To put the economics of how this all works into context, if you are a larger customer of DRAM, the likelihood is that you can secure better terms or contracts for pricing; the smaller you are, the less leverage you'll ultimately have. When combined with the fact that the AI industry is not only sucking up demand, but paying top-dollar for chips, means that smaller customers of DRAM and NAND are pulling the short straw, and might be more heavily affected.
Framework's latest update on the ongoing crisis states that DDR5 memory pricing is now between $12-16 per gigabyte, and their end product pricing has to be increased as a result of that. "The new system and Mainboard prices are 6-16% higher than before. We anticipate that here as well, costs from our suppliers are going to continue to increase over the next few months," says Nirav Patel, CEO and Founder of Framework.
So, if costs get higher, and consumer appetite for these products is lower, the future for smaller manufacturers gets called into question: How will they survive if product run-rates get lower, margins get slimmer, and there's seemingly no end in sight? The reality is, it's already too late to prepare for what's to come over the next few years, and the damage to these smaller companies has yet to be fully quantified.
IDC's latest analysis suggests that the PC market alone could shrink by up to 9%, which may not sound like much on paper, but a figure like this might be life-or-death for some businesses. If the chip supply crisis is hitting the PC market this hard alone, what about other industries, where the risk has yet to be factored in?
Now that we've laid out the effects of AI demand on the memory and storage industries, it's important to note exactly how AI is using these chips in large-scale deployments across the globe. The average Nvidia Rubin NVL72 superchip is equipped with 288 GB of HBM 4 memory, which uses vertically stacked memory ICs, bonded together to offer more density in the same physical footprint. Therefore, High Bandwidth Memory requires around three times the number of ICs on a single chip compared to a DDR5 module. That's in addition to 128 GB of GDDR7 VRAM on the Nvidia CPX GPU on any single unit. Bolstered by high-speed data interconnects like Spectrum-X Photonics Ethernet, and Quantum-CX-9 Photonics for scale out ( Photonics is another AI bottleneck , which is next-in-line after memory and packaging).
A single Nvidia Rubin NVL144 rack integrates 144 GPUs, equating to a staggering 20,736 TB of HBM 4 memory. So, if you're wondering where all the memory is going, look no further. The reasons behind these massive AI demands are the scale of AI model sizes. As models become larger, the number of parameters and weights associated with them also increases. This creates a demand for fast compute performance in loading model weights, which is why the interface width of HBM is so crucially important, with a rapid interface to keep up with demand when saturated. For example, Moonshot AI's Kimi 2.5 offers 1 trillion parameters in its latest Mixture of Experts (MoE) model, and can only be run in full-fat form on data-center-grade hardware.
Per-bit quantization is also a huge factor in AI deployment. Effectively, an AI model's weights (or 'experts' in an MoE model) are high-precision values, mapped to lower-precision data types. This results in a lower bit-density per-weight, which also affects the amount of VRAM used by the model. Nvidia's NVFP4 format can offer a substantial reduction in memory usage. But, despite efficiency gains thanks to breakthroughs like NVFP4, KVCache, or Deepseek's Engram , the race toward AGI means that frontier model developers want to get their hands on all of the compute power they can get if they want to train, develop, and run the latest and greatest models at scale.
Spending on AI infrastructure (which includes memory and storage chips at an eye-watering scale) could surpass $3 trillion over the next five years. Tech giants like Meta, Amazon , and Microsoft have also dedicated around $650 billion in CapEx in 2026 alone to facilitate these AI capabilities. The long-term outlook as a result of this level of spending remains to be seen, but one thing is clear: not every company that we know today will survive the deep product winter that we're already in.
"Consumer electronics will see a large number of failures. From the end of this year to 2026, many system vendors will go bankrupt or exit product lines due to a lack of memory." Phison CEO Pua Khein-Seng reportedly said in a recent interview. He reportedly added that the soonest we might see reprieve from the ongoing AI onslaught is by 2030 at the earliest, or another decade.
The last helicopters have already left, and the consumer electronics industry, while remaining clearly profitable for a select few, might be unrecognizable once this is all over. Wrap yourself up warm, and arm yourself with as much compute as you reasonably require; it might be a long wait until a new norm is established.
Sayem Ahmed is the Subscription Editor at Tom's Hardware. He covers a broad range of deep dives into hardware both new and old, including the CPUs,\u00a0GPUs, and everything else that uses a semiconductor. ","collapsible":{"enabled":true,"maxHeight":250,"readMoreText":"Read more","readLessText":"Read less"}}), "https://slice.vanilla.futurecdn.net/13-4-16/js/authorBio.js"); } else { console.error('%c FTE ','background: #9306F9; color: #ffffff','no lazy slice hydration function available'); } Sayem Ahmed Social Links Navigation Subscription Editor Sayem Ahmed is the Subscription Editor at Tom's Hardware. He covers a broad range of deep dives into hardware both new and old, including the CPUs, GPUs, and everything else that uses a semiconductor.
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://www.tomshardware.com/tech-industry/SPONSORED_LINK_URL
- https://www.tomshardware.com/tech-industry/ai-demand-reshapes-consumer-electronics#main
- https://www.tomshardware.com
- Rufus devs blame Microsoft for blocking latest Windows 11 ISOs, resulting in IP bans — Windows Insiders can't download latest Insider ISOs and are met with IP b
- Flight Controls Are Cleared for Takeoff on GeForce NOW
- BenQ Mobiuz EX271UZ 27-inch 240 Hz 4K OLED gaming monitor review: BenQ shines with its first gaming OLED monitor
- Corsair AI Workstation 300 review: Strix Halo sets sail in a compact and classy (but pricey) package
- High-end Android phones are now powerful enough to emulate the PC version of Cyberpunk 2077 — YouTuber gets 2020's hottest PC game running at playable frame rat
Informational only. No financial advice. Do your own research.