First electricity, then water, and now (basically) all high-tech devices – is there anything that AI isn’t making more expensive?
Don’t misunderstand us – The Fusion Report isn’t against artificial intelligence (AI). We use AI for doing basic research, contrarian research (used to avoid confirmation bias), and basic writing, but in all cases, the work is directed against the structural writing concept that we have developed for an article. We also use AI to help keep track what’s going on in the world, but we digress. This article is one in a series about how AI is negatively impacting the price of different commodities; in this case, Dynamic Random Access Memory (DRAM). DRAM is the memory that programs execute out of as they are running; it is generally the fastest class of memory used in computer systems (excepting for cache memory, which is used sparingly inside and near processors).
So like many people who have worked in our industry, I generally build my own PCs (and fortunately/unfortunately, usually those of my immediate family as well). Last weekend was ordering the parts to build a replacement PC for my oldest son, which totals roughly $1,000 dollars (FYI, not everything needed replacing). Interestingly, nearly half the cost was for the DRAM modules (about $450); three years ago, the same models would have cost (at most) half the price. Which got me interested – why are DRAM prices so high? As you might guess, the answer to that question is AI.
Why is DRAM Critical for Artificial Intelligence?
Most of the high computers today are built using NVIDIA GPUs (Graphics Processing Units) at their heart. Each NVIDIA H200 GPU consists of 16,896 CUDA cores and 528 Tensor Cores. HH200 board also has 141 GB of HBM3E (High Bandwidth Memory Generation 3 Extended), which uses vertically stacked DRAM dies. A typical hyperscale supercluster for AI typically uses 16,000 to 65,000 H200 GPU, which also includes over 2,256TB of memory.
While HBM3E memory and DDR5 DRAM sticks are not interchangeable, from a memory manufacturing perspective they compete with each other. Since memory chip manufacturers can make more money (and profit) from HBM3 e-memory than from DDR5 DRAM sticks, they have prioritized manufacturing HBM3 over DDR5 DRAM. Additionally, DRAM is not used simply for PC memory; it is also used in devices such as cell phones and PC graphics cards. Unsurprisingly, the prices of these devices has also increased significantly over the past 12 to 24 months due to memory chip prices, as have game consoles such as PlayStation 5, Xbox consoles, Nintendo Switch and the Steam Deck OLED.
DRAM Supply and Demand Over The Past 10 Years
To say that DRAM pricing fluctuates over time is a gross understatement. Over the past ten years, DRAM pricing has generally followed a cyclical pattern rather than a smooth decline: periods of sharp drops, brief shortages, and then rebounds. Long-term data show that DRAM got much cheaper for decades, but since about 2010 the pace of decline slowed noticeably, with some datasets suggesting only about a 12% to 15% annual price decrease in that later period. In the last ten years specifically, prices were relatively flat or rising at times around shortages, then fell again during oversupply, and more recently have turned upward in another strong upcycle driven by AI demand and tighter supply.
The sharp swings in DRAM prices over the past 10 years mostly come from a classic memory-chip boom-and-bust cycle: demand and supply are both highly volatile, but supply changes slowly because fabs take a long time to build or retool. When demand jumps faster than manufacturers can add capacity, prices spike; when production catches up or demand softens, the market flips into oversupply and prices fall hard. In the last few years, the cycle has been amplified by a shift toward higher-margin memory for AI systems, which has pulled capacity away from consumer DRAM and tightened supply for ordinary RAM products. Older standards being phased out, plus manufacturers’ preference for more profitable products, have also reduced available supply and made price moves bigger than in a more balanced market.
One could understandably conclude that the DRAM manufacturers don’t really understand their market or cannot control it, but that would be incorrect. In a sense, it is almost like the DRAM supply is fixed (or can only change very slowly), while DRAM demand reacts to larger macroeconomic trends, but that is only half the story. In the case of AI, it is the speculative and massive US build-out of data centers (and roughly 45% of it is in the US) that is by and large the cause of current material and skills shortages. Worse yet, half of the known global pipeline for upcoming additional hyperscale data centers is in the US, putting pressure any number of skills and material markets, including electricity, water, DRAM and other materials. Because of that, we’re in a massive demand boom, which almost certainly will be followed by a similarly-sized demand bust.
Conclusion: We Are All Paying For The AI Gold Rush
Again, there’s nothing wrong with artificial intelligence; it’s the unrestrained and uncontrolled growth of data centers that is a problem, followed by the casino-like atmosphere in which it is played out. And just like the mortgage crisis of 2008, it will be the average consumer that pays the price for all of this; the only difference is that instead of banks being bailed out, it will be tech companies and data center operators. We are already seeing signs: the large layoffs in the tech sector, followed by the rehiring of (new) people when the benefits of AI didn’t pan out as expected. AI is great technology; it is only when it is expected to solve unrealistic problems under unrealistic time frames that it becomes an issue.
