One CPU, Five Battles: Nine Years Later, Qualcomm Bets Again

June 24, New York. Qualcomm CEO Cristiano Amon announced: By fiscal year 2029, non-mobile business will reach $40 billion, with data center chips accounting for $15 billion. After-hours stock jumped 13%. Then Mark Zuckerberg appeared on the screen. "Qualcomm will become Meta's data center CPU supplier." He recorded a video, not the kind of plastic wording in press releases like "we're excited to explore opportunities," but a full subject-verb-object commitment. The contract covers multiple product generations. Sounds like a hero’s return. Eighteen years ago, Snapdragon defined mobile phones; now Qualcomm wants to define data centers. But no one should know better than Qualcomm itself—nine years ago they tried, it was called Centriq. That attempt failed spectacularly. Piled on the Ruins from Nine Years Ago November 2017, San Francisco. Qualcomm launched Centriq 2400, the world’s first 10nm server chip, Arm architecture, 48 Falkor cores. The PPT then looked much like today’s: high efficiency, low power consumption, breaking Intel’s monopoly. Less than two years later, the project was scrapped. The director left, the team disbanded. Since then, Centriq became a name nobody inside Qualcomm proactively mentioned. Why did it die? Three factors converged. 2017: Intel held more than 90% of the server CPU market. AMD’s EPYC hadn’t shipped. Arm’s cloud market share was barely 1%—not even “experimental.” Cloud providers' attitude translated: “Looks alright, but why switch?” More fatally, Centriq never had a public mass production customer. Microsoft Azure reportedly tested it, but never officially announced. For data center chips, no mass production customer commitment means no product. This industry doesn’t want “the chip works”; it wants “someone uses it for production workloads.” Most fatally, Qualcomm itself nearly fell apart those years—embroiled in a global patent war with Apple, Broadcom came out swinging with a $100 billion acquisition offer. Management woke up each morning not with “How to sell server chips?” but “Will the company still be ours tomorrow?” Centriq died at the intersection of these three things: wrong timing, no customers, internal crisis. This Time, the World Has Flipped Nine years is enough for the world to change. Arm's data center market share rose from 1% to 25%. Graviton is the watershed—Amazon started making its own Arm server CPUs in 2018, and for three consecutive years, accounted for over half of new AWS CPU capacity. Of AWS’s largest 1000 customers, 98% use it. Nvidia’s Grace CPU revenue in the cloud is now similar in scale to their GPUs. Translated to Qualcomm’s terms: In 2017 it faced two problems—“Can Arm make servers?” and “Can Qualcomm make Arm server chips?” Now, Graviton and Grace answered the first one. They just have to prove they’re better than Graviton. The challenge isn’t even in the same league. Another change, not shown in financials but clear in Zuckerberg’s video. Intel and AMD ruled server CPUs for over 30 years; no cloud CEO ever publicly said, “I will sign a multi-generation strategic contract with XX.” There’s technical cooperation—but no one put their chips on the table. Meta did this time. The reason is so straightforward it needs no analysis: Intel’s server CPU market share dropped from absolute monopoly to 62%, AMD took 46% of x86 revenue. On Arm’s side, Graviton, Grace, Ampere are lined up tearing down walls. For any cloud provider’s procurement department, to keep all their eggs in the x86 basket would be dereliction. So Meta didn’t sign “we’ll try it,” but “multi-generation.” Translation: If Gen 1 doesn’t work, Gen 2 keeps trying. No one ever got such terms in the data center chip market. CPU is a Decoy; There Are Five Battles Here, the story looks like a beautiful comeback script. But what Qualcomm aims to do is much bigger than a single CPU. The Dragonfly family Amon showed is actually an entire data center skeleton. There’s an Arm server CPU, over 5GHz, 250 cores, called C1000, to be mass-produced at Meta in mid-2028. There’s an AI inference accelerator, AI200, AI250, AI300, annual updates starting next year, designed to run real model calls post-deployment. There’s a high-bandwidth compute platform, HBC, that supposedly provides 4–8x throughput at the same cost, with samples to Microsoft by mid-next year. All these need to be connected—Qualcomm laid out a network solution from electrical to optical interconnect, so data flows smoothly between chips. Then, a software layer on top. Qualcomm spent $3.9 billion to acquire a company called Modular, aiming to build a platform where developers write code once and deploy it on any hardware. Modular’s CEO is Chris Lattner, creator of LLVM, Swift (the language), later headed Tesla’s Autopilot. Sounds solid. But doing five things at once—CPU, accelerator, HBC, network, software—means Qualcomm isn’t building one chip, it’s building a tower. Foundation, walls, plumbing, wiring, and finishes—each could be a standalone project, but if any part is unfinished, the whole tower is unstable. Historically, only one company won two or more fronts in the data center: Nvidia. Their GPU, plus acquired Mellanox for networking, plus CUDA software ecosystem. Intel’s track record is the counterexample: In the past ten years, they used Atom for mobile chips (gave up in 2016), Rialto Bridge for GPU (cancelled in 2023, Falcon Shores isn’t pushed anymore), Omni-Path for networking (stopped in 2019). Tens of billions spent, industry’s best engineers poached, none succeeded. The Nuvia team making a high-performance Arm CPU is unlikely to fail—Apple alumni, M1 lineage. The risk isn’t in design. The risk is Qualcomm never ran five independent chip product lines at once. Phone SoC, automotive cockpit, PC chip, IoT, XR—these are essentially one base, five variants. Data center is a reset, supply chain, sales, customer support—everything starts anew. The Top Floor Might Not Get Built Of the five fronts, software is the most fragile string. Nvidia’s CUDA has fifteen years of structure. Not a tool, not a platform, but a system where millions of developers have settled. Tens of thousands of papers, thousands of optimized libraries, every backend from PyTorch to TensorFlow grows from it. Modular wants to use $3.9 billion and two years to build a parallel system. The story sounds good: The MAX platform lets you write once, run anywhere—CPU, GPU, NPU, custom ASIC. Mojo language looks like Python, compiles like C. But AMD told a nearly identical story—ROCm, open, compatible, CUDA replacement. Ten years on, application support still needs to be checked list by list. Intel's oneAPI is the same. Also, Modular has a contradiction they may not even have figured out. If MAX treats all chips equally, why does Qualcomm make chips instead of just being a software company? If MAX runs best only on Qualcomm chips, it is essentially like AMD’s ROCm—a software layer claiming openness but actually tied to their own hardware. This isn’t a technical problem. CUDA’s wall isn’t built by technology, but by fifteen years and millions of habits. Modular lacks not technology, but time. Will Qualcomm’s investors wait several years? Intel Should Be the Most Nervous Qualcomm’s latest move poses the most direct threat to Intel. Arm CPUs directly target x86 server CPUs—that’s Intel’s last high-profit fortress. Q1 2026, AMD already took 46% of x86 server revenue, Intel’s shipment share is 62%, still declining. With Qualcomm coming in, Intel isn’t just fighting AMD for x86 share, but is being attacked from the Arm camp simultaneously. Qualcomm's valuation logic is also shifting. A 15–18x PE is a mobile company price. If the market accepts the “AI infrastructure company” narrative, aligning with AMD and Nvidia’s 25–30x, Qualcomm stock has another revaluation opportunity. But that hinges on everything going smoothly. Watch Three Dates The hardest part of Qualcomm’s story: they’ve announced the money, but haven’t delivered the product. The nearest checkpoint is early next year. Amon says custom chip business will generate substantive revenue from FY2027 Q1, so that quarter’s figures are the first scorecard. Over $1.2 billion, story holds. Under $800 million, everything is discounted. The second is mid-next year, HBC commercial samples to Microsoft. Before then, Qualcomm spent two years saying “We can do it”; on that day, they first show “We did it.” The last is mid-2028, C1000 goes live at Meta. Until then, every chapter in this story is “to be continued.” Qualcomm is betting on a three-year cycle. If any checkpoint misses, the market will reopen the price. June 25 intraday trading says the same—after-hours from 197 to 223, then pulling back. Wall Street’s takeaway: You told a good story. Now show us. Risk Warning and Disclaimer The market involves risks, investments require caution. This article does not constitute personal investment advice, nor does it consider the specific investment goals, financial circumstances or needs of individual users. Users should assess any opinions, views or conclusions herein against their own situation. Investing based on this is at your own risk.