Inference is splitting in two — Nvidia's $20B Groq bet explains its next act

Nvidia struck a roughly $20 billion strategic licensing deal with Groq to integrate Groq’s high-speed inference IP (SRAM-backed LPU) into its inference roadmap, a move framed as both defensive and strategic to protect CUDA dominance as inference workloads fragment. The article cites Deloitte data showing inference surpassed training in data-center revenue in late 2025 and notes Nvidia’s announced Vera Rubin family (Rubin CPX for prefill using 128GB GDDR7) to separate prefill (compute-bound) from decode (memory-bandwidth-bound) workloads. The deal, coming amid Anthropic’s multi-accelerator portability and Meta’s acquisition of Manus (stateful agents/KV cache), signals a market shift to disaggregated inference architectures that will force enterprises to route workloads by latency, context size, and model scale.

Analysis

Market structure: Nvidia (NVDA) materially strengthens a multi-tier inference stack by licensing Groq IP for low-latency decode while shipping Vera Rubin CPX for prefill; this preserves CUDA’s centrality and fragments demand into SRAM-optimized edge (8B models) and GDDR7/HBM tiers for large-context training. Winners: NVDA (defensive moat), Google (GOOGL) for TPU scale and Anthropic portability, edge/robotics vendors; losers: incumbents dependent on HBM economics and monoculture GPU stacks (INTC exposure via legacy server silicon suppliers).

Risk assessment: Tail risks include antitrust/royalty scrutiny on the $20bn license, Groq integration failure, or a faster-than-expected Pascal-like TPU/software substitution — each could erase multi-quarter guidance; supply risks center on SRAM capacity (manufacturing bulks) and GDDR7 ramp. Timelines: immediate price reaction (days–weeks), product-led revenue mix shift (quarters), structural market fragmentation and revenue reallocation to specialists (2–4 years).

Trade implications: Primary trade is convex exposure to NVDA’s inference monetization — favor 12–24 month LEAPs or concentrated equity with a 3–4% portfolio weight; hedge with modest GOOGL (2%) exposure to TPU/cloud arbitrage and a tactical short/underweight in INTC (1–1.5%) where HBM/legacy CPU demand can compress. Use options to express view: buy NVDA Jan 2027 calls (LEAPs) and sell shorter-dated calls into strength to fund cost.

Contrarian angles: Consensus understates software portability (Anthropic) and overestimates SRAM’s market breadth — much of enterprise spend will slow until orchestration (routing tokens across tiers) is standardized, extending monetization timelines by 6–12 months. Also NVDA’s defensive $20bn outlay risks compressing gross margins if it must subsidize partners; watch adoption metrics (KV cache hit rates, Vera Rubin shipments) before adding exposure beyond the initial allocation.

Key Decisions for Investors

• Establish a 3–4% portfolio position in NVDA via 12–24 month LEAP calls (e.g., Jan 2027 expiries) or equivalent delta-equity exposure; target a 30–40% upside capture horizon and use a 15% trailing stop on the equity or roll/downsize if NVDA rises >50% in 3 months.
• Implement a pair trade: long NVDA 3% vs short INTC 1.5% (cash short or buy ITM puts) to express hardware/architecture dispersion; cover INTC if it outperforms NVDA by +20% or if NVDA underperforms sector by >15% over 60 days.
• Allocate 2% to GOOGL equity to play TPU/cloud scale and Anthropic multi-platform demand; add if Google announces >100k TPU commitments or if Anthropic/Google partnership discloses >6MW incremental capacity within 90 days.
• Tactical options: sell 8–12 week covered calls on NVDA into rallies to monetize implied vol and use proceeds to buy further-dated calls; size such that total options premium funds no more than 0.5% of portfolio and set a 20% cap on annualized short-call allocation.