Original source: Wall Street View

On June 24, New York. Qualcomm CEO Cristiano Amon announced something: by fiscal year 2029, non-phone business is to reach 40 billion dollars, with data center chips accounting for 15 billion. After hours, the stock jumped 13%.

Then Zuckerberg appeared on the screen.

"Qualcomm will become Meta's data center CPU supplier." He recorded a video, not the plastic wording of "we are pleased to explore collaboration opportunities" in a press release, but a complete promise with subject, verb, and object. The contract covers multiple product generations.

It sounds like a story of a king returning. Eighteen years ago, Snapdragon defined smartphones; now Qualcomm wants to define data centers.

But Qualcomm should know better than anyone else—nine years ago, it tried, called Centriq. That time it failed thoroughly.

Piled on the Ruins from Nine Years Ago

In November 2017, San Francisco. Qualcomm released Centriq 2400, the world's first 10-nanometer server chip, Arm architecture, 48 Falkor cores. The PowerPoint back then looks quite similar to today's: high efficiency, low power consumption, breaking Intel's monopoly.

In less than two years, the project was gone. The head left, the team was disbanded. Centriq became a name no one voluntarily mentioned within Qualcomm.

Why did that one die? Three things came together.

In 2017, Intel held more than 90% of the server CPU market. AMD's EPYC hadn't shipped yet. Arm had a market share of barely 1%, not even worth calling an "experiment."

The attitude of cloud vendors can be translated as—looks okay, but why switch?

More fatal was that from launch to cancellation, Centriq had no publicly announced mass production customers. Microsoft Azure was said to be testing but never officially announced it. For data center chips, the lack of a mass production customer commitment means no product. This industry does not want "the chip has been made," but "someone is using it to run production loads."

The most critical issue was that Qualcomm itself was almost falling apart during those two years—fighting a global patent war with Apple, and Broadcom came along with a $100 billion acquisition offer. The question management faced every morning was not "how to sell server chips," but "is the company still ours tomorrow?"

Centriq died at the intersection of these three things: the timing was off, customers did not come, and there was a fire in the backyard.

This Time, the World Turned Itself Over

Nine years is enough for the world to change.

Arm's market share in data centers rose from 1% to 25%. Graviton is the watershed—Amazon began making its own Arm server CPUs in 2018, and by last year, it accounted for more than half of AWS's new CPU capacity for three consecutive years. Of AWS's top 1000 clients, 98% are using it. NVIDIA's Grace CPU is also nearing revenue levels similar to GPUs in the cloud.

Translating into Qualcomm's language: in 2017 it faced two questions—"Can Arm make servers?" and "Can Qualcomm make Arm server chips?" Now, the first question has been answered by Graviton and Grace. It only needs to prove it can do better than Graviton. The difficulty is not on the same level.

Another change, the numbers in reports may not express it, but Zuckerberg's video made it clear.

Intel and AMD have dominated server CPUs for over thirty years; never has a cloud vendor CEO jumped out to say "I will sign multi-generation strategic contracts with such-and-such company." There were no technical collaborations; no one wanted to put their chips on the table.

Meta made its move this time.

The reasons behind it are straightforward and don’t need analysis: Intel's server CPU market share has dropped from absolute monopoly to 62%, and AMD has taken 46% of revenue on the x86 side. On Arm's side, Graviton, Grace, and Ampere are lined up to break walls. Any cloud vendor's procurement department looking at this situation and continuing to put all their eggs in the x86 basket would be negligent.

So what Meta signed is not "let’s test," but "multi-generational." Translated, that means: if the first generation fails, try the second generation.

In the data center chip market, no one has ever received such conditions.

CPU is the Decoy, There Are Five Wars

At this point, the story looks like a beautiful turnaround script.

But what Qualcomm really wants to do is much bigger than a CPU. The Dragonfly family shown by Amon is actually a complete set of data center architecture.

There is an Arm server CPU with more than 5GHz and 250 cores, called C1000, scheduled for mass production at Meta by mid-2028. There are AI inference accelerators called AI200, AI250, AI300, set to be updated annually starting next year, specifically for running models after they go live. There is a high-bandwidth computing platform called HBC, which claims it can achieve 4 to 8 times more throughput at the same cost, with sample chips to be provided to Microsoft by mid-next year.

These components also need to be connected—Qualcomm has laid down a network solution from electrical interconnects to optical interconnects to ensure that data flows between chips without blockage.

Then, a layer of software was topped. Qualcomm spent $3.9 billion to acquire a company called Modular, aiming to create a platform that allows developers to write code once and run it on any hardware they choose. The CEO of Modular is Chris Lattner, who wrote the LLVM compiler and also the Swift language, and later managed Autopilot at Tesla.

It all sounds reliable. But doing five things simultaneously—CPU, accelerators, HBC, networking, software—means Qualcomm is not just making a chip, but building a building. The foundation, walls, pipes, wires, and decoration can all be independently set, but if any one of them fails, the entire building is a dangerous structure.

Historically, only one company has successfully operated more than two independent chip product lines in data centers—NVIDIA. The GPU itself, plus acquiring Mellanox for networking, along with the software ecosystem of CUDA.

Intel's résumé is a counterexample: over the past decade, it used Atom for mobile chips and abandoned it in 2016. It used Rialto Bridge for GPUs, canceled in 2023, and the subsequent Falcon Shores was also not pushed to the public. Intel stopped using Omni-Path for network interconnection in 2019. Tens of billions of dollars burned, having brought in the industry's best engineers, yet none of them passed.

The Nuvia team designing a high-performance Arm CPU is unlikely to fail—coming from Apple's core group, with M1 lineage. The risk is not at the design level. The risk lies in the fact that Qualcomm is a company that has never simultaneously operated five independent chip product lines. Mobile SoCs, automotive cockpit, PC chips, IoT, XR—these are essentially five variants built on one foundation. Data centers mean starting over, revamping the entire supply chain, sales, and customer support.

The Last Layer of the Building May Not Be Erected

Among the five lines, software is the most fragile string.

NVIDIA's CUDA is a structure that has grown over fifteen years. It is not a tool, nor a platform—it is a system where millions of developers have established themselves inside. Tens of thousands of papers, thousands of optimized libraries, and every underlying pathway from PyTorch to TensorFlow have grown out of it.

Modular wants to build a parallel thing with $3.9 billion and two years of time.

The story is appealing: the MAX platform lets you write once and run on CPU, GPU, NPU, and custom ASICs. The Mojo language resembles Python while achieving compiler-level performance.

But AMD has told a nearly identical story. ROCm, open, compatible, a replacement for CUDA. It's been ten years, yet the supported applications still require page-by-page checks for compatibility. Intel's oneAPI is the same.

Moreover, Modular has a contradiction that it may not have fully realized. If MAX treats all chips equally, why does Qualcomm want to make its own chips? Wouldn’t it work just as a software company? If MAX runs best only on Qualcomm chips, then it is no different in essence from AMD's ROCm—a software layer that claims to be open, but in reality, is tied to its own hardware.

This is not a technical issue. CUDA's wall is not built on technology but on fifteen years of time and the habits of millions of people. Modular does not lack technology; it lacks time. And how many years are Qualcomm's investors willing to wait?

Intel Should Be Most Nervous

The company most directly threatened by Qualcomm's actions is Intel.

Arm CPUs are directly competing with x86 server CPUs. That is Intel's last high-profit stronghold—by Q1 2026, AMD has already taken 46% of x86 server revenue, while Intel's shipment share is down to 62%, continuing downward. Qualcomm entering means Intel is not just competing with AMD over x86 inventory but is being attacked from another side by the Arm camp.

Qualcomm's valuation logic is also shifting. A PE ratio of 15 to 18 times is the price for a mobile company. Once the market accepts the narrative of "AI infrastructure company," compared to AMD and NVIDIA's 25 to 30 times, Qualcomm's stock price still has room for revaluation.

But this is predicated on everything going smoothly.

Focus on Three Timelines

The most troubling aspect of Qualcomm's story is that—money has been promised, but the products have not yet been delivered.

The most recent report is due early next year. Amon stated that the custom chip business is expected to generate substantial revenue starting in FY2027 Q1, and that quarterly figure will be the market's first report card. If it exceeds $1.2 billion, the narrative will stand firm. If it's below $800 million, everything gets discounted.

The second report is due mid-next year, when HBC commercial samples are delivered to Microsoft. Before that, Qualcomm has been saying for two years “we can do it,” and only on that day will it unveil “we have done it.”

The final report is due in mid-2028 when C1000 officially launches at Meta. Until that node, each chapter of this story is marked "to be continued."

Qualcomm is betting on a three-year cycle. Any misstep at any node will lead the market to reprice it.

The market movements on June 25 also say the same thing—after hours, it jumped from 197 to 223, then pulled back. Wall Street's message is clear: you’ve told a great story. Now show it to us.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。