Sadiq S BhatWashington, DC — Looking back on 2025, the year didn’t end with fireworks or a single flashy keynote that everyone live-streamed to millions.
The stuff that actually changed everything snuck up on us. Instead, they showed up in data centres, quarterly filings and technical papers, then changed the way the industry thinks about what machines can actually do.
By December, the whole industries had quietly rewired themselves around ideas that, twelve months earlier, still felt like science fiction.
Here are the five things that really moved the needle this year.
Nvidia $100 billion landmark deal with OpenAI
It wasn’t announced on a stage with smoke machines.
One Tuesday in September, a short joint press release dropped: Nvidia would supply OpenAI with as much as $100 billion in H100s, H200s, Blackwell GPUs, and the data-center space and electricity to run them.
Starting in 2026, OpenAI gets guaranteed access to something like ten gigawatts of compute, roughly the annual power budget of New Zealand.
The deal is likely to accelerate the path to superintelligence, benefiting humanity through faster, more scalable AI advancements.
OpenAI releases o3, first model that deliberates
Back in April, OpenAI pushed out a new model called o3.
The big difference? It doesn’t blurt out the first answer that pops into its head. It stops, sketches out a plan, runs little test calculations in the background, double-checks itself. Sometimes the delay is two seconds. Sometimes it’s three minutes.
The catch — and it’s a big one — is that every thoughtful answer costs ten to twenty times more juice than the old fire-and-forget models.
Not only does o3 have superior performance in coding, math, and science, but it also integrates agentic tool use for efficient, multi-step problem-solving at lower costs.
Gemini 3 changes everything for AI that handles images, video, and text
November rolled around, and Google released Gemini 3.
Text, images, video, audio, 3-D point clouds, it eats all of it in one bite and gives answers that make the old AI specialist models look antique.
Gemini 3 achieves all-around excellence by fusing state-of-the-art reasoning with native multimodal mastery that can handle text, video, audio, and code.
Tech watchers note that it has PhD-level problem-solving and agentic tools that autonomously vibe-code interactive solutions.
China’s DeepSeek V3 levels field
In May, a team most of us had never heard of, DeepSeek, uploaded the complete weights for DeepSeek V3.
It wasn’t just competitive with the closed frontier models from the West; in several reasoning and coding tasks, it flat-out won, and it did so while sipping a fraction of the electricity.
Within weeks, developers in Sao Paulo, Bangalore, and Lagos had fine-tuned it in Portuguese, Hindi, Yoruba, and Swahili, languages the big labs still treat as afterthoughts.
DeepSeek V3 delivers offers crushing benchmarks in math, coding, and reasoning, matching pricey closed models like GPT-4o, but it's fully open-source, so anyone can tweak it for free.
Plus, it trains for just $5.6 million (way below big rivals), making super-smart AI accessible to startups, developers, and everyday users without breaking the bank.
Autonomous agents cross into everyday workflows
By fall, the agent hype finally grew up.
These aren’t chatbots with a to-do list; they’re full little programmes that can open your customer relationship management (CRM) system, read the latest shipping manifest, notice the container is late, email the supplier, update the enterprise resource planning (ERP) system, and flag finance if the delay triggers a penalty clause, all without anyone telling them exactly how.
The shift is not seamless. Error rates remain in complex tasks, and every major rollout still requires human intervention.
Put it all together, and 2025 wasn’t the year the future stopped asking for permission.
The tech didn’t just get incrementally better; it crossed the line from “impressive demo” to “default AI people use when they want to get something done.”
AI is already here, running half the planet’s most important workflows.
The only question left for 2026 is who gets to steer them — and whether the rest of us will like where they’re headed.
