Qualcomm Is Spending $14 Billion in Six Months on Three RISC-V Companies. Together, They Target Every Layer of NVIDIA's $75 Billion-a-Year Software Moat.

Abstract semiconductor wafer with layered chip architectures converging into a unified stack

Qualcomm is in advanced talks to acquire AI infrastructure startup Modular Inc. for approximately $4 billion, Bloomberg reported on Monday. Two weeks earlier, The Information reported that Qualcomm was negotiating to buy Tenstorrent, a RISC-V AI chip company led by Jim Keller, for $8 to $10 billion. And in January, Qualcomm quietly closed the acquisition of Ventana Micro Systems, a RISC-V server CPU startup, for an undisclosed sum.

Three acquisitions. Six months. Roughly $14 billion in committed capital. None of these companies depend on ARM or NVIDIA.

That matters because Qualcomm's entire existing business runs on ARM. Every Snapdragon processor in every Android flagship, every Windows laptop Qualcomm has courted, every automotive platform it ships pays royalties to ARM Holdings. ARM sued Qualcomm over licensing terms in 2022. Qualcomm won the jury trial in December 2024. But winning a lawsuit does not eliminate a dependency, and dependencies make strategists nervous.

What $14 Billion Buys

Each acquisition fills a distinct layer of a complete data center stack:

Company	Role	Reported Price	What It Replaces
Ventana Micro Systems	RISC-V server CPUs	Undisclosed	ARM-based host processors
Tenstorrent	RISC-V AI accelerators	$8–10B	NVIDIA GPUs
Modular (MAX platform)	AI compiler & runtime	~$4B	CUDA software ecosystem

Ventana brings server-class RISC-V CPU designs. Its Veyron V2 chiplet features up to 32 RISC-V RVA23-compatible cores clocked at 3.85 GHz, with 128 MB of shared L3 cache. Tenstorrent brings AI accelerator hardware built on a proprietary Tensix core architecture, plus an open-source software stack, plus Jim Keller. Modular, founded in 2022 by Chris Lattner (creator of LLVM and the Swift programming language) and Tim Davis, brings the MAX inference engine, a platform designed to deploy AI models efficiently across hardware backends without requiring code rewrites.

Combined, that is something nobody else has assembled: CPUs, AI accelerators, and the compiler to program both, all built on RISC-V, all royalty-free from ARM, all positioned as a direct alternative to the CUDA-locked ecosystem that currently extracts $75 billion a year in margin premiums from customers who cannot afford the engineering cost of switching away.

For the first time, every layer of a complete alternative to NVIDIA's data center stack exists inside one company. All royalty-free from ARM. All aimed at the same $75 billion margin pool.

Quantifying the CUDA Tax

NVIDIA reported $75.2 billion in data center revenue for Q1 fiscal 2027 (ended April 26, 2026), up 92% year-over-year, according to its SEC filing. Annualized, that run rate approaches $300 billion, earned at a non-GAAP gross margin of 75.0%.

AMD, the only other company shipping data center GPUs at scale, operates at roughly 50% gross margins across its business, per AMD's Q1 2026 filings. Revenue was $10.25 billion, up 37.8% year-over-year.

That 25-point gross margin premium, applied to NVIDIA's $300 billion annualized data center revenue, implies approximately $75 billion per year in excess gross profit that exists because NVIDIA can charge prices its competitors cannot match and customers will still pay rather than undertake the multi-month, multi-team engineering project of migrating their CUDA codebases to an alternative platform.

Not all of it is software lock-in. Hardware performance, manufacturing scale, and first-mover ecosystem effects all contribute. But the core mechanism is switching cost. This is the margin pool Qualcomm is targeting.

Jim Keller Has Done This Before

At DEC in the 1990s, Keller co-designed the Alpha 21264, the fastest microprocessor of its era, aimed squarely at Intel's server monopoly. At AMD in the early 2000s, he led the K8 microarchitecture (Athlon 64, Opteron), the first x86-64 processor, which forced Intel to abandon Itanium and adopt AMD's 64-bit instruction set instead. At Apple, he led the A4 and A5 chips that powered the iPhone 4 and iPad 2, launching Apple's transition from Samsung-supplied processors to custom silicon that now underpins a $3.5 trillion company. At Tesla, he designed the Full Self-Driving computer (Hardware 3), replacing NVIDIA's Drive PX platform with an in-house inference chip.

Every engagement followed the same pattern: a dominant supplier controlled a critical silicon dependency, and Keller designed the chip that eliminated it, which is why his presence in a deal this size matters more than any product roadmap or revenue forecast, because the pattern has held across four companies, three decades, and three different monopolies that each looked permanent until they weren't.

Tenstorrent raised $693 million in a Series D at a $2.6 billion post-money valuation in December 2024, backed by Samsung Securities, Bezos Expeditions, Fidelity, and Hyundai. At $8 to $10 billion, Qualcomm would pay a 3–4x premium over that round, which sounds expensive for a startup with $150 million in signed customer contracts but starts to look reasonable if you believe that what Qualcomm is actually acquiring is not a chip company but arguably the only credible RISC-V AI accelerator team in the world led by an engineer who has successfully designed monopoly-breaking silicon at four different companies across three decades.

Why Now

Qualcomm's investor day falls on June 24, where the company is expected to reveal a major hyperscaler customer for its data center chips. J.P. Morgan analyst Samik Chatterjee projects more than $3 billion in data center revenue for fiscal 2027, potentially surging to $35 billion by fiscal 2031. Cristiano Amon, Qualcomm's CEO, has spent three years repositioning the company beyond smartphones, and the pressure to validate that strategy has intensified after revenue from the QCT handset business declined 3% year-over-year last quarter because of memory supply constraints and inventory corrections that suggest the mobile chip market may no longer reliably deliver the growth multiples shareholders now expect from AI-era semiconductor companies.

Qualcomm also completed a $2.4 billion acquisition of Alphawave Semi (high-speed chip-to-chip interconnect technology) in late 2025. Four acquisitions in under a year, totaling roughly $16 billion including Alphawave, all targeting the data center.

Limitations

Both the Tenstorrent and Modular deals remain in talks. Neither has closed. Either could collapse, and the reported price ranges could shift. The $14 billion total assumes midpoint Tenstorrent pricing ($9 billion) plus an estimated $300–500 million for Ventana, which was never officially disclosed. The CUDA tax calculation uses company-wide gross margins as proxies for segment-specific data center margins, which neither NVIDIA nor AMD reports in a directly comparable format. Annualizing a single quarter overstates the run rate if growth decelerates and understates it if the 92% year-over-year pace continues. Modular's MAX platform has limited production deployment data; its $4 billion valuation represents a 2.5x step-up from its September 2025 round just nine months prior.

The Strongest Case Against This Bet

NVIDIA's moat is not just software. Blackwell GPUs remain the fastest AI training hardware shipping today. TSMC's most advanced nodes are capacity-constrained, and NVIDIA has priority allocation. Every major AI framework, every cloud provider SDK, every enterprise MLOps tool is optimized for NVIDIA first. AMD has been shipping ROCm as a CUDA alternative for seven years and still holds single-digit data center GPU market share. If AMD, with $880 billion in market capitalization and a complete GPU product line, cannot meaningfully erode NVIDIA's lock-in, assembling a RISC-V startup stack is unlikely to do it faster. Qualcomm's $243 billion market cap means these acquisitions represent roughly 6% of the company's value. A bet this large going wrong compresses margins, distracts engineering talent from the mobile business that actually generates cash, and hands ARM an opening to court Qualcomm's own customers while management is focused elsewhere.

The Bottom Line

NVIDIA's data center dominance rests on a software moat worth approximately $75 billion a year in gross margin premium. Qualcomm is assembling the first vertically integrated alternative by spending $14 billion to acquire RISC-V CPUs, RISC-V AI accelerators, and an AI compiler platform, bringing Jim Keller (who has broken three silicon monopolies) and Chris Lattner (who built the compiler infrastructure most modern software depends on) under the same corporate roof. Nobody has attempted a full-stack challenge at this scale before. It will take years to know whether the bet works. But for the first time, every layer of the alternative exists inside one company.

What You Can Do

If you hold NVIDIA: this changes nothing in 2026 but represents the first credible architectural threat to CUDA since AMD launched ROCm. Monitor Qualcomm's investor day on June 24 for revenue targets that validate the data center strategy. If you work in ML infrastructure: track Modular's MAX platform for inference workloads on non-NVIDIA hardware. Lattner's LLVM track record suggests the tooling will be technically serious. If you're evaluating data center hardware for 2027 and beyond: request quotes from Qualcomm's server division alongside NVIDIA and AMD, because even if Qualcomm's chips launch slower, the credible threat of a third supplier gives you bargaining power against NVIDIA pricing today.