Qualcomm spent eight years staying out of the data center. That ended on June 24, 2026, at Qualcomm’s Investor Day in New York, when the company unveiled the Dragonfly C1000, its first server CPU since it abandoned the Centriq project in 2018, alongside a new AI inference accelerator called AI300. The headline customer is Meta, which signed on to put the chip into its next-generation server fleet under what Qualcomm called a “multi-generation collaboration,” with Meta’s first C1000 deployment scheduled to enter production in the second half of 2028. Microsoft is on board too, and Qualcomm says it is building custom silicon for two more hyperscalers it has not named yet.
The pitch is straightforward. Cloud operators are trying to cut the power bill on AI infrastructure, and Qualcomm is betting that the low-power engineering it perfected in smartphones translates to the server rack. Whether that bet pays off will not be clear for years, since Dragonfly C1000 does not reach production until the second half of 2028 – though Qualcomm has already disclosed that the chip uses a chiplet design built around more than 250 custom Oryon cores clocked above 5 GHz, which the company says deliver more than 2x the performance per watt of competing server CPUs. But the numbers behind the announcement, and the skepticism it has already drawn from parts of Wall Street, tell a more immediate story about how crowded the data center CPU market has become in 2026.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Dragonfly C1000 at a Glance
The story spans a June announcement and a 2028 production date, so here is the core of it in one place as of August 2026:
- Announced: June 24, 2026, at Qualcomm’s Investor Day in New York
- First named customer: Meta, for its next-generation server fleet
- Production start: second half of 2028
- Data center revenue target: more than $15 billion annually by fiscal 2029
- Also announced: the AI300 rack-scale inference accelerator
- Other customers: Microsoft confirmed; two additional hyperscalers in custom-silicon talks remain unnamed
What Qualcomm Actually Announced: Dragonfly C1000 and AI300
Qualcomm’s June 24 Investor Day in New York laid out a full data center roadmap rather than a single product. The Dragonfly C1000 is the company’s first general-purpose data center CPU, positioned to handle the scale-out compute work that hyperscalers run alongside AI training and inference; per Futunn’s June 2026 reporting, Qualcomm has said Dragonfly C1000 racks will support up to 43 TB of DRAM and PCIe Gen 7 connectivity, with engineering samples already reaching partners in fiscal 2026, well ahead of the 2028 production date. StorageReview’s June 25, 2026 breakdown of the launch, echoed by a Constellation Research note the same month, detailed that Dragonfly C1000 is planned to ship in three distinct CPU configurations – an orchestration-focused variant for agentic AI workloads, a general-purpose variant, and a dedicated AI head-node design – giving Qualcomm more than a single SKU to sell against Intel, AMD, and AWS once production begins. Sitting next to it is the Dragonfly AI300, a rack-scale AI inference platform that follows the AI200 and AI250 accelerators Qualcomm announced in October 2025.
Dragonfly AI300 integrates what Qualcomm calls HBC Gen 2 memory and is aimed squarely at inference workloads: large language models, multimodal models, and agentic AI pipelines. Qualcomm reiterated at its June 24, 2026 Investor Day that the accelerator delivers 4x to 8x higher memory bandwidth per watt per card than existing GPU-based architectures, though the company still has not published a full spec sheet for either chip. Commercial sampling for the platform is expected in 2028, the same window Qualcomm has set for Dragonfly C1000 production and for the general-purpose and AI head-node variants StorageReview described in its June 25, 2026 coverage of the announcement; a follow-up StorageReview report in July 2026 reaffirmed that C1000’s commercial availability remains targeted for 2028, with no acceleration to that timeline disclosed since the launch.
Qualcomm CEO Cristiano Amon framed the announcement as the start of a long-term relationship rather than a one-off product launch. “Our new strategic, multi-generation collaboration with Meta will bring the Qualcomm Dragonfly C1000 CPU to their next-generation server fleet, with production starting in the second half of 2028,” Amon said in Qualcomm’s announcement.
Inside the Meta Deal
Meta is the first named hyperscaler customer for a Qualcomm server CPU, and that alone makes the deal significant. Qualcomm’s own announcement described the arrangement in practical terms: the Dragonfly C1000 is planned to power Meta’s next-generation server fleet, with the chip entering production in the second half of 2028. CNBC reported the same day, June 24, 2026, that Meta’s commitment is structured as a multi-year, multi-generation agreement rather than a one-time chip order, meaning Meta is signing on for future Dragonfly generations as well once the first units start shipping in 2028.
In its official statement, Qualcomm Technologies described the deal this way: “Qualcomm Technologies’ data center CPU, the Qualcomm Dragonfly C1000, is planned to power Meta’s next-generation server fleet, underscoring the growing importance of high-performance, power-efficient compute in large-scale, scale-out environments.”
No public dollar figure has been attached to the Meta agreement, and its value and chip volumes had not been disclosed by October 6, 2026. What Qualcomm did disclose is a power-efficiency argument: “We engineered our data center CPU to achieve leading performance per core, along with a significant breakthrough in power efficiency for large-scale data center applications,” the company said. That framing matters because power, not just raw compute, has become the binding constraint on new data center capacity across the industry in 2026.
Microsoft and the Hunt for More Hyperscalers
Meta was not the only name Qualcomm attached to the June 24 rollout. The company also said Microsoft would use its new AI chips, and that it was developing custom silicon for two additional hyperscalers it declined to identify. By late July, reporting indicated Qualcomm had signed deals with three major hyperscale customers in total, a sign the June announcement was the opening move in a broader sales push rather than an isolated win. As of August 2026, those two additional hyperscalers remain unnamed publicly.
Qualcomm has put real numbers behind that push. Reporting on the company’s data center strategy indicates Qualcomm expects its hyperscaler customers to generate at least $1 billion in combined revenue within a year of signing, and the company has guided to more than $15 billion in annual data center revenue by fiscal 2029. Those are aggressive targets for a business segment that barely existed at Qualcomm a year earlier, and they only work if the unnamed hyperscalers convert from pilot deals into full production orders.
From Centriq to Nuvia: Qualcomm’s Long Road Back to the Data Center
Qualcomm has tried this before. The company entered the server CPU market with Centriq, then walked away from it in 2018 to refocus on its core smartphone business. That retreat looked final at the time. Intel and AMD controlled the x86 server market, and Arm-based server chips were still a niche bet outside of a handful of hyperscaler experiments.
The Nuvia Bet
Qualcomm’s real re-entry point came in January 2021, when it agreed to acquire Nuvia for $1.4 billion, a deal that closed that March. Nuvia was a two-year-old startup founded by three former Apple chip engineers, and Forbes framed the acquisition at the time as Qualcomm’s move to take Apple, Intel, and AMD head on. For years afterward, the Nuvia team’s custom CPU cores showed up mainly in Qualcomm’s Snapdragon laptop and phone chips, not in servers. Dragonfly C1000 is the first product to point that engineering back at the data center, and it took more than five years from the Nuvia deal to land a named hyperscaler customer.
The $15 Billion Question: Qualcomm’s Data Center Revenue Bet
Qualcomm’s fiscal 2029 target of more than $15 billion in annual data center revenue is the number that will define whether this comeback counts as a success. For context, that figure would need to rank Qualcomm alongside established server silicon suppliers within roughly three years of Dragonfly C1000 entering production, not from when it starts shipping in volume.
The near-term marker is more modest and more telling. Qualcomm’s own guidance suggests its first hyperscaler deals could generate a combined $1 billion within a year, a figure built mostly on AI accelerator sales rather than CPU shipments, since Dragonfly C1000 will not exist in production form until 2028. That gap between what Qualcomm can sell today (AI300-class accelerators) and what it is promising for the future (a full CPU platform) is the crux of the skepticism the announcement has drawn.
How Dragonfly C1000 Stacks Up Against the Field
Qualcomm is not entering an empty market. Nvidia, AMD, Intel, and Amazon all ship data center CPUs today, and each has years of production experience and hyperscaler relationships that Dragonfly C1000 does not yet have. Qualcomm has since disclosed more of the chip’s specs than it offered at launch: Dragonfly C1000 uses a multi-chiplet design built around Qualcomm’s Oryon cores, packing more than 250 cores clocked above 5 GHz, with PCIe Gen 7 interconnects rated at more than 2 TB/s of bandwidth and CXL support, and Qualcomm says the design delivers more than 2x the performance per watt of current server CPUs, per Counterpoint Research’s June 29, 2026 breakdown of the chip and StorageReview’s contemporaneous spec analysis. Qualcomm also disclosed that Dragonfly C1000 rack designs comply with the OCP ORv3 (Open Rack V3) standard, a detail meant to reassure hyperscalers that the chip can slot into existing Open Compute Project infrastructure rather than demanding bespoke rack engineering. A process node still has not been published, so a full spec comparison remains incomplete, but the core count, chiplet architecture, OCP rack compliance, and power-efficiency claims give the rest of the field its first real benchmark to react to.
| Chip | Maker | Max Cores | Process Node | Status (as of August 2026) |
|---|---|---|---|---|
| Dragonfly C1000 | Qualcomm | 250+ (custom Oryon) | Not yet disclosed | Announced June 2026, production 2H 2028 |
| Grace | Nvidia | 72 (Arm Neoverse V2) | N/A | Shipping since 2023, 2.5M+ units shipped |
| EPYC Turin (9005 series) | AMD | 192 | 4nm compute dies | Shipping since Oct. 10, 2024 |
| Xeon 6 (Granite Rapids) | Intel | 128 P-cores | Intel 3 | Shipping since 2024 |
| Graviton5 | AWS | 192 | 3nm | Announced Dec. 2025, M9g/M9gd GA June 10, 2026 |
AWS is the clearest proof that a non-traditional CPU vendor can win at hyperscale. Amazon has said Graviton chips have accounted for more than half of the new CPU capacity added to AWS for three consecutive years, with 98% of its top 1,000 EC2 customers having run workloads on Graviton in production. That track record is exactly what Qualcomm needs to build with Meta and Microsoft, and it took AWS the better part of a decade to get there.
AI300: Taking Aim at Nvidia’s Inference Business
The CPU is the long game. AI300 is where Qualcomm is trying to compete sooner, and it is arriving in a market Nvidia still dominates. Qualcomm’s pitch for AI300 is not about beating Nvidia on raw throughput. It is about cost per query on inference workloads, where memory bandwidth per watt determines how many requests a rack can serve before it hits a power ceiling.
That is also why Qualcomm’s data center chips are designed to work alongside Nvidia hardware rather than only against it. Qualcomm has said its data center CPUs will support Nvidia’s NVLink Fusion interconnect, which would let Dragonfly C1000 pair with Nvidia GPUs in the same rack instead of forcing hyperscalers to choose one vendor’s full stack. That is closer to the role Nvidia’s own Grace CPU already plays inside Nvidia’s GB300 NVL72 systems, where 36 Grace CPUs sit alongside Blackwell-generation GPUs in a single rack, than to a head-on CPU-versus-GPU fight.
Wall Street’s Mixed Verdict
The market’s reaction to Qualcomm’s data center push has not settled into a clean narrative. Some coverage framed the Meta and Microsoft wins as validation, describing Qualcomm stock gaining ground on the strength of the hyperscaler announcements. Other market summaries pointed to short-term weakness tied to the same news cycle, since Computex coverage of Nvidia’s own roadmap ran at the same time and pulled attention away from Qualcomm’s rollout.
The more pointed skepticism came a month later. On July 30, 2026, The Register published a piece headlined “Qualcomm won’t be a big datacenter player anytime soon”, capturing a view that has followed the company since the announcement: a 2028 production date is a long time to ask investors and customers to wait, especially against rivals shipping chips today.
Why Some Analysts Remain Skeptical
The skeptics’ case rests on timing and history. Qualcomm already tried and failed once in this market. Centriq had real engineering behind it and still could not build enough hyperscaler demand to survive past 2018. Dragonfly C1000 has a better starting position, with a named customer in Meta before the chip even exists in silicon, but it is also asking the market to wait two more years for proof.
The 2028 Problem
Every competitor in the table above will ship at least one more CPU generation before Dragonfly C1000 reaches production. AMD and Intel both run roughly two-year release cadences for EPYC and Xeon, and AWS has already moved from Graviton4 to Graviton5 inside a single year. Qualcomm is not just competing against today’s Grace, Turin, and Xeon 6. It is competing against whatever those companies ship in 2027 and 2028, chips that do not exist yet either. That is the standard every new entrant to this market has to clear, and it is a higher bar than the one Nuvia’s founders faced when Qualcomm bought them in 2021.
Beyond Servers: Qualcomm’s Broader Silicon Push
The data center announcement is part of a wider pattern at Qualcomm in 2026. The company has also been named BMW Group’s lead compute silicon provider, extending a diversification strategy that now spans automotive compute, PC chips, and data center silicon, on top of the smartphone business that still generates most of its revenue. The common thread across all of it is the same argument Qualcomm is making with Dragonfly C1000: that the low-power design skills built for battery-powered devices are worth more outside phones than they were a decade ago, as power has become the limiting factor in everything from electric vehicles to AI server racks.
The $40 Billion Non-Handset Target
Qualcomm has now attached a specific number to that diversification push: Barron’s and Yahoo Finance reported on June 25, 2026 that Qualcomm raised its fiscal 2029 non-handset revenue target to more than $40 billion on the strength of the Dragonfly C1000 launch, a target that rolls up automotive, PC, and data center silicon into one goal. The data center business’s $15 billion target is the largest single piece of that total, which makes Dragonfly C1000 and AI300 more than a side bet on the server market. They are the anchor of Qualcomm’s broader case, as of August 2026, that it can stop being defined almost entirely by smartphone chip revenue.
What This Means for the AI Infrastructure Buildout
Qualcomm’s entry adds a fourth serious CPU option to a market that, for most of the last decade, was really a two-vendor contest between Intel and AMD. Arm-based alternatives have already changed that math. AWS builds its own Graviton chips, Nvidia sells Grace as part of its GB300 racks, and Arm itself has said close to half of the compute shipped to top hyperscalers in 2025 would be Arm-based rather than x86. Qualcomm’s arrival, backed by named commitments from Meta and Microsoft, is another data point in that same shift.
For Intel and AMD, the immediate threat is not Dragonfly C1000 itself, since it will not ship for two more years. It is what the deal signals about hyperscaler buying behavior. Meta and Microsoft are both willing to commit engineering resources to a fourth or fifth CPU architecture years before it exists in silicon, purely on the promise of better performance per watt. That is a negotiating lever hyperscalers can now use against every incumbent supplier, whether or not Qualcomm ever ships a single Dragonfly C1000 unit.
| Date | Milestone |
|---|---|
| 2018 | Qualcomm exits the data center market and discontinues the Centriq server CPU line |
| March 2021 | Qualcomm completes its $1.4 billion acquisition of Nuvia, a server-chip startup founded by former Apple engineers |
| October 2025 | Qualcomm announces AI200 and AI250, its first rack-scale AI inference accelerators |
| June 24, 2026 | Qualcomm unveils Dragonfly C1000 CPU and AI300 accelerator at its Investor Day in New York, names Meta and Microsoft as customers |
| June 24, 2026 | Qualcomm announces its Dragonfly data-center portfolio, including a multigeneration CPU collaboration with Meta, and targets more than $15 billion in data-center revenue in fiscal 2029 |
| 2H 2028 | Dragonfly C1000 targeted for production, deploying into Meta’s next-generation server fleet |
5 Predictions for Qualcomm’s Data Center Comeback
- At least one more hyperscaler gets named before Dragonfly C1000 ships. Qualcomm has strong incentive to convert its two unnamed custom-silicon customers into public references well before 2028, since the whole pitch depends on visible momentum.
- Meta’s early deployment will be a fraction of its total server fleet. First-generation server silicon from a new vendor almost never launches at full scale. Expect a targeted rollout in specific workloads before any broad fleet-wide commitment.
- Dragonfly C1000 will ship paired with Nvidia GPUs more often than it ships alone. The NVLink Fusion support points toward a Grace-style role: complementary compute next to Nvidia’s accelerators rather than a standalone rack architecture competing directly with GPU-centric designs.
- AMD and Intel respond on power efficiency, not just price. Both companies have entrenched hyperscaler relationships and multi-generation roadmaps already in motion. The likely response is louder marketing around performance-per-watt in their next EPYC and Xeon launches rather than any single dramatic move.
- The fiscal 2029 revenue target gets revised before Dragonfly C1000 ships. Two and a half years is a long runway for a first-generation product in a market this competitive, and Qualcomm’s own history with Centriq shows how quickly data center plans can change.
Frequently Asked Questions
What is the Qualcomm Dragonfly C1000?
It is Qualcomm’s first data center CPU since the company discontinued its Centriq server chip in 2018. Qualcomm unveiled it on June 24, 2026, at its Investor Day in New York, alongside the AI300 inference accelerator, with Meta confirmed as the first customer for its next-generation server fleet.
When will Qualcomm’s data center CPU actually ship?
Qualcomm has targeted production for the second half of 2028, meaning Dragonfly C1000 will not be generally available for roughly two more years after its June 2026 announcement.
Why did Meta choose Qualcomm for its servers?
Qualcomm has pointed to power efficiency and performance per core as the core pitch, arguing its chip delivers a “significant breakthrough in power efficiency for large-scale data center applications.” Meta has not published its own reasoning beyond confirming the multi-generation agreement.
Did Qualcomm try building data center chips before?
Yes. Qualcomm built the Centriq server CPU line and exited the data center market in 2018 to refocus on smartphones. It re-entered the space in 2021 by acquiring the server-chip startup Nuvia.
How much did Qualcomm pay for Nuvia?
Qualcomm agreed to acquire Nuvia for $1.4 billion, announcing the deal in January 2021 and closing it that March. Nuvia was founded by three former Apple chip engineers roughly two years before the acquisition.
What is Qualcomm’s non-handset revenue target?
Qualcomm is targeting more than $40 billion in non-handset revenue by fiscal 2029, combining automotive, PC, and data center silicon. The data center segment’s $15 billion goal, tied directly to Dragonfly C1000 and AI300, is the largest single contributor to that total.
What is the Qualcomm AI300?
Dragonfly AI300 is Qualcomm’s third-generation rack-scale AI inference platform, following AI200 and AI250. It targets large language model, multimodal, and agentic AI inference workloads, with Qualcomm claiming 4x to 8x higher memory bandwidth per watt than existing GPU-based architectures. It launched alongside the Dragonfly C1000 CPU, which StorageReview and Constellation Research both reported will ship in three separate configurations – orchestration/agentic AI, general-purpose, and AI head-node – once production starts in 2028, with rack designs built to comply with the OCP ORv3 standard.
Is Qualcomm’s data center chip Arm-based or x86?
Qualcomm describes Dragonfly C1000 as using custom Qualcomm Oryon CPU cores in a chiplet architecture with more than 250 cores and target frequencies above 5 GHz. Given Qualcomm’s Snapdragon lineage and the Nuvia team’s Arm-based CPU core design background, an Arm-based architecture is the expected direction, though Qualcomm has not confirmed this in its public statements.
How does Dragonfly C1000 compare to Nvidia Grace?
A direct comparison is not possible yet, since Qualcomm has not disclosed core counts or process node details for Dragonfly C1000. Nvidia’s Grace, by contrast, is already shipping: a 72-core Arm Neoverse V2 design with more than 2.5 million units shipped since its 2023 launch.
For more on the AI chip race, see our AI chips and data center hardware hub.
