Regime Change: The Interregnum

A regime change is not the moment the new rules arrive. It is the interval in which the old rules stop describing reality and the new ones have not yet been written down. In that interval, the balance sheet that is already positioned for the next rulebook is worth more than the one that is compliant with the current one. That interval is open now, in four ledgers at once, and this essay is about how to read it.

Architecting Alpha, Part II

Three days after I published Intelligence Meets Resistance, the week did me the courtesy of proving the point. Andreessen Horowitz announced a $1.1 billion Machine Age Fund with the opening line "it's time to open the throttle and accelerate the physical buildout of AI." NVIDIA printed $96.2 billion of quarterly revenue, up 106 percent, and its CFO told analysts the company has supply for roughly 70 percent growth next year and demand for much more. Porsche, whose automotive operating margin fell to 0.3 percent in 2025, sold its 4,500-person IT consultancy to Tata Consultancy Services and signed a five-year, EUR1.25 billion agreement for TCS to "industrialize AI at scale" inside the company.

Part I argued that value in this cycle migrates from the model to the seam: the boundary between intelligence, bits, electrons and atoms, where the scarce inputs live. This essay is the execution half of that argument. It is about what happens to capital when the seam becomes visible to the people who write the rules, and what an operator should do in the interval before the rules catch up.

I am going to use the language of banking to do it, because banking is where regime changes are most legible. A bank is nothing but a ledger with rules about what counts. When those rules change, the bank that positioned for the change becomes more liquid, more levered and more valuable without doing anything to its assets. It simply gets recognised. The same mechanism is running right now across the physical economy, the enterprise, and the attention market. I will take them in that order.

My view

Contingent capacity is about to start counting as capacity. In the banking system, that means borrowing power prepositioned at the central bank counts as liquidity. In the physical economy, it means compute and its power supply get treated as an infrastructure asset class that insurance and private credit balance sheets are allowed to hold. In the enterprise, it means the ability to wire a model into an operating business is recognised as the binding constraint, and the firms that hold that ability get bought. In the attention market, it means owned distribution is recognised as balance-sheet equity rather than a marketing expense.

Four ledgers, one move: a claim that used to sit off the books, unrecognised, gets moved on to the books. The operators who prepositioned before the recognition are the ones the new regime pays. Everyone else pays the premium to catch up.

If that is right, three positions follow. I put them at the end.

Part I. The ledger that teaches the mechanism

A bank can be insolvent for a long time and not die. Illiquidity is what kills a bank, and specifically illiquidity on the afternoon your depositors all want their money back. Silicon Valley Bank was underwater on paper for months before March 2023. What ended it was $40 billion of withdrawals on 9 March and another $100 billion queued for the next morning, against a securities book of roughly $120 billion that was perfectly good collateral and almost none of which had been positioned at the Federal Reserve's discount window. The Fed's own review notes the bank had not tested its capacity to borrow there in 2022. The collateral existed. The claim on it did not, because nobody had done the paperwork.

That is the whole lesson, and it generalises: an institution positioned to draw on its contingent capacity is more liquid than one that is not, even if both hold identical assets.

Three years later that lesson has become policy intent. On 3 March 2026 the Treasury Secretary's prepared remarks at a roundtable on liquidity regulation described "the pressing necessity of unlocking hundreds of billions, potentially trillions, in new lending capacity to finance AI infrastructure, domestic supply chains, and the defense industrial base," and asked that liquidity rules "give appropriate capped recognition of borrowing capacity associated with collateral prepositioned at the discount window" (Treasury, sb0412). The same day the Fed's Vice Chair for Supervision said the current framework produces "liquidity hoarding" and that "by increasing the demand for reserves, it also requires the Fed to maintain a larger balance sheet" (Bowman, 3 March 2026). Three weeks later a Fed staff paper co-authored by a sitting governor put numbers on the menu: letting banks count prepositioned discount-window capacity toward their liquidity coverage ratio, with industry discussion centring on a cap of 20 percent of high-quality liquid assets, would cut structural reserve demand by $200 billion to $900 billion, and the full menu opens the door to $1.2 to $2.1 trillion of balance sheet reduction without leaving the ample-reserves regime (Anderson, Barbarino, Diercks and Miran, FEDS 2026-019).

Read the arithmetic the way a bank treasurer reads it. Say a bank holds $114 of cash and government bonds against $100 of deposits that could run in a panic. If prepositioned capacity counts for up to 20 percent of the numerator, the bank does not need a higher ratio. It needs the same ratio with less cash: about $95 of cash and bonds plus $19 of recognised capacity still shows $114 against $100. The $19 that used to be idle goes out as loans, and loans multiply. Large banks hold roughly 25 percent of their balance sheets in safe assets against roughly 10 percent before 2008; the Treasury remarks make that comparison explicitly. The gap is the lending capacity being described.

Now the part that matters for an operator. The largest bank in the country is already running its book to the future rule rather than the current one. JPMorgan's second-quarter filing shows about $310 billion of cash and deposits with banks against $5.0 trillion of assets, roughly 6 percent, the lowest of the post-crisis era (JPM 10-Q, Q2 2026). Its chief executive's April shareholder letter proposed that "the liquidity component of loans and securities should be equal to what the Fed discount window would lend against those securities" and estimated that discount-window credit alone "would increase JPMorganChase's lendable liquidity by almost $500 billion" (Dimon, 2025 letter). Six months earlier the bank had raised its ten-year financing commitment to "industries critical to national economic security" by up to $500 billion, to $1.5 trillion, with advanced manufacturing, defence, energy and frontier technology named as the sectors (JPM, 13 October 2025). The capacity the bank believes reform will unlock and the incremental lending it has committed to are the same order of magnitude. I will let you decide whether that is coincidence.

Two honesty notes before I generalise. First, none of this is enacted. The Fed's balance sheet is currently growing, not shrinking, because it has been buying roughly $40 billion of Treasury bills a month since December to keep reserves ample (New York Fed), and its balance sheet policy task force, announced in July and co-led by Jeremy Stein, Raghuram Rajan and Karen Dynan, has not reported. Second, the Treasury has not shifted issuance away from long bonds; the August refunding kept coupon sizes flat and the long-end buyback increase announced on 19 August is small (Treasury, sb0607). What exists is alignment: the Treasury asking, the supervisor agreeing, the staff pricing it, the largest bank already positioned. That is what an interregnum looks like from the inside. The rules have not changed. The ledger of who will be liquid when they do is already written.

Hold on to the shape of it: a change in what counts, sized in the hundreds of billions, transmitted through recognition rather than through new assets, front-run by the party with the best information. Now watch it happen three more times.

Part II. The machine age: compute becomes collateral

In Part I, I wrote that capital converts into intelligence almost without friction and cannot convert into transformers at all. The past three weeks show the financial system building the conversion mechanism anyway.

On 10 August NVIDIA announced memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build compute-financing platforms intended to mobilise more than $500 billion of third-party capital for AI infrastructure. The language in the release is the language of a new asset class being recognised, not of a technology being sold. Apollo's Jim Zelter: "Modern compute has emerged as a scarce, mission-critical asset class." Brookfield's Bruce Flatt: compute "is fast becoming the essential layer of infrastructure." KKR's co-chief executives: "Delivery, not ambition, is the hard part" (NVIDIA, 10 August 2026). Two months earlier Apollo and Blackstone's credit arms had closed a $35 billion private credit vehicle to finance one gigawatt of chips for a single model lab, the largest private credit deal on record. These are insurance and pension balance sheets. They can hold an asset only once the asset has been recognised as the kind of thing they are allowed to hold. That recognition is the regime change. The chips were always there.

Then the venture side. The Machine Age Fund's own announcement says AI hardware went "from a small amount of deal flow to now over 20 percent," that "machine intelligence is going vertical," and that compute density has risen 28 times from an H100 rack to a Rubin rack with one-megawatt racks expected within three years (a16z, 28 August 2026). The named scope is chips, memory, networking, storage, data centres, power and cooling, materials and electrical systems, robotics and edge devices. Every item on that list is an atom or an electron. The portfolio examples include a power company, a robotics company, a launch company and a drone company. Compare that to the H1 2026 venture picture from Part I, where 70.8 percent of $407 billion went to frontier models, and you can see the rotation starting inside the allocator that best predicted the last cycle. Its co-founder now chairs one of the Federal Reserve's new task forces. I make nothing of that beyond the observation that the people writing the rules and the people positioning for them are increasingly the same people.

Robotics capital tells the same story from the bottom up: Skild AI at $14 billion in January, Neura Robotics up to $1.4 billion in June, XPeng's robotics unit $900 million on 24 August, and robotics startups collectively raising more by mid-June than in the whole of 2025. Jensen Huang's framing, which I quoted in the Substack thesis on civilisational shifts earlier this year, is that "physical AI, as a large category, is the technology industry's first opportunity to address a $50 trillion industry," on a three-to-five-year cycle from existence proof to product (All-In, 19 March 2026). His CFO's line this week that "compute is revenue" is the same recognition event stated from the supply side: a cost line has become an income line.

So the second ledger has moved. Compute, power and the physical stack are now recognised collateral for institutional money. What was contingent capacity (a data centre you could build, a gigawatt you could interconnect) becomes recognised capacity the moment a financing platform will lend against it. And exactly as in banking, the value accrues to whoever prepositioned the collateral: the interconnection queue position, the transformer order placed in 2024, the site with power already contracted. The scarce inputs I listed in Part I did not get less scarce this month. They got a bid.

The framing I used in March, writing about a different geography, holds internationally: railroads did not win by having better technology, they won by being the only connection between two points that every commercial actor required, and the install base defines the architecture. The machine age fund is a bet on rails. So is the $500 billion financing platform. So, more quietly, is a bank running its liquidity book to a rule that has not been written yet.

Part III. Incumbents mark to market

A regime change is also visible in who is forced to sell.

Porsche's 2025 automotive operating profit was EUR90 million on EUR32 billion of revenue, a margin of 0.3 percent, after EUR3.9 billion of charges that included a EUR2.4 billion retreat from its electric product plan; first-half 2026 deliveries fell 16.5 percent and China deliveries fell 32 percent from a base that was already down 56 percent from the 2021 peak (Porsche H1 2026). Roughly one in five jobs is going by 2035. Mercedes-Benz cut its sales forecast on 28 July after China car sales fell 30 percent and it wrote down EUR704 million of Chinese assets (Euronews). These are the two most valuable automotive brands in Europe by almost any measure of desire, and they are being repriced by a cycle that combines electrification, Chinese cost curves and the arrival of intelligence in the vehicle at the same time.

What Porsche did about it is the instructive part. It sold MHP, its own management and IT consultancy, to TCS, and simultaneously bought five years of AI industrialisation from the acquirer. The chief executive's line was "we are firmly aligning our company with our core business." Translate that into the ledger: Porsche held contingent AI capacity on its own books for thirty years, could not preposition it, and has now sold the collateral to someone who can lend against it and rented back the borrowing capacity. It is the SVB move made voluntarily and in time. Whether it works depends on whether "industrialising AI at scale" can be bought as a service, which is a question I return to in Part IV.

The opposite move is vertical integration, and its most aggressive practitioner is Elon Musk. SpaceX absorbed xAI in February at a combined valuation of $1.25 trillion, then listed in June and closed its first day near $2.3 trillion of market value (TechCrunch). Tesla and SpaceX are building a joint chip fab because, in Musk's words, without it Tesla is "constrained in our ability to scale Optimus production"; Grok goes into the vehicles and, he says, will "help drive" the robot; Starlink goes into the cars; and on the Q2 call he said "we can't talk about combining companies, that kind of thing, on earnings calls," which is the kind of thing you say when you are thinking about combining companies (Electrek). In Part I the SpaceX purchase of Cursor was the tell: buy the demand because you already own the supply. The pattern since is broader. Every seam I listed (intelligence to bits, bits to electrons, electrons to atoms) is being brought inside one ownership boundary, and the engineers inside that boundary are being told to figure out the integration rather than buy it.

Two legacy responses, then: sell the seam and rent it back, or buy every seam and own the whole stack. Both are rational. Both are expensive. The interesting question for an allocator is which one the middle of the market can afford, and the answer is neither, which is the subject of the next section.

Part IV. The wiring bottleneck: implementation becomes the collateral

Here is the number that should reorganise how you think about enterprise AI. The Census Bureau's business survey put national AI use at roughly one in five firms as of May 2026, hovering between 17 and 20 percent since December (Census BTOS). Firms with more than 250 employees are at 37 percent; information and finance are near 40 and 34 percent. Four out of five businesses are not using it at all in any function. McKinsey's survey of the firms that do use it, published this week, finds that 88 percent report regular use somewhere, 44 percent are scaling it enterprise-wide, and only 37 percent can attribute any earnings impact to it, a figure that did not move in a year (McKinsey, State of AI 2026). The MIT finding from last summer that 95 percent of enterprise pilots showed no measurable return has not been contradicted by anything since.

The models are not the constraint. The wiring is. An executive search study reported by TechCrunch in July estimated the number of engineers in the United States who can reliably take a frontier model from pilot to production inside a real company at about two thousand, and the study's own emphasis was "not 2,000 available, 2,000 total" (TechCrunch, 30 July 2026). The share of companies intending to hire forward-deployed engineers went from under 10 percent in January to 70 percent by the end of June. Total compensation for the elite bench at the labs and Palantir sits around $620,000.

Watch how the capital responds to that scarcity, because it is the banking move again. In May, Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs launched an enterprise AI services firm with about $1.5 billion of committed capital (Anthropic). It bought a boutique called Fractional AI in May, named itself Ode in July, and on 20 August bought a second boutique, Casper Studios, a company that began as a Chrome extension in December 2022 and reached Netflix and PepsiCo campaigns through agency work within two years (Ode, 20 August 2026). Its chief technology officer's line is the one to remember: "Model selection matters, but it's not where the majority of calories are spent." Blackstone's operating-team head, who sits on the board, said "very few firms can take a company from one hard custom build to AI running throughout the business," and Blackstone's president said the plan is to deploy "across a range of businesses in our portfolio and beyond." OpenAI did the same thing one week later with a $4 billion deployment company led by TPG, with McKinsey, Bain and Capgemini as partners, built on the acquisition of a London consultancy (OpenAI).

Read it as a ledger entry. Implementation capacity is now recognised as the binding constraint on the value of a model, so the model companies and the private equity firms that own thousands of mid-market businesses are buying the collateral that unlocks it: small teams of people who have wired a model into a real operating company and survived. The stated targets are community banks, mid-sized manufacturers and regional health systems. That is the four-in-five. And the acquirers are not paying for the consultancies' revenue; Casper was reportedly about forty people. They are paying for prepositioned capacity: relationships inside operating businesses, and engineers who already know where the wiring goes.

Two consequences for anyone allocating capital or running a company.

For the portfolio: the liquidity of a business's AI position is its prepositioned implementation capacity, not its software spend. A company that has bought licences and run a pilot holds collateral it cannot draw on, exactly like a bank with an untested discount window. A company with three engineers who have shipped one model into one workflow that a line manager depends on has a claim it can draw on at will. When you diligence a target, ask the SVB question: not what AI do you have, but what could you draw on tomorrow afternoon if you had to.

For the operator building a business: the forward-deployed model is not just how the labs do go-to-market. It is a discovery method. Conventionally you build a product and then learn whether the pain point you guessed at exists. The forward-deployed route inverts it: you go into twenty companies in one industry, get paid to wire intelligence into them, and watch which problem shows up every time. By the time you build the product you are not guessing, and you already hold the relationships that turn it into revenue. This is what I mean when I say LNCELOT does not build the infrastructure; it builds the systems inside it. The physical economy is the largest and least-wired industry there is, and the four orders of magnitude of error budget I described in Part I mean the wiring cannot be done from outside. That is the seam we work in, and I am not going to publish the playbook.

Part V. The narrative ledger: attention becomes equity

The fourth ledger is the one that sophisticated capital still tends to treat as marketing, and I include it here precisely because it is undergoing the identical recognition event. I run growth for portfolio companies and the reason I think about content in balance-sheet terms is that the market has started to price it that way.

Start with the incumbents, because as in Part III, distress is the tell. Netflix, which built its position on owning what people watch, is now licensing video podcasts from Spotify and iHeart and pulling the video versions off YouTube; its co-chief executive told investors in January that "YouTube is not just UGC and cat videos anymore… They are TV" (Netflix Q4 2025 call). Disney invested $1 billion in OpenAI and licensed more than 200 characters for user-generated video, which is a studio conceding that the masses generate more entertainment than the studio does (Disney, 11 December 2025). NBCUniversal sent more than 25 creators to the Winter Olympics with platform partnerships from YouTube, Meta and TikTok and let advertisers sponsor the creator posts through NBC (NBC Sports). Tribeca opened its festival to social-media work. The largest owners of attention are renting it back from individuals.

Now the capital. Creator advertising in the United States reached about $37 billion in 2025, growing 26 percent against 5.7 percent for total media (IAB). A pasta-sauce company closed a round with roughly 21 creators on the cap table, individual cheques in the tens of thousands, on the explicit logic of its lead investor that "founders need attention, and creators have attention" (Fortune). Andreessen Horowitz bought a podcast network, made its founder a general partner, and a year later reports a million followers on X, a quarter of a million newsletter subscribers and a million monthly podcast downloads under the motto "own your distribution; or better yet, use ours" (a16z). The same firm that just raised a machine age fund is running a media company, and it is not doing so for fun. Distribution decides which founders it sees first. Deal flow is the asset; content is the collateral that secures it.

The rule change inside this ledger is what makes it rhyme with the other three. Content that could be produced without a person attached to it (faceless channels, recycled clips, mass-produced automation) has been derecognised. YouTube renamed its "repetitious content" rule to "inauthentic content" in July 2025 to make clear that mass-produced material is not monetisable, then in January removed sixteen channels with 35 million combined subscribers under the same policy (Social Media Today). Meta cut off monetisation for accounts that repeatedly reuse others' content. In the space of eighteen months the market's definition of what counts as an attention asset shifted from reach to recognisability: a specific person, with specific expertise, in a recognisable format, whom the feed stops for because it knows them. The practitioners who track this closely describe the shift as value in 2024, story in 2025, format in 2026: find the angle, become recognisable, then expand the mediums. I agree with that reading and would add the ledger view: value and story were income-statement items, things you produced and spent. Recognition is a balance-sheet item. It is what remains after the post is gone, and it is the only thing in the attention market that AI has made scarcer rather than cheaper.

That is why I treat distribution as capital allocation rather than marketing, and why the operators I work with are building owned content infrastructure instead of hiring an influencer. A private equity firm that controls narrative around its sector sees the deals first, prices them better, and exits into an audience it already owns. A portfolio company whose founder is recognisable in its category has prepositioned collateral for every launch that follows. And a company that is still posting a single static product image is holding the SVB portfolio: assets that are fine on paper and cannot be drawn on when the moment comes.

The three positions

I promised trades. Here they are, in the register of the ledger rather than the ticker, because the people reading this allocate capital into companies and systems rather than into baskets.

Position one: own prepositioned seam capacity in the physical stack. The financing platforms and the machine age fund have turned compute, power and the long-lead physical inputs into recognised collateral. The value in a recognition event goes to whoever positioned before it. That means interconnection queue positions, transformer and switchgear orders already placed, sites with contracted power, and verified operational data from physical assets that nobody else has instrumented. None of these are models. All of them appreciate the day a financing platform agrees to lend against them. If you run physical assets, instrument them now; the data is collateral you cannot manufacture retrospectively. If you allocate, underwrite the queue position and the operational dataset, not the pitch.

Position two: preposition implementation capacity before you need to draw on it. For a portfolio company, that means a small in-house bench that has shipped intelligence into one operational workflow that a line manager depends on, with the operational discipline (evaluation, monitoring, rollback) that lets you draw on it again tomorrow. Buying it later costs what Ode and the deployment company are paying: a premium for people rather than revenue, in a market where the elite bench numbers in the low thousands and 70 percent of companies decided in the same quarter that they need one. For a founder, the forward-deployed route into an industry nobody has wired is the cheapest discovery process available, and the physical economy is the largest such industry. The hedge: any strategy that depends on buying implementation as a commodity service is short the very thing that is being recognised as scarce. Porsche's deal will tell us in three years whether that hedge pays.

Position three: own distribution, and make it recognisable. Attention is being recognised as equity by the allocators who move first, and derecognised in its faceless, mass-produced forms by the platforms that decide what counts. The position is an owned content system with a specific person and a recognisable format in front of a defined category, run as infrastructure with a budget line and a measurement discipline, not a campaign. For a fund, the payoff is deal flow and exit demand. For a company, it is the launch that does not need to buy its audience. For the individual operator, it is the only asset in this cycle that becomes more scarce as intelligence becomes cheaper.

What does not work in an interregnum is the same in all four ledgers: holding assets you cannot draw on. The compliant balance sheet. The pilot that never shipped. The product photo. The transformer you meant to order.

How this could be wrong

This is a thesis, so it can lose. The banking leg fails if the Fed's task force reports in December and recommends nothing, in which case the alignment I described collapses into a lobbying story and the Treasury's remarks become a speech rather than a regime. The physical leg fails if the financing platforms cannot find the returns, if power and interconnection do not arrive on the timelines the buildout assumes, or if the demand from Part I simply does not materialise; the $500 billion is a memorandum, not a commitment. The implementation leg fails if wiring gets cheap faster than I expect, through better tooling, formal verification or models that integrate themselves, in which case the consultancies being bought today are stranded assets and the two thousand engineers become twenty thousand by next summer. The narrative leg fails if platform recognition turns out to be fashion rather than structure, or if the attention premium is bid away as every fund and every founder builds a media arm at once, which is what usually happens after a recognition event.

I hold all four legs as convergent evidence rather than as one causal chain, and I would rather be told which one breaks first than be agreed with. Short, specific messages to info@selfbuiltsystems.com get answers. Slop gets silence.

The rules are changing. The ledger of who is positioned when they do is being written right now, in the interval, and it will be read aloud later.

Architecting Alpha is published in the spirit of bold conjecture and ruthless criticism. Every claim is linked to its source. Where a policy is proposed rather than enacted, or a figure is a memorandum rather than a commitment, this essay says so rather than dressing it as fact.