Background: my day job is running an AI company in India whose stated ambition is to replace the screen. People occasionally ask whether that’s a joke. It isn’t, mostly. What follows is the entire plan, in public, because the plan only works if we execute it — and if we execute, knowing the plan won’t help anyone who starts later.
Almost everyone in AI is working on the same problem: make the model smarter. That work is going fine without us. Intelligence improves on a public roadmap, the price of a token falls every quarter, and any startup whose moat is “our model is smarter” is renting that moat from someone else’s release schedule. So we don’t compete on it.
The secret
Here is the thing almost nobody is building for: intelligence is trapped twice. Personally, it’s trapped behind a screen — every AI interaction begins with you typing your own life into a text box like a witness giving a statement. Organisationally, it’s trapped in silos — a company’s accumulated knowledge sits scattered across forty systems that don’t talk to each other and never will.
Everyone is racing to build a smarter brain. Almost no one is building the layer that sees your context, remembers it, and acts on it across everything you touch. That layer is the product. What follows looks like four products. It is one loop, and each turn of it pays for the next — in cash, and in context that a latecomer cannot buy at any price.
Models are rented. Memory is earned. We are in the earning business.
Phase 1 (2026–2027): IVA Glass, the beachhead
We start with smart glasses. Not because glasses are the endgame — they aren’t — but because they are the wedge: camera, microphones, private audio, a small AR overlay. Glasses that see what you see, hear what you hear, and remember it so you don’t have to. They will be expensive, on purpose, and sold to people who want an assistant rather than a demo. Amazon started with books. Tesla started with a $109,000 Roadster. Nobody’s first product is for everyone; it just has to be worth it to someone.
The business earns twice — hardware margin at purchase, an IVA OS subscription every month after. It clears the 10x bar because it deletes the most expensive step in every AI interaction today: a human re-typing their own context into a box. That subscription revenue funds Phase 2. But the cash is the smaller half. Every hour of worn context is an hour a competitor arriving in 2029 simply does not have.
Phase 2 (2027–2028): IVA OS, the personal reasoning OS
The glasses are the sensor; the OS is the point. Each user accumulates a persistent, encrypted personal knowledge graph — what you’ve read, said, decided, promised — and agents that act on it: book, draft, chase, follow up. Not a chatbot that answers. A system that executes. You hold the encryption keys; I’ll come back to why.
The strategic logic: anyone can rent a frontier model — we do. When models are commodities, proprietary context is the only durable advantage, and context can’t be scraped, licensed, or shortcut. It accrues one user, one day at a time. Meanwhile, every consumer running IVA OS is stress-testing the same reasoning engine that enterprises will pay roughly a hundred times more per seat to run. Consumer subscriptions pay for the enterprise build; consumer usage debugs it. Which brings us to Phase 3.
Phase 3 (2028–2030): one engine for your entire stack
Somewhere between 80 and 90 percent of what a company knows is unstructured and effectively unemployed — documents, email, chat, meetings, tickets, video. We connect every system through MCP servers, reason over all of it as one body of knowledge, and deploy agents that finish work across systems: close the ticket, update the CRM, draft the contract, reconcile the invoice. Not insights. Work.
Each new connector makes the engine more valuable to every existing customer, so the product compounds rather than merely scales. And once an organisation’s institutional memory lives in IVA, ripping it out means forgetting what the company has learned. Nobody signs up to forget. Enterprise revenue — the largest pool in this plan — is what pays for Phase 4.
Phase 4 (2030+): the reasoning OS for all human work
One engine, three doors — consumer, enterprise, developer — on a shared, permissioned memory fabric. We open the APIs and run a marketplace of connectors and agents, so the world builds on the platform instead of around it. There is a name for this position: last-mover advantage. Once the memory and the connective tissue are yours, further entry isn’t hard. It’s pointless.
Why the math works
I’ll spare you the spreadsheet; the logic matters more than the cells. Enterprise is the biggest line because agent execution gets priced against labour, not software budgets — a category error the incumbents will spend the next decade regretting. Consumer subscriptions plus hardware margin come second, the developer platform third. Together they pencil to roughly $150 billion a year, which at unremarkable multiples is a trillion dollars of value. No single line clears that bar. The multiplication across three does.
The honest math is simpler still. Knowledge work costs the world $40–50 trillion a year. We are not competing for the software budget; we are competing for a sliver of the work itself. The screen is a temporary accident of computing, and the entire enterprise-software industry is built on the accident. Consumer wearables and enterprise intelligence look like two businesses. They are one: context is trapped, and whoever untraps it wins both.
On trust
The obvious objection: a company whose product remembers everything had better be fanatical about who the memory belongs to. It belongs to you. Personal graphs are encrypted with keys we do not hold — not “we promise not to look” but “we cannot.” Enterprise memory never leaves the enterprise’s permission boundary. Agents act only with authority you have explicitly granted, and everything they do is auditable. This is not compliance bolted on at the end; it is the business model. The plan runs on trust compounding for a decade, and trust, like context, cannot be back-filled.
If we ever monetise your memory against you, the plan dies that quarter. It should.
So, in short, the master plan is:
- Build glasses good enough that people pay to wear a computer on their face
- Use that money — and that context — to build the personal reasoning OS
- Use that to put one reasoning engine on every system and every unstructured byte a company owns
- While doing the above, open the platform so the world builds on it, not around it
Don’t tell anyone.