I spend a lot of time thinking about potential futures and the kinds of companies that could either help us get there, or would benefit once there.
If you're thinking along similar lines, I am always interested in meeting. These themes are ones I'm very interested to invest behind. If you're thinking about starting a company in these areas, I'd be potentially interested in helping you do that.
- An app server for agent apps: every agent vendor is rebuilding the same underlying infrastructure stack to power their agents. Buyers want to deploy many agents, but currently would have to establish vendor relationships with dozens of vendors, all reimplementing the same stack, including auth, integrations, etc. Someone has an opportunity to build the AI app server, on which agent apps are deployed as configurations. I wrote more about this here, from the perspective of Claude native apps, but I don't think this theme has to be restricted to a lab. At the present moment though, it does make the most sense for one to do this, as it would commoditize the app layer and let them compete on the runtime and underlying infrastructure (including inference).
- A community-driven ecosystem for open apps: as the cost to write code goes to zero, more people will write their own apps, purpose built for their use case. However, for the same reasons SaaS emerged, there will be a lot of overlap in required functionality. This creates a potential opportunity for an ecosystem of templated, modular apps that can be built, shared, and iterated upon by a community. This could potentially be an open source ecosystem around which an infrastructure and hosting platform could be created.
- Agent memory: cheap, fast, scalable, easy memory for agents. As memory becomes more important for long running agentic systems, the opportunity to go very deep on memory semantics will emerge (supersession, decay, provenance, conflict detection, compaction). You would probably want to build this on object storage. The "x for y" simplification of this might be "turbopuffer for memory."
- An intelligence router for existing platforms: existing scaled software platforms already aggregate demand for intelligence but have no way to capture the resulting economic value. While many inference providers already exist, none is focused on this narrow, but potentially very large use case. The opportunity would be to build an embedded inference platform that enables existing vendors to hydrate inference ecosystems. You would enable model routing, grounding against customer data in their tenant on the underlying platform, metering and billing, and eventually continuous post training. I've written more in depth about why a platform like Salesforce should do this.
- The infrastructurization of everything: what happens to software business models when code is cheap and abundant? This will likely change the economics of the modern software business, resulting in commercialization entirely via infrastructure primitive consumption. This one is left intentionally vague because there are infinite directions you can take it. I'll be writing a lot more about this in the coming months, and you can expect that this will spawn more specific themes.
- New methods to finance Compute: the market requires better financing methods for compute. These will take the form of specialized debt instruments, derivatives, etc. This is a significant limiting factor in the current compute buildout. I believe it's important that we find ways to enable this, and that the West takes the lead. Financialization of AI is not a bad thing, as some would have you believe, and it's one of America's superpowers.