# Can ‘team rainbow’ loosen CUDA’s grip on AI? > Source: > Published: 2026-09-18 11:19:50+00:00 LONDON – Escaping Nvidia’s AI hardware ecosystem may not require developers to escape its CUDA parallel computing platform at all. Instead, CUDA could become a common programming model, with compilers translating its code for different accelerator architectures. Michael Søndergaard, CEO of Spectral Compute, advanced that proposition during a fireside chat at this week’s Xcelerated Compute Show. CUDA became the default foundation for accelerated computing partly because Nvidia has continued to provide stability and strong out-of-the-box performance across successive hardware generations, Søndergaard said. Its lead is reinforced by the developers who learned CUDA in academia, the resulting talent pool, extensive documentation, and Nvidia’s CUDA-X libraries. The effect is self-perpetuating. Søndergaard, who was born and raised in Denmark, compared CUDA with learning English because it provides the most effective way to work internationally. “We’re no sooner getting rid of the CUDA ecosystem than we are convincing the entirety of the world to standardize around the U.K. power socket,” he said. “It’s a superior power socket, we can all agree, but it requires ripping out physical infrastructure to make people switch. That’s not happening.” ## Port once, maintain forever The lock-in becomes apparent when organizations try to evaluate alternative hardware. Søndergaard said Spectral Compute initially developed real-time computer-vision algorithms in CUDA because it offered the shortest route to the required performance. When customers asked whether the software could run on their AMD clusters, the company discovered that it required a substantial rewrite. The work did not end once the software had been ported. Every subsequent feature then needed separate implementations for each vendor. “It’s an ongoing, never-ending maintenance burden that scales more or less exponentially with the number of hardware vendors you need to support,” Søndergaard said. That creates a circular problem for prospective buyers. An organization may need to rewrite its software before it can run a proof of concept on another accelerator, but it cannot justify the rewrite until it knows whether the accelerator will deliver the required business outcome. Faced with that uncertainty, buyers gravitate toward the familiar option. Spectral Compute’s proposed answer is to retain CUDA as the developer-facing programming model while using a compiler to target other hardware. Søndergaard compared the approach with central processing unit (CPU) vendors introducing new instruction sets without requiring customers to rewrite their applications. That would give developers baseline performance figures on new accelerators before they invest in hardware-specific optimizations. “Instead of having a conversation about whether it is possible, we’re now having conversations about what, in addition, we can do and how this extends our capabilities and roadmap,” he said. ## Join ‘team rainbow’ The approach could become particularly valuable when a neocloud operator needs additional capacity but can only obtain accelerators from a different supplier. Søndergaard argued that operators should resist treating hardware selection as a choice between Nvidia’s “team green,” AMD’s “team red,” or Intel’s “team blue.” “Let’s go team rainbow and actually combine them,” he said. Individual workloads could still be optimized for particular chips. Nvidia already optimizes CUDA across different generations of its own hardware, Søndergaard noted, and the same principle could be applied across vendors if someone is sufficiently motivated to establish a de facto standard. Portability could also give infrastructure buyers greater negotiating leverage and allow them to assess hardware according to cost, capacity, and business outcomes rather than the software ecosystem they have already adopted. Specialized accelerators will continue to have a place, Søndergaard predicted, but buyers are unlikely to stake the next five years on a chip designed too narrowly for workloads or models that may soon change. CUDA’s dominance may therefore prove to be part of the solution as well as the lock-in problem. If the programming model can travel beyond Nvidia hardware, the safest software choice may no longer dictate the underlying accelerator.