一颗CPU，五场战争：九年之后，高通再赌一次

Original Source: Wall Street CN

On June 24th, New York. Qualcomm CEO Cristiano Amon announced a goal: by fiscal year 2029, non-handset business must generate $40 billion, with data center chips contributing $15 billion. The stock jumped 13% after hours.

Then Mark Zuckerberg appeared on the screen.

“Qualcomm will become Meta’s data center CPU supplier,” he said in a recorded video. It wasn’t the typical plastic language of press releases like “we are excited to explore opportunities for cooperation.” It was a full-sentence commitment. The contract spans multiple product generations.

It sounds like a comeback story. 18 years ago, Snapdragon defined the smartphone. Now Qualcomm wants to define the data center.

But no one inside Qualcomm should understand better than anyone else that it tried this 9 years ago. The project was called Centriq. And it failed completely.

Built on the Ruins of Nine Years Ago

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

Less than two years later, the project was gone. The lead left; the team disbanded. Since then, "Centriq" has become a name no one at Qualcomm willingly mentions internally.

Why did it die back then? Three things came together.

In 2017, Intel held over 90% of the server CPU market. AMD's EPYC hadn't shipped yet. Arm’s share in the cloud was barely 1%, not even qualifying as an “experiment.” The cloud providers’ attitude, translated, was: It looks okay, but why should we switch?

More critically, from launch to cancellation, Centriq never had a single publicly known mass-production customer. Microsoft Azure was reportedly testing it but never made an official announcement. For a data center chip, having no committed mass-production customer means having no product. This industry doesn’t ask for “we made a chip”; it demands “someone is running production workloads on it.”

Most fatally, Qualcomm was on the verge of falling apart itself during those two years – fighting a grueling global patent war with Apple, while Broadcom launched a $100 billion acquisition bid. Every morning, the question management faced wasn’t “How do we sell server chips?” but “Will the company still be ours tomorrow?”

Centriq died at the intersection of these three things: the timing wasn't right, the customers didn't come, and there was a fire at home.

This Time, the World Has Flipped Itself

Nine years is enough for the world to change completely.

Arm’s market share in data centers has gone from 1% to 25%. Graviton was the watershed moment – Amazon started building its own Arm server CPUs in 2018, and for the last three consecutive years, they've accounted for over half of new CPU capacity on AWS. 98% of AWS's top 1,000 customers use them. NVIDIA's Grace CPU has also reached a revenue scale in the cloud comparable to its GPUs.

Translated into Qualcomm’s language: In 2017, it faced two questions – “Can Arm do servers?” and “Can Qualcomm make Arm server chips?” Now, Graviton and Grace have already answered the first question. Qualcomm only needs to prove it can do it better than Graviton. The difficulty is not on the same level.

There’s another change the financial reports can’t articulate, but Mark Zuckerberg’s video made it very clear.

Intel and AMD have dominated server CPUs for over thirty years. Never has any cloud company CEO come out and said, “I will sign a multi-generational strategic contract with so-and-so.” It’s not that there was no technical cooperation; it’s that no one was willing to put their chips on the table publicly.

Meta did it this time.

The reason behind it is so straightforward it doesn’t need analysis: Intel’s server CPU market share has fallen from absolute monopoly to 62%, while AMD has taken 46% of x86 revenue. On the Arm side, Graviton, Grace, and Ampere are lining up to tear down the walls. Any cloud provider’s procurement department looking at this situation would be negligent to keep all their eggs in the x86 basket.

So Meta didn’t sign a “let’s try it out” deal. It was “multi-generational.” Translated: If the first generation doesn’t work, try the second.

In the data center chip market, no one has ever gotten this kind of treatment.

The CPU Is a Diversion; There Are Five Battles

So far, the story looks like a beautiful comeback script.

But what Qualcomm is truly doing is much bigger than a single CPU. The Dragonfly family Amon showcased is essentially the entire skeleton of a data center.

There’s an Arm server CPU with over 250 cores clocked above 5GHz, named C1000, set for mass production at Meta by mid-2028. There are AI inference accelerators, called AI200, AI250, and AI300, with annual updates starting next year, specifically for running real-world calls after models are deployed. There’s a high-bandwidth computing platform called HBC, claiming 4 to 8 times more throughput at the same cost, with samples delivered to Microsoft by mid-next year.

All these pieces need to be connected – Qualcomm laid out a networking scheme spanning electrical and optical interconnects to keep data flowing smoothly between chips.

On top of all that, there’s a software layer. Qualcomm spent $3.9 billion to acquire a company called Modular, aiming to build a platform where developers write code once and run it on any hardware of their choice. Modular’s CEO is Chris Lattner, who created the LLVM compiler, co-created the Swift language, and later managed Tesla’s Autopilot.

It all sounds plausible. But doing five things simultaneously – CPU, accelerator, HBC, networking, software – means Qualcomm isn’t just building a chip; it’s constructing a building. Foundation, walls, plumbing, wiring, and decoration each could be a separate project, but if any one of them goes unfinished, the entire building is at risk of collapse.

Historically, only one company has successfully fought on more than two fronts simultaneously in the data center – NVIDIA: the GPU itself, plus Mellanox for networking, plus the CUDA software ecosystem.

Intel’s track record serves as a cautionary tale. Over the past decade, it used Atom to target mobile chips, abandoning it in 2016. It used Rialto Bridge to target GPUs, canceling it in 2023 and not proceeding with its successor, Falcon Shores. It used Omni-Path for network interconnects, shutting it down in 2019. Billions of dollars burned, and some of the industry’s best engineers hired without a single successful breakthrough.

The Nuvia team is unlikely to botch designing a high-performance Arm CPU – it's the former Apple team with the M1 pedigree. The risk isn’t in the design. The risk is that Qualcomm has never operated five independent chip product lines simultaneously. Phone SoCs, automotive cockpits, PC chips, IoT, XR – these are essentially five variants based on one core design. The data center is starting from scratch, from supply chain to sales to customer support – everything must be built anew.

The Top Floor Might Never Get Built

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

NVIDIA’s CUDA is a structure that took fifteen years to grow. It’s not just a tool or a platform – it’s a system where millions of developers have made their home. Tens of thousands of papers, thousands of optimized libraries, every underlying pathway from PyTorch to TensorFlow has grown out of it.

Modular wants to build a parallel system with $3.9 billion and two years.

The story sounds compelling: the MAX platform lets you write once and run on CPUs, GPUs, NPUs, and custom ASICs. The Mojo language looks like Python but performs like a compiled language.

But AMD told a nearly identical story once. ROCm: open, compatible, an alternative to CUDA. It’s been ten years, and users still have to comb through compatibility lists page by page to see which applications are supported. Intel’s oneAPI is the same.

And Modular has a contradiction it might not have fully thought through. If MAX treats all chips equally, why should Qualcomm build its own chips? Why not just be a software company? If MAX only runs best on Qualcomm chips, then it’s no different from AMD’s ROCm – a software layer claiming to be open but effectively tied to its own hardware.

This isn’t a technical problem. CUDA’s wall wasn’t built by technology; it was built by fifteen years of time and the habits of millions of people. Modular doesn’t lack technology; it lacks time. And how long are Qualcomm’s investors willing to wait?

The One Who Should Be Most Nervous Is Intel

The entity most directly threatened by Qualcomm’s latest move is Intel.

An Arm CPU directly targets x86 server CPUs. That is Intel’s last high-margin stronghold – In Q1 2026, AMD had already taken 46% of x86 server revenue, and Intel's shipment share was down to 62%, with a downward trend. Qualcomm entering the fray means Intel isn’t just fighting AMD for the remaining x86 market share; it’s being fired upon from the other side by the Arm camp.

Qualcomm’s valuation logic is also shifting. A PE ratio of 15-18x is the price of a phone company. If the market accepts the “AI infrastructure company” narrative, aligning with AMD and NVIDIA’s 25-30x range, there’s room for a significant revaluation of Qualcomm’s stock.

But this is predicated on everything going smoothly.

Watch Three Milestones

The trickiest part of Qualcomm’s story is that the money has been talked about, but the products haven't been delivered yet.

The first major check-in point is early next year. Amon said the custom chip business would start generating substantial revenue from FY2027 Q1. That quarter’s numbers will be the market’s first report card. Above $1.2 billion, and the narrative holds. Below $800 million, and everything gets discounted.

The second is mid-next year, when commercial samples of HBC are delivered to Microsoft. Until that point, Qualcomm has been saying “we can do it” for two years. That’s the day they show “we have done it” for the first time.

The final milestone is mid-2028, when the C1000 goes live at Meta. Until that point, every chapter of this story is marked “To be continued.”

Qualcomm is betting on a three-year cycle. If they stumble at any single point, the market will reprice them.

The intraday action on June 25th told the same story – after hours, the stock jumped from $197 to $223, then gave some back. Wall Street’s message is clear: You told a good script. Now show us the movie.