This Israeli Startup Just Raised $300M From Nvidia to Break the GPU Lock-In — Meet Decart

Nvidia Just Funded Its Own Disruption

Here’s a story that should make you question everything you know about the AI chip wars: Nvidia just invested in a startup whose core product helps AI companies reduce their dependence on Nvidia GPUs. Decart, a 25-person Israeli AI infrastructure company, raised $300 million at a $4 billion valuation — with Nvidia as a participating investor. The round was led by Radical Ventures.

That’s right. The company that controls 92% of the AI chip market just wrote a check to a startup building the technology that could erode that market share. Either Nvidia sees something we don’t, or this is the most counterintuitive investment in tech history.

What Decart Actually Does

Decart builds what it calls an “AI optimization layer” — middleware that sits between AI workloads and the underlying hardware, allowing companies to seamlessly move their AI training and inference jobs between different chip vendors. Think of it as a universal translator for GPU compute.

Currently, if you train a model on Nvidia’s A100 GPUs, moving that workload to Google TPUs or Amazon’s Trainium chips requires significant re-engineering. The code, the optimization tricks, the memory management — everything is vendor-specific. This lock-in is a massive part of Nvidia’s moat. Once you’re on Nvidia, switching costs make it prohibitively expensive to leave.

Decart’s Optimization Stack (DOS) abstracts away hardware-specific optimizations, allowing AI developers to move workloads across chips from Nvidia, Amazon, Google, and others without rewriting code. It’s the anti-lock-in layer that every cloud provider and AI lab has been dreaming about.

The $300M Funding Round

The $300 million Series B was led by Radical Ventures, with Nvidia participating as a new investor. This brings Decart’s total funding to $450 million. The $4 billion valuation for a 25-person company translates to $160 million per employee — one of the highest per-capita valuations in startup history.

The investor list reads like an AI dream team: Radical Ventures (leading AI-focused fund), Nvidia (despite the apparent conflict), Zeev Ventures, Sequoia, Benchmark, Atreides Management, former Disney CEO Michael Eisner, OpenAI co-founder Andrej Karpathy, and the Yamauchi family (Nintendo’s founding family).

Karpathy’s investment is particularly notable — he just joined Anthropic for pretraining research, and Decart’s optimization technology directly addresses the kind of hardware efficiency challenges he’ll face in that role.

Why Nvidia Would Fund a Company That Reduces GPU Dependence

This is the $300 million question. Why would Nvidia invest in something that could reduce its customers’ switching costs — and thus its competitive moat?

The most logical explanation: Nvidia is confident enough in its hardware performance that it doesn’t fear competition, and by investing in Decart, it gains visibility into which alternative chips are being used and how they compare. It’s a “keep your enemies closer” strategy.

There’s also a market expansion angle. If Decart makes it easier to deploy AI workloads across multiple chip types, it could expand the total AI compute market faster. A bigger market with slightly lower market share could still mean more absolute revenue for Nvidia. And with a $91 billion quarterly forecast, Nvidia isn’t exactly feeling threatened.

Finally, Decart’s optimization technology could actually benefit Nvidia customers by making their existing GPU clusters more efficient — extracting more performance from hardware they’ve already purchased. That makes Nvidia’s products more valuable, not less.

The GPU Lock-In Problem

The AI industry’s GPU lock-in problem is real and expensive. Companies spending billions on compute infrastructure are effectively making one-way bets on specific hardware vendors. If a better chip emerges (Google TPU v6, Amazon Trainium 3, Intel Gaudi 4), switching means:

Rewriting training code to use different APIs and libraries. Re-optimizing memory management for different architectures. Retraining staff on new development tools. Potentially rerunning expensive training jobs to validate performance. Losing months of optimization work specific to the original hardware.

For hyperscalers spending $50-100 billion annually on AI infrastructure, even a 10% efficiency gain from hardware flexibility is worth billions. Decart’s value proposition is compelling precisely because the lock-in costs are so enormous.

Decart’s Three Product Lines

Decart plans to launch new versions of three core products with this funding:

DOS (Decart Optimization Stack): The flagship AI optimization layer that enables cross-hardware portability. This is the product generating current revenue and the primary reason for the $4B valuation.

Lucy: A world model for real-time interactive video — think AI that can generate and modify video streams in real-time. Applications include game streaming, virtual production, and interactive entertainment.

Oasis: A world model designed for physical AI and robotics simulations. This enables training robots and autonomous systems in simulated environments that accurately model real-world physics.

The world model products (Lucy and Oasis) are earlier-stage but represent massive market opportunities. Real-time video generation and physics simulation are foundational technologies for the next generation of AI applications.

Profitable From Day One

Unlike most AI startups burning through cash, Decart claims to have been profitable almost from day one — generating millions in revenue through agreements to use its GPU optimization technology. This is remarkable for a company that’s raised $450 million; most startups at this stage are years away from profitability.

The profitability signal explains the sky-high valuation: Decart isn’t a research lab seeking product-market fit. It’s a company with paying customers, proven technology, and clear demand. The $300 million isn’t funding research — it’s funding scale-up of a product that already works.

In a market where AI companies burn billions before seeing revenue, Decart’s capital efficiency is a competitive advantage. It means the company isn’t dependent on continued fundraising to survive, giving it strategic flexibility that cash-burning competitors lack.

The World Model Play

Decart’s world model products (Lucy and Oasis) position it at the frontier of a technology category that could be as transformative as large language models. World models don’t just process text or images — they simulate entire environments, understanding physics, spatial relationships, and temporal dynamics.

The applications are vast: autonomous driving simulation, robotic training, game generation, architectural visualization, scientific simulation, and more. If Decart can build world models that run efficiently across multiple hardware platforms (leveraging their DOS optimization layer), they’d have a unique competitive position that combines infrastructure efficiency with cutting-edge AI capabilities.

This dual nature — infrastructure company plus frontier AI lab — is unusual and potentially very powerful. The optimization layer generates revenue today while the world models position Decart for tomorrow’s opportunities.

The Investor List Tells a Story

The combination of investors reveals how the AI industry views Decart. Nvidia sees it as an ecosystem play. Sequoia and Benchmark see it as a generational infrastructure company. Karpathy sees it as solving a real technical problem he’s encountered firsthand. The Yamauchi family sees the gaming and entertainment applications.

When investors with such different perspectives and priorities all converge on the same company, it usually means the technology is genuinely foundational rather than narrowly applicable. Decart isn’t just solving one problem — it’s building infrastructure that multiple industries need.

Conclusion

Decart is the kind of company that could reshape the AI hardware landscape from underneath. While everyone watches the chip war between Nvidia, Google, Amazon, and Intel, Decart is building the abstraction layer that makes the chip war irrelevant to AI developers. Hardware agnosticism, delivered at scale, could be worth far more than $4 billion.

And the fact that Nvidia funded it tells you everything you need to know: even the king of AI chips acknowledges that the lock-in era can’t last forever. The question is whether Nvidia invested in Decart to embrace that future or to control it. Either way, AI developers win.