A few months ago I was sitting at a Mac Mini in my home office watching two agents come online simultaneously. One is called Neo. He is my personal operational brain: running in the background, monitoring what matters, briefing me every morning, drafting, building, coordinating. The other is Vicky. She is Chief of Staff for Vecto Ventures, running 24 hours a day via WhatsApp, executing across multiple projects and a distributed team. Neither of them is a chatbot. Neither of them waits to be asked.
Neither of them lives inside someone else’s product.
That distinction is the whole point of this piece.
The reset nobody expected
For most of technological history, competitive advantage in technology was simple to explain and almost impossible to close. Large companies had access to infrastructure, capital, and engineering talent that smaller organisations could not match. If you were not a technology business, you bought technology from one. You depended on vendors, integrators, and platforms built by others, because building your own was not a realistic option. That gap between the technology haves and have-nots was structural. It compounded over decades.
What is happening right now is something that has no real precedent. The intelligence infrastructure that is reshaping every industry, the models, the frameworks, the agentic systems, is being made available to everyone simultaneously. A founder with a Mac Mini has access to the same underlying intelligence as a Fortune 500 enterprise. A small professional services firm can deploy the same quality of reasoning as a bank with a thousand-person technology team. The models do not care who is calling them.
This has never been true before. Not with computing. Not with cloud. Not with mobile. Each of those waves democratised access at the interface layer while keeping the underlying infrastructure concentrated in the hands of a small number of platform companies. AI is different because the intelligence itself, the thing that creates the value, is the infrastructure. And for the first time that infrastructure is open, accessible, and in many cases free.
Every business has proprietary knowledge, institutional context, and operational data that nobody else has. For most of history that knowledge was trapped inside people and processes. It could not be operationalised at scale without significant technology investment. That barrier has collapsed. The organisations that understand this earliest will compress decades of capability building into months. The ones that do not will watch others do it instead.
This is the reset. And it is happening whether your organisation is ready for it or not.
What most AI strategies are actually doing
Walk into almost any boardroom right now and ask about AI strategy. You will hear about Microsoft Copilot. You will hear about ChatGPT Enterprise. You will hear about teams using Claude, or Gemini embedded in Google Workspace. You will hear about pilot programmes, rollout plans, usage metrics.
What you will not hear, almost anywhere, is a clear answer to this question: where does your intelligence infrastructure actually live?
Because the truth of most enterprise AI adoption in 2025 and 2026 is this: it is not transformation. It is subscription. Companies have bought access to someone else’s interface, layered on top of someone else’s model, running inside someone else’s product roadmap. The workflows, the memory, the institutional context: all of it sits in a platform owned by a third party who can change pricing, restrict access, adjust capabilities, or simply be acquired.
That is not a strategy. That is dependency.
Understanding why requires unpacking three things that most business leaders treat as one.
The model is not the product
When people say they are using ChatGPT, they usually mean they are using a product built on top of GPT-4o or GPT-5. The model and the product are not the same thing.
The model is the raw intelligence. It is accessed via an API. Claude Opus 4.6, GPT-5, Gemini 2.5 Pro, Llama 4. These are the underlying engines: extraordinary, and, at a structural level, becoming a commodity input. This is not a slight. It is the inevitable trajectory of every enabling technology. Compute commoditised. Bandwidth commoditised. Storage commoditised. Each one was transformative when it arrived. Each one became infrastructure. Intelligence is following the same path.
The evidence is in the benchmark curves. Two years ago, open weights models, models you can download, run locally, and fine tune yourself, were meaningfully behind the frontier. Today, Llama 4 from Meta, DeepSeek R1 and V3.2 from a Hangzhou startup released under MIT licence, Qwen 3.5 from Alibaba, and Mistral Large 3 from Paris are competitive across most enterprise tasks. Even OpenAI, historically the most closed of the major labs, released gpt-oss in 2026: its first open-weight model in years, with the 120 billion parameter version running on a single GPU and matching its own proprietary models on core benchmarks. The gap is closing faster than most people expect.
The model is electricity. What matters is what you build with it, and whether you own what you build.
The strategic implication is this: if your AI strategy is built around which model you are using, you are optimising for the wrong variable. The model will change. Your competitor will have access to the same one within weeks of you. Differentiation at the model layer is temporary by design. The organisations building durable advantage are not asking which model is best. They are asking what we are building on top of it, and whether we own what we build.
The walled garden you did not notice you were inside
The products built on these models are not the same as the models themselves. Claude.ai, Copilot, ChatGPT Enterprise, Gemini Workspace. These are applications. Carefully designed, useful, commercially sophisticated applications.
They are also walled gardens.
And here is where the data question becomes important, because it is more nuanced than most people realise.
When your organisation uses AI via the API, whether directly or through an enterprise contract, the major providers, Anthropic, OpenAI, Google, and Microsoft, do not train on that data by default. API usage sits outside the training pipeline. That is a meaningful protection.
The consumer product layer is different. Anthropic, OpenAI, and Google all shifted their policies in 2025: consumer plans on Claude, ChatGPT, and Gemini now default to contributing conversation data to model training unless users actively opt out. Enterprise and API customers remain protected, but anyone using a personal subscription for work, a common reality in most organisations, is operating under different terms than they may realise.
The implication is not that these products are unsafe. It is that the terms differ significantly across layers, and most organisations have not mapped where their employees are actually using AI or under which conditions.
Beyond the data question, there is a deeper structural point. When your workflows are configured inside Copilot, the intelligence about how your business operates is expressed in Microsoft’s environment, subject to Microsoft’s roadmap. When your context lives inside ChatGPT Enterprise, it compounds inside OpenAI’s platform, not yours.
We have been here before. In the early years of social media, thousands of businesses built their customer relationships, their content libraries, their community infrastructure entirely inside Facebook. It felt like transformation. Then the algorithm changed. Then they started charging for reach that had been free. The businesses that had built on the platform discovered they did not own the relationship. They had been tenants, not owners.
The same dynamic is playing out in enterprise AI right now. The question is not whether these products are useful. They are. The question is whether your strategy ends there.
The specific risk that most boards have not yet quantified is this: every workflow your team configures inside Copilot or ChatGPT Enterprise is accumulating institutional knowledge inside someone else’s platform. The way your sales team frames proposals. The way your support function resolves issues. The context, the memory, the process logic. All of it compounds inside Microsoft or OpenAI’s infrastructure, not yours. When you switch providers, and at some point every organisation does, that intelligence does not travel with you. You start again. Meanwhile the provider you just left has learned from everything your organisation taught it.
Intelligence you actually own
This is the part of the conversation happening predominantly among developers and largely invisible to business leaders. It matters more than most people in the boardroom currently appreciate.
Open weights models are models whose underlying parameters are publicly released. You can download them, run them on your own hardware, fine tune them on your own data, and deploy them without sending a token to a third-party API. Llama 4 from Meta, DeepSeek R1 and V3.2 under MIT licence, Qwen 3.5 from Alibaba, Mistral Large 3, and now gpt-oss from OpenAI itself. The ecosystem is expanding and the capability gap with proprietary frontier models, for most enterprise use cases, is narrowing fast.
It is important to be honest about the reality of deploying these models at scale, because the argument only holds if it is grounded.
Running an open weights model is not as simple as calling an API. Meaningful deployment requires infrastructure: GPUs, operational tooling, ongoing maintenance. Research suggests organisations self-hosting open models spend 30 to 40 per cent more on operational complexity than API-based alternatives. Larger frontier-class models, the ones that compete with the very best proprietary systems, require serious hardware investment, GPU clusters running into the hundreds of thousands of dollars for the biggest deployments.
But the spectrum is wider than that framing suggests. Smaller open weights models, in the 30 to 70 billion parameter range, now run efficiently on two data centre GPUs at around thirty thousand dollars, with minimal performance loss versus much more expensive alternatives. Managed cloud services like Together AI, Fireworks, and Groq let organisations access open weights models via API, with data sovereignty guarantees that proprietary providers cannot match, at a fraction of the cost. And edge deployment is becoming viable: a quantised 30 billion parameter model runs comfortably on a single consumer-grade GPU today.
The trajectory matters as much as the current state. The cost of inference is falling. The tooling for deployment is improving. The gap between running an open model and calling a proprietary API is narrowing every quarter.
The strategic question is not whether you should deploy your own model tomorrow. Most organisations should not do that today. The question is whether you understand this layer well enough to make deliberate choices about where you sit, and whether your intelligence infrastructure is building toward sovereignty or deeper dependency.
The opportunity here is less obvious than it first appears. Open weights models are not primarily a cost play, though the economics are compelling. They are a sovereignty play. The organisations that invest now in intelligence infrastructure they actually own will have a compounding advantage that is very difficult to replicate later. Not because the models will stay better. They will not. Because the context, the fine tuning, the institutional memory embedded in your own systems will have been accumulating for years by the time your competitors start. The window for that head start is open now. It will not stay open indefinitely.
The new operating system forming underneath everything
Here is where the argument sharpens for business leaders.
On 3 March 2026, an open source project called OpenClaw surpassed React to become the most starred non-aggregator software project in GitHub history: 250,000 stars in roughly 60 days. React, the JavaScript library that powers most of the modern web, took over a decade to reach that number.
OpenClaw is not a model. It is not a product built on a model. It is an agentic harness: the framework that defines how an AI agent perceives context, forms memory, makes decisions, uses tools, hands off tasks, and runs continuously. Not a chatbot that answers questions. An agent that runs. That executes. That remembers. That coordinates.
One developer built it in a weekend. The creator described the vision simply: an AI that actually does things. Within weeks, Sam Altman hired him and the codebase was moved to an independent open-source foundation.
Part of what makes OpenClaw, and the broader class of agentic harnesses like it, so capable is also what makes it risky. These agents do not work through sandboxed, curated interfaces. They work through direct access: CLI terminals, npm packages, APIs, file systems, messaging platforms, calendars, email. That is why they can actually do things. As Steinberger put it on the Lex Fridman podcast when discussing the project: a powerful AI agent with system-level access is a security minefield, but it also represents the future. The two are not separate. The power and the risk come from exactly the same place. This is not a problem that gets solved before adoption. It is a problem that gets managed as the infrastructure matures, the same way every powerful open platform before it, the web, Linux, cloud infrastructure, went through the same cycle of capability first and governance catching up.
OpenClaw is one node in a much larger picture.
LangGraph, from the LangChain ecosystem, provides the infrastructure for stateful multi-agent workflows where agents coordinate, hand off tasks, and operate in controlled loops. 34.5 million monthly downloads. Used in production by companies like Klarna and Replit. CrewAI lets you define teams of agents with specific roles that collaborate on tasks: 44,000 GitHub stars, the most of any agent framework. AutoGen from Microsoft, Google’s Agent Development Kit, Pydantic AI, Amazon Bedrock Agents: the ecosystem is expanding at a pace that recalls the early years of cloud infrastructure tooling.
Underneath all of this, open protocols are forming that represent the connective tissue for how agents will communicate and coordinate. MCP, the Model Context Protocol, now under Linux Foundation governance, is designed to standardise how agents connect to tools and data. A2A, Google’s Agent-to-Agent protocol, standardises how agents discover and communicate with each other. Both have serious institutional backing.
The honest position is that neither is settled. Perplexity’s CTO moved away from MCP publicly in March 2026, citing efficiency problems at production scale. The debate in the builder community is live and contested. What is not contested is the direction: agents are gaining increasingly direct access to the systems they operate within, APIs, terminals, services, data. That is where the real capability lives. It is also where the real risk lives. The two are inseparable, and every serious organisation building in this space is navigating that tension in real time.
At Nvidia’s GTC 2026, Jensen Huang stood on stage and said: “Every company in the world today needs to have an OpenClaw strategy and an agentic systems strategy.” Nvidia built NemoClaw directly on top of OpenClaw. Tencent launched a full product suite on top of it integrated with WeChat. Local governments in Shenzhen announced subsidies for OpenClaw-based projects. Gartner forecasts that 40 per cent of enterprise applications will include task-specific AI agents by end of 2026, up from less than five per cent in 2025.
This is not hype. The infrastructure is forming in real time.
The decision facing your organisation right now is not whether to use AI agents. That question is already answered. It is whether the agents you deploy will be yours or someone else’s. Built on open infrastructure you control or rented from a platform you do not. That choice, made in the next 12 to 18 months, will shape your AI architecture for the better part of a decade. Infrastructure decisions of this kind are rarely made consciously. They accumulate through convenience, one product subscription at a time, until the cost of changing becomes prohibitive. The organisations that end up with sovereign intelligence infrastructure in 2030 will largely be the ones that understood this layer in 2026.
Why the infrastructure layer always wins
Every major platform transition follows the same structural pattern. Proprietary products dominate early adoption. Open infrastructure forms underneath. The open layer becomes the foundation everything else builds on. The businesses that understood the infrastructure early built advantages that compounded for decades. The ones that only ever operated at the product layer stayed permanently dependent.
TCP/IP transformed global communication not by being a product but by becoming an open protocol anyone could build on, anywhere, without permission. Linux spread because it was open, forkable, and improvable, and became the foundation of cloud infrastructure, mobile operating systems, and most of the world’s server infrastructure precisely because nobody owned it. Kubernetes transformed container orchestration because it became the open standard all the cloud providers, all the tooling vendors, and all the infrastructure teams converged on.
The agentic harness layer is in the same moment that Linux was in the late 1990s. Early. Rough in places. With real security concerns that honest participants acknowledge openly. But the trajectory is unmistakable, and the structural pattern is one we have seen before.
The organisations that understood those earlier infrastructure moments early did not necessarily move fastest. They moved most deliberately. They understood what they were looking at before the consensus caught up. That is the only advantage this moment is offering.
What this means in practice
I am not arguing that every organisation should immediately deploy OpenClaw or build their own agentic infrastructure. The point is to understand the terrain clearly enough to make deliberate choices.
The most useful frame I have found for explaining what building at the harness layer actually feels like is this: you are not buying software. You are designing employees.
Think about the difference between hiring a contractor from an agency and building your own team. The contractor is capable. Convenient. But the agency owns the relationship, sets the terms, and can change the arrangement. You get the output but you do not own the capability. When you build your own team member, you design their role, give them the tools they need, define what they have access to, shape how they think about problems, and own everything they produce and everything they learn. Nothing about that relationship is owned by a third party.
Neo and Vicky work exactly like the second model. Neo is my personal operational brain for the business. He monitors schedules, health data, and communications, runs a briefing every morning before the day starts, and operates quietly in the background unless something needs attention. When an idea or opportunity comes in he translates it into a commercial playbook: deal structure, next actions, who does what. He drafts in our voice, builds tools when we need them, and delegates routine processing to cheaper models while handling strategy and orchestration himself. He knows who we are, what we are working on, and what matters. That context persists and compounds across every session.
Vicky is Chief of Staff for Vecto Ventures. She runs 24 hours a day via WhatsApp, connected to the team, coordinating across multiple ventures simultaneously. She executes tasks rather than describing them. Research, content, scheduling, campaign execution, first drafts, competitor intelligence. She knows the business, the team, the history. She compresses work that would require several people into a single always-on system that costs a fraction of the headcount it replaces.
Both were designed and built by us. We chose their tools, their memory structure, their skill sets, their access permissions. We can switch the underlying model powering either of them in minutes, the way you might give a team member a better laptop. Everything they know, every piece of context they hold, every configuration that makes them useful: all of it belongs to us, not to any platform. Building them required patience and some technical fluency. It does not require a data science team or an enterprise budget. A Mac Mini and a clear sense of what you are trying to build is enough to start.
Here is what makes this distinction urgent right now. The model companies have noticed the agentic layer forming and are building products on top of it. Claude Code, GitHub Copilot’s agentic mode, Cursor, Devin, OpenAI’s operator products. These are excellent tools. Genuinely impressive. And they are, structurally, exactly what they look like: another product layer built on top of the infrastructure, owned by the same companies whose products we already discussed. Using them is the contractor-from-an-agency version of agentic AI. You benefit from their capability. You do not own the worker, the memory, the accumulated context, or the infrastructure. When they change pricing, restrict access, or alter how the product works, you feel it. The lock-in is identical in kind to every walled garden that came before. It is just better dressed.
The implication for organisations is not that you need to do exactly what we have done. It is that this is now possible, and the people building the layer underneath your industry are doing it right now, regardless of whether you are paying attention.
The questions worth sitting with
This piece is not a prescription. Smart people will draw different conclusions based on their sector, their risk profile, and where they currently sit. But there are a few questions every business leader should be able to answer clearly.
Do you know the difference between the model, the product built on the model, and the infrastructure layer underneath both?
Do you know which of your employees are using which AI products, under which data terms?
Is your current AI strategy building something your organisation owns, or deepening dependency on platforms you do not control?
Do you have a point of view on open weights models and what happens to your intelligence strategy when frontier capability is available at minimal cost via managed services?
What is your organisation’s position on the agentic layer: aware, experimenting, building, or not yet looking?
The companies that will have durable AI advantages are not necessarily the ones spending the most on the most expensive models today. They are the ones that understood the infrastructure layer early enough to build on it rather than be permanently dependent on it.
The infrastructure is forming right now. It is open. It is moving fast. And unlike most things in technology, the pattern of how this plays out is not actually hard to read. If you know where to look.
Craig Hepburn is an AI strategist and builder, Perplexity Fellow, former CDO at Art Basel and UEFA. Works across tech, business and system design shaping how AI operates responsibly.
If this gave you a useful lens, subscribe. The next piece goes deeper on what it actually means to build intelligence infrastructure that compounds, and why context is the most undervalued asset in the AI era.
Switching costs changed how I think about this. Moved my setup off shared cloud and onto a dedicated Mac Mini in March. The muscle memory of using vendor products like Claude.ai and ChatGPT is strong enough that most people never build the infrastructure layer you are describing here.
Once you own the machine and the agent process the calculus changes entirely. Different reliability profile, different cost structure, different level of control. The boring infrastructure decisions end up mattering more than model selection for anything running longer than a single session.
Great post Craig, I like this analogy "Think about the difference between hiring a contractor from an agency and building your own team"