Elon Musk Doesn’t Run Six Companies. He Runs One Router.

On Wednesday morning, Andrej Karpathy — the man who taught a generation of engineers to build neural networks — told everyone to stop writing code. Manage the agents that write it, he said. The guy who wrote the playbook just rewrote it. We covered the implications in our Signal/Noise briefing: who builds the Cisco for agents, what happens when agents get wallets, why the Fastenal vending machine is the best metaphor for the AI economy.

But the more we pulled on the thread, the more a different question emerged. Not about code. Not about models. About organizations. Specifically: what does a company need to look like when AI agents handle the work that middle management currently does?

And the answer might already exist — in analog form, running across six companies, sleeping on factory floors, and shitposting at 2 AM.

One Vision, Multiple Cap Tables

The conventional read on Elon Musk is that he’s running six companies simultaneously, which is either proof of superhuman capacity or reckless ego depending on whom you ask. Both readings are wrong.

Musk isn’t running six companies. He’s running one project — making humanity a multiplanetary species — across multiple execution vectors that happen to have separate cap tables. SpaceX is transport. Tesla is energy and automation. The Boring Company is underground infrastructure. Neuralink is the human-machine communication layer. xAI is the intelligence layer. Each company is a function within a single coherent architecture. The “context window” isn’t fragmented across six unrelated businesses. It’s unified because every packet belongs to the same network.

X is the outlier. Maybe it’s the communications bus — distribution at worldwide scale, a direct channel to 200 million people. Maybe it’s $44 billion in marketing spend. Either way, it’s the weakest fit, and we should be honest about that rather than retrofit it into the thesis.

But the other five? They’re not a portfolio. They’re a system.

The Million-Token Context Window

Here’s what most people get wrong about Musk. The popular narrative is that he’s a genius — some singular intellect operating beyond normal human capacity. That framing is both flattering and useless, because it makes the model non-replicable. “Be a genius” isn’t a strategy.

The better framing: Musk is a router with an extraordinarily deep context window.

A professional CEO manages by headers. They read quarterly reports, attend board presentations, rely on skip-levels and staff summaries to understand what’s happening four levels down. They’re essentially running a Google search on their own organization — they get an answer, but they’re sifting through noise, and they probably have to search again to get the real picture.

Musk operates like an LLM with a million-token context window loaded with SpaceX propulsion engineering, Tesla battery chemistry, Boring Company tunnel-boring mechanics, and xAI model architecture. When an engineering team presents a problem, he doesn’t route it to a committee for analysis. He routes it directly — because he understands the packet well enough to know where it goes, right now, without intermediaries.

He’s not going to win a Nobel Prize for metallurgy. But the stories from SpaceX engineers about him interrogating Raptor engine alloy decisions, or at Tesla getting into cell chemistry tradeoffs — that’s not the behavior of a pure orchestrator. That’s a router that can do deep packet inspection. He doesn’t just move information. He understands it well enough to know when someone’s routing it wrong.

And critically, that context window isn’t just deep in each domain. It’s wide across domains. SpaceX manufacturing innovations migrate to Tesla. Tesla’s energy storage knowledge feeds the Megapack grid business. xAI gets compute priority because Musk understands the energy bottleneck from running Tesla’s infrastructure. The cross-pollination between domains is where the compound advantage lives. It’s not genius. It’s accumulated, interconnected context — and it’s the reason the routing decisions are good, not just fast.

The Router Test

Musk isn’t the only one who’s demonstrated this. The pattern shows up wherever you find a founder who refuses to let the organization become the routing layer.

Steve Jobs was a router with a completely different context window — not engineering depth, but taste. Design, consumer psychology, marketing, the intersection of technology and liberal arts. He couldn’t debug a kernel panic, but he could look at a prototype and know in his bones that the radius on the corners was wrong. His context window was aesthetic, not technical. But the routing principle was identical: minimal hops, deep inspection in his domain, no tolerance for information loss between layers. Even Jobs used ad agencies — Chiat\Day, TBWA — because they were better at the craft. But he routed the creative direction personally. The agencies were agents executing within parameters he set.

Tobi Lütke at Shopify is the current-era version. A founder who still runs AI bug-fix software overnight to make his own product better. Not because the engineering team can’t handle it — because staying in the code is how he maintains the context window that makes his routing decisions good. The moment a CEO stops inspecting packets is the moment they start managing by headers. And managing by headers is how you get Nokia.

Then there’s Sergey Brin at Google. A founder who came back, declared a code red, and rerouted the entire organization toward AI. It worked — Google has done remarkable things since the founder returned to the routing table. But Google is 180,000 people. Even a founder-router can only turn an oil tanker so fast. The bureaucracy at that scale isn’t a bug. It IS the company. Every node in that network has its own objectives, its own metrics, its own political incentives. Brin can reroute the big packets. But the thousands of smaller routing decisions that determine actual execution speed? Those still move through the bureaucratic layer. That’s why OpenAI and Anthropic — smaller, leaner, founder-driven — continue to ship faster despite Google having orders of magnitude more resources and talent.

The Antibody Response

You want proof that the bureaucratic routing layer knows it’s replaceable? Look at what happened when Musk applied his model to the U.S. government.

Strip away the politics entirely. What DOGE attempted, at its core, was deep packet inspection on the largest bureaucracy on earth. Walk in, examine the routing tables, identify where information was being lost, duplicated, or deliberately misrouted, and restructure accordingly. Agree or disagree with the execution — the antibody response tells you everything. The routing layer didn’t fight back because the inspection was wrong. It fought back because the inspection was right, and every node in the network understood what that meant for them. When you try to remove a routing layer and the routing layer screams, that’s your signal that the routing layer knows it’s replaceable. Code doesn’t have a mortgage. Code doesn’t have a government pension to protect. Code self-optimizes overnight — that’s Karpathy’s self-healing agents. The humans in the routing layer have every incentive to resist optimization. The software replacing them has every incentive to pursue it.

Bureaucracy Was the First Router

So what’s the actual insight here? It’s not “be like Elon Musk.” Nobody should try. The human cost of the Musk model — sleeping at factories, burning through executives, working hours that would hospitalize most people — is unsustainable for normal humans and arguably unsustainable for Musk himself. The man is a proof of concept, not a template.

The insight is this: bureaucracy was humanity’s first corporate routing layer. And it was never very good. It was just the only option.

Think about why bureaucracies exist. Humans can’t route at scale. One person can hold meaningful context about maybe a dozen direct relationships, a few product lines, a limited set of technical domains. Beyond that, you need layers. Middle managers to aggregate information upward and distribute decisions downward. Committees to resolve conflicts between departments. Quarterly reviews so the C-suite can do header-level inspection without ever looking at the actual packets. Standups, all-hands, skip-levels, town halls — all of it is routing infrastructure built for human bandwidth limitations.

The entire apparatus of corporate middle management is a routing protocol designed around the constraint that humans are slow, forgetful, political, and limited in how much context they can hold at once.

Musk proved that you can get dramatically better results by eliminating that layer and routing directly — if you’re willing to pay the human cost. Jobs proved it in a different domain. Tobi is proving it with AI assistance. Brin proved that even a founder can only partially override it at massive scale.

Now AI enters the picture.

The Second Router

The question isn’t whether AI can replicate Elon Musk. That frames it wrong. The question is whether AI can replace the bureaucracy — the routing layer that every company uses instead of Musk, because they don’t have a Musk.

AI agents that can aggregate information, filter signal from noise, resolve resource conflicts, escalate genuine decisions to humans, and do all of it at millisecond speed rather than meeting-cadence speed. Not replacing the founder’s vision. Not replacing the strategic judgment calls that require taste, conviction, and willingness to be wrong. Replacing the ten thousand small routing decisions per day that currently require an army of middle managers forwarding emails, scheduling alignment meetings, and producing slide decks that summarize other slide decks.

Can the Musk model be replicated in software? Partially. You still need someone who’s right about the vision and has enough domain knowledge to inspect the important packets. That’s the hard problem — and possibly the permanently human one. But the sub-routing? The day-to-day traffic management that currently requires hundreds of well-paid humans sitting in the routing layer? That’s solvable. And solving it means a founder who today can run one company because they’re out of bandwidth could potentially run two or three. Not because AI replaced their judgment, but because it freed their judgment for the decisions that actually require it.

Here’s the twist: Musk himself has xAI. He’ll be the first to plug an AI routing layer into his own workflow. He can spin up agents that know more than he ever could about metallurgy, liquid sodium cooling, orbital mechanics — expanding his already enormous context window by orders of magnitude. AI doesn’t replace the human router. It makes the best human router dramatically more powerful. The fast get faster. As Munger would say: the first rule of compounding is never interrupt it unnecessarily.

The Incentive Problem

So if the model is obvious, why won’t most companies adopt it?

Charlie Munger, as always, has the answer: show me the incentives and I’ll show you the outcome.

Professional CEOs are incentivized to not do deep packet inspection. If you look too closely and something blows up, you knew. Better to stay at the header level and maintain plausible deniability. The entire management consulting industry exists to provide executives with expensive, beautifully formatted headers so they never have to examine the actual data. Bureaucracy isn’t a dysfunction. It’s a feature — for the people inside it. It distributes responsibility so thoroughly that nobody is responsible. It creates so many routing hops that no single node can be blamed for a dropped packet.

Founders operate under different incentives. Their name is on the building. There is no plausible deniability. When SpaceX has a launch failure, nobody blames the middle management layer — they blame Musk. All the risk, all the reward. That incentive structure produces fundamentally different routing behavior: faster, more direct, more willing to eliminate layers that aren’t carrying their weight.

This is why the agent transition will be led by founders, not professional managers. No hired CEO is going to automate the layer that insulates them from accountability. That’s not a failure of will. It’s a perfectly rational response to incentives. The people most capable of ripping out the bureaucratic routing layer are the ones least incentivized to do it. And the people most incentivized — founders with skin in the game and a vision that demands speed — are the ones who’ll move first.

What This Means

The companies that win the agent era won’t be the ones with the best models. They’ll be the ones willing to rip out the bureaucratic routing layer and replace it with software. Lean organizations where AI handles the sub-routing and founders handle the deep packet inspection. Where the distance between a strategic decision and its execution is measured in seconds, not quarters. Where the org chart looks less like a hierarchy and more like a network — and the routing protocol is code, not consensus.

Musk has been running the beta version of this operating system for twenty years. In analog. At enormous personal cost. The agent era makes that architecture available to every founder willing to adopt it — without the factory floors and the 3 AM shitposts.

The agent economy doesn’t need more coders. It doesn’t need more managers. It needs more people who can think like routers — who understand the packets well enough to know where they go, and who have the conviction to send them there without asking a committee first.

Bureaucracy was humanity’s first corporate routing layer. It was never very good. It was just the only option.

It’s not the only option anymore.

This essay builds on themes from our Signal/Noise briefing of March 12, 2026: “Karpathy Says Stop Coding. A Fastenal Vending Machine Explains Why He’s Right.” If you’re not subscribed, you should be.

— Harry