Nvidia and Mira Murati’s Thinking Machines announce gigawatt-scale AI partnership

NVIDIA and Mira Murati’s Thinking Machines Lab agreed a multi-year strategic partnership to deploy at least 1 gigawatt of next-generation Vera Rubin systems and includes a significant investment from NVIDIA. Thinking Machines — founded in early 2025 and previously backed by a $2.0B seed round that valued it at $12.0B — will use the infrastructure to train frontier models and expand its Tinker fine-tuning API; NVDA stock was up ~1.7% in late morning trade. The deal is a material infrastructure commitment that should boost NVIDIA data-center demand and accelerate enterprise access to large-scale AI training.

Analysis

The market is pricing an extended cycle of outsized hardware demand into NVDA’s equity multiple, which creates a flow-through boost for upstream server OEMs, power and cooling suppliers, and memory vendors. Expect orderbook-driven margin expansion for a 12–36 month window, but that window is sensitive to delivery cadence: a single quarter of inventory digestion at hyperscalers can erase forward guidance and compress multiples quickly.

A key tail risk is algorithmic efficiency — a 20–30% improvement in model compute-per-token (via sparsity, quantization, or distillation) materially reduces incremental hardware TAM versus consensus. Equally important are non-market constraints: grid capacity, data-center permitting, and specialized packaging lead times mean supply elasticity is months-to-years, not weeks; these widen the dispersion between winners who have supply contracts and laggards who must buy on the spot market.

Consensus is underweight two convergent risks: (1) commoditization of training stacks as system-level differentiation narrows (puts more revenue into lower-margin services), and (2) policy/export constraints that can bifurcate addressable markets. That argues for concentrated, duration-aware trades: own dominant exposure with defined downside protection, and take the other side on peripheral, sentiment-driven names that have less direct capture of the hardware-to-software capture chain.

Key Decisions for Investors

• Buy NVDA equity sized 3% of portfolio for a 12–24 month horizon; set target +40% and a hard stop at -15%. Rationale: asymmetric payoff from sustained data-center demand, but protect against a single-quarter inventory shock.
• Buy NVDA Jan-2027 LEAP call spread (buy 25-delta, sell 10-delta) sized 1–1.5% of portfolio to lever long-term secular exposure while capping premium outlay; exit if NVDA premium compresses >40% or implied vol drops >6 vol points.
• Go long SMCI (server OEM) for 6–12 months sized 1.5% as a levered play on build-outs; set target +60% and stop -20%. If SMCI reports order softness or margin erosion, trim to zero — this is conditional on persistent hardware backlogs.
• Establish a pair: long NVDA / short APP sized 1% each (dollar-neutral) for 3–9 months to express hardware/software capture divergence. Target pair return +30% if NVDA re-rates while APP languishes; cut pair if sector breadth turns uniformly positive.
• Maintain a 0.5–1% hedge via ATM NVDA puts 3–6 months out to protect against event risk (export controls or large macro drawdown); treat as insurance — accept premium as cost of preserving asymmetric upside.