Generative AI Landscape
Transforming Technology: The State of Generative AI After Two Years of Rapid Growth
I started writing about generative AI articles (1, 2, 3) 2 years ago. What is going on recently? Let’s see through articles below.
Numbers
Spending on GenAI initiatives will increase by
50% in 2025 compared with 2024.
The use cases in 2024 are horizontal use cases, like customer service support, IT testing, HR support, marketing research/content management. Don’t see many vertical, industry-specific use cases. For 2025, still not many industry use cases, but some interesting use cases come up, e.g. Planning, budgeting and forecasting, Supply chain optimization, Regulatory documentation/compliance.
Only 15% enterprises have GenAI in full production.
GenAI in software development: % Improvement with Introduction of GenAI: coding 11%, testing 7%, maintenance 4%, deployment 2%
Top areas GenAI making impact: IT (AI landing zone, AIOps), Customer service/ contact center, Product/service development, Sales/business development, Marketing/PR/media
Buy vs. Build: 65% buy MSP due to Expertise and Knowledge Acquisition, Speed and Time Efficiency, while 35% build due to Specialized In-House Knowledge/Expertise, Cost Considerations, Data Privacy and Security, Regulatory Compliance
What would you do differently?: Better governance/ coordination (having clear structures, policies, and frameworks to guide the development, deployment, and use of AI technologies.), Faster learning/failing, External support, More research/evaluation
Generative AI infrastructure
AI data center is power-hungry: NVIDIA H100 chips (sold at 30 times of cost and hoarded by big tech buyers like Microsoft, Meta, Google, Amazon) annual electricity consumption is on par with small nation like Georgia, Guatemala, Costa Rica. Demand for sustainable/renewable energy like nuclear
Free internet text is going to exhaust in 2026. Impact: vendors with proprietary content will become hot licensing and acquisition targets, demand for synthetic datasets
GPU shortage: forces Nvidia’s biggest customers push their own chips (e.g. Amazon Trainium, Google TPU, Microsoft Maia, Meta MTIA).
LLM security, 2, 3: prison break, data leaks, data poisoning, vendors listed
MLOps consolidation: fragmented market, while buyers want a one-stop shop for their AI needs, vendors listed, e.g. databricks, H2O; while some other companies are feeling the pain in an overcrowded market. End-to-end platforms are mentioned to dominate, e.g. Databricks, traditional cloud platform/Google, AWS, Microsoft.
Foundation models
2024 starts to see native multimodal model, e.g. GPT-4o, Gemini, Claude 3. Use cases: healthcare/patient clinical image analysis, autonomous driving, consumer devices/ray ban meta smart glasses, robotics/multimodal with sensory data
Small models, e.g. Phi-2, Gemini Nano, Mistral, posting strong performance on common benchmarks. Sectors with sensitive data will choose small models at the edge, as well as narrow tasks across finance, healthcare, and law
Open-source LLM developers are seeing rising investor & commercial interest. Smaller open models are outperforming GPT-3.5 and reporting comparable performance to some larger closed models on the MMLU benchmark
Model architecture, transformer-based models are leaving a massive carbon footprint, and users want longer-context, together.ai stripedhyena, hazy research monarch mixer are mentioned
Generative AI applications: in software development, developer spends 56% less development time on average in completing task with copilot (GitHub Copilot X, code llama, starcoder, CodeWhisper, replit, tabnine, Gemini code assist). Other apps: documentation automation in healthcare, drafting contracts, summarizing documents, & optimizing research in law; AI vs. AI, e.g. Deepfakes, ChatGPT email phishing attacks, Gen AI cybersecurity counter attacks (WRAITHWATCH, Jericho security, Brightside AI, Nexusflow.ai), journalism, media (news-gathering, production, and distribution, AI news anchors here/reporters next), text-to-image/text-to-video (Runway, Meta, Stability, Pika Labs, Moonvalley.ai)
GenAI value chain: AI infrastructure (compute, expert, data), providers of GenAI models ( foundation models and fine-tune), AI deployment (made available to end users through stand-alone applications or
interfaces, or integrations into applications)
The article lists these types of GenAI models: text, image, audio/music, 3D, video, protein/DNA
Microsoft Copilot
Google Gemini
AWS Bedrock
Anthropic Claude
OpenAI
Perplexity.ai
Jasper.ai
Meta AI
Huggingface
Databricks
Mistral.ai
Adept.ai
Figure.ai
Foundation Model Ops comprises of: Prompt Engineering, Prompt Templates & Marketplaces, Prompt Management, Prompt Chaining, Data & Embeddings Management, Fine-tuning, Deploy, Optimize & Monitor
OSS LLM wins in
Speed of Innovation & Customization
Accessibility & Cost (GPT-4o $2 per million token input 1, Llama-3–70B 60 cents per million token input 1, 2)
Security and control (own control of security and compliance on your own infra)
41% of interviewed enterprises will increase use of open-source models, and 41% will switch from closed to open if the open-source model matches the closed model’s performance.
Closed Source LLMs are appropriate when you don’t have in-house AI expertise.
Appendix
https://go.i4cp.com/hubfs/2025%20Priorities%20&%20Predictions.pdf
https://hkifoa.com/wp-content/uploads/2024/12/2025-tech-trends-cbinsights.pdf
