Qualcomm walked into Manhattan on June 24 and named two anchor customers most analysts did not expect. Meta signed a multi-generation deal to deploy the new Dragonfly C1000 server CPU in its fleet, and Microsoft confirmed it will run Qualcomm’s HBC architecture on Azure. The shares jumped about 12 percent after hours [TradingKey, 2026-06-24]. The story for our desk is not the equity move. It is what the chip roadmap, the customer list, and the rack architecture imply for the materials underneath.

What’s happening

Qualcomm used its 2026 Investor Day at the Ziegfeld Ballroom to unveil a full Dragonfly data center portfolio and confirm a $3.92 billion all-stock acquisition of AI software company Modular [Qualcomm press release, 2026-06-24; TechStartups, 2026-06-24].

The headline product is the Dragonfly C1000, a custom Oryon-core server CPU with more than 250 cores at sustained frequencies above 5 GHz, more than 2 TB per second of PCIe Gen 7 and CXL connectivity, and full enterprise RAS. Qualcomm targets more than 2x performance per watt versus benchmarked server CPU competitors. Commercial availability is expected in 2028, in OCP ORv3 compliant racks with both air and liquid cooling support [Qualcomm news release, 2026-06-24].

The inference accelerator line, AI200 and AI250 (announced October 2025), ships first. The AI200 pairs with LPDDR5X memory and targets commercial availability in 2026. The AI250 introduces Qualcomm High Bandwidth Compute Gen 1, a 3D-stacked near-memory architecture that the company says delivers 133 TB per second per card, an 18x effective bandwidth increase over AI200. AI250 commercial sampling is expected in mid-2027 [Qualcomm news release, 2026-06-24]. Both target 160 kW rack-level power draw, with liquid cooling, PCIe scale-up, and Ethernet scale-out [DCD, 2025-10-28]. Saudi Arabia’s Humain has already committed to deploying 200 MW of Qualcomm AI200 and AI250 hardware in the Kingdom [Qualcomm release, 2025-11].

The Modular deal is the missing software piece. Modular’s stack lets a single model run across CPUs, GPUs, NPUs, and custom ASICs without per-architecture rewrites, which is the gap Nvidia has used CUDA to defend [TechStartups, 2026-06-24; CNBC, 2026-06-24]. Closing is expected in the second half of 2026.

Qualcomm projects data center chip revenue of about $5 billion by fiscal 2027 and more than $15 billion by 2029 [TradingKey, 2026-06-24]. For scale, Broadcom is sitting on a roughly $73 billion AI backlog and targeting $100 billion in annual AI chip revenue by 2027 [Tom’s Hardware, 2026-05]. Custom ASIC shipments are forecast to take 27.8 percent of AI server share in 2026, growing 44.6 percent year-over-year against 16.1 percent for merchant GPUs [SNS Insider, 2026].

Brazil angle

Brazil does not make these chips. Brazil supplies the substrate underneath them. The June BNDES and Finep announcement of 56 pre-selected critical mineral projects, totaling R$ 45.8 billion, is roughly nine times the program’s original R$ 5 billion envelope. The first approved disbursement is a R$ 280 million credit line to WEG for a battery energy storage (BESS) factory in Itajaí, Santa Catarina, with first output at the end of 2027 [BNDES Agência de Notícias, 2026-06; Finep, 2026-06]. The program names lithium, graphite, copper, and rare earths as the priority chains.

The connecting line to Manhattan is power density. A 160 kW rack is roughly four to five times the average enterprise rack today, and the copper bill scales with it: cold plate liquid cooling, which now holds about 65 percent of the data center liquid cooling market, depends on copper micro-channel blocks bonded to each die [KAD, 2026]. Power distribution, busbars, and on-site transformers also pull more copper as rack density climbs. LME copper traded near $13,200 to $13,500 per tonne in late June, with a 2026 refined balance running at roughly a 600 kt deficit [Trading Economics, 2026-06-24; Crux Investor / ING, 2026]. Brazil owns the second-largest known rare earth reserves and a meaningful copper position through Vale’s expansions. The BNDES queue is the country’s bet on absorbing the AI materials wave that announcements like Qualcomm’s accelerate.

United States angle

For US hyperscalers, the Dragonfly customer list is a hedge against Nvidia concentration risk and CUDA lock-in. Nvidia still holds 75 to 80 percent of enterprise AI accelerator share [BingX, 2026]. Meta committed in April to deploy 1 GW of its MTIA custom accelerators co-designed with Broadcom through 2029 [Broadcom investor release, 2026-04-14]; layering Qualcomm CPUs on top of that fleet is a second-silicon hedge, not a replacement. Microsoft pairing Qualcomm HBC with its own Maia and Nvidia GPUs is the same playbook on Azure.

The constraint is not the chip. It is power. ERCOT’s large load interconnection queue reached 226 GW as of November 2025, nearly quadruple the 63 GW from the prior year, with about 77 percent of that demand coming from data centers aiming to connect by 2030 [Latitude Media, 2025-12-03 / ERCOT board filing]. PJM took in 95 large-load adjustment requests totaling about 54 GW through November 2025, and Wood Mackenzie projects 55 GW of new large-load growth across the PJM footprint by 2030 [Utility Dive, 2026]. Timelines from application to commercial operation in PJM have moved from under two years in 2008 to over eight years in 2025 [PJM Cycle 1 reform documentation, 2026]. If Qualcomm’s $5 billion by FY2027 ramps as guided, those racks need substations Meta and Microsoft will not have if they wait in line.

China angle

The competitive read from Beijing is sharp. Huawei is targeting roughly 600,000 Ascend 910C units in 2026 on SMIC’s enhanced 7 nm process, about twice the prior year’s output, with total Ascend dies potentially reaching 1.6 million next year (Reuters reporting, summarized in [SCMP and RCR Wireless, 2025-09]). The 910C runs at 60 to 80 percent of H100 inference performance at a fraction of the cost. A Huawei-led team has post-trained DeepSeek’s largest model on 910C silicon [Tom’s Hardware citing SCMP, 2026]. The strategic point is that Qualcomm’s pitch (a viable non-Nvidia, non-AMD alternative for inference) and Huawei’s pitch (a viable non-Western alternative for the same workload) describe two halves of a market that is decoupling, not converging.

What it means

The agentic AI era Qualcomm is selling is a story about more inference, longer context windows, and CPU-bound orchestration. The materials economy reads that as more LPDDR, more HBM, more copper per kilowatt, and a steeper power curve. For Tantalum’s coverage, the announcement is less about Qualcomm specifically and more about the diversification of silicon buyers, which lengthens the demand cone for the inputs that show up in TAI-M and TAI-P.

What to watch