Thoughts on AI Conference San Francisco 2026

Our work is helping office tenants negotiate better office leases. So why would we spend two days at an AI conference?
 
Because we use AI every day. It has become an increasingly important resource in our business, streamlining workflows, accelerating research and analysis, and improving the way we present our findings to clients.  Also, over 50% of the current demand for office space in San Franscisco is comprised of AI companies - - - they are the key engine driving this market.
 
We attended the first AI Conference in 2023. It was a relatively small affair. This year, roughly 5,500 people attended, with about 75 companies presenting products and services. The growth of the conference itself says something about how quickly this industry is developing.
 
It should come as no surprise that one of the largest groups of companies was focused on AI testing, security, governance and evaluation. There was considerable discussion about jailbreaks involving OpenAI and Anthropic models, along with the broader challenge of monitoring and controlling increasingly agentic behavior.
 
Another major category was infrastructure and productivity.
 
Compute is massively expensive. A growing number of companies are building tools that measure AI productivity, track usage and spending, and determine which models are best suited to particular tasks. Others are focused on optimizing inference, the process that takes place when a trained model is actually called upon to perform a task. As AI use scales across organizations, managing the cost and efficiency of all this activity is becoming an industry of its own.
 
There was a smaller contingent of cloud and hardware companies, including GPUs, servers and related infrastructure, and a handful of robotics companies. The dancing robots, predictably, were a big hit.
 
One thing that stood out was how many companies had raised substantial amounts of capital while maintaining relatively small teams. That makes sense when you consider the nature of many of these products, which are highly technical, highly scalable and, once built, not necessarily people intensive.
 
Most of us have probably yet to fully grasp the implications of agentic AI.
 
We still tend to think of large language models as search engines on steroids, systems that answer questions, summarize information and generate content. But the next generation of use cases goes much deeper. These systems will increasingly be given goals and the ability to execute multiple steps on our behalf.
 
Over the next five years, for example, e-commerce may be substantially rebuilt around AI agents that research products, compare alternatives, negotiate transactions and make purchases on behalf of humans. Similar changes are likely across finance, healthcare, professional services and countless other industries.
 
Some argue that forms of AGI have effectively already arrived. However one defines it, the gains in model capability over the past five years have been extraordinary. The harder question is what happens if, or when, models become capable of substantially improving the systems that come after them, without human intervention.
 
Lately, I’ve been reading Karen Hao’s Empire of AI, a decidedly more pessimistic examination of the AI industry. It is a useful counterweight to the optimism surrounding much of the technology.
 
AI is an extraordinarily powerful and increasingly polarizing force. Whatever one thinks of it, the technology appears likely to reconstitute significant aspects of how society and the economy function.
 
And slowing or stopping its development seems unlikely, absent some truly catastrophic event at scale. Even then, the technology may already be too widely distributed to put back in the box.
 
Which leaves us with perhaps the most important challenge of all: building increasingly powerful AI systems in ways that keep their objectives aligned with those of humankind.