Elon Musk just followed this AI report.

What will artificial intelligence look like in 2030?

Commissioned by Google DeepMind, Epoch released a new report, conducting a detailed analysis from aspects such as computing power, data, and revenue.

The key points are summarized as follows:

• The training cost may exceed $100 billion;
• It will consume several gigawatts (1 gigawatt = 1000 megawatts) of electricity;
• The publicly available text data will be exhausted by 2027, and synthetic data will fill the gap;
• Artificial intelligence is expected to drive comprehensive breakthroughs in the scientific field.

In this regard, Elon Musk is also very concerned and directly commented: "It's obvious."

Let's take a detailed look below.

Costing over $100 billion

The report points out that if artificial intelligence continues to expand at the current trend until 2030, the cost of the computing power clusters for cutting - edge AI will exceed $100 billion.

Such clusters can support training tasks of approximately 10^29 FLOPs, which is equivalent to running the world's largest AI computing power cluster in 2020 continuously for three thousand years.

Meanwhile, the computing power consumption of AI models trained with such clusters will reach thousands of times that of GPT - 4, and the required electricity will be as high as the gigawatt level.

Although these challenges are daunting, they are not insurmountable.

If AI can generate corresponding economic returns by improving productivity, it will be sufficient to justify the large - scale investment of hundreds of billions of dollars, and these investments will be worthwhile.

However, can this trend really continue? The answer is yes.

As can be seen from the report, the views that the development speed will slow down actually lack sufficient basis.

In terms of large - scale expansion and revenue

Although some views suggest that this large - scale expansion may encounter bottlenecks, the report points out that recent AI models have made significant progress in various benchmark tests and revenue.

This bottleneck may appear, but there is currently no obvious evidence that it has occurred.

Taking GPT - 5 and GPT - 4 as examples, they have both achieved significant leaps compared with their previous generations in benchmark tests.

In terms of revenue, in the second half of 2024, the revenue growth rates of OpenAI, Anthropic, and Google DeepMind all exceeded 90%, equivalent to an annualized growth rate of more than three times.

According to the revenue forecasts of OpenAI and Anthropic, the two companies will still maintain a growth rate of more than three times in 2025.

In terms of data

Will the data really be exhausted?

The report points out that the currently publicly available human - generated text can support at least until 2027.

With the emergence of inference models, synthetic data can not only be generated on a large scale, but its effectiveness has also been further verified.

For example, AlphaZero and AlphaProof learned to play chess and solve geometric problems only through self - generated data, and their performance reached or even exceeded the level of human experts.

In terms of electricity

There are now many ways to quickly increase electricity output, such as solar energy combined with battery energy storage or off - grid natural gas power generation.

In addition, the training tasks of cutting - edge AI have begun to be geographically distributed across multiple data centers, which will relieve some of the pressure.

In terms of capital

Some people worry that if the expansion cost is too high, AI developers may choose to slow down their investment.

However, Epoch indicates that if the revenue of AI developers continues to grow at the recent trend, its scale will be sufficient to match the predicted investment of over $100 billion required in 2030.

Although it seems extreme for AI revenue to grow to hundreds of billions of dollars, if AI can significantly improve the productivity of a large number of work tasks, its potential value may be as high as trillions of dollars.

In terms of algorithms

Some views suggest that artificial intelligence development may shift to more efficient algorithms.

In fact, against the background of continuous growth in computing power, the efficiency of algorithms has been continuously improving.

There is currently no particular reason to expect a sudden acceleration in algorithm progress. Moreover, even if this happens, it may instead stimulate a further increase in computing power demand.

In terms of computing power allocation

Some people propose that AI companies may "transfer" the computing power originally used for training to the inference stage.

But the report shows that currently, the scale of computing power consumed by training and inference is comparable, and there are sufficient reasons to suggest that the two should be expanded synchronously.

Even if there is a tilt towards inference tasks, the expansion of the inference scale is unlikely to delay the development process in the training field.

Therefore, Epoch believes that the prediction of extrapolating the current development trend to 2030 is quite convincing, which prompts them to further infer the future ability level of AI.

AI will accelerate scientific research and development in multiple fields

The progress of existing benchmark tests shows that by 2030, artificial intelligence will be able to use natural language to implement complex scientific software, assist mathematicians in formalizing proof sketches, and answer complex questions about biological solutions.

Many scientific fields will have artificial intelligence assistants comparable to the programming assistants of today's software engineers.

Epoch visually demonstrates the increasingly enhanced scientific research capabilities of AI in 4 fields in the form of a chart:

Software engineering

The report points out that according to the current trend, by 2030, artificial intelligence will be able to independently fix problems, implement functions, and solve complex (but well - defined) scientific programming problems.

The two benchmark tests in the above figure are as follows:

SWE - Bench - Verified: A programming benchmark test based on solving real GitHub problems and accompanied by relevant unit tests, which also includes models using private methods, such as Claude Sonnet 4.

RE - Bench: A research engineering benchmark test based on tasks similar to job - seeker homework, which takes about 8 hours for humans to complete.

Mathematics

It can be seen that AI has performed excellently in high - difficulty math competitions such as AIME, USAMO, and FrontierMath.

Therefore, Epoch predicts that AI may soon become a research assistant, helping to perfect proof drafts or mathematical intuitions.

Molecular biology

The report points out that public benchmark tests on protein - ligand interactions (such as PoseBusters) are expected to make breakthroughs in the next few years.

Meanwhile, an AI desktop research assistant in the field of biological R & D is about to appear. The existing benchmark tests on answering questions about biological experimental schemes are expected to be fully resolved before 2030.

Weather forecasting

Currently, AI can already outperform traditional prediction methods in the time range from a few hours to a few weeks. However, with the increase in data volume, its prediction ability is expected to be further improved.

The future challenge lies in further improving the accuracy of existing predictions, especially for predictions of rare or extreme events, and applying the improved predictions to real - world scenarios, thereby bringing more extensive social and economic benefits.

Generally speaking, in 2030, AI is likely to be everywhere, affecting all aspects of our work, study, and life.

So, want to win in the future? First, learn how to use AI.

Reference links:

[1]https://epoch.ai/blog/what-will-ai-look-like-in-2030

[2]https://x.com/elonmusk/status/1968323077315649853

[3]https://www.reddit.com/r/singularity/comments/1niqrsx/epochs_new_report_commissioned_by_google_deepmind/

[4]https://epoch.ai/files/AI_2030.pdf

This article is from the WeChat public account "QbitAI", author: Shiling. It is published by 36Kr with authorization.