Paris Startup Bets on Multi-Chip Inference to Break AI Hardware Lock-In

The Inference Bottleneck Gets Crowded

Steeve Morin's Paris-based AI infrastructure company has shipped software that runs large language models at full speed on chips from five different manufacturers. The new inference server, ZML/LLMD, supports Nvidia GPUs, AMD processors, Google TPUs, Apple Metal, and Intel Arc silicon without forcing enterprises to pick a single vendor ecosystem.

The release arrives as organizations deploying AI face mounting compute bills and fragmented tooling. Inference workloads now account for the majority of AI spending at scale, yet software often ties users to specific hardware platforms. Morin argues this creates unnecessary cost and limits technical choices. His answer is a layer that extracts maximum performance from whatever silicon a company already owns or can access affordably.

At DailyTechWire, we've tracked the inference layer as it transitions from an engineering footnote to a strategic chokepoint. Model training dominated headlines through 2024, but prompt processing now drives more revenue and operational pain. The startups competing here are building the invisible plumbing that will determine which AI applications can run profitably at scale.

Breaking the Silicon Silo

ZML's approach centers on compatibility and speed. The software is designed to deliver peak throughput on each supported architecture, sometimes exceeding what chip vendors' own tools achieve. Morin frames this as returning control to engineering teams who want to mix hardware based on workload, cost, or energy consumption rather than accepting a single vendor's roadmap.

The company has also begun co-designing silicon with chipmakers, a deeper engagement than typical software vendors pursue. Morin points to European chip startups like Axelera, Fractile, Kalray, and VSORA as potential partners where novel architectures might benefit from ZML's abstraction layer. These firms face the cold-start problem of software support; a credible inference stack that treats their hardware as a first-class citizen could accelerate adoption.

Morin maintains a pragmatic stance toward Nvidia, which still commands the majority of AI training and inference deployments. He describes the relationship as collaborative, noting that Nvidia itself has invested heavily in inference optimization. ZML's thesis is not that Nvidia will lose, but that the market will support multiple winners and that software enabling that diversity will capture value.

Lean Team, Dense Capital

Twenty people work at ZML, a headcount Morin credits for the company's velocity. The startup raised twenty million dollars in 2024, led by Harry Stebbings' 20VC, with participation from commit, AALVC, Drysdale Ventures, Kima Ventures, Kindred Capital, LocalGlobe, and Puzzle Ventures. Morin previously served as VP of engineering at Zenly, the social mapping app Snap acquired for a nine-figure sum in 2017, giving him both credibility and a network of technical operators.

The cap table includes founders from adjacent infrastructure companies. Solomon Hykes, who created Docker and Dagger, has backed the company, as have Hugging Face co-founders Clément Delangue and Julien Chaumond. Turing Award laureate Yann LeCun, now with AMI Labs, is also an investor. These names signal that ZML is being watched by the open-source and model-deployment communities, where performance and portability are existential concerns.

The first ZML project, an inference-focused machine learning framework, launched as open source in 2024 and received updates through March 2025. The new LLMD server diverges from that model. It will be free at launch, with no immediate monetization, as the team gathers usage data and learns which workloads and chip combinations matter most to users. Morin describes the strategy as avoiding premature pricing that could throttle adoption before the product proves its value.

The Inference Gold Rush and Its Contenders

ZML enters a market segment that has attracted significant capital and talent. Baseten, which offers managed inference infrastructure, reached a thirteen billion dollar valuation in its most recent funding round. Inferact, built by the team behind the open-source vLLM project, and RadixArk, the commercial entity behind SGLang, both compete in overlapping parts of the stack. vLLM and SGLang are widely deployed inference engines; LLMD will need to demonstrate clear performance or operational advantages to displace them in production environments.

Morin positions ZML's scope as broader than a single-purpose inference server. The co-design work with chip manufacturers and the ambition to support emerging hardware suggest a platform play rather than a point solution. If the company can establish itself as the compatibility layer that makes novel accelerators viable, it gains leverage with both hardware vendors seeking software ecosystems and enterprises hedging against single-vendor risk.

The broader trend Morin is betting on is real. As generative AI moves from prototype to production, inference costs are becoming a line item that finance and procurement teams scrutinize. Enterprises that locked into a single cloud provider or chip architecture during the experimental phase are now looking for optionality. Software that delivers performance portability without requiring re-architecture has a clear buyer.

Paris as an AI Infrastructure Hub

Morin insists ZML could only have been built in Paris, citing the city's combination of technical talent, investor appetite for deep-tech bets, and regulatory environment. Europe's AI startup ecosystem has historically lagged Silicon Valley in both funding velocity and exit outcomes, but infrastructure and tooling companies have found traction by focusing on interoperability, efficiency, and sovereignty concerns that resonate with European enterprises and governments.

The presence of European AI chip startups gives ZML a natural customer and partner base. These companies need software that proves their hardware can run real workloads without requiring users to rewrite applications. If ZML can credibly claim to deliver Nvidia-class performance on alternative silicon, it becomes a key enabler for any chip vendor trying to break into data centers currently dominated by incumbent accelerators.

The risk is execution. Inference optimization is technically demanding, requiring deep knowledge of both model architectures and hardware microarchitectures. Achieving peak performance across five different chip families, each with distinct memory hierarchies and instruction sets, is a moving target as both models and silicon evolve. ZML's small team will need to maintain parity with vendor-optimized stacks while also supporting new entrants.

What Adoption Looks Like

ZML has not disclosed usage numbers or customer names for LLMD, which is launching now. The free tier is explicitly designed to generate data on which models, chips, and workloads users prioritize. That information will inform both product development and eventual pricing, but it also means the company is operating without revenue from this product in the near term.

The path to monetization likely involves tiered pricing based on scale, support, or access to proprietary optimizations. Enterprises running inference at high volume are accustomed to paying for performance gains measured in milliseconds or percentage points of latency reduction. If ZML can document cost savings or throughput improvements against baseline deployments, it has a straightforward value proposition.

The harder question is whether the market will reward a multi-chip abstraction layer or continue to consolidate around a few dominant platforms. If Nvidia and the major cloud providers extend their leads in inference performance and tooling maturity, the case for portability weakens. If hardware diversity increases and cost pressure intensifies, ZML's bet looks prescient.

Morin's timeline is patient. He describes the fundraise as sufficient for the team's size and the free launch as a deliberate choice to prioritize learning over early revenue. That posture reflects confidence that the inference market will remain fragmented and that enterprises will pay for software that reduces lock-in risk. The next twelve months will test whether that confidence is justified by adoption and whether ZML can scale its engineering effort to match the pace of change in both models and silicon.