Qualcomm bets $14 billion to break NVIDIA monopoly with RISC-V and open compiler

Qualcomm announced on June 24 the acquisition of Modular — creator of the Mojo programming language and the MAX inference engine — for approximately $3.92 billion in stock. Simultaneously, advanced negotiations with Tenstorrent, a RISC-V chip startup led by Jim Keller, point to an additional deal of $8 to 10 billion. Combined, the moves represent up to $14 billion invested in a strategy to break NVIDIA's monopoly in the AI ecosystem — attacking not just hardware, but the software layer that keeps developers locked into CUDA.

The problem Qualcomm is trying to solve

NVIDIA does not dominate the AI accelerator market by having the fastest GPUs. It dominates because it has CUDA — an 18-year programming ecosystem including libraries, compilers, tutorials and thousands of models trained with optimizations specific to NVIDIA hardware.

For an AI developer to migrate to alternative hardware, they must rewrite or recompile code for a new software ecosystem. That migration cost is what keeps NVIDIA on top even when hardware alternatives exist.

What Modular solves

Modular created two products that attack this problem directly. Mojo is a programming language combining Python ergonomics with C performance — and compiles to heterogeneous hardware including CPUs, NVIDIA GPUs, ARM chips and, in the future, Qualcomm chips. MAX is an inference engine that automatically optimizes AI models for available hardware.

In practical terms: a developer can write code in Mojo once and run it on NVIDIA, AMD or Qualcomm without rewriting. This removes the CUDA lock-in barrier.

What Tenstorrent and Jim Keller add

Jim Keller is the most respected chip engineer in the industry — he designed or co-designed architectures at AMD (Zen, K8), Apple (A4, A5), Intel and Tesla (computer vision chip). Tenstorrent develops RISC-V AI chips with open, licensable and modifiable architecture — the opposite of NVIDIA's proprietary architecture.

If Qualcomm acquires Tenstorrent, it has hardware (Dragonfly, Tenstorrent RISC-V), software (Mojo + MAX) and an existing customer base (Microsoft, Meta, Samsung for mobile chips). That is the most complete NVIDIA ecosystem alternative ever assembled by a single player.

Why this matters now

NVIDIA dependence has created real systemic risk for the entire technology industry. NVIDIA servers have 12 to 18 month lead times. Prices rose 300% in two years. A single vendor controls the bottleneck of all global AI infrastructure.

Any technically and economically viable alternative that reduces this dependence is good for compute buyers — which includes everyone building AI applications.

For 10Dobro

Our 26 systems run on cloud and are not direct GPU buyers. But more competition in AI hardware and software is structurally good: API token costs tend to fall in the medium term when cloud providers have alternatives to NVIDIA hardware. We are watching closely.