What will computer chips look like in 2045?

Full analysis

Intel's chief technology officer, Pushkar Ranade, told Dwarkesh Patel the thing nobody selling AI wants said out loud: the world's chip factories can roughly double leading-edge output this decade, no matter how much money you throw at them. AI's appetite for compute has grown about 3× a year for ten years. Two does not equal three. That gap is the whole story, and it decides what the models you rent actually cost in 2027 and 2028.

This is barely a decision for the reader, because there's no button to press. It's a planning fact. The question underneath is whether capable AI gets cheaper on the curve everyone has penciled in, or whether a physical ceiling bends that curve up. Nothing here is undoable by you. Everything here is a reason to check your assumptions about pricing before you sign a two-year commitment.

The Compute Pragmatist. Follow the machines. Arande's ceiling isn't about fabs or capital. It's about ASML, the one Dutch company that makes the extreme ultraviolet lithography machines that print the smallest features. Fewer than 100 of those get built a year. Order a batch today and the lead time is about 18 months, so meaningful new capacity lands around 2028. That's the hard number. And he says memory is tighter than compute right now: the high-bandwidth memory stacked next to every AI chip, made mostly by SK Hynix and Samsung, is the squeeze that bites first. Your next GPU cluster may be gated by memory you can't buy, not silicon you can't afford.

The Skeptic. Before anyone panics: a ceiling on leading-edge wafers is not a ceiling on useful AI. Most of the 3×-a-year demand is for training ever-bigger frontier models. Inference, the part the reader actually runs, keeps getting cheaper through smaller models, quantization, and better serving, none of which needs the newest node. The scary quote is also a sales quote. Arande runs a foundry that wants AI labs to pre-pay for capacity, and his $100-billion-of-model-revenue-from-a-few-billion-of-capex line is a pitch for exactly that. True or not, it's designed to make you commit. Treat the ceiling as real and the urgency as motivated.

The Enterprise Buyer. This reshuffles who gets guaranteed silicon, and you are not at the front of the line. When output can't grow fast and hyperscalers can eat a whole fab's output alone, the scarce supply goes to whoever pre-commits with long-horizon capital: Google, Microsoft, Amazon, OpenAI, xAI. Elon Musk's "Terafab" with Intel is that logic made concrete. He'd rather own the factory than queue for it. For everyone renting from the big clouds, that means the people you buy from are vertically integrating their supply and you're on the spot market behind them. Price and availability of frontier capacity will move with their allocation decisions, not yours.

The Builder. What changes on Tuesday? Nothing, and that's the point. If you're serving features on mid-tier models, this news should push you harder toward efficiency you control: smaller models, caching, and squeezing more out of each token, because the cheapest path to more capacity is needing less of it. The one concrete bright spot is packaging. Hybrid bonding, stacking chips with connections under one micron apart instead of the old 25-plus, cuts the energy cost of moving data between chip and memory. For inference, where memory bandwidth is usually what's slow, that helps, and it's shipping now rather than a decade out. NVIDIA's Blackwell already does it.

Where they split

The real disagreement is whether the ceiling matters to you or only to the labs. The Compute Pragmatist sees a physical wall through 2028 that caps the whole industry. The Skeptic says that wall sits almost entirely on frontier training, while the inference you run keeps falling in price on older nodes and better software. Both can be right: training frontier models gets more supply-constrained and more expensive, while serving good-enough models gets cheaper. Those two curves pulling apart is the thing to plan around.

The second split is who pays for scarcity. The Enterprise Buyer reads vertical integration as a warning: the clouds secure their own silicon first. The Builder reads it as a nudge to stop depending on raw frontier capacity you can't guarantee and get efficient instead.

What it hinges on

One fact decides it: does leading-edge wafer and memory supply actually stay capped near 2× through 2028, or does ASML ramp faster and break the ceiling? Arande, who has every reason to overstate scarcity, says it holds. ASML's own machine output is the thing to track, because it's public and it's upstream of everyone. The council leans toward the constraint being real but narrow: it bites frontier training and HBM-hungry clusters. The everyday inference most readers buy is a different market. Before you sign a long GPU commitment, price in the chance that top-tier capacity gets tighter and more expensive into 2028, and keep a fallback to smaller models you can serve on hardware that isn't rationed.

The call

Prediction: At least one of the three largest HBM makers (SK Hynix, Samsung, Micron) will report on an earnings call that its high-bandwidth memory output is fully sold out or allocated through 2026, reported by the end of the Q1 2027 earnings season in early February 2027.

Confidence: Medium. Arande names memory as tighter than compute, and HBM capacity is pre-booked years out.

Why: Arande, who exited both DRAM and NAND and watched both explode, says the memory crunch is worse right now than the compute crunch, and HBM is the specific part that sits next to every AI accelerator. Unlike logic, where new nodes add some headroom, HBM capacity is set by how many stacking and packaging lines these three makers have, and those are booked far ahead. For HBM to be freely available, AI accelerator demand would have to cool sharply or a maker would have to bring on a line nobody has announced. Neither is in view. SK Hynix has already signaled multi-year sold-out positions, so the mechanism is visible, not speculative.

Revisit by 2027-02-15: We're right if SK Hynix, Samsung, or Micron states on a late-2026 or Q4 earnings call that HBM is sold out or fully allocated through 2026. We're wrong if all three report available HBM capacity or soft pricing for 2026.