On Wednesday June 24, 2026, at its Investor Day in New York, Qualcomm walked on stage as the fourth name in a data-center chip race already being run by three — Nvidia, AMD, and now OpenAI-with-Broadcom — and made the case that being late is a feature. The company unveiled the Dragonfly C1000 CPU, the AI300 inference accelerator, and a multi-generation supply agreement with Meta for the CPU. Both chips ship in 2028. The C1000 will power Meta's next-generation server fleet starting in the second half of that year.

Underneath the silicon roadmap sat the announcement the equity desks actually traded on: a ~$4 billion all-stock acquisition of Modular Inc., expected to close in H2 2026. Modular is the company that has spent the last four years building the most credible alternative to Nvidia's CUDA software moat. Qualcomm did not buy Modular to ship a Mojo compiler on a phone. It bought the only inference stack that runs the same model binary across Qualcomm, AMD, Intel and Nvidia silicon without a rewrite. The chips ship in 2028. The bypass ships in eighteen months.

The Modular sentence inside the roadmap

The most quoted line of the morning came from CEO Cristiano Amon, per the Data Center Dynamics writeup: "Agentic AI is driving a significant increase in demand for AI inference." The companion line, repeated by StorageReview, was crisper: "We believe the future belongs to developer-friendly, horizontal platforms that can run across diverse compute environments and give customers real choice in how and where they deploy AI."

Horizontal. Across diverse compute environments. Real choice. That is the exact pitch Modular has been making since 2022, when Chris Lattner — the architect of LLVM, Clang, Swift, and MLIR — left Google with Tim Davis and raised $100 million for Modular to build the Mojo language and an inference engine designed for one job: make a model trained on Nvidia run, at competitive speed, on something that is not Nvidia. CUDA's lock-in is the moat. Modular's stated mission is to cut that moat.

Qualcomm paying $4 billion for the bypass before its own silicon arrives reads two ways. The optimistic read is that Qualcomm is buying the software story it will need to make Dragonfly competitive in 2028, when the market will already be flooded with second-source training pools (AMD), custom inference ASICs (Jalapeño, Trainium, TPU) and whatever Nvidia is shipping that year. The cynical read is that Qualcomm is admitting it cannot beat Nvidia on raw silicon and is paying $4 billion for a story it can tell hyperscalers about portability. Both reads are compatible. That is a sentence that costs Qualcomm $4 billion before the chip exists.

The deal is all-stock — roughly 19.2 million Qualcomm shares per Yahoo Finance's read of the filing. Lattner stays on. Modular's headcount, which sat around 150 the last time the company disclosed it, becomes Qualcomm's data-center software group.

What the chips actually are

Per the StorageReview teardown of the Investor Day deck, three products were on the slide:

• Dragonfly C1000 CPU. Multi-chiplet, 250+ cores, clocks above 5 GHz, PCIe Gen 7 at 2 TB/s, CXL for memory disaggregation. Three SKU variants — agentic, general-purpose, AI head node. Custom Oryon cores, the same architecture Qualcomm has been refining for the Snapdragon X laptop line. Stated claim: 2× better performance per watt than current competitive server CPUs. Commercial: 2028.
• AI250 accelerator. Inference part with High Bandwidth Compute Gen 1. Claim: 133 TB/s effective memory bandwidth per card, 18× the AI200. Commercial sampling: mid-2027.
• AI300 accelerator. Third-generation inference platform with HBC Gen 2. Claim: 54× memory-bandwidth increase versus AI200 and 4-8× better performance-per-watt versus GPU-based architectures. Commercial sampling: 2028.

Two things to flag on those numbers. The "2× per-watt CPU" and "4-8× per-watt accelerator" claims are vendor benchmarks, not third-party MLPerf results. Until somebody runs Llama 4 or GPT-5 inference on shipping silicon, they belong in the directional-vendor-PR column. And the bandwidth ratios — 18× and 54× — are measured against Qualcomm's own AI200, a 2025 part that nobody outside its launch customers ever benchmarked. Stacking ratios against your own first-generation chip is how a roadmap looks better than the market.

What is harder to argue with is the chiplet count. 250+ cores on a single CPU package in 2028 is plausible — TSMC's N2 family supports it — and would put the C1000 in the same neighborhood as AMD's Turin-2 and the rumoured Granite Rapids-AP refresh. Meta's interest is consistent with that read: hyperscalers buy CPUs by the core-per-watt-per-rack number, and Meta has been the most aggressive hyperscaler at running its own custom silicon program (MTIA) in parallel with merchant CPUs.

Meta is the validation, not the volume

The Meta deal does the same job Microsoft's name did inside last week's Broadcom press release: it tells the market that a hyperscaler with real capex looked at the roadmap and signed. Tony Pialis, Qualcomm's EVP and GM of data center, told Data Center Dynamics that enterprises need "orchestrating multiple types of compute across distributed, always-on infrastructure" — corporate for we are selling you a CPU because the other guys keep changing their training stack.

Financial terms of the Meta deal were not disclosed. The supply window starts H2 2028, two and a half years from the announcement. Two and a half years, in 2026 AI-infrastructure time, is geologic. Stargate's first GW lands in H2 2026. Jalapeño's first wafer is already in the foundry. Vera Rubin is ramping. Qualcomm is competing for the rack space that exists after the current cycle of capex commitments has already burned through.

That is the structural problem with the Dragonfly story. The chips are good. The Modular acquisition is the cleverest software move in the AI-infrastructure stack since Microsoft's Inflection license. But the ship date is too far out for the company to participate in the inference-cost compression that is happening now. By the time the C1000 reaches Meta's racks, the cost-per-token curve will already have moved twice.

The number that moved the stock

Pre-market, Yahoo Finance reported Qualcomm spiked, then gave most of it back inside the first hour of trading — closing the morning around +1.9%. The number the analysts circled was not in the chip deck. It was the revenue line: Qualcomm raised its 2029 non-handset revenue target to $40 billion, up from $22 billion — a near-doubling of the company's stated diversification ambition. Of that, $15 billion is data centers.

That is the contract Qualcomm is now signing with its own shareholders. $15 billion of new data-center revenue, on chips that do not ship for two years, against a market where the incumbent has 90% gross margin and a software lock-in. The Modular acquisition is the only line item on the slide that gives any of it a chance of working.

What to watch

1. The Mojo roadmap inside Qualcomm. Modular's value is the inference stack, not the company. If Qualcomm absorbs Mojo and slows the open-source cadence, the bypass loses credibility within twelve months. If it commits to keeping Mojo open and cross-vendor — even running on Nvidia — the acquisition is the most important software move of the year. The early signal will be the first post-close commit graph.

2. The second hyperscaler. Meta is the launch CPU customer. Microsoft, Google, Amazon and Oracle were not on the slide. The Dragonfly economics start to compound only when a second one signs. Watch the Azure Hot Chips agenda in August, and Re:Invent in December.

3. The benchmark. Qualcomm's claims — 2× CPU perf-per-watt, 4–8× accelerator perf-per-watt, 54× memory bandwidth — are vendor numbers measured against a vendor part. MLPerf Inference v6.0 results land in the spring. Until then, the silicon is a slide.

4. The Nvidia counter-move. Nvidia's response to a credible CUDA bypass funded by a $4 billion acquisition is the question that decides the next decade of inference economics. Watch GTC 2027. The CUDA roadmap has been the moat. The moat just had a vendor walk into the room with a bridge.

The chip-naming pattern of the season has settled. OpenAI named its first chip after a pepper. Qualcomm named its first server CPU after an insect with 30,000 lens facets in each eye and a flight efficiency that has not been improved on in 300 million years. The branding works. The roadmap is more interesting than the branding. The Modular line on the cap-table is the most interesting thing on the slide.