2025-The Year AI ‘Grew Hands’: Agents, Atoms, and the End of the Pilot Phase
2025: The year AI stopped talking and started doing—while China quietly won the open-source war and nuclear power became tech's unlikely savior.
Welcome to Dharma of AI. Find here essay on AI’s true nature, the world it’s quietly building, and our place within that new order. A journey beyond hype and fear into what AI really is. By Jaspreet Bindra—The Tech Whisperer.
Rise of Agentic AI: The buzzword shifted from ‘Generative’ to ‘Agentic.’ AI stopped just chatting and telling you what to do and started doing it—executing complex, multi-step workflows autonomously across enterprise software. OpenAI Browsing and Shopping Agents, the agentic browsers from Perplexity (Comet) and OpenAI (Atlas), and Agentic Workflows from n8n, Zapier, and Make, to name a few, created waves as conversations in enterprises shifted to building AI agents from adopting AI tools. More new buzzwords like Service as a Software being the new SaaS and A-commerce (agentic commerce) started being whispered. (My article on A-commerce here)
Claude claimed the coding crown: Anthropic’s Claude 4/Opus 4 positioned itself as the best coding and long-context reasoning model, pushing assistants closer to genuine “co-worker” status. While Google and OpenAI hogged the airwaves, Anthropic and its thoughtful CEO Dario Amodei quietly donned the mantle of the serious, enterprise-focused AI company, amidst all the hype. As software coding became the first real, at-scale use case of Generative AI, it was Claude’s models that became the benchmark to emulate.
Going beyond the Pilot: 2025 was the “Put Up or Shut Up” year. C-suites slashed vanity AI projects, focusing strictly on use cases with proven, measurable ROI. Enterprises want to scale; they have waited three years now. Surveys showed AI adoption becoming the default: over three-quarters of firms used AI somewhere, and many began measuring real revenue and cost impact, not just pilots.
Coding is Dead; Long Live Coding: Vibe coding with Claude Code, Lovable, Replit, Cursor, Github copilot etc. made AI generate over 40% of new enterprise code. The role of “software engineer” seems to be shifting to “systems architect” and “AI supervisor.” Or even the “HR for software agents” as Jensen Huang memorably put it in his 2025 CES address.
The “Atoms for Algorithms” Movement: Big Tech signed historic deals with nuclear energy providers to power gigawatt-scale data centers, validating nuclear as AI’s battery. Microsoft struck a landmark $1.6 billion deal with Constellation Energy to restart Pennsylvania’s Three Mile Island Unit 1 by 2027. Amazon moved aggressively with $1 billion in commitments, including a 1.9-gigawatt power purchase agreement with Talen Energy’s Susquehanna plant. Meta committed to sourcing 1.1 gigawatts from Constellation Energy’s Clinton Clean Energy Center in Illinois. Google partnered with startup Kairos Power to deploy small modular reactors. Meanwhile, China went all in, building more nuclear capacity than the rest of the world combined: as of 2025, China had 32 reactors under construction—more than half of the total worldwide—and 102 total nuclear units (operating, under construction, and approved) with 113 GW of installed capacity, ranking first globally in overall nuclear scale. My take on this is that AI is bringing nuclear power out of the cold, and even if the AI bubble sputters out, as some pessimists claim, it will leave behind a rejuvenated, strong nuclear energy infrastructure.
2025 belonged more to China than any other country in the world. China focused on applications, while its US competitors trumpeted AGI. It created massive clean energy capacity, to make energy ‘free’ or a non-issue for its AI companies, while the US scrounged for power. It made AI chips a national mission, and is fast catching up with the US. More than anything else, it quietly became the leader in open source AI, with models like DeepSeek V3/R1 and Qwen 2.5 Max which are credible frontier-class alternatives. History tells us that the one who owns open-source rules the world, and China is best positioned here (my article on China winning AI).
Deepfake Democracy: 2025 elections globally faced a tsunami of hyper-realistic deepfakes, forcing social platforms to implement “provenance” watermarking standards that largely failed to stem the tide. India, UK, US, and many others fear AI subverting democracy, even as AI Slop became mainstream. I believe countries have been slow to counter this insidious threat with legal and societal measures, and hopefully 2026 brings in a raft of them.
The Loneliness Cure: AI Companions saw massive adoption in 2025, sparking fierce ethical debates about human-AI attachment and psychological impact on social skills. Apps like Replika, Character.AI, and newcomers like Nomi reached tens of millions of users, with some spending hours daily in conversations with AI personalities. The phenomenon raised questions if AI companionship was filling genuine human connection voids. Mental health professionals noted both therapeutic benefits for lonely individuals and concerning patterns of withdrawal from real-world relationships. Now major tech platforms have quietly integrated such features, while regulators struggle to define appropriate boundaries for these intimate digital relationships (see my article on Social AI).
The Data Wall: We officially ran out of high-quality public human text. Labs pivoted aggressively to “Synthetic Data”—AI teaching AI—to continue model scaling. This represented a fundamental shift in AI development methodology. Where previous models trained on human-written text from books, websites, and conversations, frontier labs now generate vast synthetic datasets using their own models, then train next-generation systems on this machine-created content. The approach raises fascinating questions about whether this creates a “Habsburg AI” problem (like the genetic issues from royal inbreeding), where models become increasingly detached from human patterns of thought and communication. Some researchers worry about model collapse—where quality degrades over generations of synthetic training. (refer to my Habsburg Internet article for deeper analysis).
The King is Dead, Long Live the King: 2025 started with most experts writing Google off as too slow and a has-been in the AI race, even as OpenAI, Anthropic, Microsoft and Meta were taking leaps forward. The year is ending with Google donning the crown of AI King. Google has always been the ONLY full stack AI player (chips to data to models to research to applications). It has unrivalled access to the Internet data, crushing distribution through its mail, search, mobile and video properties, the best lab in the business with Demis Hassabis and DeepMind, and now it has arguably the best AI model (Gemini 3), product (NotebookLM), and image/video generator (Nanobanana Pro). With the DoJ unshackling its regulatory chains, the AI war is Google’s to lose. What a difference a year makes!
AI Funding, Cross-Deals and Data Centres Propped Up the US Economy: By 2025, AI startups captured roughly a third of global VC funding, with valuations and round sizes staying frothy despite macro-jitters. The real story was AI infrastructure becoming an economic pillar. Tech companies were projected to invest $250 billion in AI infrastructure in 2025 alone, with Microsoft allocating $80 billion. This spending spree created a cascade effect: semiconductor manufacturers expanded fabrication capacity, cloud providers built gigawatt-scale data centers, and energy companies invested billions in nuclear and renewable power generation. The construction boom generated hundreds of thousands of jobs—from chip fab technicians to data center electricians to nuclear engineers. Wall Street analysts noted that AI infrastructure spending had become a measurable component of US GDP growth. By year’s end, projections suggested trillions in cumulative AI infrastructure investment by 2030, fundamentally reshaping global capital flows.
Smart glasses quietly became the next screen and saved Meta’s AI blushes. Even as its LLaMA open source model sank into relative irrelevance, and its AI star Yann LeCun moved away from the mothership, it is the Rayban AI glasses that became a sleeper hit, hinting at AI-first wearables as a post-smartphone category. Zuckerberg dangled hundred million dollar salaries and billions of dollars of acquihire money to corral the world’s AI talent. He has been successful in getting an AI A-team, now his Meta Superintelligence Labs has to deliver. He is doubling down on devices, with the acquisition of Limitless. Meanwhile, there are more devices coming in, the most eagerly awaited one being the OpenAI- Jony Ive one.
Meanwhile in India: Mission “Make in India” AI: The IndiaAI Mission (₹10,000+ Cr) disbursed its first major tranches in 2025, funding local GPU clusters and subsidizing compute for domestic startups like Sarvam AI, which focuses on India-centric language models, and Gnani.ai, specializing in voice AI for Indian languages. The mission provided critical infrastructure access to hundreds of Indian AI startups that previously couldn’t afford expensive compute, democratizing AI development across the country.
2025 brought the AI world to India in unprecedented fashion. OpenAI opened its first India office in New Delhi, with a focus on education initiatives. Anthropic also announced plans for its own office in Bangalore, targeting enterprise customers. Perplexity gave its paid model free to Airtel’s 350+ million users, instantly creating India’s largest AI user base. Then came the blockbuster: Google’s $15 billion announcement for a 1GW AI data center in Visakhapatnam in partnership with Adani and Airtel. Satya Nadella swung by, promising investments in cloud and AI infrastructure. Amazon upped the ante with data center commitments across multiple Indian states. The big tech AI announcements will fundamentally reshape India’s position in the global AI infrastructure landscape.
Chip Fabs Rising: Meanwhile, Tata Electronics commenced production runs at its semiconductor fabrication facility in Gujarat, marking India’s tangible entry into the chipmaking supply chain essential for AI sovereignty. Micron broke ground on its $2.75 billion assembly and test facility in Gujarat. The government noted, “With a vision to build a self-reliant semiconductor ecosystem by 2030, the government has rolled out dedicated incentives under the India Semiconductor Mission, complementing the broader PLI (Production-Linked Incentive) scheme framework.”
DPDP Rules 2025 Made Data the New Compliance Battlefield: The Digital Personal Data Protection Rules fully operationalized India’s data law, forcing AI firms to re-engineer consent, retention and training data practices. The rules created particular challenges for AI companies using web scraping, as they had to demonstrate a legitimate basis for collecting and processing Indian users’ publicly available data. Some complained the regulations created barriers to innovation, while privacy advocates argue they were essential safeguards.