At Snapdragon Summit, Qualcomm’s annual product event held in Maui from September 22 to 24, the company launched its next two Snapdragon 8 flagships for Android phones, the Snapdragon 8 Elite Gen 6 and 8 Elite Extreme Gen 6, then a second day of personal AI, PC, and software news. Google, Mastercard, Xiaomi, Motorola, and Honor took the Day 1 stage; Microsoft, Meta, HUMAIN, and Liquid AI featured on Day 2.
Qualcomm now sells compute at every point from the smallest IoT device to the datacenter. At the low end, Snapdragon and Dragonwing parts run earbuds, watches, glasses, and sensors at milliwatts. In the middle, Snapdragon powers hundreds of millions of phones a year, more than 170 PC designs, and cars that produced a record $1.59 billion in automotive revenue last quarter, up 61%. At the top, the Dragonfly portfolio announced in June reaches the rack. When I sat down with CEO Cristiano Amon and Qualcomm exec (and Modular CEO) Chris Lattner in August, Amon put the range at under 2 milliwatts to 600 kilowatts.
Qualcomm’s Datacenter Changed Most in 90 Days
The datacenter is the part of that span that changed most in 90 days. Dragonfly spans the AI200 and AI250 accelerators, High Bandwidth Compute (the near-memory architecture Qualcomm shortens to HBC), the C1000 server CPU on Oryon cores, custom silicon, and Alphawave connectivity. Meta signed a multi-generation collaboration under which the C1000 is planned to power Meta’s next-generation server fleet, Satya Nadella endorsed HBC at Investor Day, HUMAIN is targeting 200 megawatts of AI200 and AI250 racks, and two unnamed hyperscalers are each expected at above $1 billion in custom-silicon revenue in fiscal 2027. In September, Amazon signed a multi-generation agreement for customized inference silicon and optical connectivity. Qualcomm’s targets are $5 billion in datacenter revenue in fiscal 2027, and $15 billion in fiscal 2029.
Why Qualcomm’s Two Ends Now Feed Each Other
That business arrives as the AI buildout reshapes the device markets below it. Datacenters now face the power limits that phones have always lived with, and their memory appetite pushed DRAM and NAND prices up more than 300%; IDC expects 2026 smartphone shipments to fall 16.7% while average prices rise 27.6% to $581. The strategic news at Snapdragon Summit sat underneath the product news: the two ends of Qualcomm’s range started to feed each other, with Oryon, the NPU, and low-power memory expertise going up into Dragonfly, and HBC, built for the datacenter, coming back down into Snapdragon. I wrote in July that the open question was execution timing, not credibility, and in August that AI infrastructure faces three constraints that more accelerators alone can’t fix: power, memory, and software portability. What remains to prove is independent validation of HBC’s economics.
The Agent Becomes the Phone’s Second User
Amon’s Day 1 keynote made one claim I think matters more than any spec. The phone isn’t going away, he said, and apps and agents will coexist, but the phone now has two users: the person holding it and an agent that will use the phone on that person’s behalf even when the person isn’t. My read is that this decouples compute demand from screen time and raises the value of silicon per device even as units fall. As I said on Yahoo Finance in July, the unit decline is real, but Qualcomm’s premium Android position has held.
Both chips are 2nm, and Qualcomm says it has the first mobile CPU to reach 5GHz, 30-billion-parameter models served from flash storage, and, on the Extreme, 50% more NPU shared memory. Qualcomm’s late-August design previews described all three engines as built to cut system memory access, the right priority when DRAM is scarce. Qualcomm’s gains are measured against last year’s Gen 5. My colleague Anshel Sag, who was in Maui, benchmarked the Extreme on reference devices and found that the GPU and multi-core gains hold up.
HBC and Oryon Now Flow Both Ways
Technology used to flow one way at Qualcomm, out of the phone. Summit was the first time I saw it flow back. Amon said that HBC now has a Snapdragon co-processor version, with details due at Mobile World Congress, and framed the reason in a way datacenter operators will recognize: compute outrunning available power has been the phone’s constraint from day one.
Oryon is making a round trip. Nuvia designed the core for servers, Qualcomm shipped the core first in PCs and phones, and the C1000 is expected to bring the design back to servers in 2028, with cores designed for more than 5 GHz. The NPU goes the other way; Lattner said the Dragonfly accelerators build on Qualcomm’s scalable NPU platform. Modular is bringing the stack to Snapdragon, and Lattner said the work has just started. Mojo is open source under Apache 2.0. I called the acquisition a platform play for Qualcomm’s own silicon and wrote after ModCon that openness and portability were the point.
Customers Now Buy from Both Ends of Qualcomm’s Range
The loop shows up in the customer list. Microsoft launched the Surface Pro 12-inch and Surface Laptop 13-inch on Snapdragon X2 Plus at Summit. Meta has shipped Snapdragon in its glasses since Ray-Ban Stories in 2021 and is the first named C1000 customer. HUMAIN is deploying AI200 and AI250 racks, unveiled the Horizon Ultra AI PC at LEAP in August on Snapdragon X2 Elite, and showed a Linux-based HUMAIN OS version at Summit for the first half of 2027. Amazon is datacenter only, but the September agreement is the strongest validation so far: AWS has offered EC2 instances on Qualcomm’s AI 100 since 2023, and I said in July that one of the two unnamed custom-silicon wins was probably Amazon. The agreement fits, though Qualcomm hasn’t said so.
What the Datacenter Ramp Means for Investors
Management expects non-handset revenue to pass half of the chip business in fiscal 2027 on the way to $40 billion by fiscal 2029. Fiscal 2027 is mostly custom silicon, plus Alphawave connectivity and AI200 deployments; the two custom-silicon hyperscalers are neither Microsoft nor Meta, with revenue starting in the December quarter. The loop products, HBC accelerators and the C1000, land in 2027 and 2028, so the loop underwrites the back half of the targets. Custom silicon dilutes gross margin in QCT, Qualcomm’s chip segment, by 1.5 to 2 points, though management says custom silicon is accretive at the operating line. And the Amazon warrant’s $60 billion is a ceiling on what counts toward vesting, not an order.
Why Partners Should Consider Building More with Qualcomm
Oryon, the NPU, HBC, and Modular’s software are being extended across devices and datacenters, so work a partner does in one category carries into the next. The software is mostly open: Mojo’s Apache 2.0 license means a developer isn’t betting the codebase on one chip, although MAX, as my colleague Matt Kimball wrote after ModCon, stays under the Modular Community License, source-available rather than open-source, with the cluster layer closed
New categories are getting faster to enter: Snapdragon START, launched at AWE in June, took Inspecs from kickoff to a working device in four months. The PC ecosystem is past the app question, with more than 7,300 apps running on Windows on Snapdragon across 170-plus designs, and Googlebook, on Snapdragon X Elite, and Linux on the X2 series opening new channels. And flash-served models and HBC on devices are Qualcomm’s answer to the memory squeeze, more AI per gigabyte of DRAM, provided OEMs validate the claims on their own designs first.
What I Would Like to See Qualcomm Do Next
Across the table from Amon and the business unit leaders, I’d push four things: publish independent HBC economics before mid-2027 sampling, since the cost-per-token claim is the whole bet; get Microsoft to put a date and a scale on Azure; move MAX to the same license as Mojo or explain why the cluster layer stays closed; and tell OEMs plainly which designs the Extreme is for and which the Gen 6 is for, so the two-chip split reads as strategy rather than creating potential confusion.
One Roadmap Makes Qualcomm’s 2029 Targets More Credible
Qualcomm’s device and datacenter businesses now share one technology roadmap. For partners, that means one supplier whose roadmap runs from wearables to the rack, which they can build on without giving up control of their own products. For investors, it makes the fiscal 2028 and 2029 targets more credible, while fiscal 2027 depends mainly on custom silicon. I wrote in July that the question was timing, not credibility, and I’ll hold to that. The proof points are dated above, plus an independent benchmark of HBC’s economics.
