Five times. That is the number Qualcomm attached to its first high-bandwidth computing technology on Nov. 4, and the comparison it chose is the one that matters to the people who buy AI hardware.
The company said it had completed initial hardware verification of the technology, measured on two-layer DRAM samples. Bandwidth per watt came in at five times that of HBM4, the memory standard that sits beside the accelerators in AI servers. Total bandwidth measured 18 times that of LPDDR5X, the low-power memory used in phones and, increasingly, in machines that need capacity without a large power bill.
The two comparisons point in opposite directions, and the reasons are the substance of the claim.
HBM is the expensive option. It stacks DRAM dies vertically and wires them through silicon, which buys enormous bandwidth at a cost in manufacturing complexity and heat. LPDDR is the cheaper option: mature, widely made, designed for a battery. A memory subsystem that borrows the second and claims an advantage over the first is a way of saying the constraint has moved from raw bandwidth to the power required to move a bit.
That is a fair reading of where data centers are. Operators have spent three years discovering that the limit on AI capacity is electricity, not floor space. A rack of accelerators draws as much power as a small neighborhood, and a meaningful share of it goes to shuttling data between memory and compute. Anything that improves bandwidth per watt changes how many tokens a building can produce.
Qualcomm’s answer is a compute architecture rather than a memory chip. It calls the approach high-bandwidth computing, and the first product built on it is Dragonfly AI250, a data-center part scheduled to sample in the first half of 2027. Sampling is the stage where customers receive silicon, run their own workloads, and decide whether a benchmark survives contact with a real model.
The timing sets the competitive frame. A 2027 sample means volume in 2028 at the earliest, two product cycles after the parts shipping now. Qualcomm is not trying to win the current generation. It is trying to arrive when the power bill has become the procurement problem, and when buyers are willing to look at a second source.
The company has tried this before. Qualcomm spent years and a great deal of money on Arm-based server processors, launched a part aimed at cloud operators, and withdrew when the customers did not follow. The lesson it took away, executives have said since, was that a chip is not a business: software, memory supply, and the willingness of a customer to rewrite code all have to line up.
The 2027 plan faces the same questions, and the memory claim is only the first. Bandwidth per watt measured on a two-layer sample is a component result. A complete system has to feed that memory from a package, cool it, and connect it to a fabric that other chips can talk to. Whether the five-to-one advantage holds once the surrounding parts are counted is what customers will test.
There is also the supply question. HBM is scarce and allocated years ahead by a small number of manufacturers. LPDDR is not scarce, which is part of its appeal, but a data-center product built on mobile memory has to convince buyers that the supply chain can support a fleet rather than a pilot.
And there is the software question. The accelerator market is decided less by specifications than by which programming model a generation of engineers already knows. A new entrant has to offer a path that does not require rewriting a training stack, or a workload narrow enough that rewriting is worth the cost.
Qualcomm has been positioning itself around inference, the serving of trained models, where workloads are more standardized and where power per token is the figure customers carry into a budget meeting. That is a narrower target than training, and a larger market than it was two years ago.
The claim itself will be tested in public. Customers who take samples in 2027 will publish results, competitors will answer with comparisons of their own, and the industry will argue about which baseline was fair. Such arguments are routine in memory, where every generation arrives with a slide that measures it against the last.
What is different here is the denominator. For two decades the memory industry sold bandwidth and let the customer worry about power. Qualcomm is selling the ratio, which is an admission that its customers now count in megawatts. Whether it can deliver that ratio on a rack, and not merely on a sample, is the question the 2027 parts exist to answer. If it cannot, the five-to-one number will be remembered as a measurement of a die rather than of a machine.
