The "Game Within a Game" of Memory Chips
In October 2025, OpenAI signed an agreement with Samsung and SK Hynix under the name "Stargate", locking in a supply of up to 900,000 DRAM wafers per month—roughly 40% of global DRAM output, all claimed by a single buyer in one swoop, as part of its four-year infrastructure plan with a total investment of $500 billion.
Everyone following the memory market saw what was coming next.
Contract DRAM prices jumped 171% after the deal was announced. The impact on the retail market was even more dramatic. A Team Delta RGB 64GB DDR5-6400 memory kit, which sold for $190 in August, now? The same model is listed online at $700. In less than three months, prices surged 268%. Such a steep climb is virtually unheard of in peacetime commodity markets.
Even DDR4—which should be in surplus—has more than doubled in price. Storage devices were also hit: Western Digital’s WD Blue SN5000 1TB SSD rose from $64 to $111, and the 2TB version from $115 to $154.
Demand is only part of the reason.
OpenAI's "Stargate" expansion requires a staggering amount of high-bandwidth memory. Data centers and consumer electronics manufacturers are now competing for the same foundational resources, but the former are bidding much higher—far higher. Factories are shifting advanced DRAM capacity toward AI because data centers' offers dwarf those of consumer device makers.
If supply had stayed stable, maybe the situation could be controlled. Unfortunately, it did not.
Samsung, SK Hynix, and Micron have all shifted production lines toward HBM—high-bandwidth memory specifically for AI accelerator chips. Profits there are far greater than for consumer DDR5. Micron’s entire 2026 HBM output is already booked out. Wafer starts for conventional memory modules are steadily declining, even as PC makers continue to purchase.
So we see: As demand surges, production is simultaneously shifting. This means a short-term deadlock—a “colluded” memory price spike that won’t be solved soon.
This cycle in memory is different from before: the upgrade path for AI infrastructure is fundamentally distinct from that of consumer hardware. Once these data centers are built, as model sizes expand, they will continue to devour memory. The demand curve only goes up.
Faced with this storm, tech giants' responses have sharply diverged, forming two distinct camps.
Sony stockpiled ample memory for PlayStation 5 before prices took off. Industry insiders say its inventory will suffice to maintain console pricing for months if not the entire shortage period. Apple likewise locked in supply ahead of time, and its profit margins are enough to absorb cost increases which would crush thinner-margin businesses. Lenovo went even further—with Bloomberg reporting its memory inventory is about 50% higher than usual, enough to last until 2026.
Microsoft's position is much more passive. Industry analysts believe Xbox consoles will probably have to raise prices again—even though the company already did so earlier this year.
Valve’s timing is even worse. Its once highly anticipated living room gaming PC, "Steam Machine", hit the market when core component costs were two to three times budgeted amounts. The company refused to commit to a final price and directly blamed memory shortages. Any planned value proposition was wiped out by memory costs. Modular laptop company Framework has removed all standalone memory kits from its store.
This division by “foresight” is carving out long-lasting effects in the competitive landscape—even if prices stabilize in the future, those consequences won’t fade quickly. Sony and Lenovo retained their market positions; Microsoft and Valve didn’t escape. Just a few months of supply chain management compressed what should have been years of market share evolution.
A natural question: Seeing prices triple, why don’t Samsung, SK Hynix, and Micron ramp up production to capture more profit?
Ask anyone who lived through the 2018 oversupply—they’ll give you the answer.
Late 2016 to early 2017: memory supply squeezed, prices rose. The three big producers responded—by breaking ground on new wafer fabs in Korea and China, expanding capacities at existing plants. On paper, chasing high prices by expanding supply is totally logical.
But when 2018 arrived, the math failed. Demand stabilized, but all the new capacity came online simultaneously. Chips piled up in warehouses. Samsung’s semiconductor division—once its profit engine—shocked investors with quarterly results. SK Hynix was in equally dire straits. The shadow of oversupply persisted until it gradually lifted in 2019.
That experience deeply shapes every capacity decision these companies make today. It takes years to build and ramp up a memory wafer fab. The wafer allocation choices made in November 2025 will decide the supply landscape of 2027 or even 2028. By then, will AI infrastructure spending still be this hot? No one knows. What if AGI development stalls? What if hyperscale customers realize they overinvested and start canceling orders?
Manufacturers would rather stick with the current shortage than repeat the oversupply. They call it "discipline". But as consumers watch $89 memory kits turn into $310, they might use a different word.
Industry forecasts suggest DRAM and NAND shortages will persist through 2026; optimism for normalcy in 2027 is only partial. Orders for large-capacity nearline hard drives are already backlogged for two years. Distributors have started bundling memory with motherboards just to manage inventory allocation.
Zooming out, a discomforting picture comes into focus:
PC builders and gamers are funding AI infrastructure development—not directly, not by choice, but that's how the system works. OpenAI, Microsoft, Google, Amazon—all hyperscale customers—have driven up memory prices to stockpile training clusters. Manufacturers prioritize these clients. Retail availability keeps shrinking. The teenager saving up for a gaming PC has to pay $400 more for memory that cost just $130 three months ago.
There has never been a public policy debate weighing whether “accelerating AI development” is worth the price of “higher consumer computing costs.” The market simply sends resources to whoever bids higher. And when a single customer can lock in 40% of global supply, the process moves swiftly and violently.
The three companies produce nearly all the world’s DRAM. This level of concentration brings systemic fragility that regulators haven’t adequately addressed. To be frank: it’s time to start examining this situation.
Source: Jinduan
Risk Warning and DisclaimerThe market involves risk; invest cautiously. This article does not constitute personal investment advice, nor does it consider individual users’ special investment objectives, financial status, or needs. Users should consider whether any opinions, views, or conclusions in this article fit their own situation. Invest accordingly at your own risk.