Hot French startup ZML releases free product to speed inference across lots of AI chips

French AI startup ZML launched ZML/LLMD, a newly released LLM inference server aimed at running open-source models across chips (including Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc) at peak or sometimes faster performance. The pitch targets reduced vendor lock-in and lower AI inference costs/energy use via multi-chip deployment for enterprises and clouds, potentially pressuring “inference gold rush” incumbents as Nvidia ramps inference demand. ZML raised $20M and is launching the product for free while learning usage, with competition including Baseten ($13B valuation), Inferact (vLLM creators), and RadixArk (SGLang).

Analysis

This is less a near-term revenue threat to NVDA than an attempt to move the bargaining power in AI inference from the silicon vendor to the software layer. If broadly adopted, the economic winner is whichever platform can re-route workloads to the cheapest acceptable compute, which compresses CUDA-style lock-in and improves procurement leverage for hyperscalers and large enterprises. The first-order market reaction should be modest; the more important signal is whether production usage grows enough to matter for enterprise architecture decisions.

Relative winners are the alternative accelerators and integrated platform owners: AMD gets a larger addressable market if inference stacks normalize heterogeneous deployment, GOOGL can use TPU more aggressively if software portability lowers adoption friction, and AAPL could benefit at the margin if on-device inference becomes easier to optimize on Metal. The second-order loser is not just NVDA, but the economic rent embedded in single-vendor tooling; if workloads become portable, pricing power migrates toward buyers and cloud operators, which can show up later as mix shifts rather than outright unit loss.

The contrarian view is that inference performance is usually constrained by kernel tuning, memory bandwidth, reliability, and ops maturity more than by abstract portability. A free, 20-person tool from a startup is enough to influence sentiment, not enough to prove procurement conversion. The thesis is falsified if NVDA keeps taking share in hyperscaler capex, if Blackwell ramps with no margin compression, or if AMD/TPU design-win chatter does not translate into booked revenue over the next 1-3 quarters.

Key Decisions for Investors

• Stay flat/avoid chasing NVDA weakness on this headline; treat it as a sentiment hedge only. Reassess if NVDA data-center growth or gross margin shows any 1-3 quarter slowdown.
• Relative value: long AMD / short NVDA on rallies, sized modestly, targeting 3-6 months. Thesis is that heterogeneous inference tooling improves AMD’s wallet share faster than it dents NVDA’s absolute demand; stop if AMD design-win conversion disappoints or NVDA reaccelerates.
• Construct a small long GOOGL vs NVDA pair if you want exposure to heterogeneous inference adoption. GOOGL benefits if TPU becomes easier to deploy across mixed-model stacks; thesis breaks if Google Cloud AI monetization does not improve over the next 1-2 quarters.
• Watch AAPL as an optionality name, not a primary trade: if Metal-based inference tooling starts showing real developer adoption, it supports on-device AI differentiation over 6-18 months. No action until there is evidence in WWDC/product telemetry.
• Set an alert on enterprise inference benchmarks and cloud procurement language over the next 1-3 months; if ‘vendor-neutral’ becomes a repeated buying criterion, the first durable multiple compression should hit NVDA before it hits the broader AI basket.