heading · body

YouTube

CTO Insights Ep. 1 | the infrastructure crunch | Data#3

Data#3 Limited published 2026-06-02 added 2026-06-05 score 5/10
tech hardware ai supply-chain memory infrastructure
watch on youtube → view transcript

ELI5/TLDR

The world is running short on computer memory, and the reason is AI. AI needs a special, fast kind of memory, so the factories that make memory chips are switching their production lines over to make that fast kind instead of the ordinary kind that goes in everything else — laptops, phones, servers, even fridges. Now both kinds are scarce, prices are up several times over, and nobody, not even the big cloud providers, has a clean way out.

The Full Story

Two chief technology officers — Graham Robinson of Data#3 and Carl Salder of Cisco for Australia and New Zealand — sit down to talk about why hardware suddenly got expensive and hard to find. The whole thing is short and the answer is simple.

One kind of memory is eating the other

There are two flavours of memory in this story. The ordinary kind, called DRAM (the chips that hold data while a computer is working), goes into nearly everything. And a faster, premium kind, called high bandwidth memory, or HBM, which is what AI systems crave because they need speed and low delay.

The catch is that the same factories make both, and a factory can only run so many lines. HBM sells for more and earns the manufacturers more profit, so they are quietly shifting their lines toward it.

“They’re starting to shift their manufacturing lines to produce more HBM… at the expense of the more generic DDR memory that we use.”

The result is a double shortage. There was already not enough HBM. Now, by chasing it, the makers have starved the ordinary memory too — and they still can’t make enough HBM to satisfy demand anyway.

What it does to buyers

Prices on the same gear have gone up three to five times. Budgets that used to buy a lot now buy a little. So the conversation inside companies becomes triage: which part of the business genuinely needs the hardware first, and where does the money go.

“You might be now paying 3x or 5x for the same gear… You have to then be very particular about where you’re going to be applying that investment.”

No clean escape

The obvious instinct is to flee to the cloud — let someone else own the hardware. But the cloud providers buy from the same constrained supply. If everyone runs to the cloud at once, the cloud runs short too.

“You’re kind of caught between a rock and a hard place here because everyone is constrained, even the cloud providers.”

Salder’s practical advice is unglamorous: stay close to your vendors and check in often, because availability changes week by week. Even Cisco’s own delivery plans shift constantly with supply.

Key Takeaways

  • The shortage traces back to one root cause: AI demand for high bandwidth memory (HBM).
  • Memory factories are converting production lines from ordinary DRAM (DDR) to higher-margin HBM, creating a shortage of both at once.
  • HBM supply still can’t meet demand even after the shift.
  • DDR memory sits in almost everything — laptops, phones, servers, network gear, even appliances — so the squeeze is broad.
  • Prices for the same hardware have risen roughly 3x to 5x.
  • Cloud is not a guaranteed escape; providers draw on the same constrained supply.
  • Vendor relationships matter more than usual right now — availability and delivery timelines shift week by week.

Claude’s Take

This is episode one of a corporate interview series, and it shows. Two vendors talking shop, no slides, no numbers beyond a hand-wavy “3x to 5x,” and it wraps in about five minutes. Treat it as a clear, honest framing of a real problem rather than a deep analysis.

That said, the framing is genuinely useful and correct. The HBM-cannibalizes-DRAM dynamic is the actual story behind the 2025-26 memory crunch, and Salder explains it in plain terms without pretending Cisco has a magic answer. The admission that even the cloud is constrained, and that his own delivery plans change weekly, is more candid than most vendor talk.

What’s missing is everything that would make it actionable: how long the shortage lasts, which workloads to deprioritize, whether to buy now or wait, any view on alternative architectures. It diagnoses the weather without forecasting it. A 5 — accurate and well-explained, but thin, and structurally an ad for a series.