Hardware Ceilings and Chip Intake Limits (Part 1)
Hardware Ceilings and Chip Intake Limits (Part 1)
The sense that AI progress is accelerating is grounded in data. METR's time horizon measurements — the length of task an AI can finish at 50% success where a human expert would take that long — have been compressing. The long-term trend was roughly 7 months per doubling. More recent windows show intervals closer to 4 months, with some internal lab signals even shorter.
That compression is real. It comes from better models and, increasingly, from the harnesses wrapped around them.
Agent + Harness, Then Meta-Harness
Capability is no longer just the model. The reusable scaffolding — tools, memory policies, sandboxes, recovery, evaluation hooks — multiplies what a given model can do reliably. When that scaffolding itself becomes editable by the agents running inside it, you get a meta-harness loop.
Early results from self-modifying harness work show large wall-clock speedups on complex tasks even with frozen weights. At AI-native teams, large fractions of merged code are now AI-authored, and experienced groups report 5-8× effective developer multipliers.
These are localized, measurable gains in software engineering velocity.
Compression Math
Project a constant ratio forward (each doubling interval ~4/7 the last). The sequence of intervals drops quickly:
| Step | Doubling interval | Approx. days |
|---|---|---|
| 0 | 7 months | 213 |
| 1 | 4 months | 122 |
| 2 | ~2.3 months | 70 |
| 3 | ~1.3 months | 40 |
| 4 | ~3 weeks | 23 |
| 5 | ~13 days | 13 |
| 6 | ~1 week | 7 |
The sum of the infinite geometric series starting from a 7-month baseline at 4/7 ratio converges in roughly 16 months. That is finite-time convergence to extremely short doubling intervals for the measured tasks, not an instantaneous jump.
This is still software-task doubling. It does not automatically translate to general, novel, cross-domain intelligence.
Main comparison (identical recreation using Recharts npm library from original Plotly data ported from upcoming/):
Bounded S-curve (identical recreation using Recharts):
Original interactive HTML versions available in /blog/charts/*.html (ported from upcoming/ directory).
Evaluation Wall
Self-improvement loops hit a structural limit. Novel architectures generated by the loop are hard for the same (or lightly varied) judge models to score accurately. Reflexion-style single-model loops re-apply their own biases. Research on automated failure attribution and two-player evaluation setups shows that judge quality degrades on truly new output.
Around the 4-to-6 month mark under these curves, progress either slows while better judges or diversity mechanisms are built, or the curve bends earlier.
Physical Intake Constraint
Even if the software curve keeps compressing for a while, the total stock of usable compute does not. Production of advanced AI accelerators is measured in the low millions of high-end equivalents per year globally, with the largest share going through constrained advanced packaging (CoWoS-class capacity). New fab capacity and packaging lines take years to bring online at scale. Energy delivery, cooling, and grid connections add further lags.
Installed base grows at the rate of new production plus replacement, minus retirements. Under current trajectories, that is a steady but not explosive increase in raw substrate — perhaps low-to-mid single-digit multiples over a few years at optimistic ramps, not orders of magnitude overnight.
A system whose capability per FLOP (or per deployed accelerator) improves only modestly will therefore see its overall reach limited by that physical stock and the energy budget attached to it. The logistic ceiling L in the growth model is set by:
- Annual chip intake rate
- Energy available for training and inference at scale
- Rate at which diverse, high-quality interaction data (especially physical) can be collected
- Speed of real-world actuator and robotics loops
What the Hardware-Bound Path Delivers
By roughly month 6-8, software engineering agents are likely to be near-superhuman inside well-instrumented codebases for generation, refactoring, debugging, and migration work. Productivity multipliers for teams that adopt strong harnesses will be large.
General reasoning on novel problems, long-horizon scientific discovery, and fluid physical-world competence improve much more slowly. They remain gated by the rate at which new silicon and new real data enter the system.
Pure speculation here: this path still produces enormous value in digital domains. It also clarifies some of the practical priorities worth watching: making every deployed accelerator count through better scaffolding, explicit memory, supervision, and evaluation that stays reliable as output distribution shifts.
Guessing at the implications, the primitives that seem to matter are durable external memory that doesn't bloat the core, verifiable loops, efficient execution on available hardware, and interfaces that keep humans in the loop at the right level.
These are the same kinds of things that would let systems degrade gracefully when the absolute frontier is gated by production schedules.
Part 2 examines the alternative: what changes if scaling down (smaller, more efficient reasoning cores) and energy wins compound inside the same chip intake and fab reality.
Numbers draw from publicly discussed METR time-horizon trends and rough public estimates of advanced AI accelerator production and packaging capacity. The trajectory framing is scenario analysis, not a forecast. Actual outcomes depend on whether evaluation and efficiency mechanisms clear the next set of walls.