DeepMind重磅报告《AI in 2030》：5年后AI成本飙升，数据不再是瓶颈

Key points:

• By 2030, the cost of AI training is expected to reach hundreds of billions of dollars, the computing power demand will be thousands of times that of GPT - 4, and the power consumption will reach the gigawatt level.
• Despite facing six major challenges including performance ceilings, data exhaustion, power supply, cost pressure, algorithm efficiency, and computing allocation, current trends indicate that the expansion of AI will continue.
• In four major fields - software engineering, mathematics, molecular biology, and weather forecasting - AI is expected to achieve breakthroughs such as generating code from natural language, assisting in mathematical proofs, predicting protein interactions, and improving weather forecasting accuracy, which is expected to increase scientific research productivity by 10 - 20%.

What will artificial intelligence (AI) look like by 2030? Epoch AI, a non - profit AI research institution, released an 119 - page research report titled "AI in 2030" commissioned by Google DeepMind, specifically answering this question.

The report points out that if the current trend of AI expansion continues until 2030, global investment in AI will reach hundreds of billions of dollars, and the required power will be measured in gigawatts. However, these investments are expected to bring significant productivity improvements in high - value areas such as scientific research.

Computing power consumption will be thousands of times that of GPT - 4! Training costs will reach hundreds of billions of dollars

Epoch AI analyzed in detail in the report the core elements required for the large - scale development of AI: computing resources, capital investment, data reserves, hardware facilities, and energy consumption, and looked forward to the new generation of AI capabilities spawned by them, especially in the field of scientific research, which is also the focus of leading AI developers.

Epoch AI believes that although the expansion of AI depends on unprecedented infrastructure support, its large - scale development is likely to continue until 2030 and bring disruptive changes to science and more fields.

According to the current development trajectory, by 2030, the top - tier AI models will cost hundreds of billions of dollars and consume gigawatt - level power. Despite the huge challenges, the report believes that these obstacles can be overcome. As long as AI can bring corresponding economic returns through productivity improvement, such large - scale investment is reasonable. If the revenue growth rate of AI laboratories remains at the current level, their output benefits will be sufficient to support an investment of hundreds of billions of dollars.

Figure: The upfront investment in purchasing AI cluster hardware is increasing at a rate of 1.9 times per year, and ultra - large - scale clusters worth billions of dollars are under construction.

It is expected that by 2030, AI will be able to generate complex scientific software based on natural language instructions, assist mathematicians in proving intuitive conjectures, and even accurately answer open - ended questions in biological experiments.

These inferences are all based on the continuous progress of existing AI benchmark tests, and relevant tasks are expected to be solved before 2030. Epoch AI predicts that AI capabilities will bring fundamental changes in multiple scientific fields, but full deployment and full utilization of its effectiveness may not happen until after 2030.

Based on current trends, by 2030, the cost of computing clusters for training cutting - edge AI will exceed $100 billion. These clusters can support a computing scale of up to 10^29 FLOP, which is equivalent to running the world's largest AI cluster in 2020 continuously for more than 3000 years. The computing power consumption of AI models trained on this basis will be thousands of times that of GPT - 4, and the required power will also reach the gigawatt level.

Figure: Since 2010, the training computing power of well - known AI models has increased by about 4 to 5 times per year, and leading models show a similar trend.

Six key findings: Will investors be discouraged?

Although the large - scale development of AI faces many challenges, Epoch AI believes that it is very likely that the current trend will continue until 2030. The report systematically evaluated six potential bottlenecks and analyzed them as follows:

1. Will the model performance hit a ceiling?

There is indeed a possibility that the performance of AI systems will stagnate due to scale expansion. However, from the recent breakthroughs in multiple benchmark tests and the significant increase in commercial revenue, there is no clear evidence of a "ceiling". The models are still continuously improving with the increase of parameters and training scale.

2. Will the training data be exhausted?

Human - generated text data is expected to support model training until around 2027. More importantly, the technology of synthetic data generation is becoming increasingly mature. Especially after the emergence of inference models, the effectiveness of synthetic data has been verified. Although the data bottleneck cannot be completely ruled out, it is no longer a fundamental limitation.

Figure: The amount of training data for language models increases by about 2.7 times per year.

Figure: The stock of publicly available human - generated text is very large. However, according to the current trend, if the models are over - trained, the training of cutting - edge AI may exhaust the data before 2030.

3. Can the power supply keep up?

If the scale of AI continues to expand, the power required for training the top - tier AI models in 2030 may reach the gigawatt level. Although the challenge is huge, rapid power deployment can be achieved through solutions such as solar + energy storage and off - grid gas power generation. At the same time, cutting - edge AI training has begun to be geographically dispersed across multiple data centers, which will help relieve local power consumption pressure. It is expected that power will not be a major bottleneck before 2028, and the problem is also expected to be solved later.

4. Will the high cost discourage investors?

The training investment of tens of billions of dollars seems astonishing, but if the revenue of AI companies continues to grow at the current rate, they will be fully capable of covering such investments. Once AI truly achieves a productivity revolution, even if it only penetrates some work processes, the economic value it brings may reach the scale of trillions of dollars, which is sufficient to support continuous investment.

Figure: Large - scale AI developers can already achieve annual revenues of billions of dollars, and their revenue has increased by about 2 to 3 times per year in the past few years. If this trend continues until 2030, the revenue is expected to reach hundreds of billions of dollars.

5. Will the improvement of algorithm efficiency reduce the demand for computing power?

Although algorithm optimization has been ongoing, the improvements have been incorporated into the current computing power growth curve. There is no sign that algorithm breakthroughs will suddenly accelerate. Moreover, efficiency improvements often stimulate the demand for more computing power rather than replacing it.

6. Will computing resources be tilted towards inference?

Currently, AI companies may indeed allocate more computing resources to the model inference stage, for example, to support the actual operation of various inference models and product services. However, at present, the scale of computing power occupied by model training and inference is still relatively balanced.

Figure: The expenditure on computing resources and energy consumption allocation of large - scale AI developers show that the scale of training computing and inference computing is roughly the same. In addition, the overall deployed AI computing power increases by about 2.3 times per year, which is similar to the growth trend of cutting - edge AI training clusters.

Logically, training and inference should expand and grow in tandem. Because more advanced models usually require large - scale training, and these models can support higher - value and lower - cost inference tasks. Therefore, the proportion of inference may increase, but it is unlikely to squeeze the resources required for training.

Epoch AI therefore emphasizes that if the technological evolution and investment rhythm continue along the existing path, AI capabilities are also expected to make corresponding leaps, especially in key areas such as scientific research.

AI will become a "magic assistant" in the scientific research field, and productivity may increase by 20%

Epoch AI pointed out in the report that many leading AI companies have regarded scientific research as a key development direction and analyzed through specific cases how AI can substantially improve scientific research productivity.

Research shows that AI will achieve significant breakthroughs in the scientific research field, especially in disciplines suitable for pure computer training such as software engineering and mathematics. Based on the progress of existing benchmarks, by 2030, AI will be able to generate complex scientific software based on natural language descriptions, assist mathematicians in formal proofs, and even accurately answer complex questions related to biological experiments.

Epoch AI predicts that in the future, most scientific fields will have AI assistants similar to the coding assistants used by programmers today. Although scientific AI assistants need to pay more attention to the sorting and integration of massive heterogeneous literature, rather than being mainly limited to work within a single project like current programming AIs, they will still have three core similar functions: intelligent option recommendation based on context, rapid retrieval of relevant information, and independent completion of small - scale closed - loop tasks.

Taking software engineering as an example, Epoch AI expects that AI can bring a 10 - 20% productivity improvement to scientific research tasks. Even in fields such as mathematics and theoretical biology where automation is more difficult, relevant benchmark tests have shown signs of continuous progress and are expected to make further breakthroughs in the next few years. Although it may take a long time for AI to be fully implemented and have a profound impact, even beyond 2030, Epoch AI firmly believes that multiple scientific fields will witness fundamental changes brought about by AI.

From code generation to protein prediction, AI reshapes the R & D boundaries of four major fields

Epoch AI focused on analyzing four typical fields - software engineering, mathematics, molecular biology, and weather forecasting - in the report to illustrate the progress of AI capabilities. Although these benchmark tests do not fully represent the complexity of each field, they clearly show the progress trajectory of AI technology and reveal which tasks may be automated soon.

The following analyses are all based on the public results of current leading models:

▍Software Engineering

AI has significantly changed software engineering practices through programming assistants and intelligent Q&A tools. It is expected that by 2030, AI will be able to independently fix bugs, meet functional requirements, and even solve well - defined but challenging scientific programming problems.

So, what kind of development scenario does the existing progress reveal for artificial intelligence in the field of software engineering in 2030? Epoch AI integrated three aspects of evidence: the actual application of current AI in software engineering, the development status of benchmark tests, and the unsolved problems and research directions summarized by domain experts.

Overall, these evidences indicate that AI will surely bring profound changes to the face of software engineering, and its impact is already evident. However, there are still major doubts about whether AI truly has the ability to independently complete complex tasks end - to - end in a real - world environment. Although the benchmark test results show that we are rapidly approaching this goal, experts still hold different opinions.

Figure: SWE - Bench - Verified: A code repair benchmark based on real GitHub issues and corresponding unit tests, used to evaluate the model's ability to solve actual programming problems. RE - Bench: A research engineering benchmark simulating engineers' home - based tasks, which usually takes about eight hours for humans to complete.

▍Mathematics

AI is gradually becoming a research assistant for mathematicians, helping to enrich proof ideas or formalize intuitive conjectures. Existing cases show that AI has provided practical help to mathematicians. However, there are still significant differences in the mathematical community regarding the actual relevance of current AI mathematical benchmarks and when AI will be able to independently (rather than assist) produce mathematical results.

Figure: AIME: The qualifying test for the US Mathematical Olympiad, with high - school math questions whose answers must be integers. USAMO: The US Mathematical Olympiad, with high - school math questions that require constructing proof processes. FrontierMath: A benchmark covering mathematically challenging problems of expert - level difficulty, with answers in the form of numbers or simple expressions, which are easy to verify automatically.

▍Molecular Biology

Protein - ligand interaction benchmarks such as PoseBusters are expected to be solved in the next few years, while reliable prediction of arbitrary protein - protein interactions will take longer, and the outcome is still uncertain.

At the same time, an AI literature assistant for biological research is about to emerge, and the