SpaceX, Tesla to continue ordering Nvidia chips at scale- Musk

Tesla and SpaceX will continue ordering NVIDIA chips at scale, while Tesla is progressing on its in-house AI5 chip optimized for Optimus and robotaxi applications; reports say volume production of AI5 has been pushed to mid-2027. Tesla signed a $16.5 billion deal with Samsung in mid-2025 to build its AI chips, xAI/SpaceX compute (Colossus) primarily uses NVIDIA H and Blackwell chips, and SpaceX is targeting a large IPO and plans for AI data centers in space.

Analysis

The immediate market structure is one of overlapping demand windows: legacy GPU suppliers continue to capture near-term training and inference spend while a subset of large OEMs pursue bespoke, workload-specific silicon that only meaningfully reduces GPU demand if it reaches high-volume production and feature parity. That creates a two-tier opportunity set for foundries and packaging firms — steady high-margin GPU orders now, potential step-function ramp for specialized wafer demand later — which supports revenue visibility for GPU vendors for the next 12–24 months even as long-term substitution risk grows.

Key supply-side constraints (EUV tool cadence, advanced node allocation, and OSAT packaging throughput) act as both a cap and a pricing floor. If macro AI capex softens, these constraints will flip from scarcity premium to margin compression via inventory builds within 3–6 quarters; conversely, an acceleration in bespoke chip tapeouts could push advanced-node lead times beyond 12 months and keep margins elevated for incumbents.

From a competitive-risk perspective, specialized chips optimized for narrow inference tasks can achieve 2–4x efficiency gains per watt versus general-purpose GPUs, but only in constrained workloads — that limits total addressable displacement to a subset of datacenter spend unless software ecosystems migrate rapidly. This asymmetry implies incumbents retain pricing power on general-purpose GPUs while facing gradual erosion on commoditized inference segments over a multi-year horizon.

Consensus underestimates the optionality embedded in dual-sourcing strategies: firms that both buy external GPUs and develop in-house silicon create durable, overlapping demand that fronts revenue for GPU suppliers while buying time for internal designs. The market should price a smoother revenue path for GPU names near-term and a conditional, binary hit to customers that successfully scale in-house silicon beyond their own fleets over 24–48 months.

Key Decisions for Investors

• Long NVDA via 6-month call spread: buy NVDA 6-month 1:1 call spread (long nearer-term ATM call, short higher strike) sizing 2–3% portfolio notional. Rationale: captures continued datacenter demand while limiting premium decay; target 30–50% upside, max loss = premium (~100% of spread cost).
• Pair trade (directional hedge): long NVDA equity (or calls) and short TSLA equity sized to neutralize beta for 6–12 months. Rationale: capture asymmetric upside in GPU exposure vs execution/long-horizon chip risk in EV/robotics OEM; target 2:1 reward-to-risk with stop-loss if pair diverges >25% intra-horizon.
• TSLA downside hedge: buy 12-month 30% OTM puts (small position, 0.5–1% portfolio). Rationale: protects against a delayed/failed in-house chip ramp and continued EV volume weakness; payoff is convex if market de-rates optionality.
• Event-driven: buy NVDA 3–6 month call skew into major earnings/releases and IPO/data-center partnership milestones (size 1–2% notional). Rationale: positive surprise on contracted GPU orders or pricing sustains, implied vols peak but upside catalysts are binary — use vertical spreads to cap premium risk.