The race to artificial general intelligence just hit a significant reality check. According to Tim Dettmers, a research scientist at the Allen Institute and assistant professor at Carnegie Mellon University, the processors powering today’s AI boom simply aren’t up to the task and our ability to make them better is running out of steam.
Dettmers pulls no punches in his recent blog post, calling the thinking around AGI and superintelligence “not just optimistic, but fundamentally flawed.” His argument? While everyone’s busy philosophizing about what AGI could do, they’re ignoring a critical problem: it has to actually run on something.
Why Future GPUs Will Deliver Incremental Progress at Extravagant Costs?
That something primarily GPUs may have already peaked. Dettmers predicts we have maybe one, possibly two more years of meaningful scaling before we slam into physical limitations that no amount of engineering cleverness can overcome.
Here’s the uncomfortable truth: GPUs maxed out their performance-per-cost sweet spot back in 2018. Since then, the industry has been surviving on what Dettmers calls “one-off features that exhaust quickly.”
Think about Nvidia’s progression over the past several years. The gains we’ve celebrated, Ampere’s BF16, Hopper’s FP8, Blackwell’s FP4, have come from using lower precision data types and tensor cores. Each time precision was halved, computational throughput roughly doubled. Impressive on paper, but it masks a less inspiring reality.
Strip away these precision tricks and look at raw computational power generation-over-generation, and the picture changes dramatically. From Ampere to Hopper, BF16 performance increased by 3x, but power consumption jumped 1.7x. The leap from Hopper to Blackwell? Performance went up 2.5x, but required double the die area and another 1.7x increase in power.
These aren’t the exponential improvements the AI industry has been banking on. They’re incremental gains that come at increasingly steep costs.
Hardware, AGI, and the US-China Divide
Dettmers acknowledges that stitching GPUs together more efficiently might buy us a bit more time. Nvidia’s GB200 NVL72 system demonstrates this approach, expanding from eight GPUs in a 10U box to 72 in a rack-scale system, delivering a 30x boost in inference performance and a 4x improvement in training over comparable Hopper setups.
But this isn’t a long-term solution. Better rack-level hardware optimizations might extend the runway, but Dettmers estimates that advantage will evaporate somewhere between 2026 and 2027.
Despite his skepticism about AGI timelines, Dettmers doesn’t think the hundreds of billions being poured into AI infrastructure is inherently unreasonable. The growing demand for inference actually using AI models justifies significant investment.
The catch? If model improvements can’t keep pace with hardware deployment, all that expensive infrastructure could become a massive liability.
Even if we somehow solve the hardware scaling issue, Dettmers points to another fundamental challenge: true AGI would need to operate in the physical world, not just the digital one. That means robotics, which faces its own scaling nightmares.
“Data in the physical world is just too expensive to collect, and the physical world is too complex in its detail,” he writes. Training AI to manipulate objects, navigate spaces, and perform physical tasks requires resources that dwarf what’s needed for language models or image generation.
Is the American Fixation on AGI a ‘Compelling Narrative’ Masking Practicality?
The divergence between American and Chinese approaches to AI couldn’t be starker. While US labs remain fixated on whoever builds AGI first winning some imagined arms race, China has taken a more pragmatic route, focusing on applications of AI that exist today and deliver measurable value.
Dettmers sees China’s strategy as far more sensible. “The key value of AI is that it is useful and increases productivity,” he notes. Chasing a potentially unattainable goal does nothing to generate real economic benefits.
Perhaps most tellingly, Dettmers suggests that predictions of imminent AGI persist “not because they are well founded but because they serve as a compelling narrative.” It’s a story that attracts funding, talent, and attention even if the underlying physics and engineering realities tell a different tale.
As we hurtle toward 2026 and 2027, we’ll find out whether Dettmers is right about hitting the wall. If he is, the AI industry will need to recalibrate its expectations and, more importantly, its strategy. Building useful AI might not be as sexy as promising AGI, but it could be the only realistic path forward.
