Andrew Ng's Year-End Summary: 2025 May Be Remembered as the Dawn of the AI Industrial Era
```
Key Points:
Dawn of the AI Industrial Era: 2025 marks the formal transition of AI from "academic exploration" to the "industrialized infrastructure" era. AI investment has become the core driving force for US GDP growth, with global annual capital expenditure surpassing $300 billion.
Trillion-Dollar Investment and Energy Anxiety: Tech giants (such as OpenAI, Microsoft, Amazon) have launched colossal data center projects like "Stargate," where single investments run into hundreds of billions of dollars. Power supply is now a hard constraint, prompting tech companies to restart nuclear plants (like Three Mile Island) to ensure computing power requirements.
Reasoning Models and Agentic Coding: Reasoning models, represented by OpenAI o1 and DeepSeek-R1, have become mainstream, giving AI "multi-step thinking" abilities. "Agentic coding" has exploded, with AI agents now able to independently handle complex software development tasks, greatly improving programming efficiency.
Sky-High Salaries Reshape the Talent Market: Top talent commands compensation comparable to sports stars, with giants like Meta even offering compensation packages worth up to $300 million over four years.
On the 26th, renowned AI scholar Andrew Ng pointed out in his annual letter and the special issue of "The Batch" that 2025 may be remembered as the dawn of the AI industrial era. This year, model performance achieved new heights through reasoning capabilities, infrastructure became a key factor in driving US GDP growth, and top tech companies engaged in unprecedented salary wars to compete for talent. Ng believes that as technology becomes more deeply integrated into daily life, the new year will further consolidate these transformations.
Trillion-Dollar Capital Expenditures and Energy Challenges
Ng notes that in 2025, tech giants led by OpenAI, Microsoft, Amazon, Meta, and Alphabet announced a series of stunning infrastructure investment plans. According to disclosures, each gigawatt of data center capacity costs about $50 billion to build. OpenAI and its partners announced the $500 billion "Stargate" project, planning eventually to build 20 gigawatts of capacity worldwide. Microsoft's 2025 global data center budget reached $80 billion, with a 20-year contract to restart Pennsylvania's Three Mile Island nuclear reactor by 2028 to ensure continuous electricity supply.
Massive investment also faces real challenges. Bain & Co. estimates that to support this scale of construction, annual AI revenue will need to reach $2 trillion by 2030, which is more than the combined profits of major tech giants in 2024. Additionally, inadequate grid capacity has left some Silicon Valley data centers idle. According to the Financial Times, Blue Owl Capital withdrew in mid-December from negotiations on providing $10 billion in data center financing for Oracle and OpenAI, due to concerns about debt levels.
Sky-High Salaries Reshape the Talent Market
As AI shifts from an academic interest to a revolutionary technology, top talent now commands compensation at the level of professional sports stars. Ng states that in 2025 Meta broke with traditional compensation structures, offering salary packages—including cash bonuses and massive equity stakes—worth up to $300 million over four years to researchers from OpenAI, Google, and Anthropic.
Mark Zuckerberg personally joined the talent battle, 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, eventually also joined Meta. In response, OpenAI provided a more aggressive equity vesting schedule and retention bonuses of up to $1.5 million for new employees.
The Popularization of Reasoning Models and Agentic Coding
Ng notes that 2025 is considered the first year of widespread reasoning model adoption. Starting with OpenAI’s o1 model and followed by DeepSeek-R1, these models demonstrated the ability to use reinforcement learning (RL) for chain-of-thought reasoning. This allows models to perform multi-step thinking before producing outputs, significantly improving performance in math, science, and programming. For example, OpenAI o4-mini achieved 17.7% accuracy in a multimodal understanding test after combining tool use.
This technical advancement directly propelled 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. In the SWE-Bench benchmark, code agents based on the latest large models completed over 80% of tasks. Though research from Apple and Anthropic indicates reasoning models still have limitations in some complex logic and that the reasoning process increases inference cost, this has not stopped companies from leveraging AI for automatic code generation and reducing development costs.
Below is the full English translation, with some parts abridged:
Dear friends,
Another year of rapid AI progress 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 will too. This helps me polish old skills, learn new ones, and can help you build your career in technology.
To be skilled at building AI systems, I recommend:Take AI coursesPractice building AI systems(Optional) Read research papers
Let me share why these are all important.
I've heard some developers advise others to skip learning and just start building projects. This is bad advice! Unless you’re already immersed in a community of experienced AI developers, building without understanding AI fundamentals means you may reinvent the wheel—or, more likely, reinvent it badly!
For example, when interviewing candidates, I've met developers who have reinvented standard RAG document chunking strategies, copied existing agent AI evaluation methods, or written messy LLM context management code. If they had taken a few relevant courses, they'd have better understood the existing building blocks. They could still rebuild these from scratch or invent something better, but could have avoided weeks of unnecessary work. Structured learning is important! Also, I find classes genuinely fun. I prefer watching knowledgeable AI instructors to Netflix!
At the same time, just taking classes isn’t enough. Many lessons can only be learned through practice. Understanding the theory behind how planes work is important for becoming a pilot, but no one ever learned to fly just by sitting in class. At some point, you’ve got to sit in the pilot’s seat! The good news is, with highly intelligent coding tools, building things is easier than ever. Also, understanding AI’s building blocks can spark new ideas for what to build. If I’m lacking inspiration for projects, I usually take a class or read papers; after a while, I always gain ideas. Plus, I personally find building fun, and hope you will too!
Finally, although not everyone has to do this, I’ve found that many of the strongest job market candidates regularly read research papers. While I find papers harder to understand than classes, they contain knowledge that hasn’t yet been translated into more accessible formats. I’d put this after classes or building things, but if you can, I urge you to build up the ability to read papers too (see my previous video with tips on reading papers). I find classes and building fun; reading papers can be more of a chore, but the moments of insight are rewarding.
Wishing you a wonderful winter break and happy New Year. In addition to learning and building, I hope you'll also spend time with your loved ones—it's just as important!
With love,
Andrew NgTop AI Stories of 2025
Dawn of a New Era
2025 may be remembered as the start of the AI industrial age. Innovation has pushed model performance to new heights, AI-driven applications are indispensable, top companies are competing fiercely for skilled practitioners, and infrastructure construction is boosting US GDP. As in past winter seasons, this special issue of The Batch reviews the past 12 months’ major themes. The coming year is poised to consolidate these changes as we weave this technology more tightly into daily life.Thinking Models Solve Bigger Problems
Step-by-step thinking. Explain your reasoning. Backward from answers. In early 2025, models only did these reasoning strategies when prompted. Now, most new LLMs do this by default, boosting performance across a wide range of tasks.
What happened: At the end of last year, OpenAI launched the first reasoning or "thinking" model o1, built with agentic reasoning workflows. In January, DeepSeek-R1 showed the world how to build such capabilities. The results: better performance on math and coding, more accurate question answering, more capable robots, rapid progress in AI agents.
Drivers: Early forms of reasoning began with the paper “Large Language Models are Zero-Shot Reasoners,” which introduced prompts like "let's think step by step." Researchers soon realized they could train this ability into models with RL, so models use such strategies even without explicit instruction. The key: RL fine-tuning. Reward a pretrained LLM for correct output, training it to "think" before generating answers.The first reasoning models were trained by RL to solve math problems, answer science questions, and/or produce code passing unit tests. This made o1-preview score 43 points higher than its non-reasoning predecessor GPT-4o on AIME 2024 (contest math) and 22 points higher on GPQA Diamond (PhD-level science problems), with Codeforces coding at the 62nd percentile among human contestants (GPT-4o: 11th percentile).When models learned to use tools like calculators, search engines, or bash, performance improved. For example, on a challenging multimodal test across 100 domains, OpenAI o4-mini with tools achieved 17.7% accuracy, over 3 points higher than without tools.Robot action models have been trained with RL for reasoning. For example, ThinkAct gets rewarded for reaching a goal, yielding about 8% improved performance on robotics tasks over models like OpenVLA without reasoning skills.Reasoning models help agents solve hard problems. AlphaEvolve uses Google Gemini to iteratively generate, evaluate, and alter code, producing faster algorithms for real-world tasks. Similarly, the AI Co-Scientist uses Gemini to generate, review, rank, and improve scientific proposals, leading to a hypothesis addressing a long-standing issue of antibiotic resistance in microbes—an insight independently discovered and verified by human researchers at almost the same time.
But: Reasoning models may not be as rational as they appear.In a controversial paper, Apple concluded reasoning models can’t solve puzzles beyond certain complexity even when given algorithms, because they can’t apply the algorithm—casting doubt on the apparent similarity to human reasoning.Anthropic found reasoning steps can help explain a model’s conclusions, but may omit key information contributing to the answer. For example, a prompt could steer a model to an output, but its reasoning steps may not mention that prompt.
Status: Reasoning significantly improves LLM performance, but at a cost. Gemini 3 Flash with reasoning used 160 million tokens (scoring 71) on the Artificial Analysis IQ benchmark, versus 7.4 million tokens (score 55) for the non-reasoning version. Generating reasoning tokens can also delay outputs, increasing pressure on LLM providers to serve tokens faster. Researchers are searching for more efficient approaches; Claude Opus 4.5 and high-reasoning GPT-5.1 reached the same score using 48 million versus 81 million tokens, respectively.Major AI Firms Woo Talent with Sky-High Salaries
Leading AI firms engaged fierce talent wars, offering salaries usually only seen in professional sports to poach top talent from rivals.
What happened: In July, Meta launched a hiring spree for its Meta Superintelligence Lab, offering researchers from OpenAI, Google, Anthropic, and others pay packages worth hundreds of millions. These included large cash bonuses and compensation for forfeited equity. Meta’s rivals struck back, poaching key Meta employees and pushing AI talent’s value to new heights.
Drivers: Meta upended compensation norms with salaries worth up to $300 million over four years, far exceeding what other companies’ equity vests in the same time. After hiring key members from Scale AI’s CEO Alexandr Wang’s team, Mark Zuckerberg drew up a “wish list,” according to the Wall Street Journal.Zuckerberg personally visited and persuaded talent to join, sometimes bringing homemade soup. His efforts landed OpenAI’s Jason Wei and Hyung Won Chung, two researchers in reasoning-model work.Andrew Tulloch, co-founder of Thinking Machines Lab with ex-OpenAI CTO Mira Murati, initially turned down a Meta offer (reportedly with $150 million in bonuses, per the WSJ), but a few months later, he joined Meta.Meta hired Ruoming Pang, who oversaw AI models at Apple; Bloomberg reports a compensation package worth hundreds of millions over several years. Meta’s offer was higher than any at Apple except for the CEO, and Apple refused to match it.Microsoft AI CEO Mustafa Suleyman poached over 20 Google staff, including VP of engineering Amar Subramanya.Elon Musk's xAI hired over a dozen AI researchers/engineers from Meta. Musk criticized rivals’ “crazy” offers and promoted his firm’s "super-elite" culture and higher equity growth potential.
The backstory: AI engineer salary trends reflect the evolution from academic curiosity to revolutionary technology.In 2011, when Google Brain started under Andrew Ng, AI talent gathered in academia. As neural nets entered products like search and AI assistants, "machine learning engineer" became a standard role in industry.In 2014, Google’s DeepMind acquisition brought AI pay well above generic software engineering—DeepMind’s personnel costs averaged $345,000 per worker, per the New York Times. By 2017, when transformers launched at Google, top pay hit $500,000.Around 2023, with ChatGPT's rise, compensation surged again. Some reports cite top software engineers’ offers exceeding $700,000.
Status: As 2026 begins, the AI hiring landscape has changed greatly. Defensively, OpenAI sped up new hires’ equity vesting and issued up to $1.5 million retention bonuses, per the Wall Street Journal, in face of recruiting. Despite talk of an AI bubble in 2025, for companies spending billions on AI data centers, high salaries seemed rational: If you’re spending so much on hardware, why spare talent?
Top AI companies announced data center plans expected to cost trillions of dollars and consume thousands of gigawatts in coming years.
How it happened: AI industry capex topped $300 billion this year, mostly for new data centers to handle AI. This is just the start—companies have plans to build facilities the size of small cities, with the energy demands of mid-sized towns. McKinsey forecasts that by 2030, the race to build enough capacity for AI inference/training could cost as much as $5.2 trillion.
Drivers: Top AI firms announced a range of global data center projects. Each gigawatt costs about $50 billion in capacity investments.In January, OpenAI launched its "Stargate" project, worth $500 billion, with partners like Oracle, SoftBank, and UAE’s MGX. The company plans to build a global capacity of 20 GW and sees demand potentially five times higher. Sam Altman said he hopes to eventually add 1 GW per week.Meta spent about $72 billion on infrastructure in 2025, mostly in the US, with promise for more in 2026. Its Hyperion project in rural Louisiana is a $27 billion, 5 GW data center, funded so assets/liabilities don’t appear on Meta’s books.Microsoft spent $80 billion globally on 2025 data center projects (including Wisconsin and Atlanta), networked by dedicated fiber into a huge supercomputer. For power, it signed a 20-year deal to restart Three Mile Island’s reactor (835 MW from 2028). Microsoft also pledged to expand European cloud and AI to 200 data centers continent-wide.Amazon expects to spend $125 billion on infrastructure in 2025, more in 2026. Its $11 billion "Raingod" project in Indiana is a 2.2 GW data center running 500,000 Trainium 2 chips. Amazon will also spend around $14 billion expanding in Australia and $21 billion in Germany from 2025–2029.Alphabet expects to spend $93 billion on infrastructure in 2025, up from an earlier projection of $75 billion. The company announced a $40 billion plan adding three data centers in Texas before 2027, committed to $15 billion in India, $6 billion in Germany, and new/expanded projects in Australia, Malaysia, and Uruguay.
But: Can the US economy and infrastructure support such huge investments? There is cause for skepticism.Consultants at Bain say, to fund all this data center construction through 2030 would require about $2 trillion per year in AI revenue, more than Amazon, Apple, Alphabet, Microsoft, Meta, and Nvidia will together earn in 2024.The existing grid may not be able to power these centers. Bloomberg reports two Silicon Valley facilities have been idle because local utilities can’t connect them to the grid.In mid-December, per the Financial Times, Blue Owl Capital walked away from a $10 billion data center financing deal for Oracle and OpenAI, due to worries about Oracle’s rising construction debt. Blue Owl is still financing other Oracle-OpenAI center projects.
Status: Despite AI bubble worries, the infrastructure boom is creating real jobs and revenue in a soft economy. Harvard economist Jason Furman says data center and AI investments accounted for almost all US GDP growth in early 2025. At this stage, there’s credible evidence 2025 has kicked off a new industrial era.Agents Write Code Faster and Cheaper
Software creation has evolved from autocomplete-style code assistance to agent systems managing broad swaths of development tasks.
How it happened: Coding has become the most immediately commercially valuable agent workflow. Apps like Claude Code, Google Gemini CLI, and OpenAI Codex made coding agents a new competitive battleground for major AI firms. Smaller rivals launched their own agent models to try to keep up.
Drivers: When the pioneering code agent Devin launched in 2024, it raised the SWE-Bench benchmark from 1.96% to 13.86%. By 2025, agent coders with latest LLMs commonly complete over 80% of the same tasks. Developers adopted more complex agent frameworks enabling planners, critics, search/web, terminal emulation, and control of whole codebases.When reasoning models emerged at the end of 2024, they immediately improved coding ability and lowered costs: reasoning enables agents to plan tasks for cheaper models to perform. Agentic variable-reasoning budgets make it easier to use a single model, allocating more tokens for planning and fewer for simple edits. By late 2025, Gemini 3 Pro, Claude Opus 4.5, and GPT-5.2 were the top choices for coding/agent workflows.Open-weight models closely followed. Z.ai GLM-4.5 and Moonshot Kimi K2 became popular open choices, helping automated coding startups slash costs. The July release Qwen3-Coder featured 480 billion parameters, trained on 5 trillion+ code tokens—approaching Claude Sonnet 4’s performance.Anthropic built an agent framework around Claude, launching an app: Claude Code. It was immediately popular when launched in February and set expectations for what agentic code systems should do. OpenAI launched Codex, based on special GPT-5 coding editions, in response. Claude Code initially ran locally; Codex operated in the browser, popularizing cloud-based agent coders. By late year, such agents could use multiple sub-agents to manage long-running jobs—typically an initializer tracking progress, and various coding sub-agents for tasks—each with its own context window.A tug-of-war between model providers and IDE devs led leading IDEs like Anysphere (Cursor) and Cognition AI (Windsurf) to build their own models. Google, in turn, built its own IDE, Antigravity, debuting in November.
Context: Agents steadily raised the bar on the popular SWE-Bench code benchmark, and researchers searched for alternative evaluation methods.
This yielded benchmarks like SWE-Bench Verified, SWE-Bench Pro, LiveBench, Terminal-Bench, ????-Bench, CodeClash, etc. As vendors trusted (or cherry-picked) different tests, comparing agents got harder. Choosing the right agent remains a challenge for specific tasks.
However: In early 2025, most observers agreed agents excel at generating routine code, docs, and unit tests, while experienced engineers and product managers still outperform them on higher order strategic issues. By late year, companies reported automating advanced tasks. Microsoft, Google, Amazon, and Anthropic said their self-generated code volumes continued to rise.
Status: In a short time, agentic coding took “vibe-coding” from meme to new industry. Startups like Loveable, Replit, and Vercel enabled users with little or no coding experience to build web apps from scratch. While some feared AI would replace entry-level developers, those adept with AI tools proved better and faster at prototyping. Soon, AI-assisted coding may simply be considered coding—just as spellcheck and autocomplete are now part of writing.
Risk DisclaimerThe market is risky; invest cautiously. This article does not constitute individual investment advice and does not consider any user's specific investment objectives, financial situation, or needs. Users should judge whether any opinions, viewpoints or conclusions in this article fit their own situation. Investing based on this is at your own risk. ```