CXL Fails to Dent HBM as OpenAI and Intel Flag Bandwidth Gap While Mobile DRAM Prices Plateau

The clearest read on the AI memory hierarchy this week came from the AI Infrastructure Summit in California. Daniel Morris, an OpenAI researcher working on accelerator design, said he has yet to find a compelling use case for CXL in real model workloads, limiting it largely to storing inactive data that large models rarely touch. Vidhya Thyagarajan, Intel's head of AI SoC architecture, was blunter on physics: data transfer between GPUs over CXL is 'by no means as fast as HBM', which puts the technology alongside secondary storage rather than in place of the high-bandwidth tier. Both remarks were reported by Chosun Biz and Financial News, and relayed by TrendForce on September 18.

That does not stop the memory makers from building the tier out. SK hynix unveiled its Inference Tier Memory Expansion architecture in June, inserting CXL hybrid memory between DDR and SSD in the conventional HBM-DDR-SSD stack and reporting a 35.7% improvement in AI inference efficiency. Samsung plans to start mass production of its CMM-D 3.0 module, based on the CXL 3.2 specification, by the end of this year, while SK hynix has begun sampling a second-generation CMM-DDR5 and is in supply talks without a firm mass-production date. Chosun Biz's conclusion is that the competitive position of Samsung and SK hynix in HBM stays largely intact even as CXL adds a layer underneath.

The near-term pricing picture is about conventional DRAM, not CXL. TrendForce's 3Q26 Mobile DRAM analysis says the growth rate of mobile contract prices is narrowing this quarter because of a high base and rising inventory at buyers, and that prices are projected to enter a period of high-level consolidation by year-end. The mechanism is capacity allocation: original manufacturers are accelerating their exit from legacy parts and shifting wafers toward server DRAM and HBM, so supply stays tight into next year even as the price curve flattens. TrendForce also flags that surging memory prices triggered early phone replacements this year, which lifts 2026 production but lengthens the replacement cycle and worsens the demand picture for 2027.

For buyers, the practical consequence is that the AI memory tier and the consumer tier are now moving on separate tracks. Server and HBM demand keeps absorbing incremental bits, which is why supply remains tight into 2027 while mobile contract prices stop accelerating; the upward shift in price bands is already squeezing the entry-level segment, with brands responding by cutting specifications or raising outsourcing ratios. Anyone budgeting 4Q26 and 1Q27 hardware should treat plateau rather than relief as the base case: TrendForce's language is consolidation at a high level, not a reversal.

The one thing that could change the structure is qualification, not architecture. HBM4, custom HBM and CXL modules all remain in specification and sampling phases, and the reports this week describe negotiations and sampling rather than volume pricing. Until those products clear qualification at the hyperscalers, the economics of AI memory continue to run through HBM3e and high-capacity server DDR5, where suppliers retain pricing power.