France’s ZML wants to break Nvidia lock-in with free cross-chip AI software

Paris startup ZML launched a free software tool aimed at running open-source AI models quickly across multiple hardware platforms (Nvidia, AMD, Google, Apple, and Intel), reducing vendor lock-in. While Nvidia’s AI hardware dominance remains, the move highlights growing ecosystem competition that could gradually pressure its software-level moat. Near-term price impact is likely limited as this is a new tooling release without disclosed revenue or customer traction.

Analysis

This is less a near-term product launch than a bargaining-power shift. If a free abstraction layer makes workloads portable across accelerators, the first economic impact is not lower Nvidia unit demand but higher customer willingness to dual-source, which compresses pricing power at renewal and weakens the moat of proprietary runtimes. The beneficiaries are the buyers of compute — hyperscalers and large model operators — because even modest portability improves procurement leverage and reduces the cost of capacity shortages over the next 1-3 quarters.

The second-order effect is selective: AMD, Intel, and Google silicon gain credibility only if the tool survives real production workloads, while Nvidia’s risk is mostly in inference-heavy deployments where switching costs are lower and utilization is more price-sensitive. Training remains the harder moat to attack, so any revenue displacement is likely gradual; the more immediate risk for NVDA is multiple compression if investors start underwriting a lower terminal share of AI capex rather than an actual revenue step-down.

Contrarian view: the market may be overstating the competitive threat because portability layers often shift complexity rather than remove it, and performance parity is usually lost in the last mile of kernel tuning and memory bandwidth optimization. If that happens, the story becomes a software convenience feature, not a hardware regime change. Falsifiers are simple: if NVDA data center growth and gross margins stay elevated through the next two quarters, or if third-party benchmarks show no sustained production adoption on non-NVDA silicon, this is noise rather than thesis-changing evidence.

Key Decisions for Investors

• No immediate standalone trade; treat this as an alert item until independent benchmarks show real production adoption across non-NVDA silicon over the next 4-8 weeks.
• Conditional relative-value trade: long AMD / short NVDA only if inference workloads on AMD close to within ~10-15% of Nvidia on cost-per-token in independent tests; target 8-12% relative outperformance over 1-3 months, cut if NVDA data center guide re-accelerates.
• Small tactical long GOOGL vs NVDA on the thesis that portable software increases TPU utilization and Google’s bargaining leverage in AI infrastructure; use a 2-3 month horizon and exit if TPU adoption metrics do not improve by next earnings.