CEO Cristiano Amon unveiled an aggressive five-year data center strategy at Qualcomm's Investor Day, backed by hardware nobody had seen before, two of the world's most valuable tech companies as anchor customers, and an acquisition that targets the one thing Nvidia competitors have failed to crack for fifteen years.
Qualcomm walked into the Ziegfeld Ballroom in Manhattan on Wednesday with a specific message for Wall Street: the smartphone chip company is gone, and in its place is something that intends to compete for Nvidia's data center business.
The announcements that followed covered a new server CPU, a proprietary memory architecture, an AI inference accelerator roadmap, endorsements from Mark Zuckerberg and Satya Nadella, and a $3.92 billion acquisition of an AI software startup whose entire reason for existing is to make Nvidia's software platform unnecessary. Qualcomm shares surged as much as 15% on the day. By Friday, those gains had mostly unwound as broader Nasdaq selling pressure dragged the stock back down.
The stock movement told a short story. The Investor Day told a longer one.
A Target That Would Have Seemed Implausible Two Years Ago
Qualcomm laid out a financial target that doubles its prior long-range forecast: $40 billion in annual revenue from businesses excluding handsets by fiscal 2029. Within that number, the company is projecting more than $15 billion specifically from AI components in data centers.
That projection is ambitious on its own terms. For context, Qualcomm reported total revenue of $10.6 billion in its most recent quarter. The data center business it is projecting for 2029 does not yet exist at meaningful scale.
CEO Cristiano Amon addressed that gap directly. "A lot of people ask, 'Oh, in this crowded market, is it too late?'" he told a packed room of Wall Street analysts. "It's never too late for Qualcomm."
Amon has made that claim before, across automotive and PC chips, and it has so far held up. Qualcomm's automotive revenue reached a record $1.3 billion in its most recent quarter, up 38% year over year. The company entered that market when it was also viewed as late and crowded.
The global AI accelerator market is forecast to reach $680 billion by fiscal 2029. Qualcomm is targeting $15 billion of that. The math on market share is modest enough to be plausible. The execution required to get there is not.
The Hardware: Dragonfly, HBC, and a Memory Claim Worth Testing
The centerpiece of Qualcomm's data center hardware push is a portfolio called Dragonfly, which spans a server CPU, an AI inference accelerator, and a proprietary memory architecture that the company is positioning as a structural alternative to the existing industry standard.
The Dragonfly C1000 CPU is an Arm-based chiplet design with more than 250 cores. Qualcomm says it delivers more than twice the performance per watt of existing server CPUs. It is built for AI agent workloads and optimized for power efficiency, which Amon identified as the wedge Qualcomm intends to use to get a foothold in a market where energy costs have become a material concern for data center operators.
Meta has signed a multi-generational partnership to deploy the C1000 in its next-generation server fleet. That makes Meta Qualcomm's first formal CPU customer at hyperscale. The significance of that endorsement is hard to overstate. Hyperscalers do not sign multi-generational agreements with unproven silicon suppliers.
The Dragonfly AI300 is an inference accelerator that sits atop the existing AI200 and AI250 on Qualcomm's annual roadmap. Commercial sampling for the AI300 is expected to begin in 2028. Microsoft Azure has confirmed it will deploy Qualcomm's AI250 and AI300 accelerators using HBC, the company's new memory technology.
HBC, which stands for High-Bandwidth Computing, is the technical bet that separates Qualcomm's pitch from what every other Nvidia competitor has offered. Conventional AI accelerators connect processors to stacks of high-bandwidth memory across a silicon interposer, an arrangement that burns power moving data across that physical distance. Qualcomm's approach places compute cores directly beneath a DRAM stack, collapsing that distance to near zero.
The performance figures Qualcomm cites are significant if they hold under independent testing. The company claims HBC delivers six times the bandwidth per watt of existing high-bandwidth memory solutions and 200 times the memory capacity per watt compared to SRAM. The AI250 equipped with HBC Gen 1 delivers effective memory bandwidth of 133 TB/s per card, an 18-fold improvement over the AI200 using conventional LPDDR5X. Qualcomm also claims the architecture delivers eight times more tokens per watt than traditional GPU configurations.
Those numbers will face scrutiny. Independent benchmark verification remains the standard, and Qualcomm has yet to ship this technology at scale.
The $3.92 Billion Bet That Isn't Really About Chips
The Modular acquisition is the most strategically significant announcement Qualcomm made on Wednesday, and it has almost nothing to do with hardware.
Nvidia's dominance in AI infrastructure is not a chip story. AMD, Intel, and others have produced competitive silicon for years without making a meaningful dent in Nvidia's market position. The actual reason developers and enterprises stay on Nvidia hardware is CUDA, a software platform built over fifteen years that has accumulated roughly 4 million developers and the entire toolchain of modern AI development. Leaving Nvidia means rewriting code, retraining engineering teams, and betting on a less mature software ecosystem. Most organizations decide the hardware savings do not justify that disruption.
Modular was built to eliminate that switching cost.
The company's two flagship products are the Mojo programming language and the MAX inference engine. Together, they let developers write AI inference code once and run it optimized across CPUs, GPUs, NPUs, and custom ASICs without a rewrite for each chip architecture. The platform was built deliberately without any Nvidia vendor libraries, making it structurally hardware-neutral rather than just rhetorically so.
Modular's co-founders are not newcomers to this problem. Chris Lattner created LLVM, the compiler infrastructure that underpins most modern programming languages, and Apple's Swift programming language. Tim Davis co-created TensorFlow Lite, which brought machine learning to mobile devices at scale. They are among the few engineers in the world with the credentials to build a credible CUDA alternative.
Lattner described the acquisition's purpose plainly: "Joining Qualcomm gives us the scale and platform reach to accelerate that mission. Together, we can make AI development more accessible and performant for developers."
Amon framed the broader intent in structural terms. "As agentic AI scales across data centers and edge environments, the industry is moving toward disaggregated, multi-vendor architectures that demand a more open and modern software foundation," he said.
Qualcomm also announced a partnership with Hugging Face, the AI development platform with 16 million registered developers, which will route its community directly into Qualcomm silicon support from experimentation through production deployment. If Modular can pull even a fraction of that developer base away from CUDA-first workflows, the downstream hardware opportunity is substantial.
The deal is expected to close in the second half of 2026, pending regulatory approval. Qualcomm's previous major chip acquisition attempt, the $47 billion bid for NXP Semiconductors in 2018, collapsed when Chinese regulators declined to approve it. The Modular deal is smaller and a software acquisition, which carries a different regulatory profile.
The Tenstorrent Talks That Could Add Another $10 Billion
Modular may not be the final piece.
Reports indicate Qualcomm is in advanced talks to acquire Tenstorrent, the AI chip startup led by Jim Keller, a chip architect previously responsible for Apple's processors and Tesla's Autopilot silicon. The reported valuation range is $8 billion to $10 billion.
If both the Modular deal and the Tenstorrent deal close, Qualcomm will have committed approximately $14 billion to reshaping its AI portfolio within a matter of weeks.
Tenstorrent's significance is architectural. Its Blackhole chip, which reached general availability in April 2026, is built on RISC-V rather than Arm. That distinction matters to Qualcomm for reasons beyond performance. Qualcomm won its lawsuit against Arm in December 2024 and secured a final dismissal of Arm's remaining claims in October 2025, but the underlying dynamic, a licensor that now competes directly with its own licensees, creates long-term structural exposure that a patent victory does not fully neutralize. Tenstorrent would give Qualcomm a RISC-V accelerator architecture independent of the Arm license relationship entirely.
Qualcomm already acquired RISC-V server chip designer Ventana Micro Systems in December 2025 and completed a $2.4 billion purchase of high-speed interconnect supplier Alphawave Semi earlier this year. Tenstorrent would add the accelerator layer that turns those pieces into a coherent end-to-end stack.
What Qualcomm Is Walking Into
The competitive environment Qualcomm is entering is harder than any market it has previously targeted.
Nvidia's most recent quarter produced data center compute revenue of $60.4 billion, up 77% from a year ago. The company's market capitalization sits at $4.7 trillion, making it the most valuable company in the world. Its software ecosystem has fifteen years of compounded developer investment behind it. AMD, the most credible existing alternative, has been trying to crack the same market for years and is now entering the AI data center server segment with its Helios rack server. Neither outcome from AMD's efforts to date is particularly encouraging for challengers.
The hyperscaler landscape is also shifting in ways that complicate straightforward market share math. Amazon, Google, and Microsoft have all been developing proprietary AI chips for internal use. Meta has its own in-house silicon efforts alongside its new Qualcomm CPU partnership. These companies are buying Qualcomm products while simultaneously building what may eventually compete with them.
Tony Pialis, Qualcomm's Executive Vice President and General Manager for the data center business, offered a framing that stood out during the investor presentations. He described customer demand as "customers pulling us in, rather than us forcing our way in." The Meta and Microsoft commitments appear to bear that out. Those are not companies that can be talked into chip partnerships by a vendor's investor day presentation.
Amon's Bet on Agentic AI as the Opening
The technical argument Qualcomm is making goes beyond raw performance figures.
Amon is betting that the shift from AI training workloads, where Nvidia's GPU architecture is most dominant, to AI inference and agentic AI workloads, where models run continuously and efficiency per token matters more than peak throughput, plays to Qualcomm's strengths. Power efficiency in training is a secondary consideration when a company is building a model once. Power efficiency in inference is a primary consideration when a model is running billions of queries a day at cost.
The HBC architecture is specifically designed for inference. The C1000 CPU is optimized for AI agent workloads. The Modular platform specializes in inference serving. The product strategy coheres around a single thesis: the AI market is moving toward a workload profile where Qualcomm's architectural choices become advantages rather than constraints.
Whether that thesis is correct, and whether Qualcomm can execute the acquisitions, ship the hardware, and build the software community in time to matter, will take years to answer. The first hard data point arrives in Q1 fiscal 2027, when Qualcomm says its custom silicon business will begin contributing meaningful revenue.
Amon's track record in automotive, a market he entered against skepticism and where Qualcomm now posts record revenue quarter after quarter, means investors are at least willing to wait for those results. The remaining question is whether the AI chip market gives him the same runway that automotive did.
Nvidia, which has been given that question before about previous challengers, is not waiting to find out. Its data center networking revenue in the same quarter was $14.8 billion, up 199% year over year.
