Four positions against consensus.

Coase explained the firm as the region where internal coordination beats the transaction costs of the market. Cybernetics (the study of control systems and informational-decision propagation) supplies the better metaphor for what the firm actually is: an information-processing organism that senses, decides, and acts. For a century the substrate of that organism was purely human — meetings, memos, hierarchy. It is now becoming a blend of human and silicon intelligence, and the blend ratio is about to move violently.

AI agents collapse the cost of coordination itself. Work that required a department becomes a workflow; a workflow becomes a standing process that runs while the humans sleep. The firms that internalise this do not merely cut costs — they change shape. The org charts of a cybernetic AI-enabled firm become wiring diagrams of human judgment and machine execution, and their revenue per employee decouples from headcount entirely. For a traditional digital-native firm to step up to the cybernetic vision, it has to enable the context graph: a machine-readable record of not just what the firm decided, but why.

Every system the enterprise has built to date — the ERP, the CRM, the ledger, the document store — captures the final state of a decision and discards the reasoning that produced it: the evidence weighed, the policy version applied, the precedent invoked, the exception granted, the override and who made it. The stack can answer what is true right now; almost nothing records what happened, in what order, and for what reason. An agent grounded only in the what must confabulate the why from general world-knowledge — it hallucinates a generic firm in place of the actual one, and where governance is poor it automates bad precedent at scale. The cybernetic corporation closes this gap by making its own judgment legible to its silicon half: decisions written down before they are final, decision rights codified so that every human override of an agent becomes a captured training signal rather than vanished tacit knowledge, a single governed definition of every entity that matters. The context graph is the organism's memory and nervous system at once — and it cannot be bought, because it is produced as a byproduct of how work already gets done. This is why AI-readiness is an organisational-design problem before it is a technical one, and why the winners will look like culture stories long before they look like product stories.

Three traits predict who makes the transition. Strategic autonomy: founder or owner control that can rewire the organisation without asking permission from a committee. Collaborative culture: workflows and people that accept agents as colleagues rather than threats — every human–agent collaboration is also a context-capture event, so the culture is the data pipeline. Capital-allocation discipline: the will to reinvest operating leverage into further silicon leverage rather than distributing it away. These traits are observable before they are priced — in internal agent deployment, in software spend per employee, in decision latency, in whether decisions live in documents or in heads.

The market still values companies as if headcount were destiny. The cybernetic corporations will make that assumption look as dated as valuing railways by the number of horses they retired. The falsifier: if agent capability plateaus at the level of autocomplete, the incumbency of process survives. Note what does not falsify it: infinitely large, infinitely cheap context windows would compress the premium on engineered context plumbing, but a model can only ingest a why that someone wrote down — the organisational leg of the thesis survives every technical scenario. We do not believe the evidence points to a plateau.

The share of economic profit captured by the top decile of firms has been concentrating for two decades, and the distribution beneath it is the more telling fact: the median firm earns roughly its cost of capital — it works, in effect, for its creditors and its landlords. The mechanisms driving the wedge are compounding, not cyclical: network effects, default status, proprietary data, ecosystems that price the exit, and regulatory burdens that function as moats for whoever can afford the compliance department. The intelligence transition accelerates all of them at once, because the price-makers own the data, the distribution, and increasingly the compute.

The unit of analysis matters. Value is not concentrating in companies so much as in layers. The modern economy is a stack, and the winning structure is a focused near-monopoly over one functional layer that everything above it must pay to traverse — leading-edge fabrication, card-network rails, accelerator silicon, the default app-store gateway, the file format an industry cannot open its own work without. Each such layer tends toward a single owner for structural reasons: the product is information-like, so the next unit costs nothing to serve and no diminishing return ever rescues the follower; share is quality, because networks make the biggest node the best node; usage feeds proprietary data loops that competitors cannot buy; a controlled standard taxes everything built in its grammar; and compliance is a fixed cost the incumbent amortises across enormous revenue while it breaks the sub-scale entrant. None of these mechanisms mean-revert. Each compounds.

Below the lords sit two kinds of serf. The first is the layer-renter: the firm that sits on top of someone else's dominated layer and pays rent to traverse it — the app on the store, the wrapper on the model, the model on the compute, the merchant on the rails. Its margin is not its own; it is a policy decision made one layer down. The second is the diversified operator, spread across functions and decisive in none. Cheap coordination rewards owning one layer and renting the rest; breadth without a dominated core is a cost centre wearing a strategy. The one legitimate form of sprawl is platform envelopment — annexing adjacent layers from a dominated core, on terms no standalone entrant can match — and it is how lords enlarge their estates.

Machine intelligence does not soften this structure; it steepens it at both ends simultaneously. Down the stack, the infrastructure layers carry the natural-monopoly signature — massive fixed cost, near-zero marginal cost, fed by data and capital the leader already owns — and cheap intelligence concentrates them harder: a single designer supplies the great majority of AI accelerators, and roughly nine-tenths of enterprise model spend accrues to three laboratories. Up the stack, the application layer dissolves, because intelligence collapses the cost of the capabilities that used to require headcount, and radically smaller challengers now assemble competitive products in months. The bottom becomes more monopolistic and the top more contestable at the same time — and the surplus generated by that fragmenting application-layer competition flows down-stack to whoever owns the scarce complements. The only question that matters per name: is this the toll-collector on the road the challengers must travel, or is it the traffic?

This is why the market's mean-reversion instinct fails here. Mean reversion was calibrated on an economy of rivalrous goods and diminishing returns, where profit attracted entry and entry compressed margins. Increasing-returns layers invert the mechanism: entry requires replicating the installed standard, the network, the regulatory position, the capital — not merely matching the product. The emerging incumbents — firms that reached scale in the digital era and are entrenching through the silicon one — do not need to out-innovate startups, which large organisations are structurally bad at; they need only own what the startups must rent. The proof is realised, not narrated: rising take-rates and gross margins, decade after decade, where textbooks predicted erosion.

For an allocator, the implication is structural rather than tactical. Index-level returns will increasingly be an average of a small number of extraordinary outcomes and a long tail of rent-payers; owning the average means owning mostly the tail. The portfolio's task is to own the toll roads, not the traffic — with one discipline attached: a layer monopoly the market already names as such is quality, not alpha. The edge lives in the emerging ones — positions turning decisive while the price still treats the company as one contestable player among many, moats that will persist longer than the duration consensus is discounting. If you cannot name the layer and the variant perception, you own a good company, not a position.

The standing risk to this thesis is political, as it has always been: the binding constraint on Standard Oil, on Bell, on the East India Company was the state, never the market. Concentration invites the antitrust cycle, and the cycle is real. Even successful antitrust has tended to ratify incumbency economics while redistributing their surface area: Standard Oil's fragments enriched their holders, the Bell breakup minted seven regional monopolies, Microsoft's remedies left the operating-system layer untouched. The sharper political force is geopolitical, and it does not dissolve layer monopolies so much as duplicate them — one champion per layer per bloc: Visa and UnionPay, TSMC and SMIC, Boeing and COMAC. Fragmentation multiplies the lords. It does not restore competition, and the serfs remain serfs in every bloc.

The traditional research organisation is a hierarchy of lossy channels. An analyst compresses a company into a note; a sector head compresses the note into a view; a portfolio manager compresses the view into a position — three hops and several weeks from the primary evidence. At every hop, information decays and incentives distort. The hierarchy is also inverted with respect to knowledge: the juniors closest to the primary evidence hold the least authority, while everyone above them consumes second-hand information and calls it judgment. A single CIO adjudicates the recommendations of many PMs in an environment of rapid hiring and firing, bonuses tied to book performance, and careers that depend on placating one person's idiosyncrasies — so information is curated for its recipient, and the CIO ends up seeing the world through a tinted lens while the middle of the organisation plays status games. The career-safe call is the one that cannot be singled out for being wrong. The structure was rational when information was scarce and coordination was expensive. Neither is true any longer.

The incentive distortions compound with an epistemic one. Institutional ideas must pass through a quantitative translation layer — the DCF, the comp sheet, the model handed around before the meeting — not because that is how businesses create value, but because it is the only way to align a committee of idiosyncratic priors. Qualitative judgment, where the actual edge lives — organisational design, layer position, cultural readiness — is dismissed as fluffy. The output is false precision on forecasts and no rigour on the assumptions that matter. The disciplined inversion is the reverse DCF: do not forecast forward; ask what the business must look like in ten years to justify today's price, then interrogate whether that endgame is plausible. Plausibility is a qualitative verdict, which is precisely why the institution cannot render it: it does not fit in a cell.

The third defect is structural and lives in the flows. Capital arrives after performance and departs after drawdowns, so the manager is best resourced to be aggressive when markets are expensive and forced conservative when they are cheap — antimatched to opportunity by construction. At industry scale, this antimatching is itself a driver of the capital cycle. A firm whose revenue is assets under management cannot repair it, because the repair means refusing the flows.

Strip the industry to first principles and the root defect is the separation of skin: the professional manages his own assets under one incentive and other people's under another. A structure that removed the defect would look like this. A council of analyst-partners co-owns the firm's equity, each running a personal book with his own capital in it. The client book is assembled, by weighted vote, from components of those personal books — a partner may only pitch what he already owns in size and has researched in depth. The CIO seat rotates annually, awarded on personal-book returns heavily discounted by the risk taken to earn them; voting power is ranked by track record, not title. Associates are hired by partners and paid from the partner's own compensation, so junior leverage is bought, not budgeted, and every partner remains an analyst. Last-quartile partners exit regardless of the assets they attract; top-quartile partners earn portfolio economics subject to clawback. Every mechanism is doing the same work: forcing information and authority to travel on merit rather than hierarchy, and collateralising every recommendation with the recommender's own capital.

Now add machine intelligence, and the redesign collapses to a single operator. A research process built as structured, versioned context — mental models, industry maps, positions, and falsifiers maintained as living documents — can be consulted, stress-tested, and extended by AI agents continuously. Cycle time collapses from weeks to hours. Coverage stops being a function of headcount. There are no hops for information to decay across: primary evidence, thesis, and falsifier live in the same versioned file, and the agents read it whole. The personal book supplies the skin in the game the industry structurally lacks; the written falsifier supplies the accountability its investment committees perform. This is the context-graph thesis applied reflexively — the research organisation is itself a firm, and the research vault is its nervous system. A disciplined individual operating this machine fields the throughput of an institution, without the institution's information loss or its incentive to hedge every sentence. The partnership structure above remains the blueprint should the machine ever need more hands; until then, headcount is a choice, not a constraint.

This thesis is not an observation about the industry; it is a description of how Laniakea Partners operates. The Research Vault is the live artifact of that machine — the context files, industry structures, and position logic, maintained daily and consulted by agents as a matter of course. The falsifier here would be evidence that hierarchy adds alpha rather than subtracting it. Two decades of active-management performance data argue otherwise.