The question worth sitting with right now is not whether AI is improving. It is whether the business models built on AI capability scarcity will survive long enough to matter.
Saagar Pateder, writing on spateder.com on June 11, argues that frontier-class open-weight models capable of running on consumer hardware are roughly three years away. The essay is not a forecast in the technical sense. It is an extrapolation from two compounding trends that have not bent yet, and the honest reader has to acknowledge that the recent track record makes the directional argument harder to dismiss than it was a year ago.
The first trend is that open-weight models currently lag frontier models by roughly four months on benchmarks, not two years. The gap that once looked like a structural moat is now a gap that closes on a quarterly schedule. The second trend is that efficiency techniques, specifically quantization, distillation, and architectural improvements, keep compressing the hardware footprint needed to run a given capability level. Frontier performance is moving up. Hardware requirements per unit of capability are moving down. The intersection of those two lines lands, per Pateder’s analysis, at a Claude Fable 5-class model running locally on a device with 16GB of RAM by early 2029.
This week provided supporting evidence on both sides of that intersection. MiMo Code, Xiaomi’s open-weight coding model, ships competitive performance at a fraction of the parameter count that would have been considered necessary twelve months ago. DiffusionGemma, Google’s latest open-weight release, continues the pattern of frontier labs publishing capable models that run outside their own infrastructure. A purpose-built LLM assembled from scratch for roughly $80 in compute costs was reported this week, a data point that quantifies just how aggressively efficiency is compressing the floor on model production cost.
Pateder’s frame for the enterprise implication is direct: if the 90th percentile of businesses are currently spending $7,200 per employee per year on AI, the ROI calculation for switching to a free local model or a substantially cheaper open-weight alternative becomes very easy to run. The unknowable part is which workloads will continue to justify frontier pricing. Life sciences, advanced legal analysis, and high-stakes engineering decisions are candidates. The other 80 percent of enterprise white-collar work may not be.
The labs are aware of this. The counter-strategy is to stay far enough ahead that the capability gap never fully closes on the tasks that matter most. OpenAI, Anthropic, and Google are all investing in the next generation of frontier capability precisely because today’s frontier is becoming tomorrow’s open-weight baseline. The subscription model survives as long as the lead time is long enough to charge for it.
What the Pateder essay adds is a rough timeline. If 2029 is approximately correct, the window for building a durable capability moat is narrower than most enterprise AI contracts assume. A three-year infrastructure commitment signed today could expire into a world where the capability it was built around is freely available locally, with no API dependency and no per-token cost.
This is a directional argument, not a certified prediction. Trends bend. Hardware constraints could reassert themselves. Regulatory friction could slow open-weight distribution in key markets. But the direction, frontier capability diffusing onto consumer hardware through compounding efficiency gains and a shrinking open-weight lag, is well-supported by the observable evidence of the past eighteen months.
Any product team currently pricing on capability scarcity rather than workflow integration, reliability, or data-handling guarantees should run the math on what their pricing model looks like if the core capability becomes a commodity by the end of the decade.
Analysis by Saagar Pateder, published on spateder.com on June 11, 2026.
