As demand for AI grows, so does the complexity beneath the hood. Many startups are trying to climb above the hardware wars by building software that frees developers from being tied to a single chip ecosystem. One of the most visible is Modular, co-founded by veteran compiler engineer Chris Lattner, which has secured $250 million in funding to build a software abstraction layer that lets AI applications run across different GPU and CPU architectures, tells Wired.com.
Lattner’s vision is rooted in his history; he helped create Apple’s developer tools, led parts of Google’s AI infrastructure, and laid the groundwork in compiler design with LLVM and Swift. He argues that a fragmented hardware stack throttles innovation: developers often pick one vendor and build for it. Modular wants to change that by letting a piece of code compiled for one chip run on another without manual rewrites.
To make that shift possible, Modular has designed a Python-based language and platform that adapts AI workloads to different chip architectures. The startup is already working with major players—chip giants such as Nvidia and AMD, cloud providers, and GPU cluster services—to test interoperability. Its support now spans Nvidia, AMD, and Apple’s silicon.
But this path has challenges. Nvidia’s CUDA ecosystem is deeply entrenched, and porting code to alternative platforms (like AMD’s ROCm) isn’t seamless. The complexity of optimizing kernels, memory mappings, and execution across hardware is enormous. Some competing startups, such as Mako, think the answer lies in building AI agents that auto-generate optimized hardware code rather than reinventing compilers.
Modular sees itself as both collaborator and challenger: today it complements Nvidia and AMD, tomorrow it might disrupt them. The company’s mission is structural, i.e., to solve the software fragmentation that underlies the AI hardware arms race.