I spent three days in New York for Qualcomm’s 2026 Investor Day, then walked off the floor and straight into a Yahoo Finance studio. I knew the reflexive take would be that Qualcomm is showing up late to a datacenter market NVIDIA owns and they tried it before. I don’t buy it. What I saw was a company that arrived with named Tier-1 customers and a defensible read on where AI infrastructure economics are heading. This isn’t a roadmap hunting for validation. It’s an inference-economics play with real hyperscaler commitments behind it, and the open question is execution timing, not credibility.
Qualcomm Showed Up with Real Customers, Not a Dog with a Roadmap
The differentiator at this event wasn’t just the silicon. It was who stood on the virtual stage to back it. Microsoft CEO Satya Nadella put Azure behind Qualcomm’s High Bandwidth Compute (HBC) accelerators, and Meta CEO Mark Zuckerberg committed to a multi-generation roadmap built on the new Dragonfly C1000 CPU. Companies roll out datacenter roadmaps all the time. Getting two of the largest AI operators on the planet to publicly commit to parts that don’t ship until fiscal 2027 and 2028 is the much rarer signal. I’d read the Meta arrangement as a memorandum of understanding rather than a signed contract, so I’m not overweighting it, but it’ll happen if Qualcomm hits their goals. As my colleague Matt Kimball wrote, that kind of customer validation lowers the barrier for every other hyperscaler now sizing up Qualcomm.
The Real Bet Is Inference Economics, and HBC Is the Differentiator
This is a bet on inference total cost of ownership, not training. Cristiano Amon was blunt that agentic workloads generate far more tokens than the chat prompts enterprises started with, which moves the metric from peak benchmark performance to tokens per watt and per dollar. He’s right. McKinsey pegs AI datacenter capital spending at roughly $5.2 trillion by 2030, with about 60% of it flowing to chips and hardware, and it’s power, not money, that is the binding constraint. That’s the seam Qualcomm is going after. The technology that matters most is HBC. It stacks compute under the LPDDR memory to hit the decode-phase memory wall at far lower cost than HBM on paper and, as I explained on Yahoo Finance, without TSMC’s CoWoS packaging. The other real differentiators are the Modular software buy for cross-silicon portability and the Alphawave connectivity that’s already shipping. Of everything Qualcomm showed, the CPU cores had the least details, maybe because they are the farthest off. On Modular I’m cautious. Qualcomm is strong at edge software, but datacenter software is a different muscle, and the key will be in the datacenter execution but also its ability to spread the love across edge to cloud.
Why I Think One of the Custom Wins Is Amazon
Qualcomm said it landed two custom-silicon wins with hyperscalers but didn’t name them. My read is that one is Amazon. The deduction isn’t hard. Google already runs three to four TPU partners and has no room for another. Nadella was on stage for the accelerator and Zuckerberg for the CPU, so neither of them is the unnamed custom customer. Qualcomm described a “largest-class” hyperscaler on its last earnings call, which effectively leaves Amazon, because Oracle has no meaningful chip IP to integrate (it buys Ampere). Amazon was already an Alphawave customer before Qualcomm bought it, so a custom Graviton-plus-Trainium derivative stitched together with Alphawave connectivity, maybe even an NVLink port, fits cleanly. The connectivity is the tell here. (I’ve followed the custom-silicon connectivity trail long enough to weight it heavily.) Someone will raise Humain, the Saudi sovereign AI buildout Qualcomm signed in May 2025, and it’s a fair flag, but that’s a sovereign engagement rather than one of the two hyperscaler custom wins Qualcomm called out at this event. It’s also where the publicly reported NVLink-capable custom CPU traces back, which is why I lean on the Alphawave relationship, not the NVLink angle, as the real tell on Amazon. I walked through the full logic on X. I’d put ByteDance as a plausible second win cited in a few pubs, which Qualcomm wouldn’t want to showcase for obvious, geopolitical reasons.
The $15 Billion Target Is Conservative, and the Risk Is Timing
Now the number. The $15 billion fiscal 2029 datacenter target, building from about $5 billion in fiscal 2027 and close to nothing today, reads conservative to me, not aspirational. Most of that early revenue is the two custom wins. For scale, AMD did roughly $5 billion in the first full year of its MI300 accelerator, so this isn’t a wild number. I keep hearing there’s no room for another datacenter silicon vendor. That misreads the supply picture entirely. The XPU inference march is real, and I was saying so years ago, back when chip suppliers were taking 75%-plus design margins and every hyperscaler started hunting alternatives. Supply is short. No single architecture fits every workload. The real risk is timing. Most of the marquee silicon ships in 2027 and 2028, HBC still has to clear high-volume packaging, and Modular only pays off if it stays open. Net-net I believe the numbers Qualcomm put up were all conservative.
What Qualcomm Still Has to Prove
Qualcomm didn’t ‘win the datacenter at this event’, it isn’t going to kill NVIDIA, and nobody should pretend it did. It didn’t have to, either. What it walked in with was a credible entry and real customers willing to attach their names to it, with real dollar commitments, which is further than most challengers in this space ever get. From here it rides entirely on execution: turning two known anchor accounts and two unknown, and a strong architecture into silicon that actually ships and hits the cost-per-token numbers a hyperscaler will confirm on its own benchmarks or with Signal65 testing rather than Qualcomm’s slides. Clear that bar and Qualcomm is a legitimate force in inference silicon next to GPUs and the custom-ASIC builders in the datacenter. Miss it and this was a very well-produced set of promises. I’d bet on the former. 2027 and 2028 are the years that decide it.
Sources
- Patrick Moorhead, Patrick Moorhead Discusses Qualcomm and Micron on Yahoo Finance, June 25, 2026, Broadcast Analysis, Moor Insights & Strategy
- Patrick Moorhead, X thread on Qualcomm’s custom-silicon wins and the Amazon deduction, June 27, 2026
- Patrick Moorhead, X post on the XPU inference march, June 24, 2026
- Matt Kimball, Qualcomm Is Betting the AI Datacenter on Inference Economics, Moor Insights & Strategy
- Qualcomm, Qualcomm Accelerates Diversification with Comprehensive Strategy for Data Center, Investor Relations, June 24, 2026
- Qualcomm, Qualcomm Unveils Comprehensive Data Center Roadmap for the Agentic AI Era with New Qualcomm Dragonfly Portfolio, June 24, 2026
- McKinsey & Company, The cost of compute: A $7 trillion race to scale data centers
