What happened
Kerman Kohli published an essay titled "We're going to run out of compute" on May 20, 2026, arguing that aggregate compute demand is accelerating far faster than commonly modelled. Kohli reports that AI research labs are roughly doubling compute year-over-year and that infrastructure and application workloads can increase 10-100x every 2-3 years, which he describes as a parabolic trajectory. He references Micron's recent earnings cadence and analyst estimates to show expectations of flat memory revenue beyond FY2027, and he notes that SK hynix opened its M15X plant in Cheongju in February 2026 in the context of increased supply-side activity.
Technical details
Per Kohli's essay, the core claim is demand growth rather than a single technical bottleneck; he links the pressure to both large-scale training workloads and rapidly expanding production workloads in applications. The piece does not present new benchmarking or primary measurements; it synthesizes industry anecdotes, market charts, and public corporate milestones.
Editorial analysis
Industry context: Companies and labs that rapidly scale training and serving fleets tend to drive compounding demand for GPUs, memory, interconnect, and power. Observed patterns in comparable periods of hardware scarcity show lead times, bidding pressure, and accelerated capex for fabs and data centers, which can amplify market volatility.
What to watch
Indicators an observer can track include GPU spot pricing and lead times, memory utilization and pricing, fab ramp schedules (for example, new plants like M15X), and public capex guidance from major cloud and chip vendors. Kohli does not provide an operational plan from any company; his essay is a synthesis of trends and market signals.
Key Points
- 1Compute demand is likely compounding faster than many forecasts, driven by both training and production workloads.
- 2Supply responses take years, so short-term mismatches between demand and fab/data-center capacity can cause price and availability shocks.
- 3Practitioners should monitor GPU and memory spot prices, cloud capex announcements, and fab ramp schedules as leading indicators.
Scoring Rationale
The essay raises a notable operational risk for ML practitioners: sustained, faster-than-expected compute growth can affect cost and scheduling. The piece is an opinion synthesis rather than new data, so its importance is notable but not industry-shaking.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
