deployment, though individual order sizes remain smaller than hyperscaler deals.
Supply: Blackwell Ramp
GB200 / GB300 yield improved from 60% at end-2025 to 80%. CoWoS-L packaging capacity at TSMC doubles in H1 2026. Supply tightness eases but doesn't disappear.
Risks
- ASIC alternatives: Meta MTIA v3, Google TPU v6 showing clear TCO advantages on inference workloads; ASIC share of inference expected to rise from 8% to 20% by end-2026
- Export controls: H20 successor products likely facing tighter restrictions; China revenue share may drop further from 17% to below 10%
- Customer concentration: Top 4 clients accounting for ~55% of revenue; any single capex pullback means non-linear impact
Conclusion
Short-term (1Q): 15% upside reasonable on earnings catalyst. Medium-term (6-12M): Need GB300 mass production + sovereign AI orders GAAP recognition for second leg of Davis play. Long-term (24M+): Inference workload migration is the real threat, but training side maintains 18+ month moat.
Positioning: Maintain 8-10% allocation within growth sleeve; add on pullback to support after earnings.