The biggest investment in tech history is not about AI models. It is about your filing cabinets, your systems of record, your company's total memory, and ultimately your organisation's accumulated understanding of what has happened, is happening, and might happen next. OpenAI wants to replace all of those systems and most of the people who operate them.
A few weeks ago, OpenAI engineers pushed internal code to a public GitHub repo. Twice, in five days. The internet immediately started speculating about ChatGPT 5.4. Prediction markets moved. The usual hype cycle kicked in.
Wrong thing to focus on.
OpenAI is spending $600 billion on infrastructure. Not to build a better chatbot. But to own something that makes Salesforce lock-in look like a magazine subscription. The model is a component. What matters is what the model is a component of.
The Filing Cabinet No One Can Open
Think about what organisational knowledge actually looks like today. Code in GitHub. Architecture decisions in Confluence pages no one updates. Customer context in Salesforce. Project status in Jira. And then the email history, years of decisions, discussions, and agreements buried in unstructured inboxes that no one searches and no one can be bothered to migrate.
Some vendors claim that email history is "included in context" when you search. Technically true. But it consistently turns out to be a hollow claim. An email being indexed is not the same as the system understanding what it means, why it was written, or how it relates to the decision made three months later.
And the informal layer, why things were built a certain way, what crashed that time, the decision made in a meeting with three attendees, two of whom have since left. That lives in Slack threads no one reads, or in the head of a senior person thinking about a better offer elsewhere.
Every system is a filing cabinet. The problem is not that information is missing. There is an abundance of it. The problem is the synthesis layer. Today, the synthesis layer is human brains, limited, poor at context-switching, and they resign when they get a better offer.
When a senior engineer leaves, the filing cabinets are still full. What leaves is the person who knew which drawers to open and how the contents connect.
What OpenAI Is Actually Building
Imagine a system that continuously reads from all your tools, maintains a coherent picture of your organisation's decisions, and can reason from them at a depth no single person can match. Not a search engine. Not a chatbot. A new synthesis layer.
OpenAI describes it openly in their AWS press release: they are building a *stateful runtime environment*: a system with persistent memory that can act, not just respond. That is why they are spending $600 billion. Not for a better chatbot. To own that layer.
But it requires four things to work simultaneously. If any one fails, the whole bet collapses.
Intelligence and context are multiplicative. A weak model given vast organisational history drowns in it, finds a discussion that sounds relevant but concerned a different system in a different context, and synthesises confidently from there. A strong reasoning model finds signal in the noise. Every improvement in reasoning capability makes context proportionally more valuable. That is why each new model release matters. Not as an end product, but as the engine powering everything above.
Memory cannot go stale. The decision that was right six months ago may have been superseded. The architecture recommended last quarter may have been abandoned after performance tests. A system that preserves outdated context without flagging it is worse than nothing. It is institutional hallucination.
Retrieval at a scale not yet solved. This is the crux. RAG works for fact lookup. It does not work for finding 2,000 relevant tokens in a corpus of ten trillion, where relevance is defined by causal chains across eight months. Existing approaches degrade as the corpus grows. The company that solves this first has an advantage competitors cannot even benchmark from the outside.
Agents must run without crashing. When a system operates autonomously for weeks, a five percent error rate per task compounds rapidly into system failure. The target is 99.5 percent or above, including when context is incomplete or contradictory.
Why It Breaks the Software Industry
Salesforce is worth $250 billion because it owns customer data. The company that owns the synthesis layer above all that data is worth more. Much more.
The difference: Salesforce locks you in with data. This locks you in with understanding. Data is portable. A year of accumulated organisational synthesis is not.
Take a concrete example. A product manager asks: should we build the real-time analytics feature enterprise customers are requesting? Without a context layer, that is a simple question. With twelve months of organisational history, the system knows that three enterprise customers have requested this with different constraints, that the infrastructure team assessed it as impossible six months ago due to pipeline limitations, that last month's infrastructure update removed exactly that constraint, that two competitors shipped the same feature in Q4, and that the CFO requires payback within two quarters. No individual holds all of that. The system synthesises it in seconds.
And the flywheel: month one, the system knows roughly as much as a capable new hire who has read your wiki. Month three, it has processed hundreds of code reviews, project meetings, and decisions. Month six, it connects things across teams that have never met. Every day the switching cost rises. Not because data is hard to move, but because the understanding cannot be exported.
A Handful of Companies with Real Chances
OpenAI is not alone. A handful of companies share the same end goal, and all of them have a genuine shot at getting there first.
OpenAI is building top-down: large infrastructure bets, enterprise contracts sold to CIOs, the reassurance of running on AWS.
Google provably has the most intelligent models for solving problems previously considered unsolvable. And they are currently selling access to those models at a fraction of the real cost per token. Not because they need to. But because distribution is their game. The same logic that made Android free to handset makers. The value is in the ecosystem, not the price per API call.
Anthropic got there a completely different way. Claude Code is the primary tool for more than half of all developers who use AI assistance. It is, without exaggeration, the single strongest tool humanity has produced. Every day it accumulates workflows, decisions, code history, context that reflects how people actually work. Not architected strategically. Grown organically. And potentially more valuable than context designed in from day one.
All three are racing for the same prize: the synthesis layer on top of your organisation. That is what AGI actually means in practice. Not a machine that thinks like a human, but a system that holds your entire organisation's understanding and can reason from it faster and more completely than any team.
Three Questions to Ask Now
Where is your organisation's understanding accumulating? Not where data is stored, where understanding lives. If different teams are using different AI tools with no shared memory, you are building valuable assets in silos that will never compound.
Are you building a flywheel? Are your AI systems getting smarter over time, or does every session start from scratch?
What is your switching cost in a year? If you start capturing organisational context now, how portable is it later? Who owns the understanding?
It is happening now. Most people are watching the wrong pieces.