The Five-Layer Cake

An Austrian-born developer created forty-three open-source projects without much success. That was until his most recent creation: OpenClaw. Since its late 2025 release, it has seen a historic rise, taking the global development community by storm. OpenClaw became the fastest growing open-source project in history – surpassing what took Linux 30 years to do in only 60 days. Not only are individual developers purchasing their own Apple Mac Minis to create 24/7 virtual AI assistants (creating backlogs for new Macs), but enterprise adoption is increasing rapidly: “Every firm needs an OpenClaw strategy” as the Agentic Computing Era is underway. Analyzing its architecture and implications represents another massive positive driver of compute demand, both at the cloud hyperscalers and local edge devices level.

OpenClaw captures what I have believed for months: we are entering a new growth phase in AI, one that increasingly looks less like a standalone text-based chatbot and more like the dawn of a new operating system that gets work done. NVIDIA’s founder called OpenClaw a ‘profound development’ because it has four key elements which fundamentally define a computer: a memory system + skills + I/O subsystems + scheduling. To briefly explain these four: the first is memory, which allows for personalization and context management with both short-term and long-term memory files. The second is skills (or resources), which are API-driven tools and applications that let an agent perform tasks, manipulate files, and browse the web to take actions. Third involves scheduling (or task management), which is the ability to manage resources, handle cron jobs (to execute scripts & commands), decompose complex tasks, and run autonomously. The fourth is I/O subsystems, the input/output mechanisms that allow the computer to communicate with external systems such as messaging services (which make interacting with your system of agents the same as working with a person on your team).

What drives the surge in compute demand is not only better models in isolation. It is the applications and orchestration frameworks that put those models to work continuously, autonomously, and at scale. This is precisely the role OpenClaw was designed to spawn. Each workflow is not a one-shot inference; it is a sustained loop of reasoning, retrieval, tool invocation, and output generation that consumes meaningfully more compute per task than a traditional chat interaction.

To dig deeper into this new architecture, and to hone in on the technical side when conducting expert-level due diligence calls, I have purchased a dedicated new Mac Mini and configured my own OpenClaw system to test out several hypotheses in real-time. I am working with developer experts to better customize skills that matter most to accelerate the firm’s growth and further expand our research capabilities. To ensure safety, the computer is sandboxed, has no client information, and runs the latest models from Anthropic’s Claude and OpenAI’s ChatGPT.

AI’s five-layer cake

Every agent runs on a physical stack. Five layers stand between electricity and intelligence, and each one gates the layer above it.

Energy

To increase the processing performance required to run larger workloads, AI datacenters are power-hungry by design with increasing power consumption requirements for future generations. The buildout of AI infrastructure is creating a generational demand surge for electricity — a secular tailwind that transcends any single company’s fortunes and reaches grid operators and power producers. Natural gas turbine engines are seeing a surge in demand and backlogs have grown out to the end of the decade as datacenters require ‘behind-the-meter’ local power generation capabilities.