John Collison is the co-founder of the Stripe platform and has a podcast called Cheeky Pint where senior executives from tech companies stop by to talk about their stuff. A recent example is Susan Lee, Mater’s CFO. Sam Altman, for his part, needs no introduction, and a few days ago he stopped by Collin’s program to talk about a variety of topics. Among them, one that keeps you up at night: Each user can set up a local AI.
It indicates a waste of resources, but in the background “I am starving.”
In short. They talked about a number of things in the podcast, but what didn’t go unnoticed in the network was the OpenAI boss’ comments about current technology where users are adopting the open model and running it on their own hardware rather than relying on someone else’s cloud. For Altman, local assumptions are a terrible waste of scarce resources like DRAM memory:
“Local estimation is a terrible allocation of scarce resources. A lot of people want local estimation and somehow assume that the problem is solved because we have open source models that we can run on our computers and then don’t need all the things that data centers allow.
But from a global allocation point of view, it’s actually pretty terrible that everyone has their own computer with DRAM that’s only used 5% of the time when we live in a time where it’s a finite resource.
In environments where that critical supply is so limited, you want maximum density in the cloud. But, somehow, people’s perception is the opposite.”
Long story short: According to Altman, if you have local AI, you’re wasting resources that could go into a data center’s capacity.
DRAM stuff. The 5% figure is Altman’s and that’s why he speculates that it’s a waste to have something like this in everyone’s home, when in a very low supply environment, what’s wanted is maximum usage that can only be achieved in a data center. Of course, this is their point of view, since on the other side of the coin there are people who have their “own AI” thanks to the fact that open models already solve the problems they want to solve, and in addition, they want more privacy than what is achieved using ChatGPT.
From the perspective of OpenAI’s CEO, this statement makes perfect sense because (and it’s something that affects all of us and especially consumers) in an environment where Micron, Samsung or SK Hynix, which dominate global DRAM production, can’t increase that production at the rate that hyperscalers need, all global capacity should currently go to data centers, and not home-deemed equipment.
No one could have guessed it five years ago when prices were rock bottom, but DRAM is a critical and scarce commodity right now, with prices rising between 90% and 95% in the first quarter of 2026, according to TrendForce… and showing no signs of abating in the short term. Quite the opposite.
context. And that comment came at a point in the conversation that discussed computing deficits. Altman maintains that the market is not responding as it should and that the large memory manufacturers, despite the huge gains they have made over the past twelve months, have maintained a conservative investment relative to the situation.
The big three have moved to enable new production plants, even limiting production of consumer chips so that all equipment is equipped with the HBM memory needed for data centers, but this is exactly one segment where it’s not easy to change things overnight and open new high-performance memory plants.
Let them say China, which has been running as an alternative in the DRAM segment for months, but right now more than HBM, they’re positioned in the consumer memory segment because of those demands for cutting-edge chips that I mentioned a moment ago.
solution. In addition to being generous and not buying components that OpenAI could use, Altman suggests that new actors enter the supply chain as a remedy.
For example, OpenAI is building its own chip with Broadcom (Jalapeño) and, although they will still need DRAM, building a custom chip allows everything to be much more optimized. SpaceX is in the same situation, as are Amazon, Meta, Microsoft or Google which have somewhat more experience with TPUs.
the niche. Despite everything, OpenAI launched gpt-oss in August 2025, its first open weight since GPT-2 designed to run natively, and Altman himself said in January of that year that OpenAI was wrong about open source. Of course, there was no current DRAM crisis at the start of 2025.
Ultimately, my question here is whether Altman considers the segment of users using native AI to be so large and whether they really hurt the supply chain so much when we see people not paying for AI because what is offered for “free” already meets their needs.
Anyone who has a machine capable of running local models knows very well why they do it, and it could be because of specific needs, being able to use it offline, or simply because they don’t trust the security of a company like OpenAI.
Picture | Cheeky Pint (edited)
In Xataka | Nvidia wants RTX Spark to mark the before and after on PC: after seeing it in action, it doesn’t seem crazy
