Nvidia preparing Groq chips that can be sold in Chinese market, sources say

Nvidia is preparing a China-compatible variant of Groq inference chips following a $17 billion licensing deal with Groq and has restarted production of H200 chips after obtaining U.S. export licenses and Chinese purchase orders. The company plans to pair Groq chips (available around May) with Vera Rubin chips (which cannot be sold in China) for inference workloads, positioning Nvidia to address China’s sizable inference market amid local competition from firms like Baidu.

Analysis

Western AI hardware suppliers are likely to pursue modular “split‑stack” product strategies that separate training from inference so they can monetize IP even when full-stack exports are constrained. Because inference represents the bulk of deployed model cycles in production applications, modularized, exportable inference solutions can scale revenue faster than one‑off training system deals and create a recurring upgrade cadence for datacenter customers over 3–12 months.

That dynamic raises a two‑front competitive pressure: incumbents that sell cloud services or in‑house silicon will see margin erosion on inference-heavy workloads, while ecosystem players that provide software stacks, adapters, and server integration gain leverage. Second‑order winners include middleware and OEM partners who can certify and bundle these modular inference elements quickly; losers are specialist local inference chips that cannot match the combined software certification and global supply resilience of licensed designs.

Main tail risks are policy volatility and rapid local substitution. An abrupt tightening of controls or a large Chinese push to import substitute designs would reprice prospects within weeks; conversely, strong benchmark data and meaningful purchase orders would drive re‑rating in 1–3 quarters. Watch for (1) third‑party benchmark releases, (2) large OEM certification announcements, and (3) any public reversal in export policy — each is a clear catalyst that materially changes expected revenue trajectories.

Key Decisions for Investors

• Long NVDA via defined‑risk options (buy 6‑9 month call spread, size 2–4% of NAV). Thesis: modularized, exportable inference component monetization lifts consensus margins; reward asymmetry ~2–3x if market re‑rates on licensing wins. Hedge: buy 9‑month 25–30% OTM puts (smaller notional) to protect against sudden policy reversal.
• Relative trade — long NVDA / short BIDU equal‑dollar 6–12 month position (size 1–2% NAV each). Rationale: hardware licensing and appliance economics capture more gross margin than cloud software monetization in the near term; target 20–30% relative outperformance. Risk: faster monetization of models by cloud providers or regulatory/market access shocks to hardware vendors.
• Event‑driven options: buy a calendar straddle around the next major earnings or product availability window (1–3 months before event). Position small (1% NAV) to capture asymmetric volatility if benchmarks or Nav/OEM order flow surprises materially in either direction.
• Tactical: accumulate exposure to datacenter OEMs and middleware integrators off weakness (pullbacks of 10–15%) as they are second‑order beneficiaries of modular inference deployments. Keep position sizes modest and re‑rate targets tied to concrete certification/partner announcements.