The global RAM shortage isn’t coming from where many gamers think it is. Prices are rising, availability is tight, and rumours are flying about delayed GPUs and next-gen consoles. But according to Dave Eggleston, a veteran memory industry executive who previously worked at SanDisk and Micron, the real causes are structural and cyclical, and are amplified by panic.
Speaking on the Broken Silicon Podcast from Moore’s Law Is Dead, Eggleston offered a rare inside look at how the DRAM industry actually works, why AI has upended it, and why some of the loudest headlines don’t mean what they seem to.
For decades, most of the tech industry pulled from the same general pool of memory. PCs, servers, and mobile devices all relied on similar DRAM generations. That changed over time.
First, phones moved to LPDDR. Then, data centres, driven by AI workloads, moved hard toward HBM (High Bandwidth Memory). HBM is now essential for AI accelerators and GPUs, especially those sold by Nvidia.
That shift matters because HBM is very different from standard DDR4 or DDR5. It’s far more profitable, but it also uses more wafer space, requires complex stacking, and is more complicated to manufacture. When Micron, Samsung, and SK Hynix dedicate wafers to HBM, those wafers aren’t making consumer RAM.
Eggleston stresses that this isn’t about greed or poor planning. It’s basic economics. NVIDIA and large AI customers demand massive volumes of HBM and are willing to pay for it. Memory makers follow the money, and the result is less capacity left for RAM in PCs, consoles, and graphics cards.
A lot of frustration has been aimed at GPU vendors or console makers. Eggleston says that’s misplaced.
The bottleneck isn’t Nvidia, Sony, Microsoft, or even retailers. It’s the fact that HBM consumes so much silicon and manufacturing effort that every wafer shifted to AI tightens supply elsewhere. Even a small percentage change in allocation can ripple through the entire market.
This is why DDR5 pricing feels stubborn and why availability swings so fast. The DRAM market is extremely sensitive. Small changes in supply or demand don’t balance gently. They overshoot.
Recent reports claimed that OpenAI locked up as much as 40% of the world’s HBM supply through deals with Samsung and SK Hynix. Eggleston is openly sceptical.
He points out that long-term agreements in the DRAM industry rarely work the way press releases suggest. Memory is cyclical. When shortages hit, companies sign big, dramatic deals. When the market softens, those deals get renegotiated, delayed, or quietly abandoned.
Eggleston compares it to high-profile fab announcements that sound impressive but never fully materialise. Announcements are easy. Following through is harder.
There’s also a key detail often missed: OpenAI didn’t buy finished HBM modules. They bought wafers.
Those wafers can sit in nitrogen-purged storage for long periods. OpenAI can decide later how to package them, who will assemble them, or even resell them if demand cools. This is standard practice in the semiconductor world and gives buyers flexibility rather than absolute control over supply.
One theory floating around is that OpenAI could use its wafer stockpile to pressure Nvidia. Eggleston dismisses that outright.
Samsung and SK Hynix have deep engineering and executive relationships with Nvidia. They wouldn’t allow a customer to weaponise memory supply against one of their most important partners. If OpenAI tried, deliveries would be renegotiated or stopped.
In short, the leverage doesn’t work the way social media imagines it does.
You may have seen claims that RAM lead times are stretching to 13, 18, or even 24 months. Eggleston says those numbers are misleading.
Under normal conditions, DRAM lead times are about three to six months. In a real crunch, they might double or triple. When suppliers quote something like 52 weeks, it’s often a polite way of saying, “You’re not a priority.”
The real damage comes from panic ordering. Companies place multiple overlapping orders, hoard what arrives, and cancel the rest later. That destroys visibility for memory makers, who then react defensively. The cycle feeds itself and makes the shortage feel worse than the physical limits alone would suggest.
Probably not.
Internal planning points to PS6 manufacturing around mid-2027, and Eggleston agrees that it’s impossible to predict shortages that far out. The industry doesn’t have meaningful visibility beyond a few quarters.
Sony and Microsoft will secure memory supply well in advance of production ramps. Even if 2025 is painful, it doesn’t automatically spill into hardware launching years later.
Partner guidance for Nvidia’s RTX 5000 Super series has slipped several times, now floating around a tentative 2026 window. Eggleston cautions against reading too much into that.
Those dates are placeholders, not firm commitments. NVIDIA adjusts timelines constantly based on market conditions, competition, and internal priorities. Slipping guidance doesn’t mean Nvidia is staring at a guaranteed RAM wall years in advance.
Eggleston is cautiously optimistic.
Between 2026 and 2030, Samsung, Micron, and SK Hynix will be building eight to ten major DRAM fabs. Unlike some headline-friendly announcements in other parts of the industry, these fabs are real. The challenge is time.
Fabs take years to reach volume. Equipment from ASML, Tokyo Electron, and Applied Materials is in short supply. DRAM scaling has slowed from roughly 30% gains per year to closer to 10%.
The bigger long-term shift is 3D DRAM. Like NAND before it, DRAM will eventually scale vertically. That won’t immediately boost bandwidth, but it will dramatically increase density and lower costs for high-capacity memory. As cheaper, denser RAM moves into mainstream devices, traditional 2D DRAM capacity frees up for higher-performance uses.
Eggleston expects that combination of new fabs and 3D DRAM to ease the shortage well before mid-2027.
The RAM crisis is real, but it’s not a permanent disaster. AI and HBM have reshaped the memory industry faster than many expected, and panic behaviour has made the situation feel worse than it is.
OpenAI’s big memory headlines are more PR than ironclad control. Console launches years away aren’t doomed. GPU timelines are fluid, not prophetic.
The pain through 2025 is real. But relief is already being built, one wafer and one fab at a time.
Speaking on the Broken Silicon Podcast from Moore’s Law Is Dead, Eggleston offered a rare inside look at how the DRAM industry actually works, why AI has upended it, and why some of the loudest headlines don’t mean what they seem to.
Why RAM suddenly feels scarce
For decades, most of the tech industry pulled from the same general pool of memory. PCs, servers, and mobile devices all relied on similar DRAM generations. That changed over time.
First, phones moved to LPDDR. Then, data centres, driven by AI workloads, moved hard toward HBM (High Bandwidth Memory). HBM is now essential for AI accelerators and GPUs, especially those sold by Nvidia.
That shift matters because HBM is very different from standard DDR4 or DDR5. It’s far more profitable, but it also uses more wafer space, requires complex stacking, and is more complicated to manufacture. When Micron, Samsung, and SK Hynix dedicate wafers to HBM, those wafers aren’t making consumer RAM.
Eggleston stresses that this isn’t about greed or poor planning. It’s basic economics. NVIDIA and large AI customers demand massive volumes of HBM and are willing to pay for it. Memory makers follow the money, and the result is less capacity left for RAM in PCs, consoles, and graphics cards.
Why gamers are blaming the wrong companies
A lot of frustration has been aimed at GPU vendors or console makers. Eggleston says that’s misplaced.
The bottleneck isn’t Nvidia, Sony, Microsoft, or even retailers. It’s the fact that HBM consumes so much silicon and manufacturing effort that every wafer shifted to AI tightens supply elsewhere. Even a small percentage change in allocation can ripple through the entire market.
This is why DDR5 pricing feels stubborn and why availability swings so fast. The DRAM market is extremely sensitive. Small changes in supply or demand don’t balance gently. They overshoot.
The OpenAI memory deal isn’t what it sounds like
Recent reports claimed that OpenAI locked up as much as 40% of the world’s HBM supply through deals with Samsung and SK Hynix. Eggleston is openly sceptical.
He points out that long-term agreements in the DRAM industry rarely work the way press releases suggest. Memory is cyclical. When shortages hit, companies sign big, dramatic deals. When the market softens, those deals get renegotiated, delayed, or quietly abandoned.
Eggleston compares it to high-profile fab announcements that sound impressive but never fully materialise. Announcements are easy. Following through is harder.
There’s also a key detail often missed: OpenAI didn’t buy finished HBM modules. They bought wafers.
Those wafers can sit in nitrogen-purged storage for long periods. OpenAI can decide later how to package them, who will assemble them, or even resell them if demand cools. This is standard practice in the semiconductor world and gives buyers flexibility rather than absolute control over supply.
No, OpenAI can’t strong-arm Nvidia with HBM
One theory floating around is that OpenAI could use its wafer stockpile to pressure Nvidia. Eggleston dismisses that outright.
Samsung and SK Hynix have deep engineering and executive relationships with Nvidia. They wouldn’t allow a customer to weaponise memory supply against one of their most important partners. If OpenAI tried, deliveries would be renegotiated or stopped.
In short, the leverage doesn’t work the way social media imagines it does.
About those scary lead times
You may have seen claims that RAM lead times are stretching to 13, 18, or even 24 months. Eggleston says those numbers are misleading.
Under normal conditions, DRAM lead times are about three to six months. In a real crunch, they might double or triple. When suppliers quote something like 52 weeks, it’s often a polite way of saying, “You’re not a priority.”
The real damage comes from panic ordering. Companies place multiple overlapping orders, hoard what arrives, and cancel the rest later. That destroys visibility for memory makers, who then react defensively. The cycle feeds itself and makes the shortage feel worse than the physical limits alone would suggest.
Will the PS6 or next Xbox be delayed?
Probably not.
Internal planning points to PS6 manufacturing around mid-2027, and Eggleston agrees that it’s impossible to predict shortages that far out. The industry doesn’t have meaningful visibility beyond a few quarters.
Sony and Microsoft will secure memory supply well in advance of production ramps. Even if 2025 is painful, it doesn’t automatically spill into hardware launching years later.
What about Nvidia’s RTX 50 Super delays?
Partner guidance for Nvidia’s RTX 5000 Super series has slipped several times, now floating around a tentative 2026 window. Eggleston cautions against reading too much into that.
Those dates are placeholders, not firm commitments. NVIDIA adjusts timelines constantly based on market conditions, competition, and internal priorities. Slipping guidance doesn’t mean Nvidia is staring at a guaranteed RAM wall years in advance.
When does relief actually come?
Eggleston is cautiously optimistic.
Between 2026 and 2030, Samsung, Micron, and SK Hynix will be building eight to ten major DRAM fabs. Unlike some headline-friendly announcements in other parts of the industry, these fabs are real. The challenge is time.
Fabs take years to reach volume. Equipment from ASML, Tokyo Electron, and Applied Materials is in short supply. DRAM scaling has slowed from roughly 30% gains per year to closer to 10%.
The bigger long-term shift is 3D DRAM. Like NAND before it, DRAM will eventually scale vertically. That won’t immediately boost bandwidth, but it will dramatically increase density and lower costs for high-capacity memory. As cheaper, denser RAM moves into mainstream devices, traditional 2D DRAM capacity frees up for higher-performance uses.
Eggleston expects that combination of new fabs and 3D DRAM to ease the shortage well before mid-2027.
The bottom line
The RAM crisis is real, but it’s not a permanent disaster. AI and HBM have reshaped the memory industry faster than many expected, and panic behaviour has made the situation feel worse than it is.
OpenAI’s big memory headlines are more PR than ironclad control. Console launches years away aren’t doomed. GPU timelines are fluid, not prophetic.
The pain through 2025 is real. But relief is already being built, one wafer and one fab at a time.