What this is

A Reddit user on the LocalLLaMA subreddit ran the numbers: from 2009 to 2025, NVIDIA's consumer-grade GPUs gained an average of 50% per generation. Extrapolating that curve, a next-generation GPU delivering 3x the performance of the RTX 5090 wouldn't land until 2029 at the earliest (assuming 70% annual gains) or as late as 2038 (assuming 30%), with the historical mean pointing to 2030–2031.

He added one telling detail: from the 2080 Ti to the 5090, performance scaled roughly 3.1x while power consumption also doubled. If the next generation triples again, power draw could push 1200 watts—the heat output of a small air-conditioner condenser sitting in your home.

Industry view

For the AI compute market, this back-of-the-envelope projection carries two implications. First, the consumer GPU upgrade cadence is steady but slow, which means the cost of running large models locally is not dropping as quickly as the hype suggests. Data-center-grade AI silicon like the H100 and B200 sits on a different, more aggressive upgrade curve—pricier, more power-hungry, and explicitly tuned for training.

The second implication is power. 1200W sounds like a punchline in a living room, but for a data center it's a very real line item on the electricity bill. We've noted repeatedly that Anthropic and xAI have framed electricity—not chips—as the new bottleneck; this post reinforces that same trend from the consumer edge.

There's a counterargument, though. Optimists argue that surging AI demand will accelerate TSMC's advanced-node roadmap and HBM (high-bandwidth memory, the critical AI chip companion) iteration cycles, dragging consumer GPUs forward in the slipstream. And 50% is just a 16-year historical mean—the AI era could break that average, though semiconductor history offers little precedent for such a step-change.

Impact on regular people

  • For enterprise IT: On-premise deployment of 70B-parameter-class models remains a fringe experiment this year; it only enters "realistic budget evaluation" territory in 2027–2028. Don't park too much capex on self-built inference clusters in the near term.
  • For individual professionals: The "buy a 5090 and run AI at home" window has just stretched out—but cloud API pricing will keep falling, which is good news for business users and a delay for individuals hoping to unplug from the cloud.
  • For the consumer market: The line between gaming GPUs and AI GPUs will keep blurring. If the next flagship really does push to 1200W, ordinary PC cases and power supplies will fail first—a problem that arrives ahead of any performance gain.