AI Compute Power

Extracted Attributes

• Colossus Hardware Cost

  $7,000,000,000

• Energy Efficiency Trend

  1.34x annual increase in computational performance per watt

• Performance Growth Rate

  Doubling every 9 months

• Primary Measurement Unit

  FLOPs (Floating Point Operations per Second)

• Colossus Power Requirement

  300 megawatts (MW)

• Power and Cost Growth Rate

  Doubling every 12 months

• Colossus Supercomputer Chip Count

  200,000 AI chips

• Projected Global AI Power Demand (2030)

  327 gigawatts (GW)

Timeline

• Earliest recorded AI models in compute trend tracking. (Source: MIT FutureTech)

  1957-01-01

• National AI Initiative Act (US) charges the National Science Foundation with developing shared research infrastructure for compute. (Source: AI Now Institute)

  2020-01-01

• Global data center capacity reaches 88 GW. (Source: RAND)

  2022-01-01

• United States passes the CHIPS and Science Act to bolster domestic semiconductor fabrication. (Source: AI Now Institute)

  2022-08-09

• Several companies deploy AI supercomputers with over 100,000 chips. (Source: Epoch AI)

  2024-01-01

• Global AI data centers projected to require 10 GW of additional power capacity. (Source: RAND)

  2025-01-01

• xAI's Colossus identified as the most performant AI supercomputer globally. (Source: Epoch AI)

  2025-03-01

• Projected global AI data center power demand reaches 68 GW. (Source: RAND)

  2027-01-01

• Individual AI training runs projected to require up to 1 GW in a single location. (Source: RAND)

  2028-01-01

• Leading AI supercomputers projected to require 9 GW of power and 2 million chips. (Source: Epoch AI)

  2030-06-01

XAI (company)

X.AI Corp., doing business as xAI, is an American company working in the area of artificial intelligence (AI), social media and technology. Founded by Elon Musk in 2023, the company's flagship products are the generative AI chatbot named Grok and the social media platform X (formerly known as Twitter), the latter of which they acquired in March 2025.

Web Search Results

• AI's Power Requirements Under Exponential Growth - RAND

  Larger training runs and widespread deployment of future artificial intelligence (AI) systems may demand a rapid scale-up of computational resources (compute) that require unprecedented amounts of power. In this report, the authors extrapolate two exponential trends in AI compute to estimate AI data center power demand and assess its geopolitical consequences. They find that globally, AI data centers could need ten gigawatts (GW) of additional power capacity in 2025, which is more than the total power capacity of the state of Utah. If exponential growth in chip supply continues, AI data centers will need 68 GW in total by 2027 — almost a doubling of global data center power requirements from 2022 and close to California's 2022 total power capacity of 86 GW. [...] Given recent training compute growth, data centers hosting large training runs pose a particular challenge. Training could demand up to 1 GW in a single location by 2028 and 8 GW — equivalent to eight nuclear reactors — by 2030, if current training compute scaling trends persist. The United States leads the world in data centers and AI compute, but exponential demand leaves the industry struggling to find enough power capacity to rapidly build new data centers. Failure to address bottlenecks may compel U.S. companies to relocate AI infrastructure abroad, potentially compromising the U.S. competitive advantage in compute and AI and increasing the risk of intellectual property theft. [...] More research is needed to assess bottlenecks for U.S. data center build-out and identify solutions, which may include simplifying permitting for power generation, transmission infrastructure, and data center construction. ## Key Findings ### Exponential growth in AI computation is driving unprecedented power demands that could overwhelm existing infrastructure Global AI data center power demand could reach 68 GW by 2027 and 327 GW by 2030, compared with total global data center capacity of just 88 GW in 2022. Individual AI training runs could require up to 1 GW in a single location by 2028 and 8 GW by 2030, although decentralized training algorithms could distribute this power requirement across locations.

• What drives progress in AI? Trends in Compute - MIT FutureTech

  To understand how this surge in computing power is possible, it is essential to consider the key hardware innovations that have driven these advancements. Central to the rise in AI model performance are specialized processors like GPUs and TPUs, which are specifically designed to handle parallel computations. GPUs, in particular, originally developed for graphics, have become the backbone of AI development, offering a significant performance boost over traditional CPUs by allowing many computations to be processed simultaneously. [...] ‍ ## Progress in compute and its effect on AI As Figure 3 shows, the computing power used by AI models has increased dramatically over time. ‍ The horizontal axis represents the years, from 1957 to 2021, when different AI models were published. The vertical axis shows the amount of computing power these models use, measured in "FLOPs". These are shown in Logs: Log 0 is 1 FLOPs (1 calculation per second), Log 10 is 10 billion FLOPs, Log 20 is 100 quintillion FLOPs. The green dots represent individual AI models. As we move to the right (more recent years), these dots climb higher and higher, showing that newer models generally require much more computational resources. [...] ‍ ## Could AI contribute to hardware progress? Recent developments in AI have both been caused by computing power improvements and have increased the pace of computing power growth. As these models grow in complexity and usefulness, we can expect them to also help accelerate the growth of compute. For instance, AI algorithms can analyze and optimize chip layouts, potentially discovering more efficient designs than human engineers. At the same time, the need for more computing power reshaped the microarchitecture of various computing devices due to the introduction of AI-oriented compute units, such as Tensor Cores in the NVIDIA GPUs.

• Computational Power and AI - AI Now Institute

  Compute power is a key facet of the emerging industrial policy frame in AI. Nations seeking a competitive advantage in AI are investing heavily in semiconductor development and undercutting their adversaries through strict export control regimes that seek to limit access across the compute supply chain on national security grounds. United States The United States CHIPS and Science Act of 2022 was the first major industrial policy measure passed in tech in recent history, focused on growing a national US-based semiconductor fabrication industry. Prior to the passage of the act, the US produced about 10 percent of the world’s supply of semiconductors. The new Act includes measures such as these: [...] Compute is also an important element in many countries’ national strategies on AI research and development—not just the United States’. National governments have made investments on the order of hundreds of millions of dollars to increase access to compute for academic researchers and home-grown start-ups (by contrast, Amazon recently announced a $35 billion investment in its data centers in the US state of Virginia alone).50 At present, even these massive investments remain insufficient to compete with those of leading industry players, which are drawing on reserves orders of magnitude larger. But this push to “democratize” access to compute proceeds from the knowledge that compute-intensive research is largely dominated by industry, even in academic settings: in recent years, the [...] The National AI Research Resource is another compute-related policy proposal. The National AI Initiative Act of 2020 charged the National Science Foundation with developing a proposal for a shared research infrastructure that would expand access to computational power, high-quality data, and other support to facilitate more widespread AI development.166 The final NAIRR report articulates that a minimum of 18 providers should be part of the NAIRR to ensure resilience and prevent capture. However, under the current conditions of compute scarcity, the design of the NAIRR allows for—and would likely necessitate—some kind of licensing contract with one of the large cloud infrastructure providers, leading to critiques that the NAIRR falls short of its promise to “democratize” AI development

• Trends in AI supercomputers | Epoch AI

  Power requirements and hardware costs of leading AI supercomputers have doubled every year. Hardware cost for leading AI supercomputers has increased by 1.9x every year, while power needs increased by 2.0x annually. As a consequence, the most performant AI supercomputer as of March 2025, xAI’s Colossus, had an estimated hardware cost of $7 billion (Figure 2) and required about 300 megawatts of power—as much as 250,000 households. Alongside the massive increase in power needs, AI supercomputers also became more energy efficient: computational performance per watt increased by 1.34x annually, which was almost entirely due to the adoption of more energy-efficient chips. [...] ## Computational performance, energy, and cost trends The computational performance of leading AI supercomputers has doubled every 9 months, driven by deploying more and better AI chips (Figure 1). Two key factors drove this growth: a yearly 1.6x increase in chip quantity and a yearly 1.6x improvement in performance per chip. While systems with more than 10,000 chips were rare in 2019, several companies deployed AI supercomputers more than ten times that size in 2024, such as xAI’s Colossus with 200,000 AI chips. [...] If the observed trends continue, the leading AI supercomputer in June 2030 will need 2 million AI chips, cost $200 billion, and require 9 GW of power. Historical AI chip production growth and major capital commitments like the $500 billion Project Stargate suggest the first two requirements can likely be met. However, 9 GW of power is the equivalent of 9 nuclear reactors, a scale beyond any existing industrial facilities. To overcome power constraints, companies may increasingly use decentralized training approaches, which would allow them to distribute a training run across AI supercomputers in several locations. ## Locations and public/private sector share of AI supercomputers

• Why AI uses so much energy — and what we can do about it

  ## What makes AI model training so resource-intensive? AI model training involves training, or adjusting, billions of parameters through repeated computations that require immense processing power. This process demands high-performance computing (HPC) infrastructure, consisting of thousands of GPUs and TPUs (tensor processing units, a specialized chip that improves the speed of machine learning tasks) along with CPUs, all running in parallel. Each training session can take weeks or months, consuming massive amounts of electricity.