Washington, DC —
If you look back 2025, the year no end with fireworks or one flashy keynote wey everybody live-stream for millions.
Na the tin dem wey really change everything sneak up on us.
Dem come show for data centres, quarterly filings and technical papers, then change how industry dey reason wetin machines fit really do.
By December, whole industries don quietly rewire around ideas wey twelve months before still dey feel like science fiction.
Na these five things wey really move the needle for this year.
Nvidia $100 billion big deal wit OpenAI
Dem no announce am for stage with smoke machines.
For one Tuesday inside September, dem release short joint press statement: Nvidia go supply OpenAI up to $100 billion worth H100s, H200s, Blackwell GPUs, plus data-center space and electricity to run dem.
From 2026, OpenAI go get guaranteed access to about ten gigawatts of compute, wey be roughly New Zealand yearly power budget.
The deal fit speed the road to superintelligence, and e fit benefit humanity through faster, more scalable AI progress.
OpenAI release o3, di first model wey dey deliberate
For April, OpenAI push out new model named o3.
Wetin make am different? E no go just shout the first answer wey enter im head. E go stop, draw small plan, run small test calculations for background, and check again. Sometimes e take two seconds, sometimes three minutes.
The catch — and na big one — be say every thoughtful answer dey cost ten to twenty times more juice than the old fire-and-forget models.
o3 no only get better performance for coding, math, and science; e also dey integrate agentic tool use to solve multi-step problems efficiently at lower costs.
Gemini 3 change everything for AI wey handle images, video, and text
November come, Google release Gemini 3.
Text, images, video, audio, 3-D point clouds — e chop all at once and give answers wey make old AI specialist models look ancient.
Gemini 3 achieve all-around excellence by join state-of-the-art reasoning with native multimodal mastery wey fit handle text, video, audio, and code.
Tech watchers note say e get PhD-level problem-solving and agentic tools wey autonomously vibe-code interactive solutions.
China’s DeepSeek V3 levels field
For May, one team we most of us never hear before, DeepSeek, upload the complete weights for DeepSeek V3.
E no just compete with the closed frontier models from the West; for several reasoning and coding tasks, e flat-out win, and e do am while e dey use small fraction of the electricity.
Within weeks, developers for Sao Paulo, Bangalore, and Lagos fine-tune am for Portuguese, Hindi, Yoruba, and Swahili, languages the big labs still dey treat as afterthoughts.
DeepSeek V3 dey deliver crushing benchmarks for math, coding, and reasoning, matching pricey closed models like GPT-4o, but e fully open-source, so anybody fit tweak am for free.
Plus, e train 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 don finally grow up.
Dem no be chatbots with to-do list; dem full small programmes wey fit open your customer relationship management (CRM) system, read the latest shipping manifest, notice say container late, email the supplier, update the enterprise resource planning (ERP) system, and flag finance if the delay trigger penalty clause — all without person telling dem exactly how.
The shift no seamless. Error rates still dey for complex tasks, and every major rollout still need human intervention.
Put everything together, 2025 no be the year we future stop to ask for permission.
The tech no just small improve; e cross the line from "impressive demo" to "default AI people use when dem want to get something done."
AI don dey here already, dey run half the planet’s most important workflows.
Only question left for 2026 na who go steer them — and whether we go like where dem dey headed.
