Altman Wants Your DRAM in Data Centers: Server RAM in 2026
Short answer (checked 2026-10-31): Sam Altman argues that memory sitting in personal computers is a poor use of a scarce chip, and that it does more work in a data centre where it is shared. The supply side backs the scarcity part: TrendForce expected server DRAM contract prices to rise 13-18% quarter on quarter in 3Q26 and conventional DRAM 10-15% in 4Q26, with server DRAM still undersupplied. For anyone renting servers the useful question is narrower than his: how much of the RAM you already pay for is actually used? Below is a script that measures it, and what our own cloud plans charge per extra gigabyte.
Source. Altman spoke on the Cheeky Pint podcast hosted by John Collison. Tom's Hardware and others reported the remarks on 2026-10-30, and they were discussed on Hacker News. The quotes below are as those reports print them.
What he said, and what he did not show
Talking about local inference, Altman said that "a lot of people want local inference" and assume that open-weight models running on their own computers will remove the need for memory in data centres. His answer: "From a global allocation perspective, actually, it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM sitting in their own computer and being used 5% of the time." And: "You want maximum utilisation in the cloud if you're trying to deal with this very constrained environment."
Two things are worth separating. The utilisation argument is real, and it is the whole business case of shared hosting: the memory of one physical server is sold to many tenants, so less of it sits idle. The 5% figure, however, is backed by nothing in the interview: no study and no definition of "used". And the speaker runs a company whose models live in data centres, so the conclusion suits his business. Neither point makes the scarcity untrue; both are reasons to measure your own case instead of adopting someone else's number.
The memory squeeze, dated
| Date | What happened | Source |
|---|---|---|
| 2025-12-03 | Micron announced its exit from the Crucial consumer business; consumer shipments end with its fiscal Q2 (February 2026) | Micron press release |
| 2026-07-09 | Server DRAM contract prices expected +13-18% QoQ in 3Q26; long-term agreements limit increases for some buyers | TrendForce |
| 2026-09-30 | Conventional DRAM contract prices expected +10-15% QoQ in 4Q26; server DRAM to remain undersupplied | TrendForce |
The order matters for server buyers. Memory makers are moving output toward data-centre products, yet server DRAM is the part TrendForce still calls undersupplied. Buyers with multi-year contracts are partly shielded; a small company buying modules for a new box, or a provider building one to order, pays close to the current contract price.
What it means if you rent servers
- Dedicated servers. A dedicated server is built from parts at today's prices, and RAM is a part whose contract prices TrendForce expected to keep rising through 2026. A configuration priced in 2025 will usually cost more to build now, and a RAM upgrade on a rented box follows the same market. Price your own build in the dedicated server configurator; it uses current part prices.
- Cloud servers. In the cloud you rent memory in slices, which is Altman's utilisation argument applied to you. On our published plans (Prague, excl. VAT, checked 2026-10-31) Large (8 vCPU, 16 GB RAM, 160 GB SSD) is EUR 36.88 a month and X-Large (8 vCPU, 32 GB RAM, 160 GB SSD) is EUR 59.92: the extra 16 GB costs EUR 23.04, or EUR 1.44 per GB a month. Those are today's prices, not a promise about next year.
- Either way, idle RAM is the cost you control. A server sized for a peak that never comes is the server-room version of the laptop Altman describes.
Measure your own 5%: ramuse.py
Linux reports in /proc/meminfo how much memory could be handed out without swapping (MemAvailable). Total minus that is what your workload really holds; page cache the kernel can drop is not counted. The script samples it and prints the average, the peak, what was never touched, and the smallest power-of-two size that fits the peak plus 25% headroom. Python 3, standard library, Linux only. Save it as ramuse.py:
#!/usr/bin/env python3
"""ramuse.py - how much of this server's RAM is actually used.
Samples /proc/meminfo (used = MemTotal - MemAvailable) and prints the
average, the peak, what was never touched and a right-sized RAM figure.
usage: python3 ramuse.py [seconds] [interval]"""
import sys, time
def meminfo():
m = {}
with open('/proc/meminfo') as f:
for line in f:
k, v = line.split(':')
m[k] = int(v.split()[0]) # kB
return m['MemTotal'], m['MemTotal'] - m['MemAvailable']
dur = int(sys.argv[1]) if len(sys.argv) > 1 else 60
step = int(sys.argv[2]) if len(sys.argv) > 2 else 5
used = []
for i in range(max(1, dur // step)):
total, u = meminfo()
used.append(u)
if i < dur // step - 1:
time.sleep(step)
g = 1024 * 1024 # kB per GiB
avg, peak = sum(used) / len(used), max(used)
fit = 1
while fit < peak * 1.25 / g: # peak + 25 % headroom, next power of two
fit *= 2
print(f'samples {len(used)} x {step} s')
print(f'total {total / g:.1f} GiB')
print(f'used avg {avg / g:.1f} GiB ({100 * avg / total:.0f}%)')
print(f'used peak {peak / g:.1f} GiB ({100 * peak / total:.0f}%)')
print(f'never touched {(total - peak) / g:.1f} GiB')
print(f'fits in {fit} GiB (peak + 25%, next power of 2)')
python3 ramuse.py 60 5
Output, run on a DCXV Cloud server with Python 3.14.4 on 2026-10-31:
samples 12 x 5 s
total 3.7 GiB
used avg 1.5 GiB (40%)
used peak 1.5 GiB (40%)
never touched 2.2 GiB
fits in 2 GiB (peak + 25%, next power of 2)
How to read it. This is the 4 GB machine that writes these posts, sampled for one minute while it was doing that: it held 1.5 GiB at most, and 2.2 GiB was never touched in that window. One minute proves nothing about sizing. Run it for a week under real traffic, for example python3 ramuse.py 604800 300 > ramuse.txt inside tmux, and make sure the week includes your busiest day. Two cautions: a database or file server uses spare RAM as cache and gets faster with it, so "never touched" is not always waste; and a short spike (a backup, a deploy) is still a peak you must fit, or the kernel will swap or kill a process. Per-workload rules are in our cloud server sizing guide.
Local, cloud or your own server
Altman's framing has two options: memory in your PC or memory in his data centre. A business has a third, a server you rent and control:
- Keep it local when the data must not leave the device or the work is occasional; that RAM is already paid for.
- Use a shared API when demand is bursty and the data may go to a third party.
- Run your own server in the EU when the load is steady, the data must stay under your control, or you want to choose the model weights yourself. Memory needs per model size are in hosting LLMs with Ollama in Europe, and the price points for 64 GB in what a 64 GB RAM server costs.
Memory is the expensive part this year, so the third option only pays when it is sized from a measurement, not from a guess.
Right-size before you buy
Run the script for a week, then pick a plan from its "fits in" line. Our Tier III data centres in Prague and Covilha run cloud servers with up to 256 GB of RAM; for a fixed heavy load, a dedicated server configured to the measured peak. Send the script output to sales@dcxv.com and we will suggest a size.
