By: Baris Gultekin, Head of AI, Snowflake
As 2025 approaches, the AI landscape is set for dynamic transformations, driven by creativity and resilience. With a balanced focus on innovation and impact, the coming year promises significant opportunities, laying the groundwork for long-term success in AI. Snowflake’s Head of AI, Baris Gultekin, shares his insights on the GPU market, AI observability, the rise of AI-driven initiatives, and other key trends anticipated for 2025.
AI observability will go mainstream, and be the catalyst for driving AI to production.
In enterprise systems, observability often refers to the ability to see and understand the state of the system. The emerging field of AI observability examines not only the performance of the system itself, but the quality of the outputs of a large language model – including accuracy, ethical and bias issues, and security problems such as data leakage. I view AI observability as the missing puzzle piece to building explainability into the development process, giving enterprises faith in their AI demos to get them across the finish line.
Although AI observability is a fairly new conversation, 2025 is the year it goes mainstream. We’ll see more and more vendors come out with AI observability features to meet the growing demand in the market. However, while there will be many AI observability startups, observability will ultimately end up in the hands of data platforms and the large cloud providers. It’s hard to do observability as a standalone startup, and companies that adopt AI models are going to need AI observability solutions, so big cloud providers will be adding the capability.
A lot of AI “backlash” or negativity will be mitigated one successful use case at a time.
AI hallucinations are the biggest blocker to getting generative AI tools in front of end users. Right now, a lot of generative AI is being deployed for internal use cases only because it’s still challenging for organisations to control exactly what the model is going to say, and to ensure that the results are accurate. However, there will be improvements, especially in terms of keeping AI outputs within acceptable boundaries. For example, organisations can now run guardrails on the output of these models to constrain what generative AI can or can’t say, what tone is or isn’t allowed, etc. Models increasingly understand these guardrails, and they can be tuned to protect against things like bias. In addition to establishing guardrails, access to more data, to diverse data, and to more relevant sources will improve AI accuracy.
The GPU market will self-correct (in most places), allowing companies to better manage their AI-related costs and goals.
The problem with AI and GPU usage is that the “super-chip” future will not be evenly distributed at first. Europe is more worried about the GPU shortage than companies in the United States, where there is greater capacity. Regional availability will be a longer-term problem, and even offering organisations options to route traffic across deployments to places where there is GPU capacity gives some organisations pause. For example, there may be regional data laws to consider, and even without that, it’s still a mental shift for security architects. Sending your data to a different region, even though a secure connection and within the same platform, is something that security architects will need to vet and get comfortable with. As a result, we’ll have to get creative with cloud service providers and new players creating new chips to meet demand, alongside leveraging new models that are both highly capable and cheaper to run.
AI leaders will have to learn to pick the right battles and triage priorities to avoid team burnout and retain talent.
In the countless conversations I have with industry peers and leaders, there’s one commonality: everyone is working harder and faster than they ever have in their entire careers. AI teams are facing immense pressure to keep up with the incredibly high rate of progress and are often finding themselves on a hamster wheel. What they did ten days ago can now be better, and they’re forced to continually iterate to meet the evolving needs of the landscape. As a result, leaders must develop relentless focus to pursue the ideas that will reap the most reward, versus trying to boil the ocean. You can’t keep chasing the shiny new object, and instead you must have conviction in what really matters to your business and customers. AI leaders must be mindful not to fall into the trap of sending their team down constant rabbit holes of short-term wins. We must have a larger vision and clear priorities as our north star.
Agentic systems will emerge as the leading force for high-value applications.
2025 is when we will start seeing the hype of agentic systems start to bear fruit, with the first set of high-value agentic use cases going into production — think handling customer service problems, identifying cyberthreats, and project management. Agentic AI will extend the capabilities of AI-powered applications to take action, sometimes autonomously, but in most cases still with a human in the loop. The shift toward agentic AI will enable more sophisticated automation capabilities, and help enterprises see real ROI of their AI initiatives. Agentic AI has the potential to drive more tangible business value by carrying out tasks independently through autonomous decision-making, which generative AI is not capable of doing on its own.
