Intel and SoftBank Target HBM With New Memory Tech

Intel and SoftBank are jointly challenging the dominance of High Bandwidth Memory (HBM) through a new venture called Saimemory, which is developing an alternative memory technology promising higher bandwidth and greater capacity for AI accelerators. The move signals a potential disruption in one of the semiconductor industry's hottest markets, where SK hynix, Samsung, and Micron currently hold near-total control.

Saimemory, established by SoftBank in December 2024, has been quietly working on memory modules designed specifically for next-generation AI chips. Intel's involvement marks a dramatic return to the memory arena for a company that once pioneered DRAM technology before exiting the business decades ago.

Key Takeaways

• Intel and SoftBank are collaborating through subsidiary Saimemory to develop an HBM alternative
• The new technology targets higher bandwidth and capacity than current HBM solutions
• Saimemory was founded by SoftBank in December 2024
• Intel's participation represents a return to the memory business after decades away
• Current HBM faces challenges including packaging complexity, thermal management, and supply constraints
• The global HBM market is projected to exceed $25 billion by 2026, drawing new challengers

Intel's Dramatic Return to Memory

Intel's history with memory is one of the most famous pivot stories in tech. The company literally invented DRAM in 1970, but fierce competition from Japanese manufacturers in the 1980s forced it to abandon the business entirely. That decision led Intel to focus on CPUs, building the 'Blue Giant' empire that dominated computing for decades.

Now, with AI reshaping the semiconductor landscape, Intel appears eager to re-enter the memory game. The company's involvement in Saimemory suggests it sees an opening to disrupt the current HBM oligopoly. Unlike its original exit from memory, this return is driven not by retreat but by the enormous opportunity AI presents.

Intel's existing expertise in advanced packaging technologies — including its Foveros 3D stacking and EMIB interconnect solutions — could prove invaluable. These capabilities are directly relevant to building next-generation memory solutions that need to integrate tightly with AI processors.

Why HBM Faces Growing Pains

HBM has become the memory technology of choice for AI workloads, powering chips from NVIDIA, AMD, and other accelerator makers. Its advantages are clear: ultra-high bandwidth, relatively low power consumption, and compact form factors achieved through vertical DRAM die stacking.

However, HBM's rapid ascent has exposed several critical weaknesses:

• Manufacturing complexity: Stacking multiple DRAM dies using through-silicon vias (TSVs) requires extremely precise advanced packaging
• Yield challenges: Higher stack counts (HBM3E uses 8-12 layers) drive down manufacturing yields significantly
• Thermal constraints: Dense stacking creates heat dissipation problems that limit performance
• Supply bottlenecks: Only 3 companies — SK hynix, Samsung, and Micron — currently produce HBM at scale
• Cost pressure: HBM chips cost roughly 4 to 5 times more per gigabyte than standard DRAM
• Capacity limitations: Current HBM generations max out at 36GB per stack, insufficient for emerging AI model sizes

These constraints create a genuine opening for alternative approaches. As AI models grow from billions to trillions of parameters, the demand for memory bandwidth and capacity is outpacing what incremental HBM improvements can deliver.

What Saimemory Is Building

While specific technical details remain limited, reports indicate that Saimemory is developing memory modules that aim to surpass HBM in two critical dimensions: raw bandwidth and total capacity per package. This suggests the company may be exploring architectural approaches fundamentally different from HBM's stacked-DRAM design.

Several potential technical paths could achieve these goals. One approach involves optical interconnects between memory and processor, which could dramatically increase bandwidth beyond what electrical TSV connections allow. Another possibility is a novel chiplet-based memory architecture that distributes memory across a larger area rather than stacking it vertically, potentially solving the thermal challenges that plague HBM.

SoftBank's involvement adds strategic depth to the venture. The Japanese conglomerate has been aggressively positioning itself in the AI chip space through its subsidiary Arm and its planned investments in AI infrastructure. CEO Masayoshi Son has publicly discussed spending over $100 billion on AI-related ventures, and a proprietary memory technology would give SoftBank significant leverage in the AI hardware stack.

The combination of Intel's semiconductor manufacturing know-how and SoftBank's financial firepower makes Saimemory a credible threat to incumbents, even in a market with extremely high barriers to entry.

The Crowded Field of HBM Challengers

Saimemory is not the only venture attempting to dethrone HBM. The memory technology landscape is seeing unprecedented innovation driven by AI demand.

Samsung has been developing its own alternatives internally, including Processing-in-Memory (PIM) technology that embeds computation directly into memory chips. TSMC is exploring System-on-Wafer approaches that could integrate memory and logic on a single massive die. Several startups are also pursuing novel memory architectures:

• Enfabrica is building networking chips that enable disaggregated memory pooling
• d-Matrix uses in-memory computing to reduce the need for separate HBM entirely
• Cerebras sidesteps HBM with massive on-chip SRAM in its wafer-scale processors
• Tenstorrent, backed by Hyundai, is designing AI chips with alternative memory strategies

This wave of challengers reflects a broader industry recognition that HBM, while dominant today, may not be the final answer for AI memory needs. The market is simply too large and too strategic to remain a 3-company monopoly indefinitely.

Industry Context: The $25 Billion Memory Race

The stakes in AI memory could hardly be higher. NVIDIA alone is expected to consume over $10 billion worth of HBM in 2025 for its Blackwell and next-generation Rubin GPU platforms. The total addressable market for AI-optimized memory is growing at roughly 30% annually, driven by hyperscaler data center buildouts from Microsoft, Google, Amazon, and Meta.

Currently, SK hynix leads the HBM market with an estimated 50% share, followed by Samsung at roughly 30% and Micron at 20%. These three companies have invested tens of billions of dollars in HBM production capacity, creating formidable competitive moats.

But the AI industry's insatiable appetite for memory bandwidth creates natural pressure for disruption. Each new generation of AI models demands roughly 2 to 3 times more memory bandwidth than its predecessor. GPT-4 class models already push the limits of current HBM3E technology, and next-generation multimodal AI systems will demand even more.

This exponential growth curve means that evolutionary improvements to HBM — adding more layers, increasing per-die density — may eventually hit physical limits. Revolutionary approaches like what Saimemory appears to be pursuing could become necessary rather than merely desirable.

What This Means for the AI Hardware Ecosystem

For AI chip designers like NVIDIA and AMD, a viable HBM alternative would be welcome news. Supply constraints from the current 3-vendor ecosystem have created bottlenecks that delay product launches and inflate costs. More memory options mean more negotiating leverage and supply chain resilience.

For cloud providers and enterprises building AI infrastructure, alternative memory technologies could eventually reduce the cost of deploying large AI systems. HBM's premium pricing is a significant contributor to the high cost of AI training and inference hardware.

For developers and researchers, more memory bandwidth and capacity per chip means larger models can run on fewer GPUs, potentially democratizing access to frontier AI capabilities. Today's memory constraints are a primary reason why cutting-edge AI research remains concentrated among a handful of well-funded organizations.

Looking Ahead: Timeline and Expectations

Developing production-ready memory technology is a multi-year endeavor. Even with Intel's manufacturing expertise and SoftBank's deep pockets, Saimemory likely faces a 3 to 5 year timeline before any product reaches volume production. Memory technology requires extensive qualification cycles with chip designers, and building manufacturing capacity at scale demands billions in capital investment.

The critical milestones to watch include Saimemory's first technical disclosures, partnerships with AI chip companies, and any announcements regarding fabrication facilities. Intel's own foundry ambitions through Intel Foundry Services could provide a natural manufacturing pathway.

What is clear is that the memory industry is entering its most dynamic period in decades. The convergence of AI demand, technical limitations of current solutions, and the entry of deep-pocketed new players like Intel and SoftBank suggests that the HBM monopoly's days may be numbered — even if the disruption takes years to materialize.

The question is no longer whether alternatives to HBM will emerge, but which approach will prove commercially viable first. Saimemory's Intel-SoftBank partnership just made that race considerably more interesting.