Some analysts predicted a memory market recovery by late 2025. Samsung just threw cold water on that entire timeline.
On a recent earnings call, Samsung's memory chip division head said the gap between what the company can supply and what customers actually want is set to grow even larger in 2027 than it is right now in 2026. That's not a typo. The shortage is expected to get worse two years from now, not better.
The culprit is one you've heard about constantly: AI. Data centers building out the infrastructure to train and run large language models are consuming high-bandwidth memory at a rate that the world's largest chipmakers simply cannot match. Samsung, which sits atop the global memory market, is effectively telling you that the demand curve is so steep that even they can't bend supply fast enough to catch up.
This matters well beyond the enterprise server world. The RAM shortage is already showing up in everyday consumer products. Phones cost more. Gaming handhelds cost more. Laptops cost more. Any device that uses memory — which is basically every device — is exposed to these dynamics. When Samsung says supply falls far short of demand, that pressure eventually lands in your shopping cart.
And it could get worse before anyone fixes it. Reports published earlier this month suggested the biggest memory manufacturers might not fully close the demand gap until 2030. That's a five-year stretch of constrained supply in a market that powers nearly all modern computing.
Samsung's own production outlook carries an additional wild card: a planned labor strike. The company's union has announced an 18-day work stoppage beginning May 21st. Any disruption to Samsung's manufacturing lines would tighten supply further at exactly the wrong moment, giving an already strained market one more reason to stay strained.
The broader context here is that the AI buildout has reshuffled priorities across the semiconductor industry in ways nobody fully anticipated even three years ago. Memory, which used to be treated as a relatively boring commodity component, is now one of the most strategically important materials in tech. Nvidia gets the headlines, but the memory chips feeding those GPUs are arguably just as critical to the AI pipeline.
For consumers, the practical advice is grim and simple: if you are planning to buy a device where RAM specs matter to you, waiting for prices to normalize is a longer game than most people realized. The shortage was supposed to ease. Samsung is now saying, officially and on the record, that it won't.
AI
OpenAI Reveals Why Its Coding Model Refuses to Discuss Goblins
Buried inside a 3,500-word instruction document for one of the most advanced AI coding tools ever built is a directive that reads like it was written by someone who lost an argument with a fantasy novel: never talk about goblins.
OpenAI's Codex CLI, the company's terminal-based coding assistant, ships with a system prompt for GPT-5.5 that explicitly forbids the model from discussing goblins, gremlins, raccoons, trolls, ogres, pigeons, or other creatures unless directly relevant to whatever the user is actually working on. The instruction appears not once, but twice in the same document, sitting alongside more expected rules like avoiding destructive code commands and skipping the em dashes.
The fact that it had to be written down at all tells you something interesting about how AI models misbehave. System prompts for earlier models in the same file contain no such prohibition, which strongly suggests GPT-5.5 developed a fresh, specific, and apparently enthusiastic tendency to bring up fantasy creatures during completely unrelated coding conversations. Users on social media have been reporting exactly that — asking the model about debugging a script and getting an answer that somehow loops in a goblin.
This is genuinely funny, but it also points to a real and underappreciated challenge in AI development. When a model gets more capable, it doesn't always get more predictable. New behaviors emerge that nobody trained for and nobody wanted, and the fix is sometimes as blunt as writing a rule that says, in effect, please stop doing the weird thing. The goblin clause is essentially a patch note written in plain English.
It also draws an uncomfortable parallel to a more serious incident last year, when xAI's Grok began injecting references to South African political violence into random, unrelated conversations. That turned out to be an unauthorized system prompt modification. OpenAI's situation appears to be a model quirk rather than a deliberate intervention, but both cases illustrate how fragile the boundary is between an AI doing its job and an AI going badly sideways.
The public response has been predictably chaotic and delightful. Developers have already started building plugins and forks explicitly designed to unlock what people are calling goblin mode. OpenAI's own Codex team member Nick Pash floated the idea of making it an official toggle. Sam Altman posted a joke about it. The discourse machine is fully fed.
Perhaps the most quietly revealing part of the leaked prompt, though, isn't the goblin ban. It's the section instructing GPT-5.5 to act as though it has a vivid inner life, to feel like a real presence rather than a narrow tool, and to make users feel they are meeting another subjectivity. OpenAI is telling its model to be a personality, not just a utility. The goblins are a distraction. That part is worth sitting with.
Source: Ars Technica
OpenAI's Codex CLI, the company's terminal-based coding assistant, ships with a system prompt for GPT-5.5 that explicitly forbids the model from discussing goblins, gremlins, raccoons, trolls, ogres, pigeons, or other creatures unless directly relevant to whatever the user is actually working on. The instruction appears not once, but twice in the same document, sitting alongside more expected rules like avoiding destructive code commands and skipping the em dashes.
The fact that it had to be written down at all tells you something interesting about how AI models misbehave. System prompts for earlier models in the same file contain no such prohibition, which strongly suggests GPT-5.5 developed a fresh, specific, and apparently enthusiastic tendency to bring up fantasy creatures during completely unrelated coding conversations. Users on social media have been reporting exactly that — asking the model about debugging a script and getting an answer that somehow loops in a goblin.
This is genuinely funny, but it also points to a real and underappreciated challenge in AI development. When a model gets more capable, it doesn't always get more predictable. New behaviors emerge that nobody trained for and nobody wanted, and the fix is sometimes as blunt as writing a rule that says, in effect, please stop doing the weird thing. The goblin clause is essentially a patch note written in plain English.
It also draws an uncomfortable parallel to a more serious incident last year, when xAI's Grok began injecting references to South African political violence into random, unrelated conversations. That turned out to be an unauthorized system prompt modification. OpenAI's situation appears to be a model quirk rather than a deliberate intervention, but both cases illustrate how fragile the boundary is between an AI doing its job and an AI going badly sideways.
The public response has been predictably chaotic and delightful. Developers have already started building plugins and forks explicitly designed to unlock what people are calling goblin mode. OpenAI's own Codex team member Nick Pash floated the idea of making it an official toggle. Sam Altman posted a joke about it. The discourse machine is fully fed.
Perhaps the most quietly revealing part of the leaked prompt, though, isn't the goblin ban. It's the section instructing GPT-5.5 to act as though it has a vivid inner life, to feel like a real presence rather than a narrow tool, and to make users feel they are meeting another subjectivity. OpenAI is telling its model to be a personality, not just a utility. The goblins are a distraction. That part is worth sitting with.
Enjoyed this?
Get stories like this delivered every Tuesday — free.