Image source: Generated by Unbounded AI

A year ago, Sequoia Capital defined 2024 as the "primitive chaotic period" of AI, a judgment widely recognized by the industry. Recently, they released their predictions for the AI ​​industry in 2025. As a top investment institution, every prediction of Sequoia deserves attention.

However, after careful study, I found that although this report is insightful, it may be biased in some key judgments. To this end, I fully translated this report and attached my own observations and thoughts at every important node. Let us take a look at what the AI ​​landscape will look like in 2025.

Introduction: From Chaos to Order

In 2024, AI is in The original state of chaos. In 2025, the foundation of AI will gradually become solid.

In January last year, we compared ChatGPT to the "Big Bang" in the history of AI development and predicted that 2024 would be AI's "original chaotic period." The AI ​​ecosystem at that time was full of innovative ideas and potential momentum, and it was a golden period for entrepreneurs. We wrote: "While possibilities abound, they remain murky and require vision to transform these possibilities into tangible, ultimately impactful results."

Today , the AI ​​ecosystem has gradually taken shape. Five "final contenders" have been born in the large model competition. Nvidia's much-anticipated Blackwell chips are coming this month. Many data centers planned for early 2024 are under full construction. TSMC is building new wafer capacity and Broadcom is developing custom AI chips: the entire supply chain is firing on all cylinders. New AI projects are launching in every industry, from medical to legal to insurance.

If 2024 is a chaotic period for AI, then its infrastructure has basically taken shape now. The potential of AI is becoming a tangible reality—in the physical data centers that are popping up across the United States, from Salem, Pennsylvania, to Round Rock, Texas, to Mount Pleasant, Wisconsin. If 2024 was the year of creative ideas, 2025 will be the year of testing whether they actually work.

The following parts are translator's notes

Sequoia's definition of 2024 as a "chaos period" is questionable. In fact, 2023 will be the real period of chaos: at that time, countless large model R&D companies will emerge around the world, and the computing power will be extremely scarce, making it difficult to find a card.China has launched a vigorous "Battle of Hundreds of Models".

In contrast, 2024 shows obvious signs of "clearing out" - the US market has converged to five "final competitors", and the Chinese market has gradually focused on the "Six Little Dragons of AI" . This transformation from "a hundred schools of thought contending" to "survival of the fittest" just shows that the industry is maturing. Rather than saying that 2024 is a period of chaos, it is better to say that it is the beginning of the AI ​​industry entering a rational development stage. [End of translation]

The following are our three major predictions for 2025:

1. LLM (Large Language Model) Suppliers each develop unique strengths – which will lead to a more differentiated competitive landscape in 2025

2024, the heart of the big model competition The goal is to reach GPT-4 level. Five companies achieved or came close to achieving this goal to become the "final contenders": Microsoft/OpenAI, Amazon/Anthropic, Google, Meta, and xAI. Other companies, notably Inflection, Adept and Character, have withdrawn from the competition.

To reach GPT-4 levels, these companies have adopted similar strategies: collect as much data as possible, use the most GPUs for training, and optimize pre-training/post-training architectures to improve performance. In 2024, talent moves frequently between companies, leaving few technical secrets at all.

As companies prepare for the next round of LLM scaling—expected to be 10 times the scale of existing computing—these AI labs are cultivating their unique core strengths. It can be said that they have "selected their weapons" for the upcoming competition. In 2025, these different strategies will lead to different outcomes, with some companies coming out on top while others may fall behind.

- Google – vertical integration: Entering 2025, Google’s advantage lies in the integration of the entire industry chain. Google is the only company with a state-of-the-art chip developed in-house: its TPU is expected to compete with NVIDIA GPUs in 2025. Google also builds its own data centers, trains its own models, and has a strong internal research team. Unlike Microsoft (OpenAI) and Amazon (Anthropic), which choose their partners, Google wins by controlling the entire value chainprofit.

- OpenAI – Brand Advantage: According to multiple surveys, the gap in brand awareness between ChatGPT, Claude and Gemini is quite obvious. OpenAI has undoubtedly the strongest brand in AI. This has helped it become the highest-grossing among the big AI companies, with a reported revenue of over $3.6 billion. If the success of AI ultimately depends on consumer awareness and enterprise promotion capabilities, OpenAI may further widen the gap with competitors.

- Anthropic – Talent advantage: In 2024, a large number of research talents will switch from OpenAI to Anthropic. With the addition of heavyweights such as Jon Schulman, Durk Kingma, and Jan Leike, Anthropic's influence in the research field has grown significantly. The company has also brought on key executives, including Instagram co-founder Mike Kreiger, who serves as chief product officer. Led by GPT-3 inventor Dario Amodei, Anthropic has become the go-to place for AI scientists.

- xAI – Data Center Construction Advantage: As we stated in the article "Steel, Servers, and Power," data center construction is key to the next stage of AI competition. xAI deployed a 100,000-GPU Colossus cluster in record time, setting the industry benchmark for data center expansion. The next goal for xAI and its competitors is to build 200,000 and then 300,000 GPU clusters. If the saying "scale is everything" holds true, xAI will likely maintain its rapid development momentum.

- Meta – Open Source Advantage: Meta already has strong distribution advantages on Instagram, WhatsApp and Facebook, and now it is fully betting on the open source strategy. Meta is the only company among the major competitors to take this route. Meta's Llama model has a large number of loyal users, and the discussion about open source vs. closed source continues. If cutting-edge technology development starts to slow down, Meta will be well-positioned because of its open source model to quickly democratize these technical capabilities.

In the competition of large models, strict execution is crucial. The competitive landscape and the strategic positioning of each player have been clarified. In 2025, we will see which strategies are visionary and which ultimately prove to be the wrong choice.

The following parts are translator's notes

Sequoia's "top five" forecast is too US-centered and too optimistic.

First of all, the global AI competition landscapeIt's far more complicated than this. In the Chinese market, the "AI Six Little Dragons" camp has been formed, represented by GLM, MINIMAX, Dark Side of the Moon, Baichuan Intelligence, Zero One Wish, and Step Star, as well as Baidu Wenxin, Alibaba Tongyi, Tencent The camp of major manufacturers represented by Hunyuan, iFlytek Spark, and Huawei Pangu have unique advantages in local application scenarios and data advantages.

Mistral AI has suddenly emerged in the European market, and its influence in the open source field is rapidly expanding around the world.

More importantly, the competitive landscape in 2025 is likely to shrink further. Even among the current "Top Five" or "Six Little Dragons", not all players can have the last laugh. There are three reasons:

1. Computing power pressure continues to rise: Although each company is expanding computing power, the training cost of top models is still growing rapidly. Taking the rumored scale of GPT-5 as an example, the investment in computing power that may be required will deter some players.

2. Business realization dilemma: Currently, most large model companies are still in the cash-burning stage, and the difference in cash reserves will be highlighted in 2025. Players that fail to find a good business model may be forced to withdraw from the competition, as was the case with Inflection in 2024.

3. The technological generation gap is widening: As cutting-edge players make breakthroughs in specific fields (such as multi-modal, intelligent agents, etc.), the technological gap may further widen. Companies that fail to keep up with the pace of innovation risk being eliminated from the market.

Specifically, the major players may present the following trends in 2025:

North American market:

- OpenAI: Its advantages have been lost to other players in 2024 Competitors quickly caught up with or even surpassed it, and some features even appeared to be copied from Anthropic. Coupled with the loss of a large number of top talents, the founding team has almost disintegrated, and the role of a leader may change from a leader to a follower in 2025.

- Anthropic: Not only has a large number of top talents joined, but its product capabilities have surpassed OpenAI. The launch of innovative features such as Computer Use demonstrates strong R&D strength and innovation capabilities.

- xAI: After completing the infrastructure layout in 2024, it may redefine industry standards with Musk’s signature disruptive innovation attitude in 2025.

Chinese market:

-Zhipu, MINIMAX, Alibaba Tongyi may continue to maintain their leading position

-Dark Side of the Moon, Baichuan Intelligence, Zero Ten Thousand There have been negative rumors about funding, team and other issues in 2024 (some are true and some are false), product iteration has slowed down, and may gradually fall behind

- ByteDance and Tencent, two sleeping giants, are beginning to wake up. With their strong technology accumulation and commercialization capabilities, they may suddenly emerge in 2025

Therefore, it is expected that by 2By the end of 2025, the number of truly competitive large model companies globally may be further reduced to 5-6, which are likely to be:

- North America: Microsoft/OpenAI (focus on commercialization) and Anthropic (The focus is security)

- China: Only 2-3 companies may be able to survive under the dual pressure of computing power and funds

- Europe: Mistral AI may remain through the open source route Competitiveness

- Potential dark horses: There may be 1-2 companies that are currently inconspicuous and suddenly emerge through breakthroughs in specific vertical fields [End of Translation]

2. AI Search It is becoming a killer application - and will be widely popular in 2025

Since the advent of ChatGPT, the industry has been looking for killer application scenarios for AI. Which new user habits will stand the test of time?

In 2024, a variety of applications are being tested on the market, from AI virtual companions to AI rental assistants to voice assistants and AI accounting.

We believe that AI search will become one of the most widespread applications in 2025. Perplexity has grown rapidly since its launch, reaching 10 million monthly active users. OpenAI launched ChatGPT Search in October, further expanding its search capabilities. The Wall Street Journal recently published an article stating that "Google search is a habit of the elderly." Ironically, the challenge comes just as Google is embroiled in antitrust litigation.

AI search is a revolutionary reinvention of the killer application of traditional Internet search. Traditional Internet search is mainly a navigation tool based on web indexing, while AI search is an information tool based on LLM, capable of reading and understanding the semantics of knowledge. This would be a huge efficiency gain for white-collar workers.

AI search may break the current dominance of the search market. In the future, it’s possible that every industry will have its own specialized AI search engine—perplexity for analysts and investors, Harvey for lawyers, and OpenEvidence for doctors. Under the same logic, Midjourney can be regarded as a search engine for the "image world", Github Copilot is a search engine for the "code world", and Glean is a search engine for the "document world". Compared with traditional search, AI search can perform deeper semantic understanding, so the performance is improved by an order of magnitude, bringing significant productivitypromote.

Textual response as a product interface is more complex than it appears. Different text responses have different characteristics. We believe that LLM can achieve true product differentiation in multiple dimensions. Entrepreneurs can build unique product experiences for specific user groups around these characteristics:

- Intention understanding: Through professional field customization, you can Match user needs more accurately. For example, when doctors and patients ask the same question, they need to see different types of answers.

- Professional database: In the professional field, unique data resources are crucial, such as case law needed by lawyers, financial data needed by analysts, weather data needed by insurance companies, etc. In a business environment, accuracy is a fundamental requirement.

- Presentation format: including the level of detail of the answer, whether to use a bullet point list, whether to include multimedia content, whether to indicate the source, etc. For example, accountants and journalists receive information very differently.

- Interface design: code search should be integrated into the IDE, and accounting policy search should be embedded in the accounting software. Semantic search needs to consider the user's workflow and data environment. Different fields require different interface designs.

The new generation of professional AI search engines will be as close as possible to the "thinking mode" of target users. Doctors, lawyers and accountants all think differently. When we become experts in a field, the way we acquire knowledge and make decisions begins to diverge. Doctors need to study the medical literature, lawyers need to study case law, and investors need to analyze financial reports. Each field parses, analyzes information, and makes decisions differently.

AI search is likely to divide the consumer market and the enterprise market. As consumers, our needs are largely the same, which is why ChatGPT is such a success. But as professionals, our needs are different. It is foreseeable that every knowledge worker will use at least two AI search engines every day - one for work and one for other aspects of life.

The following parts are translator’s notes

I have reservations about Sequoia’s judgment that AI search is the next killer application. Judging from the market performance in 2024, AI search seems to have strong demand mainly in the North American market, and the response in the global market has been mediocre. OpenAI's Search product was short-lived, and products such as Secret Towers in the Chinese market also had little buzz. On the contrary, other application directions with greater potential will actually emerge in 2024.

The most striking thing is that AI programming has become the first super APP. Cursor has quickly become the tool of choice for developers with its powerful code generation and real-time assistance functions, while Windsurf has pushed the AI ​​programming experience to a new level. Its innovative interaction methods and accurate code understanding capabilities have won unanimous praise from developers. . These products not only significantly improve programming efficiency, but more importantly, they haveA clear business model and monetization path have been formed. This success has verified the monetization potential of AI in the professional field.

At the same time, the field of AI pictures and videos will show explosive growth in 2024. Chinese companies have performed particularly well in this field. For example, ByteDance successfully solved the technical problem of embedding Chinese fonts, and products such as Kuaishou Keling and Conch AI have received wide acclaim in the global market. Technology in this field is iterating at an alarming rate, product experience continues to improve, and the user base is rapidly expanding.

What deserves more attention is the rise of multi-modal creative tools. Products represented by Jimeng/Jianying integrate AI pictures, videos, audios, editing and other functions, which not only open up the usage threshold for ordinary users, but also redefine the possibilities of content creation. This type of comprehensive tool is likely to be the real killer app of 2025, rather than a single search function. They will reshape the paradigm of content creation, lower the threshold for creation, and unleash the creativity of ordinary users. Traditional search needs may be partially replaced by these smarter, more comprehensive tools. [End of Translation Note]

3. Return on investment is still a problem, but capital expenditures will stabilize in 2025

We have previously discussed the US$200 billion and US$600 billion problems of AI, analyzing the huge capital investment of big technology companies and the dilemma of end-user income being insufficient to cover these investments.

In early 2024, Big Tech companies worried that AI might threaten their oligarchic position in the cloud computing business. As we describe in our article "The Game Theory of AI CapEx," these companies have no choice but to invest aggressively to ensure they stay ahead in the AI ​​era. If they don't invest, other companies will and they will fall behind.

As we enter 2025, the situation has changed significantly. Big Tech has a firm grip on the AI ​​revolution. Not only do they control the vast majority of the data centers that power AI, they also hold significant stakes in major AI model companies and are the largest investors in emerging AI startups.

We expect capex in AI to stabilize in 2025 as confidence among Big Tech companies grows. If 2024 is the year of competition for land and power resources, then 2025 will be the year of project implementation. Construction has already begun and the companies will focus on completing the new project on time and on budget. They will then need to promote these new capabilities to customers and help enterprise customers take full advantage of the new AI capabilities.

Capital expenditures have nearly doubled since before the emergence of ChatGPT and are expected to gradually return to normal in 2025. The latest capital spending data for the third quarter shows that Microsoftand Google investments have leveled off. Amazon and Meta are still ramping up their investments, but may plateau in early 2025. (While Meta appears to be flat in the data, the company has forecast increased capital spending in the fourth quarter).

The market structure of oligopoly will also gradually take shape. Major technology companies are paying close attention to what their competitors are doing. If the industry begins to move into a "new normal," that's good news for everyone involved. This will help the market reach a new equilibrium in 2025, rather than continuing to expand uncontrollably.

As new data centers come online in 2025, AI computing costs are expected to continue to fall significantly. This is great news for startups and will spur new innovation. As we said before, startups are primarily consumers rather than producers of computing resources and therefore benefit from overbuilding. Big Tech is effectively creating a subsidy effect that benefits the entire AI ecosystem.

Cloud computing is often compared to the railroad monopolies of the Gilded Age. If data centers are the rails of the digital economy, then by the end of 2025, the new AI infrastructure will be fully in place. The key questions are: what "goods" will be carried on these "rails", and how can we use this new technology to create value for customers and end users.

Let’s look forward to a year in which AI infrastructure is used to create amazing new capabilities that change people’s lives.

The following part is the translator's note

This prediction is very rational, but it may underestimate the impact of several key factors. The first is geopolitical factors. In the context of increasingly fierce competition in science and technology, governments of various countries may increase investment in AI, thus pushing up the overall investment scale. Especially in the context of the United States further tightening export controls on China, China's domestic computing power supply will face greater uncertainty, which may cause domestic companies to have to increase investment in infrastructure.

Secondly, the development rules of large models are undergoing important changes. As Scaling Law gradually moves from the training stage to the inference stage, the computing power required may further increase. This means that even if the data center has been built, its computing power configuration may need to continue to be upgraded to meet the growing inference demand.

In addition, while infrastructure spending by large technology companies may be stabilizing, AI-related investments in vertical areas, especially in key industries such as healthcare and finance, may only be beginning to accelerate. The focus of investment will gradually shift from infrastructure to application scenarios, and this shift deserves attention. 2025 may see more traditional industry giants begin to deploy AI solutions on a large scale, driving a new round of investment boom.

Therefore, the overall investment trend in 2025 may show "infrastructure investment is relatively stable, but investment at the application level is accelerating.""Characteristics, and affected by geopolitics, the investment pace of different regions may diverge significantly. [End of translation]

Reference: https://www.sequoiacap.com/article/ai-in-2025/