Gagan Tandon, Telus Digital Solutions
Just a few years ago, large-scale language models (LLMs) had to be trained, with critical resources, an expanded timeline, and coordinated efforts of hundreds of individuals. However, in 2025, Openai It claims that the same work can be performed by less than 12 people.
How do you explain this amazing change? Advances in AI democratization, including AI chip design, dramatic cost reductions, and the availability and accessibility of powerful open source tools.
What factors can help improve AI model training?
Remember them Will Smith's AI-generated video Do you want to eat spaghetti, which went viral in early 2023? It was clear what the AI platform was trying to portray, but the video looked abstract and fictional. But just two years later, AI platforms can generate the same video very clearly and accurately, and viewers are often fooled by thinking they are real.
This advancement is primarily due to advances in how AI models train. Early models focused primarily on large quantities of loosely curated text or images. Today's systems are trained using multimodal inputs and combine text, images, video and audio to produce more accurate and context-enabled results. This change in training approach is just one of three major changes that change the way AI models are built.
1. Improved chip speed and design
With a more powerful computer chip, AI can learn faster, so that you can move faster with a more powerful engine. Following Moore's Law that Chip's transistor count doubles about every two years, rapid advances in semiconductor technology have driven a consistent increase in processing power. However, the real game changer is the rise of chips designed specifically for AI workloads. These include Graphic Processing Unit (GPU) was originally made for video games, but is now widely used for AI training. Tensor Processing Unit (TPU) Google's custom-built AI chips, etc. Custom AI chips developed by companies such as NVIDIA, dedicated to AI tasks.
These special chips make everything faster and more efficient. It is estimated that in 2020, training for the GPT-3 took several months between $4 million and $12 million to use real-world examples. Today, a dedicated design (not a basic platform) can be trained in a few weeks at a fraction of the cost.
2. Reducing training and inference costs
Every part of training an AI model involves costs, from storing and labeling data to the time and computing power needed to process it. Fortunately, many of these costs have been reduced. For example, companies currently pay around $2.50 for every million token. This is a small text unit that AI models use to understand and generate languages. This has been down from $10 last year. According to the lamp.
Costs for manipulating AI models inferenceIt's down again. According to Stanford's 2025 AI Index Report, the cost of running the model at GPT-3.5 level performance has dropped 280 times since the second half of 2022. Hardware improvements have driven this trend.. Nvidia's 2024 Blackwell GPUfor example, 105,000 times less energy per token More than the 2014 version of the company, known for its Internet trends report that defines her era, according to a report by venture capitalist Mary Meeker.
3. Open Source Software
The expansion of open source software adoption will also help speed up AI innovation. By creating model components that have the freedom to use tools, frameworks, and model components, developers and researchers can experiment, iterate, and train models more efficiently. This approach speeds innovation by lowering entry barriers, reducing costs and promoting knowledge sharing across institutions and industries. This trend shows that companies continue to change to more open collaboration in the AI industry, despite maintaining their own versions of the most advanced models.
Additionally, open source projects help to form a more comprehensive and collaborative AI ecosystem. Private companies, academic institutions, and independent developers are increasingly contributing to these efforts by addressing real challenges, improving accessibility and building tools to support responsible development. These collaborations help ensure that a broader perspective and needs are reflected in the way AI systems are built and applied.
What does lower AI training costs mean for businesses?
As the cost of training AI models continues to decline, including the costs of developing and fine-tuning small models, organizations no longer need to rely entirely on pre-built, general-purpose models. This opens the door to a more customized, domain-specific solution that is more suitable for the individual business needs that can be tailored to specific use cases.
One of the most important changes is the ability to connect pre-trained models to internal sources of knowledge. These models already include a broad understanding of the world. When combined with organizational unique domain expertise and operational data, the results become a system that can produce much more relevant, context-enabled output. The ability to correlate public and private knowledge in this way is becoming easier and more powerful thanks to the growing ecosystems of open technology.
In areas such as research, healthcare, and customer service where accuracy and context are important, organizations can now develop smaller, more intensive models without starting from scratch. Instead of relying on large-scale data labeling, many apply expert-driven tweaks supported by synthetic and structured data.
These advances make complex AI models more accessible across the industry. As adoption grows, it is essential to combine technological advancements with responsible and responsible development by paying attention to long-term impacts such as clear governance frameworks, ethical protection measures, and environmental sustainability. By making AI capabilities more accessible, we approach true democratization and ensure that the benefits of advanced technology are shared more broadly and equitably.
About the author:
Gagan Tandon is the Managing Director of AI & Data Services. Telus Digital Solutions.
