Stack Constraints · L0 · Full Supply Chain
Wafers bind now, memory confirms it, power is mid-term, and EUV becomes the 2030 ceiling
NVIDIA's decision to ship Vera Rubin at half its planned memory configuration is the clearest live signal that wafer and HBM capacity is already rationing product today, not in some future cycle. Power generation and electrical gear are real but trailing constraints: turbine backlogs and transformer lead times of 128+ weeks bite hardest in 2027-28, not now. EUV lithography is the structural ceiling that closes in around 2030, when ASML's roughly 100-machine annual output can no longer keep pace with accelerator wafer demand.
◆ The move this quarter
Now Jun 2026HBM blinked. Power didn't.
The two chokepoints everyone watched in 2025 — memory and packaging — both eased to 'Watch.' Power generation and electrical gear got worse. The constraint is now firmly physical and downstream. See the full picture →
Binding now
binding Now Jun 2026
Three layers rose this quarter; constraint is travelling.
Open constraints dashboard →Constraint matrix · then→now · click to open constraints
Four lenses — each opens with a claim, not a label
Models
Active parameter counts have plateaued while scale-up domains are set to expand 8x, signaling a shift from model size to interconnect as the frontier lever.
Frontier models have held active parameters per token near 280–288B since 2023, even as NVLink domain sizes jump from 72 GPUs today to 576 in 2027. For lab and app teams, this means the next wave of capability gains is more likely to come from larger coherent compute domains than from raw parameter scaling—and hardware roadmap decisions made now will determine who can exploit that shift.
Open lens →
Benchmarks
Coding benchmarks tripled in 18 months. Compute constraints did not.
The capability curve, measured. 86 models, 647 sourced scores across 7 labs — self-reported and independently verified side by side. The data that answers: how fast is the frontier actually moving?
Open lens →
Chips
Seven generations, one trajectory: more memory, bigger racks.
H100 to Rubin Ultra in five years. The number to watch isn't teraflops — it's the coherent rack domain, the GPUs that behave as one. It's what makes a trillion-parameter MoE trainable at all.
Open lens →
Power
Solving the chip shortage created the power shortage.
Every constraint the industry clears downstream shoves the bottleneck one layer closer to the wall socket. Watch it travel from packaging (2024) to power generation (2026) — and see what's standing in line behind it.
Open lens →
Demand
Is the spend being earned back? That's the whole ballgame.
Hyperscaler capex is a ~$340B/yr wager that AI demand is real. The bubble tell was never the spending — it's whether app-layer revenue is closing the gap. Here's that gap, over time.
Coming soon