Key Takeaways:
The dawn of the AI industrial era: 2025 marks the official transition of AI from "academic exploration" to an "industrial infrastructure" era. AI investment has become a core driver of US GDP growth, with global annual capital expenditure surpassing $300 billion. Trillion-dollar investments and energy anxieties: Tech giants (such as OpenAI, Microsoft, Amazon) have initiated super data center plans like "Stargate," with individual investments often reaching hundreds of billions of dollars. Power supply has become a hard constraint, leading tech companies to restart nuclear power plants (like Three Mile Island) to secure computing demands. Reasoning models and agentification: Reasoning models represented by OpenAI's o1 and DeepSeek-R1 have become mainstream, endowing AI with "multi-step thinking" capabilities. "Agentic Coding" has exploded, with AI agents now capable of independently handling complex software development tasks, significantly improving programming efficiency. Sky-high compensation reshapes the talent market: Top talent's value rivals that of sports stars, with giants like Meta offering compensation packages as high as $300 million over four years.
On the 26th, renowned AI scholar Andrew Ng stated in his annual letter and a special issue of "The Batch" that 2025 may be remembered as the dawn of the AI industrial era. This year, model performance reached new heights through reasoning capabilities, infrastructure construction became a key force driving US GDP growth, and top tech companies engaged in an unprecedented compensation war to compete for talent. Ng believes that as technology becomes more tightly integrated into daily life, the new year will further solidify these transformations. Trillion-Dollar Capital Expenditure and Energy Challenges Andrew Ng indicated that in 2025, tech giants led by OpenAI, Microsoft, Amazon, Meta, and Alphabet announced a series of staggering infrastructure investment plans. According to various disclosures, the construction cost for each gigawatt of data center capacity is approximately $50 billion. OpenAI and its partners announced the $500 billion "Stargate" project, with plans to eventually build 20 gigawatts of capacity globally. Microsoft's global data center expenditure reached $80 billion in 2025 and signed a 20-year agreement to restart the Three Mile Island nuclear reactor in Pennsylvania by 2028 to ensure a continuous power supply. These massive investments also face real-world challenges. Bain & Co. estimates that to support construction on this scale, annual AI revenue would need to reach $2 trillion by 2030, exceeding the combined profits of major tech giants in 2024. Furthermore, insufficient grid capacity has left some data centers in Silicon Valley idle. Reported by the Financial Times, Blue Owl Capital withdrew from negotiations in mid-December to provide $10 billion in data center financing for Oracle and OpenAI due to concerns about debt levels. Sky-High Compensation Reshapes the Talent Market As AI transforms from an academic interest into a revolutionary technology, the value of top talent has skyrocketed to levels comparable to professional sports stars. Andrew Ng stated that Meta shattered traditional compensation structures in 2025, offering packages to researchers from OpenAI, Google, and Anthropic that included cash bonuses and massive equity grants, with some four-year contracts valued as high as $300 million. Mark Zuckerberg personally participated in this talent war, successfully recruiting key researchers like Jason Wei and Hyung Won Chung from OpenAI. Andrew Tulloch, who previously co-founded Thinking Machines Lab with Mira Murati, ultimately also joined Meta. In response, OpenAI offered new employees more aggressive stock option vesting schedules and retention bonuses of up to $1.5 million. The Proliferation of Reasoning Models and Agentic Coding Andrew Ng stated that 2025 is seen as the first year of widespread adoption for reasoning models. Starting with OpenAI's o1 model and followed by DeepSeek-R1, these models demonstrated the ability to perform "chain-of-thought" reasoning through reinforcement learning (RL) fine-tuning. This allows models to engage in multi-step thinking before generating output, significantly improving performance on mathematical, scientific, and programming tasks. For example, OpenAI's o4-mini, when combined with tool use, achieved a 17.7% accuracy rate on a challenging multimodal understanding test. This technological advancement directly fueled the explosion of "Agentic Coding." By the end of 2025, tools like Claude Code, Google Gemini CLI, and OpenAI Codex could handle complex software development tasks through agent workflows. On the SWE-Bench benchmark, coding agents based on the latest large language models could complete over 80% of the tasks. Although research from Apple and Anthropic pointed out that reasoning models still have limitations in certain complex logics and that the reasoning process increases inference costs, this has not stopped companies from leveraging AI to automatically generate code and reduce development costs. The following is the full Chinese translation of the article, with some parts abridged:
Dear friends, Another year of rapid AI development has created unprecedented software development opportunities for everyone—including those just entering the field. In fact, many companies simply cannot find enough skilled AI talent. Every winter break, I spend some time learning and building projects, and I hope you can do the same. This helps me hone old skills and learn new ones, and it can also help you advance your career in technology. To become proficient in building AI systems, I recommend you:
Take AI courses Practice building AI systems (Optional) Read research papers
Let me share why these are all important. I've heard some developers suggest that others shouldn't worry about learning and should just dive into building projects. This is terrible advice! Unless you are already immersed in a community of experienced AI developers, diving into building without understanding AI fundamentals means you might reinvent the wheel—or, more likely—reinvent it poorly! For example, when interviewing candidates, I've encountered developers who reinvented standard RAG document chunking strategies, replicated existing agent AI evaluation techniques, or ended up writing messy LLM context management code. If they had taken a few relevant courses, they would have a better understanding of the building blocks that already exist. They could still rebuild these modules from scratch, or even invent something superior to existing solutions, but they could have avoided weeks of unnecessary work. So structured learning is important! Moreover, I find taking courses really enjoyable. I prefer watching courses from knowledgeable AI instructors over watching Netflix! At the same time, taking courses alone is not enough. Many lessons can only be learned through practice. Learning the theory behind how airplanes work is very important for becoming a pilot, but no one ever learned to be a pilot just by taking classes. At some point, jumping into the pilot's seat is crucial! The good news is that the building process is easier than ever by learning to use highly intelligent coding tools. And understanding AI building blocks might spark new ideas for what to build. If I feel uninspired about what project to do, I usually take a course or read research papers; after doing this for a while, I always come up with many new ideas. Additionally, I find building really fun, and I hope you will too! Finally, not everyone has to do this, but I find that many of the strongest candidates in today's job market read research papers at least occasionally. Although I find research papers harder to understand than courses, they contain a lot of knowledge that hasn't been translated into a more digestible format. I prioritize this after taking courses or practicing building, but if you have the opportunity to strengthen your ability to read papers, I urge you to do so. (You can also watch my old video with advice on reading papers.) I find taking courses and building fun; reading papers might feel more like a chore, but the moments of insight I gain from reading papers are delightful. Wishing you a wonderful winter break and a Happy New Year. Besides learning and building, I hope you also spend time with loved ones—that's equally important! With love, Andrew Ng Top AI Stories of 2025 Dawn of a New Era 2025 may be remembered as the beginning of the AI industrial era. Innovation pushed model performance to new heights, AI-driven applications became indispensable, top companies waged a fierce battle for skilled practitioners, and infrastructure construction drove US gross domestic product growth. Like past winter break seasons, this special issue of The Batch traces the major themes of the past 12 months. The coming year promises to consolidate these changes as we weave this technology more tightly into the fabric of daily life. Thinking Models Tackle Bigger Problems Think step by step. Explain your reasoning. Work backward from the answer. At the start of 2025, models performed these reasoning strategies only when prompted. Now most new large language models do it as a matter of course, improving performance on a wide range of tasks. What happened: Late last year, OpenAI launched the first reasoning or "thinking" model, o1, which built agentic reasoning workflows into it. In January, DeepSeek-R1 showed the rest of the world how to build this capability. The result: Math and coding performance improved immediately, question-answering became more accurate, robots grew more capable, and AI agents advanced rapidly. What drove the story: Early forms of reasoning arose with the paper "Large Language Models are Zero-Shot Reasoners," which introduced the prompt appendage "Let's think step by step." The authors found that manually adding these words to a prompt improved the model's output. Researchers soon realized they could train this capability into models so they would use this and other reasoning strategies without explicit prompting. The key: Fine-tuning via reinforcement learning (RL). Give a pre-trained LLM a reward for producing a correct output, and train it to "think" about a problem before generating an output.
The first few reasoning models were trained specifically via RL to solve math problems correctly, answer science questions accurately, and/or generate code that passes unit tests. This enabled o1-preview to outperform its non-reasoning predecessor GPT-4o by 43 percentage points on AIME 2024 (contest math problems) and 22 percentage points on GPQA Diamond (Ph.D.-level science questions), while it completed Codeforces coding problems at the 62nd percentile relative to competitive human coders, compared to GPT-4o's 11th percentile. Performance was even better when reasoning models learned to use tools like calculators, search engines, or bash terminals. For example, on a challenging test of multimodal understanding and technical expertise across 100 domains, OpenAI's o4-mini with tools achieved 17.7% accuracy, more than 3 percentage points higher than without tools. Robotic action models have been trained via RL to reason. For instance, rewarding ThinkAct for reaching a target location yielded roughly an 8% performance improvement on robotic tasks compared to non-thinking models like OpenVLA. Reasoning models also helped agents tackle difficult problems. For example, AlphaEvolve used Google Gemini to repeatedly generate, evaluate, and alter code, eventually producing faster algorithms for real-world problems. Similarly, AI Co-Scientist used Gemini to generate scientific research proposals, then reviewed, ranked, and improved them. Among other achievements, it proposed a hypothesis to answer a long-standing question about microbial antibiotic resistance. Human scientists independently proposed and validated the same hypothesis nearly simultaneously.
But: Reasoning models may not be as rational as they appear.
In a controversial paper, Apple concluded that reasoning models can't solve puzzles beyond a certain complexity, even when given the algorithms to solve them. The models' inability to apply the algorithms calls into question the apparent similarity between machine and human reasoning. Anthropic found that while a model's reasoning steps can help explain how it reached a conclusion, they may also omit key information that contributed to the conclusion. For example, including a prompt within a prompt can steer a reasoning model to produce a particular output, but its reasoning steps might not mention that prompt.
State of play: Reasoning significantly improves LLM performance. Yet, better output comes at a cost. Reasoning-enabled Gemini 3 Flash used 160 million tokens when running the benchmarks for the Artificial Analysis intelligence index (scoring 71), while Gemini 3 Flash without reasoning used 7.4 million tokens (scoring a much lower 55). Moreover, generating reasoning tokens can delay output, increasing pressure on LLM inference providers to serve tokens faster. But researchers are finding ways to make the process more efficient. Claude Opus 4.5 and GPT-5.1 set to high reasoning achieved the same intelligence index score, but the former used 48 million tokens while the latter used 81 million. Big AI Companies Lure Talent With Huge Pay Leading AI companies waged a fierce battle for talent, attracting top people from rivals with compensation levels typically associated with professional sports. What happened: In July, Meta launched a hiring spree to staff its newly formed Meta Superintelligence Lab, offering compensation worth up to hundreds of millions of dollars to researchers from OpenAI, Google, Anthropic, and other top AI companies. The offers included large cash bonuses and compensation for equity left behind at another company. Meta's rivals in turn poached key employees from Meta and from one another, driving the market value of AI talent to unprecedented heights. What drove the story: Meta upended traditional pay structures by offering compensation packages worth up to $300 million over four years, with its liquid compensation sometimes greatly exceeding stock options that would take years to vest at other companies. After hiring key members of the team of Scale AI CEO Alexandr Wang, Meta CEO Mark Zuckerberg drew up a wish list.
Zuckerberg made personal house calls to persuade people to jump ship, sometimes bringing homemade soup. The effort netted talent including OpenAI's Jason Wei and Hyung Won Chung, two researchers who worked on reasoning models. Andrew Tulloch, who co-founded Thinking Machines Lab with former OpenAI CTO Mira Murati, initially turned down a Meta offer that included a bonus worth $1.5 billion. Months later, he changed his mind and joined Meta. Meta hired Ruoming Pang, who oversaw AI models at Apple. The compensation package was worth hundreds of millions of dollars over a few years. Meta's offer exceeded the compensation packages for Apple's top leaders except the CEO, and Apple declined to counteroffer. Amid the churn, Microsoft AI CEO Mustafa Suleyman poached more than 20 researchers and engineers from Google, including engineering VP Amar Subramanya. Elon Musk's xAI hired more than a dozen AI researchers and engineers from Meta. Musk decried rivals' "insane" offers and touted his company's "super meritocratic" culture and greater potential for equity growth.
Behind the news: AI engineers' salary trajectory reflects AI's evolution from academic curiosity to revolutionary technology.
In 2011, when Google Brain launched under Andrew Ng's leadership, AI talent was concentrated in academia. As neural networks made their way into commercial products like search engines and AI assistants, the machine learning engineer role became a standard corporate rung. In 2014, when Google acquired DeepMind, AI salaries significantly exceeded those for general software engineering. The New York Times estimated DeepMind's personnel costs at roughly $345,000 per employee. By 2017, when Google introduced the transformer architecture, top compensation had risen to $500,000. Around 2023, with the rise of ChatGPT, compensation jumped again. According to one report, top software engineers received compensation packages exceeding $700,000.
State of play: As 2026 begins, the AI hiring landscape has changed considerably. To fend off recruiters, OpenAI offered more equity compensation than rivals, sped up the vesting schedule for stock options granted to new hires, and issued retention bonuses of up to $1.5 million. Despite talk of an AI bubble in 2025, high salaries are justifiable for companies planning to spend hundreds of billions of dollars building AI data centers: If you're spending that much on hardware, why not spend a fraction of that outlay on salaries? Top AI Companies Announce Data Center Construction Plans Expected to Cost Trillions of Dollars and Consume Gigawatts of Power in Coming Years What happened: The AI industry's capital expenditures this year alone topped $300 billion, much of it devoted to building new data centers to handle AI tasks. That's just a down payment, as companies laid out ambitious plans to build facilities the size of small towns with the energy requirements of medium-sized cities. Consulting firm McKinsey predicted the race to build enough processing power to meet inference and training demands by 2030 could cost $5.2 trillion. What drove the story: Top AI companies announced a slate of data center projects around the world. The construction cost is roughly $50 billion per gigawatt of data center capacity.
In January, OpenAI launched Project Stargate, a $500 billion effort with partners including Oracle, SoftBank, and UAE investment company MGX. The company eventually announced plans to build 20 gigawatts of data center capacity globally and predicted demand could be 5 times that figure. OpenAI CEO Sam Altman said he hopes to eventually add 1 gigawatt of capacity per week. Meta spent roughly $72 billion on infrastructure projects in 2025, mostly in the U.S., a figure executives said would rise significantly in 2026. The company's Project Hyperion includes a $27 billion, 5-gigawatt data center to be built in rural Louisiana. The project's financing agreement would keep the assets and debt off Meta's books. Microsoft spent $80 billion on global data center projects in 2025, including facilities in Wisconsin and Atlanta that will be linked by a dedicated fiber-optic network to operate as a single massive supercomputer. To power them, the company signed a 20-year agreement to restart Pennsylvania's Three Mile Island nuclear reactor, which will deliver 835 megawatts of power starting in 2028. The company also pledged to expand its European cloud and AI capacity to 200 data centers across Europe. Amazon expected to spend $125 billion on infrastructure in 2025, and more in 2026. Its $11 billion Project Pluvius is a 2.2-gigawatt data center in Indiana running 500,000 Amazon Trainium 2 chips. Additionally, Amazon planned to spend roughly $14 billion to expand data centers in Australia and around $21 billion in Germany between 2025 and 2029. Alphabet expected infrastructure spending to hit $93 billion in 2025, up from a previous forecast of $75 billion. The company announced a $40 billion project to add 3 data centers in Texas by 2027. It also pledged $15 billion in investments in India, announced roughly $6 billion in Germany, and launched new projects or expansions in Australia, Malaysia, and Uruguay.
But: Can the U.S. economy and infrastructure support such massive investment? There is reason for doubt.
According to Bain & Company consultants, the cost of data center construction would require roughly $2 trillion in annual AI revenue by 2030. That would exceed the 2024 earnings of Amazon, Apple, Alphabet, Microsoft, Meta, and Nvidia combined. Existing power grids may lack the capacity to power these data centers. Two facilities in Silicon Valley sat idle because the local utility lacked the capacity to connect them to the grid. In mid-December, Blue Owl Capital, which had been in talks to provide $10 billion in data center financing for Oracle and OpenAI, pulled out of the deal. The move was prompted by concerns about Oracle's growing debt for data center construction. Blue Owl continued to provide financing for other Oracle-OpenAI data center projects.
State of play: Despite concerns about an AI bubble, the infrastructure boom is creating real jobs and sales in a sluggish economy. Harvard University economist Jason Furman said data centers and AI investment accounted for nearly all U.S. GDP growth in the first half of 2025. At this stage, evidence supports the idea that 2025 kicked off a new industrial era. Agents Write Code Faster and Cheaper Coding applications evolved from autocomplete-style code suggestions to agent systems capable of managing a wide range of software development tasks. What happened: Coding became the application of agent workflows with the most immediate commercial value. Applications like Claude Code, Google Gemini CLI, and OpenAI Codex turned coding agents into one of the most hotly contested battlegrounds among large AI companies. Smaller rivals developed their own agent models to stay competitive. What drove the story: When the groundbreaking agent code generator Devin launched in 2024, it raised the state of the art on the SWE-Bench coding challenge benchmark from 1.96% to 13.86%. By 2025, coding agents using the latest large language models commonly completed over 80% of the same tasks. Developers adopted increasingly sophisticated agent frameworks that enabled models to collaborate with agent planners and critics, use tools like web search or terminal emulation, and operate on entire codebases.
When reasoning models arrived in late 2024, they immediately boosted coding capability and lowered costs, since reasoning enabled agents to plan tasks carried out by less expensive models. The addition of variable reasoning budgets made it easier for agents to use a single model, spending more tokens on planning and fewer on simple edits. By the end of 2025, Gemini 3 Pro, Claude Opus 4.5, and GPT-5.2 became the top models for coding and agent workflows. Open-weight models followed close behind. Z.ai GLM-4.5 and Moonshot Kimi K2 became popular open-weight choices, enabling automated-coding startups to slash costs dramatically. The July release of Qwen3-Coder offered a massive 480-billion-parameter model trained on more than 5 trillion code tokens, performing nearly as well as Claude Sonnet 4. Anthropic built an agent framework around Claude, creating an application: Claude Code. Launched in February, Claude Code was an immediate hit and set expectations for what agentic coding systems should do. OpenAI launched the Codex application based on coding-specific versions of its GPT-5 series in response. Claude Code ran locally initially, while the Codex app ran in the browser, helping to popularize coding agents that run in the cloud. By year's end, these agents were able to manage long-running problems using multiple sub-agents—typically an initializer to kick off the task and track progress, and various coding agents to accomplish different tasks—each with its own context window. A tug-of-war between model makers and integrated development environment (IDE) developers led popular IDE providers like Anysphere (Cursor) and Cognition AI (Windsurf) to build their own models. Conversely, Google built its own IDE, Antigravity, which debuted in November.
Background: Agent systems steadily raised the state of the art on the popular SWE-Bench coding benchmark, and researchers looked for alternative ways to evaluate their performance. These efforts spawned benchmarks including SWE-Bench Verified, SWE-Bench Pro, LiveBench, Terminal-Bench, ????-Bench, and CodeClash. With different vendors trusting (or cherry-picking) different benchmarks, evaluating an agent's performance grew more difficult. Choosing the right agent for a particular task remains a challenge. However: Early in 2025, most observers agreed that agents were good at generating routine code, documentation, and unit tests, but experienced human engineers and product managers performed better on higher-level strategic questions. By year's end, companies reported automating advanced tasks. Microsoft, Google, Amazon, and Anthropic said the volume of code they generated themselves was growing. State of play: In a short time, agent coding pushed vibe-coding from a confusing buzzword to a nascent industry. Startups like Loveable, Replit, and Vercel enabled users with little or no coding experience to build web applications from scratch. While some observers worried AI would replace junior developers, developers who became adept at using AI proved able to prototype applications better and faster. Soon, AI-assisted coding may simply be considered coding, just as spell check and autocomplete are part of writing.
