Qualcomm held its investor day on Tuesday and told Wall Street it was a different company. Non-handset revenue target for fiscal 2029: $40 billion, nearly double the previous $22 billion. Data center revenue alone: more than $15 billion, from zero today. The stock surged roughly 15 percent after hours and rose 9 percent during Wednesday’s regular session.
What the Chip Does
The Dragonfly C1000 drew the headlines. More than 250 cores running above 5 gigahertz, built on the Oryon architecture Qualcomm inherited from its $1.4 billion acquisition of chip startup Nuvia in 2021. Meta has committed to deploying the C1000 in production starting the second half of 2028. Microsoft’s Azure will use a separate Qualcomm design, High Bandwidth Compute, starting mid-2027. Qualcomm claims more than twice the performance per watt of existing server CPUs.
What the Money Bought
The same day, Qualcomm confirmed it would acquire Modular, an AI software company, for approximately $3.9 billion in stock. Modular was valued at $1.6 billion nine months earlier after a $250 million raise. The company builds the Mojo programming language and the MAX inference engine. Mojo compiles AI models to GPUs, CPUs, neural processors, and custom accelerators without hardware-specific rewrites. Modular was founded in 2022 by Chris Lattner, who created LLVM, the compiler infrastructure behind most of the world’s compiled software, and Apple’s Swift programming language.
Twenty Years of Switching Costs
Nvidia commands approximately 80 to 90 percent of the AI accelerator market by revenue. The advantage is twenty years of software. CUDA has accumulated more than four million developers. Every major machine learning framework is optimized for it first. Switching costs are measured in engineering-years: kernel fusions, mixed-precision tuning, distributed training paths, CI/CD pipelines all built around Nvidia’s libraries. AMD built ROCm. Intel built oneAPI. Neither achieved meaningful developer adoption because each tried to replicate CUDA’s breadth on a narrower hardware base. Each new alternative added another vendor-specific toolchain without reducing dependence on the one it was trying to replace.
The Inversion
Qualcomm is attempting something structurally different. Instead of building its own twenty-year software stack, it acquired the team trying to make all of them unnecessary. Lattner has executed this pattern before. LLVM became the dominant compiler infrastructure by making itself a modular platform that any language and any hardware target could plug into. Mojo and MAX apply the same architecture to AI inference: make the choice of chip a deployment decision rather than a development commitment.
The risk is concrete. Abstraction layers cost performance. Nvidia’s stack is fast because it is tightly coupled to specific hardware. A compiler that cannot match CUDA’s hand-tuned kernels becomes a tax on every model it runs. Lattner proved this was solvable for general-purpose compilation. Whether it holds for AI workloads, where individual kernels are optimized for specific tensor operations on specific architectures, is the $4 billion question.
Qualcomm just paid $4 billion for a compiler team. That price says more about the data center market than any chip specification could.