Personal computation is back on the market, but the word personal has become slippery.
The first personal computer was personal because a person could face it. The machine had a screen, a keyboard, a pointer, a desk-shaped metaphor, a place in the home or office. The next personal computer will not be personal because it sits near a body. It will be personal because its actions come from a model the person can correct and own.
That is the whole fight hiding underneath the current product wave.
Every major lab and platform can now say some version of the same sentence: the assistant will know you, understand your context, act for you, live across your devices, work inside your browser, help at your job, remember what matters, and eventually move through the world. The sentence is true enough to matter. The models are stronger, the tools are real, the interfaces are moving out of chat, and the old surfaces are opening: email, browser, calendar, phone, glasses, ring, operating system, workplace, home, car, robot.
The decisive variable is where the brain lives.
By brain I do not mean model weights. I mean the durable, inspectable record of what the person has let through, corrected, refused, written, trusted, delegated, bought, scheduled, and changed. A brain has address, memory, correction, voice, permissions, refusals, source trails, project state, relationships, action rules, and exit. It remembers facts about the user. The facts are evidence. Boundary decisions are the prize.
Something arrived. The system guessed what it meant. The person corrected the guess. The correction survived. The next similar arrival changed.
That is the atomic unit of personal computation.
This is why email is a serious wedge rather than a nostalgic surface. Email already has the world attached. It is the oldest still-running personal-computing boundary: people, institutions, receipts, threats, invitations, obligations, newsletters, passwords, contracts, favors, invoices, and notes-to-self all know how to arrive there. One email crosses. The machine says what it thinks it is. The user corrects it. The correction becomes part of the owned model. The next crossing proves whether the product is real.
From that loop, the market becomes legible.
There are six boundary rights in personal computation: address, memory, correction, action, custody, and distribution. Address is where the world reaches the person. Memory is what persists after the crossing. Correction is who can change the model and whether the change is readable. Action is what the system may do back into the world. Custody is whether the brain can move to another surface. Distribution is the default path that makes one system present before another can be chosen.
Each major player is trying to assemble those rights from a different starting point.
OpenAI has the strongest consumer habit around frontier intelligence and has now named personal AGI as an explicit destination. The strategic thrust is to make the account feel like the person: memory, tools, devices, voice, and a new interface family around a general assistant. That is directly competitive when the user's corrections become OpenAI account memory. It is complementary when OpenAI is rented cognition reading a user-owned graph.
Google has the densest consumer context: Search, Chrome, Android, Gmail, Calendar, Maps, YouTube, Workspace, and Gemini. Its universal-assistant strategy is mechanically correct because personal computation wants crossings, and Google owns many of them. The prediction is that Google will be formidable at context and default distribution, weaker on custody, and tempted to make account memory the self.
Apple owns the private device edge. It has the trust story, the hardware boundary, the operating-system mediation, and the bodily nearness of phone, watch, earbuds, and future glasses. The prediction is that Apple becomes the most credible consumer privacy gate for personal AI while still keeping the brain inside Apple's jurisdiction. It is a partner if the device can mediate a brain the user owns. It is a competitor if private memory becomes non-portable selfhood.
Microsoft owns the work boundary: identity, documents, email, calendar, Teams, GitHub, Windows, Azure, enterprise policy, and procurement. Its Work IQ direction will win many companies because action at work already has governance. The prediction is that Microsoft will dominate organizational personal computation before it produces personal computation in the sovereign sense. The work graph belongs to the firm.
Anthropic owns trust for high-quality work agents and computer use. It is likely to be one of the best cognition suppliers for serious tasks, especially where reliability and supervision matter. The prediction is that Anthropic is more complementary than threatening if the user-owned brain can call it. It becomes competitive only if Claude memory becomes the canonical correction layer.
Meta owns social graph, attention, identity performance, glasses, and open-model leverage. Its personal-superintelligence language points toward a social and perceptual assistant. The prediction is split: Meta will keep making capable intelligence cheaper for everyone, while trying to keep the memories that matter inside Meta surfaces. Open weights help the brain. Social custody competes with it.
Amazon owns the household, commerce, voice, devices, and logistics. Alexa-style agency is structurally suited to bounded action: reorder, remind, turn on, reserve, buy, track. The prediction is that Amazon will make useful home and purchase agents before it makes a reflective self-model. It will be powerful wherever "do the thing" matters more than "show me who I am becoming."
xAI owns X adjacency, real-time public signal, and a cultural appetite for less polished contact with the world. Its strongest role is sense organ for public volatility. The prediction is that it can be valuable to a personal brain without being a good place for the private brain to live.
Perplexity is the interesting browser case. Comet is pointed at the correct crossing because the browser is where research, shopping, forms, travel, work, and curiosity turn into action. The prediction is that an agentic browser without portable correction memory becomes another account silo. An agentic browser that can consult and update an owned brain becomes a major limb.
The S&P map tells the same story in another register. NVIDIA, Broadcom, Micron, AMD, Intel, Oracle, cloud providers, and energy suppliers sell the power plant. They win the arms race whichever personal brain wins, because intelligence and agent action consume compute, chips, storage, networking, and electricity. Walmart, Amazon, JPMorgan, Visa and payment networks, health companies like Eli Lilly and Johnson & Johnson, energy companies like Exxon, and Tesla sit closer to governed action: money, procurement, health, logistics, transport, homes, bodies, and physical work. Their personal-computation plays will not look like generic chat. They will look like permissioned action under liability.
That predicts the next phase: finance, health, and robotics will force the custody question harder than writing assistants did. When a system drafts a paragraph, the self-model can be fuzzy. When it moves money, changes medication behavior, books care, opens a door, drives a car, or moves a body in a home, the system must know what it is allowed to do and why. Audit becomes product. Refusal becomes product. Permission becomes product.
YC startups are exploring the same boundary from below. Email agents, iMessage executive assistants, calendar agents, browser agents, cloud computers for agents, sales agents, recruiting agents, and company brains are all attempts to claim a crossing before the platforms absorb it. Most will be bundled away if their value is model access plus workflow glue. The ones with a chance will own corrected local memory the platforms cannot cheaply infer: the real exceptions, the taste, the project ontology, the private refusals, the messy relation between what the user says and what she means.
That keeps the claim honest. Several adjacent projects already want user-owned memory, personal models, or self-hosted agent operating systems. The narrower claim is more demanding: the owned brain has to leave the memory product and drive action across life.
Sandbar and Optimus clarify the partner test.
Sandbar's Stream ring could be a beautiful sensory port. It can capture the half-thought before it becomes a note, the hallway commitment before it becomes forgotten, the private command before it becomes socially awkward phone behavior. I want that as an input to a brain. I do not want the ring to become the brain. If it exports crossings, accepts correction, and lets the owned graph become the place where memory coheres, it is complementary. If Inner Voice becomes the canonical self-model, it is directly competitive.
Optimus is the same test with hands. A household robot should ask the brain what matters here: what is allowed, what is fragile, what is private, what routines are sacred, what exceptions require confirmation, what help means in this house. If Tesla accepts an external action contract and writes back traces, Optimus is a limb. If Tesla owns the only household model, the limb is trying to become the self.
So the strategic line is blunt: build the brain, rent the intelligence, attach the ports.
Do not try to out-distribute Google, Apple, Microsoft, Amazon, Meta, or OpenAI. Do not try to out-model the labs. Do not try to make a ring, a browser, a robot, a payments rail, and a hospital interface all at once. Make the thing all honest surfaces need and all totalizing surfaces resist: a user-owned brain that can be read, corrected, called, audited, and carried away.
That gives practical predictions.
Any product that owns a crossing will try to turn that crossing into memory.
Any product that owns memory will try to turn memory into action.
Any platform with distribution will describe its account memory as personalization before users experience it as lock-in.
Any startup whose moat is only "we connected a model to this workflow" will be eaten by the nearest platform.
Any startup whose moat is corrected boundary memory has a chance to survive bundling.
Any wearable or robot that exports crossings and accepts external correction is a limb.
Any wearable or robot that keeps the only self-model is a competing brain.
Any lab that lets an owned graph call its models is a supplier.
Any lab that keeps the user's durable corrections only inside its account is a rival for the self.
The category will keep calling itself assistants because assistants are legible to buyers. The underlying market is stranger and more serious. It is the market for where a person's model of herself is allowed to live.
Personal computation begins when the brain is outside the skull and still drives action in a user's life.
one-email-at-a-time, with adjacent graph pressure from the-inbox-is-the-blanket, a-model-of-you, the-boundary-is-the-benefit, and workflow-owns-agent-value.