One CPU, Five Wars: After Nine Years, Qualcomm Bets Again

June 24, New York. Qualcomm CEO Cristiano Amon announced one thing: by fiscal year 2029, non-mobile business must reach $40 billion, with $15 billion from data center chips. The stock jumped 13% after hours.

Then Zuckerberg appeared on the screen.

"Qualcomm will be Meta’s data center CPU supplier." He recorded a video—this wasn’t the generic “we’re excited to explore cooperation opportunities” PR talk, but a full subject-verb-object commitment. The contract covers multiple product generations.

Sounds like a comeback story. Eighteen years ago Snapdragon defined mobile phones, now Qualcomm wants to define the data center.

But no one is more aware than Qualcomm internally—it also tried nine years ago, with Centriq. That round completely flopped.

Built on the Ruins of Nine Years Ago

November 2017, San Francisco. Qualcomm released the Centriq 2400, the world’s first 10nm server chip, Arm architecture, 48 Falkor cores. The slide deck looked a lot like today’s: high performance, low power, break Intel’s monopoly.

Within two years, the project was gone. Execs left, team was dissolved. Centriq became a name no one at Qualcomm would mention.

Why did it die? Three things collided together.

In 2017, Intel had over 90% server CPU market share. AMD’s EPYC had not shipped. Arm’s public cloud share was just 1%, not even "experimental". Cloud vendors’ attitude translated as—looks OK, but why change?

The more deadly thing: from launch to cancellation, Centriq had not a single public volume customer. Word was Microsoft Azure was testing, but never officially confirmed. For data center chips, no customer commitment means no product. What this industry wants is not “the chip is built” but “someone is running production workloads on it.”

The most fatal: Qualcomm itself was on the verge of falling apart—its global patent war with Apple was intense, then Broadcom showed up with a hundred-billion-dollar acquisition bid. Management woke up every day not thinking “how to sell server chips” but “will we still own the company tomorrow.”

Centriq died at the intersection of these three: the timing wasn’t right, no customers, and chaos at home.

This Time, the World Has Shifted

Nine years is enough for the world to change.

Arm’s share in data centers rose from 1% to 25%. Graviton was the turning point—Amazon started building Arm server CPUs in 2018, and by last year, it accounted for over half of AWS’s new CPU capacity for three consecutive years. 98% of AWS’s top 1,000 customers are using it. Nvidia’s Grace CPU also has similar cloud revenue scale as its GPUs.

Translated to Qualcomm’s perspective: in 2017 there were two questions—“Can Arm make server chips?” and “Can Qualcomm make Arm server chips?” The first is now answered by Graviton and Grace. All that’s left is to prove Qualcomm can do it better than Graviton. The difficulty is on a different scale.

There’s also a change numbers can’t capture, but Zuckerberg’s video made very clear.

Intel and AMD have ruled server CPUs for over 30 years, and no public cloud CEO has ever come out to say, “I will sign a multi-generation strategic contract with someone.” Tech collaborations existed, but no one wanted to publicly put all their chips on one table.

This time Meta did.

The reason is so plain, it needs no analysis: Intel’s server CPU share has dropped from absolute domination to 62%. AMD has taken 46% of x86 server CPU revenues. On the Arm side, Graviton, Grace, Ampere are all lining up. Any cloud procurement team seeing this, and continuing to put all their eggs in the x86 basket, would be negligent.

That’s why Meta signed not for “trial,” but “multiple generations.” In translation: if the first gen doesn’t work, try the second.

No one else has ever gotten these terms for data center chips.

The CPU Is Just a Cover—There Are Five Wars

At this point, it sounds like a perfect comeback script.

But what Qualcomm really needs to do is much bigger than just a CPU. What Amon showed with the Dragonfly family is actually a whole data center framework.

There’s a 5GHz+, 250-core Arm server CPU called C1000, to be mass-produced for Meta in mid-2028. There are AI inference accelerators—the AI200, AI250, AI300—starting annual upgrades next year, focused on post-deployment model inference. There’s a high-bandwidth computing platform called HBC, claiming 4–8x throughput increase at similar cost, and the first samples will be sent to Microsoft mid-next year.

These all have to be connected together—Qualcomm built a full-stack network solution from electrical interconnect to optical interconnect, so data flows smoothly between chips.

And on top, there’s the software. Qualcomm spent $3.9 billion to acquire Modular, seeking to build a platform where developers can write code once and run it on any hardware. Modular’s CEO is Chris Lattner—creator of the LLVM compiler, Swift language, and previously on Tesla’s Autopilot program.

All these things sound reasonable. But doing five things at once—CPU, accelerator, HBC, network, software—means Qualcomm isn’t just making a chip, it’s constructing a whole building. The foundation, walls, pipes, wires, finishing—each could be its own project, but if any one is incomplete, the whole building is at risk.