Cerebras CEO Andrew Feldman on the Future of Data Centres, Token Costs & Memory Shortages
Andrew Feldman (founder/CEO, Cerebras) with Harry Stebbings, one week after **the largest semiconductor IPO ever — priced $185, traded to $311, raised $5.5B+**. The headline thesis is the **inverse-of-a-bubble argument**: in past buildouts (fiber late-90s, 1880s rail), *'the infrastructure buildout was way ahead of demand — if we build it they will come.'* AI is the opposite: *'we can't build data centres fast enough to keep up with demand. We have a $25 billion backlog. Nvidia has a backlog, AMD has a backlog. We're building BEHIND demand, not ahead of it — that is not a characteristic of a bubble.'* On **memory** (the cross-show signal of the week): *'memory is the number-two item in the supply chain after TSMC fab space... if demand stays high we are going to continue to see memory shortages for at least the next several years'* — with prices up 4-5x in certain cases. On **Nvidia's neo-cloud strategy**: *'it has been Nvidia's strategy to create competitors for the traditional hyperscalers — they have funded and backstopped and over-allocated to the neo-clouds, creating a dependence which is probably not healthy.'* On the **OpenAI ↔ Elon compute deal**: OpenAI bought *'down-rev gear — H100s, B200s, a generation and a half maybe two generations behind. It was a good deal for Elon — he had them sitting around — but not the ideal deal OpenAI wanted.'* Endorses Gavin Baker's **metering analogy** — permitting/data-centre delays smooth demand like freeway on-ramp meters. On long-run economics: *'the history of our industry is a massive reduction in cost per unit compute for hard problems — there is no upper bound to how much faster you want to be.'*
Key points
- **The inverse-of-a-bubble thesis is the episode's spine — and the sharpest counter to this week's valuation-skeptic chorus.** Feldman: *'In past bubbles — fiber optics in the late 90s, railroads in the 1880s — the infrastructure buildout was way ahead of demand. AI is the exact opposite. We're building behind demand. We have a $25 billion backlog... that is not a characteristic of a bubble.'* **Direct rebuttal to [the '100x sales, SolarCity on steroids' skepticism in this week's 20VC roundtable](/issues/2026-05-31)** — same network, opposite conclusion, worth holding both.
- **Memory shortage confirmed from the supply side — for 'at least the next several years.'** *'Memory after TSMC, which is right after fab space, memory is number two item'* in the constrained supply chain, with costs up *'4 or 5x in certain cases.'* **This is the third independent confirmation of the memory-cohort trade this week** — alongside [Gerstner's Micron +200% YTD](/issues/2026-05-31) and [Gavin Baker's structural DRAM thesis from Issue 06](/issues/2026-05-24). For a memory long (Micron / SK Hynix / Samsung), a sell-side CEO putting a multi-year duration on the shortage is the strongest fundamental anchor yet.
- **Nvidia's neo-cloud strategy framed as a deliberate — and possibly unhealthy — dependence.** *'It has been Nvidia's strategy to try and create competitors for the traditional hyperscalers. They have funded and backstopped and over-allocated to the neo-clouds. They have created a dependence which is probably not healthy.'* **A pointed read on the CoreWeave / Nebius cohort** — the implication is that neo-cloud demand is partly Nvidia-manufactured, a risk factor under-priced when the GPU cycle eventually normalises.
- **The OpenAI ↔ Elon compute deal was a forced, second-best move for OpenAI.** *'They bought down-rev gear — H100s, B200s. They didn't get the most current, they are a generation and a half, maybe two generations behind. It was a good deal for Elon — he had them sitting around — but not the ideal deal they wanted, it was the deal that was available.'* **Reframes [the SpaceX↔xAI compute narrative from Issue 06](/issues/2026-05-24)**: the celebrated Elon-compute-access story looks, from a competitor CEO's seat, like OpenAI being capacity-constrained into older silicon.
- **Metering as a feature, not a bug — Feldman endorses Gavin Baker's framing.** *'You put meters on a freeway because it makes traffic smoother and avoids hiccups. Gavin used that analogy extremely thoughtfully.'* The permitting/power/data-centre constraints that look like friction actually prevent a demand-supply whipsaw. **Direct cross-reference to [Gavin Baker's Watts-bottleneck thesis on ILTB, Issue 06](/issues/2026-05-24)** — two infra insiders converging on 'the constraint is healthy.'
- **The OpenAI compute-foresight 'superpower' — and whether it's still rewarded.** *'Sam saw exponential growth and wasn't afraid by it — he went out and contracted for power, data centres, hardware. The ability to believe your data in an exponential growth environment out a year or two is a superpower.'* The open question Feldman poses: is that foresight still rewarded now that you can buy compute from Elon on demand? His answer: no — you can't buy the *same* (current-gen) compute on demand, so early commitment still wins.
- **Long-run cost curve: 'no upper bound to how much faster you want to be.'** *'The history of our industry is a massive reduction in the cost per unit compute for hard problems.'* Cerebras's wafer-scale bet is that inference latency/throughput keeps mattering even as per-token cost collapses — the demand for speed is structurally unbounded. **The bull case for specialised inference silicon (Cerebras / Groq) as distinct from the training-GPU cycle.**
Notable quotes
We can't build data centers fast enough to keep up with demand. We have a $25 billion backlog. If demand stays high, we are going to continue to see memory shortages for at least the next several years.
In past bubbles you had a penchant to believe that if we built it, they would come. The infrastructure buildout was way ahead of demand. That was true in railroads, true in fiber optic cabling. AI is the exact opposite. We're building behind demand. That is not a characteristic of a bubble.
It has been Nvidia's strategy to try and create competitors for the traditional hyperscalers. They have funded and backstopped and over-allocated to the neo clouds. They have created a dependence which is probably not healthy.
They bought down-rev gear. H100s, B200s. They are a generation and a half, maybe two generations behind. This was not a great deal. It was a good deal for Elon, he had them sitting around, but it was the deal that was available.
Memory after TSMC, which is right after fab space, memory is number two item in this supply chain. There is such extraordinary growth in demand that it is putting pressure on all parts of the supply chain.
The history of our industry is a massive reduction in the cost per unit compute for hard problems. There is no upper bound to how much faster you want to be.
Themes
- Memory cohort multi-year shortage
- AI bubble debate (behind-demand thesis)
- Data-centre & power constraints
- Nvidia neo-cloud dependence risk
- Specialised inference silicon