My AI predictions for 2025:
Enterprise AI Spending: The Great Reallocation
AI spends will keep rising as enterprises figure, and hopefully adopt, AI for their business context. The spends grew 6x since 2023 to reach $13.8B in 2024 (doesn’t include GPU revenue). Following a similar trajectory in 2025, we should see this ramp up to cross $100B, or at least reach within touching distance. However, the nature of these spends will shift significantly. 50% of spending in 2024 went to the foundation models while own model training and deployment, infrastructure and AI applications formed the rest. In 2025, the majority of the enterprise budgets will flow to applications delivering measurable RoI at scale.
AI-Native Growth Dynamics: Redefining Scale
Several AI native startups serving enterprises will cross $10M ARR within 18 months of founding. These startups are growing meaningfully faster than the fastest-growing SaaS antecedents - reaching $1M ARR in 11 months vs 15 months; scaling to $30M ARR 5x faster than SaaS counterparts. In 2025, we will see several AI application startups founded in 2023/24 reach $10M in ARR through rapid adoption. I would bet that the majority of these startups would be serving ‘non-tech’ native enterprises in financial services, manufacturing, healthcare etc. with “system of work” value prop as opposed to just another “system of record”.
Market Consolidation: The Natural Evolution
In the last 2 years several startups emerged across the AI tech stack. Several of them with a value prop which is more like a feature and is not comprehensive enough to stand on its own. Or operating in an area which has a dozen startups, half of which probably came out of YC. Or playing in the arena too close for current big tech incumbents to ignore. Many of these will get acquired, make a hard pivot or shut shop in 2025. In 2024, 140 AI startups were acquired by current incumbents and we should this number inch higher in 2025.
Frontier Models: Beyond Scaling Laws
There will not be a GPT5 in 2025. Or any other frontier model which is step function better than the current SOTA. Scaling laws of LLMs have started running out and pre-training has reached a point of diminishing returns. That’s why we have o3 and not GPT5. A closer look beyond OpenAI’s marketing spiel reveals that while o3 has made impressive progress in closed-domain tasks (such as coding, math), it doesn’t do better than GPT4 on open-ended reasoning tasks. Several of its previous launches are still in preview and yet to be made generally available. In 2025, we will see model improvements as a result of advancements in multi-modal algorithms, better reasoning capabilities and inference time innovation. The build out (infrastructure, data etc.) needed for the next frontier model as well as the architectural innovation required (discussed below) will shape up in 2025, and pave way for the next frontier model to come out in 2026.
Architectural Innovation: The Next Frontier
Since 2018 when Google introduced the Transformer architecture it quickly became the default architecture of modern large language models. But we have realised its limitations as well. The attention mechanism scales quadratically with sequence length, resulting in expensive computation and context length limitations. Alternative architectures such as State Space Models (SSMs) such as Mamba, S4 scale linearly with sequence length O(n) and can theoretically handle unlimited context. Over the last 12 months, these architectures have shown meaningful improvements over transformers and 2025 should see them becoming more mainstream, setting the foundation for the next frontier model.
Multi-Agent Systems: Emergent Intelligence
So far, the mainstream and at scale impact on AI has been limited to single-agent AI solutions - chatbots being the most common manifestation. While single-agents work well for focused tasks, usually with humans in the loop, they have limitations in handling complex, interconnected workflows. In 2025, we will see multi-agent systems becoming the new norm. We will see scale up of startups leveraging a network of agents collaborating together to solve complex problems requiring diverse expertise. They will find adoption among consumers (travel planning, financial planning etc.) and enterprises alike (back office ops, customer support etc.)
