# Why smarter AI models could drive up compute prices 10x **URL:** \ **Category:** Content Creators\ **Tags:** gpu, dwarkesh-patel, machine-learning, economy, investment, nvidia\ **Created:** [3 August 2026 17:36 UTC](https://www.artofsm.art/t/why-smarter-ai-models-could-drive-up-compute-prices-10x/22070 "2026-08-03T17:36:16Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [3 August 2026 17:36 UTC](https://www.artofsm.art/t/why-smarter-ai-models-could-drive-up-compute-prices-10x/22070/1 "2026-08-03T17:36:16Z")
[![](https://www.artofsm.art/uploads/default/original/3X/4/d/4d8581e01624ba85a7f5add7d9a9011182df0a34.jpeg "Why smarter AI models could drive up compute prices 10x") ](https://www.youtube.com/watch?v=oZBGAuANX6I) The video explains that while AI companies’ revenues are skyrocketing, compute capacity grows much more slowly, leading to rising compute prices driven by increased demand for training advanced models and higher profit margins. This dynamic, coupled with efficiency gains and structural limits on expanding compute, is concentrating power among top AI labs and raising concerns about market accessibility and future innovation. ---
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [3 August 2026 17:56 UTC](https://www.artofsm.art/t/why-smarter-ai-models-could-drive-up-compute-prices-10x/22070/3 "2026-08-03T17:56:50Z")
The video discusses the evolving compute landscape for AI labs over the next few years, focusing on the disparity between the rapid revenue growth of companies like Anthropic and the slower growth in compute capacity. Anthropic’s revenue has been increasing tenfold annually, reaching $9 billion last year and potentially hitting $100-150 billion this year, with a speculative target of $1 trillion by next year. However, compute capacity for labs only grows about threefold annually. To reconcile this gap, either lab profit margins must increase, compute prices must rise, or a larger share of compute must be devoted to inference rather than training. All three trends are currently occurring, with inference margins doubling and compute spot prices rising by over 40% since early 2024. Labs prefer not to increase the share of compute spent on inference because it signals a plateau in AI progress, shifting their role from innovators to cloud providers. Instead, they aim to invest most compute resources into training new, more advanced models. This leaves two main ways to bridge the revenue-compute gap: higher margins or increased compute prices. While margins could theoretically rise above 90% if a lab’s model vastly outperforms competitors, such extreme margins are unlikely to persist due to market competition. Therefore, rising compute prices appear to be the more plausible outcome, as evidenced by companies like Google paying double the spot price for specialized GPU rentals from providers like SpaceX. The video further explores the economic implications of smarter AI models that can monetize compute more effectively. If an AI model reaches human-level software engineering capabilities, the value of the compute powering it could increase dramatically—potentially over 15 times current spot prices—because the AI can work continuously and at scale. Contrary to concerns about labor oversupply reducing wages, economic principles suggest that increased innovation and specialization could maintain or even increase the marginal value of compute. This dynamic could lead to higher barriers for competitors who cannot afford the escalating costs of premium compute resources. Another key insight is that more efficient AI models that require less compute to achieve the same results effectively create “more compute” by economizing scarce resources. This efficiency allows labs to charge higher premiums, reinforcing the trend of rising compute costs. As AI capabilities improve, many current applications may become economically unviable because top labs will prioritize high-value uses like automating AI research, willing to pay more for compute tokens than casual users. This concentration of compute and revenue power raises concerns about market dynamics and accessibility. Finally, the video highlights the structural constraints limiting rapid expansion of compute capacity. The annual threefold growth in compute is driven by Moore’s Law, new semiconductor fabrication plants, and reallocating wafer production from other industries to AI. Each of these factors faces significant physical and economic bottlenecks, making it unlikely that compute growth can accelerate substantially in the near term. While future technological breakthroughs, such as robotic chip manufacturing, could eventually reduce compute costs, the current pre-singularity era is characterized by rising compute prices amid soaring AI value. This environment fosters strong economies of scale in AI model development, concentrating power among leading labs and raising important questions about the future of AI innovation and competition.