Qualcomm’s stock has held near recent highs as investors digest the company’s announcement that Microsoft and Meta will adopt its latest AI chip, a deal that positions the mobile-chip maker as a serious challenger in the market for AI inference hardware. The adoption marks the first time two of the world’s largest AI buyers have committed to a Qualcomm data-center chip, and it validates a strategy the company has pursued for years: attacking Nvidia not with a faster chip, but with a cheaper one.
The chip’s defining feature is its memory architecture. Qualcomm’s design, which the company calls HBC, does not rely on the high-bandwidth memory, or HBM, that Nvidia’s accelerators require. HBM is the fastest memory in the industry, but it is also the most expensive, and its supply is controlled by a handful of manufacturers who cannot make enough of it. Qualcomm’s chip instead uses the kind of memory found in phones and laptops — commodity DRAM that is cheap, abundant, and made by many suppliers. The trade-off is bandwidth, and Qualcomm’s argument is that the trade-off is worth it for inference, where the cost of serving a query often matters more than raw speed.
The economics are the heart of the pitch. An Nvidia accelerator with HBM can cost tens of thousands of dollars, and the memory accounts for a large share of that price. A Qualcomm chip built around commodity memory costs a fraction as much, and for workloads that are memory-bound rather than compute-bound — a category that includes much of inference — the cheaper chip can deliver acceptable performance at a dramatically lower price. If the approach works at scale, it would rewrite the cost structure of AI inference, the fastest-growing part of the AI business.
Microsoft’s decision to adopt the chip is significant because Microsoft runs one of the largest AI workloads in the world through its partnership with OpenAI, and it has been diversifying its silicon across Nvidia, its own Maia accelerators, and now Qualcomm. Meta, which runs inference for billions of users across its apps and its AI assistant, has been equally eager to cut the cost per query. Both companies have concluded, according to people familiar with their thinking, that the AI buildout’s economics depend on more than one memory architecture, and that commodity-memory designs have a role that HBM-based systems cannot fill.
Qualcomm has been here before. The company’s earlier attempts to break into data-center chips, in an era when it competed on raw performance against Intel and AMD, ended in retreat. Its current strategy is different: instead of trying to beat the leaders at their own game, it is competing on a dimension they have deprioritized — cost per query rather than peak throughput. The company has also built its AI chip to run a broad range of models, arguing that customers want hardware that can serve many architectures rather than silicon optimized for one vendor’s software.
The announcement has revived a debate about the future of AI hardware. Nvidia’s defenders argue that the industry’s workloads are growing so fast that commodity-memory chips will be relegated to the low end, while the demanding frontier models will always need HBM-class bandwidth. Qualcomm’s backers counter that the low end is where the volume is: most AI queries are short, simple, and cheap to serve, and a chip that serves them profitably at one-tenth the cost will capture the majority of the market even if it never touches the most demanding workloads.
The stock’s response has been measured. Qualcomm shares rose on the news and have held their gains, but the move has been orderly rather than euphoric, a sign that investors are waiting to see whether the deals translate into volume. The company has not said when Microsoft and Meta will deploy the chips at scale, nor has it disclosed the size of the orders; a person close to the company said production volumes will ramp through next year.
The deals also carry a message about the industry’s supply-chain psychology. The past two years taught the largest AI buyers a hard lesson about concentration: when one supplier dominates a critical component, every shortage, every allocation decision, and every price increase flows straight through to their own economics. Cloud providers have responded by building redundancy into their silicon strategies, and the willingness of Microsoft and Meta to test a commodity-memory design is part of that pattern. Qualcomm’s chip gives them an option that costs little to evaluate and could, if it works, change the baseline price of inference across the industry.
The competitive stakes extend beyond Nvidia. AMD and Intel are both developing their own approaches to cheaper inference, and the custom-chip designers who serve the biggest cloud providers are watching Qualcomm’s pricing closely. If the HBC architecture proves itself, memory supply chains, data-center designs, and the economics of every AI service will shift — and Qualcomm, long dismissed as a phone-chip company, will have done what no rival has managed: made AI inference cheap enough that the industry’s cost structure becomes a battleground.
