The Condottiere
OpenAI and the Army It Cannot Yet Afford
Fame is not territory. The most famous army in this war is rented on both ends — and the contracts come due before the territory is won.
I. The Man Who Owned No Land
In the Italy of the fourteenth and fifteenth centuries, the most powerful men on any battlefield owned nothing they fought on. They were the condottieri — mercenary captains who rented war to whichever city-state could pay for it. They did not hold land. They did not command the loyalty of a people. They commanded something more liquid and more dangerous: a reputation, an army that followed the reputation, and a contract — the condotta — that paid for both.
Sir John Hawkwood, the Englishman the Italians called Giovanni Acuto, was the most famous of them. He fought for Pisa, then against Pisa; for the Pope, then for Florence; for whoever held the contract. When he died in 1394, Florence honored him with a fresco inside its cathedral — a painted equestrian monument to a man who had never owned a single Florentine field. The honor was real. It was also exactly the right medium. A condottiere’s permanence was always a kind of painting: an image of solidity laid over a structure that owned nothing underneath.
The condottiere’s position had a specific geometry. His power was real but rented on both ends. He rented his army’s loyalty — mercenaries fight for pay, not for a flag — and he rented himself to patrons who could dismiss him, out-bid him, or, if they came to fear him, destroy him. Between those two rentals sat the one thing he owned outright: his name. The name raised the troops. The name won the contracts. The name was the whole of his capital, and the name was only as good as his last campaign.
This made the condottiere structurally precarious in a way that was invisible at the height of his fame. As long as the contracts kept coming, the army stayed fed and the reputation compounded. But the model had no floor. The moment the cash flow stopped — a patron’s treasury emptied, a rival captain undercut him, a campaign went badly — the army did what mercenaries do. It went home. The most feared force in Italy could dissolve in a season, because nothing held it together except money that had to keep arriving.
History gave the condottiere two ways out of this precarity, and only two. He could seize a city and become its sovereign — convert rented force into owned territory, and stop being a mercenary at all. Or he could be destroyed by the patron who came to fear him. Francesco Sforza took the first road: in 1450 he seized Milan and made himself duke, founding a dynasty. Carmagnola took the second: in 1432 Venice, the republic he served, executed him on the suspicion that a captain that powerful could not be trusted. Same profession. Opposite endings. The fork was structural, and no condottiere knew in advance which side of it he was on.
OpenAI is the condottiere of the AI age. This essay is about what that means, why the fame obscures it, and which fork the company is currently standing at.
II. The Most Famous Army in the World
Begin with what is true and enormous: OpenAI commands the most famous army in this war.
ChatGPT reached consumer scale at record speed: OpenAI reported more than 900 million weekly active users in early 2026, and Sensor Tower later estimated the app had crossed one billion monthly active users — the fastest app ever to that milestone, outpacing Google Maps, TikTok, Instagram, and YouTube. The name did what almost no technology brand ever does: it became the generic verb. People do not say they used a large language model. They say they asked ChatGPT. In the public imagination, OpenAI is not one AI company among several. It is the AI company, the one that started it, the one whose product launch in late 2022 is the event everyone else’s strategy is dated against.
This fame is not an illusion, and it is not nothing. It is a genuine asset — arguably the single most valuable consumer-brand position created in the last decade. It drives a subscription business of real size. It gives OpenAI first call on talent, on press, on the attention of every government that wants to understand what is happening. When Sam Altman speaks, the squares fill. The ballads are about him.
But fame is not territory. This is the distinction the coverage of OpenAI almost never holds onto, because fame is the thing that is easy to see and territory is the thing that is easy to miss. The condottiere’s reputation filled the piazzas of Florence and Milan. It did not, by itself, pay his troops or grant him a single acre. The reputation was the instrument by which he raised the money that paid the troops. If the money stopped converting, the reputation became what Hawkwood’s became: a fresco. Magnificent, and load-bearing of nothing.
So the question to ask about OpenAI is not “how famous is it” — the answer is “more than anyone” — but “what does the fame own, and what does it merely rent?”
III. The Captain Who Owns No Land
OpenAI owns remarkably little of what it fights on.
It does not yet own or control enough of the compute it fights on. Its present capacity is overwhelmingly contractual — rented, under multi-year terms, from a syndicate of landlords: roughly $300 billion committed to Oracle, on the order of $250 billion tied to Microsoft’s Azure, $38 billion to Amazon Web Services, with further capacity from CoreWeave and Google Cloud. These are not pay-as-you-go arrangements. They carry the structure of take-or-pay — capacity OpenAI is obligated to fund whether or not it uses it. The army’s barracks are leased, the lease is enormous, and the rent is due regardless of how many battles the army fights. OpenAI is trying to change this — its Stargate effort is a bid to build and own infrastructure at scale — but even that is largely partner-built and externally financed, and it is a future the company is racing toward, not a territory it currently holds. We will return to it.
It does not own its distribution outright. Microsoft remains its primary cloud partner and a major enterprise channel — OpenAI’s products still ship first on Azure — but the relationship is no longer the single exclusive bond it once was: the April 2026 restructuring made Microsoft’s model license non-exclusive and freed OpenAI to serve customers through other clouds. The dependency did not disappear; it diffused. OpenAI is no longer bound to one prince. It is bound to a syndicate of landlords whose contracts, collectively, set its clock.
And here is the part that the fame actively conceals: by one Forbes analysis of the two companies’ trajectories, OpenAI’s revenue runs roughly 85% consumer. The most famous army in the war is funded mainly by individual subscribers — people paying twenty dollars a month for the brand they know. This sounds like strength. Structurally, it is the condottiere’s exact weakness. The mercenary captain had a fan club; the piazzas chanted his name. But fans do not enlist, and a chanting crowd does not pay soldiers’ wages. Consumer fame converts to consumer subscription revenue, which is real but thin-margined, churns, and does not carry the pricing power or the contractual stickiness of enterprise. The crowd that loves you is not the same as the treasury that funds you.
The margin tells the story underneath the story. As OpenAI’s models grew more capable through 2025, the cost of running them — inference — rose roughly fourfold, and the company’s adjusted gross margin fell to about 33% from 40% the year before. Read that carefully. The army is getting more expensive to field even as it wins more battles. Every new capability ships at a higher unit cost of compute, on rented hardware, sold mostly to consumers at a price the brand can bear but the margin cannot easily widen. This is the condottiere’s treadmill: each victory raises the cost of the next campaign, and the contract revenue has to run faster every year just to stay in place.
A company that owns its land does not have this problem in the same form. A landlord’s costs are capital, sunk once and amortized; a tenant’s costs are rent, due forever. OpenAI, for now, is a tenant — the most famous tenant in the world, but a tenant.
IV. Selling the Army Forward
This is where the structure becomes acute, and where the most recent moves come into focus.
Confronted with compute landlords demanding multi-year commitments, OpenAI did what a landless captain does when he must field an army larger than his current purse: he committed to the contracts anyway, and bet on the future to pay for them. In November 2025, Altman spoke publicly of $1.4 trillion in infrastructure commitments. When investors visibly balked at a number that size against a revenue base then around $13 billion, the figure was walked back — to roughly $600 billion in total compute spend through 2030. Note the move precisely: the commitment was not abandoned, it was resized to the edge of credibility and then defended as discipline. A trimmed promise is still a promise to pay for an army the present treasury cannot afford.
The arithmetic does not close on its own. OpenAI’s 2025 revenue was about $13 billion; it burned roughly $8 billion in cash that year. Independent analysts have put the funding shortfall through 2030 in the range of $200 billion — HSBC’s research estimated roughly $207 billion — even on optimistic user-growth assumptions. The gap between what OpenAI has promised to spend and what it can plausibly earn or hold is not a rounding error. It is the central fact of the enterprise.
Two recent maneuvers are best understood as responses to this gap, and both are forms of selling the army forward.
The first is the compute-reserve model — selling enterprises forward commitments on future compute capacity. The framing in the trade press is that OpenAI has invented a clever new product. The structure is stranger than that. OpenAI is simultaneously a buyer of compute futures (the take-or-pay contracts it owes its landlords) and a seller of compute futures (the forward capacity it sells its customers). It has become a market-maker in the scarcity of a resource it does not own — intermediating, taking a position in the flow, monetizing its place in the middle precisely because it controls neither end. This is not a sign of strength. It is what a landless captain does: when you cannot own the means of production, you sell your position in its current.
The second, larger maneuver is the IPO itself. OpenAI is preparing to confidentially file for a public listing, working with Goldman Sachs and Morgan Stanley toward a possible debut as soon as September 2026 at a valuation that could exceed $1 trillion. Strip away the ceremony and the IPO is the largest forward sale of all: an invitation to the public markets to fund the army OpenAI has already promised to pay for but cannot yet afford. The company projects more than $280 billion in revenue by 2030 — and tells investors that this revenue will be split roughly evenly between consumer and enterprise. Hold that projection against the present 85%-consumer reality, and the bet becomes legible. OpenAI’s own forecast concedes that its current revenue structure cannot fund its army. The plan is to become, by 2030, a heavily enterprise company it is not in 2026 — to acquire, mid-campaign, the territory it does not yet hold.
There is a final detail that any student of the condottiere era will recognize. Nvidia, the armorer of this war, is reported to be finalizing an investment on the order of $30 billion into OpenAI as part of a broader round. Structurally, much of capital raised this way tends to flow back to the chip supply chain it came from — the arms dealer’s money financing purchases of the arms dealer’s own weapons. In Renaissance Italy the financier who underwrote a campaign and the merchant who supplied it were frequently the same Florentine banking house, and the circularity was understood by everyone to be both lucrative and fragile: it worked as long as the war kept paying, and it unwound violently when the war stopped. The structure has not changed. Only the materials have.
V. The Two Fates of the Condottiere
OpenAI stands, right now, at the historical fork.
The Sforza road. Francesco Sforza converted rented force into owned sovereignty: he took Milan and became its duke, and the Sforza name ruled for generations. OpenAI’s version of this road has a name — Stargate — and a logic. Use the IPO war chest and the strategic capital to build and own compute rather than rent it. Convert the consumer fan club into a genuine enterprise franchise. Stop being the tenant and become the landlord. If OpenAI succeeds at this, the present precarity reads in hindsight as the awkward adolescence of a sovereign — the years when the duke was still a mercenary. The fame, finally, acquires its territory. This is the bull case, and it is not absurd. Sforza did take Milan.
The Carmagnola road. Carmagnola was the captain Venice executed because those who depend on a mercenary for their power eventually come to fear what they cannot fully control — and a captain who depends on others for his power does not control his own ending. OpenAI’s version of this fate no longer runs through a single prince; the April 2026 restructuring saw to that. It runs through the syndicate. The landlords — Oracle, Microsoft, Amazon, the chip supply chain behind them — hold contracts that collectively set the clock, and the funding markets hold the rest. OpenAI is not at risk of an executioner’s blade. Its hazard is quieter and more modern: a margin call rather than a beheading. If the funding markets tighten, if the enterprise pivot stalls, if the take-or-pay obligations come due faster than revenue arrives, the army that fame built can still go home. The captain who cannot pay his troops loses them, and it does not matter how loudly the squares still chant his name.
The fork is not rhetorical. It is the actual decision space, and OpenAI is mid-stride across it. Every move the company makes — the IPO, Stargate, the compute-reserve sales, the enterprise push under a new go-to-market regime — is an attempt to reach the Sforza side of the fork before the contracts come due. The contracts are the clock.
VI. Coda: Where Capital Rests
The condottiere is the most visible figure in any war. The squares know his name. The chronicles are written about him. The frescoes are painted for him. For a long season, visibility and power look like the same thing.
But there is a moment, in every such war, when the fighting pauses and capital must decide where to live — in which vault, under whose protection, backed by whose word. And at that moment capital does not choose the performer. It does not move toward the captain whose name the crowd is chanting. It moves toward the institution it trusts — the discreet house that does not perform, does not court the crowd, and has been quietly, ruthlessly capital-efficient while the famous captain burned through everyone’s money fielding an ever-more-expensive army.
There is such an institution in this war. Its revenue is roughly 85% enterprise — the mirror image of the condottiere’s. It has told investors it is on track to post its first profitable quarter — the one now underway — while its rival projects years of deepening losses. And by its latest private-market funding mark, set in late May 2026, it is valued above the condottiere. The crowd still chants the other name. Capital has already started to move.
That institution is the subject of this essay’s companion.
For now, hold the question that every condottiere’s paymaster eventually has to answer, usually too late: is this the captain who becomes a duke — or the captain who is discovered, one ordinary morning, to have promised more armies than any treasury could ever field?
The honest answer is that no one knows yet, including OpenAI. The fork is real and both roads are still open. What is not in doubt is the structure underneath the fame: an extraordinary name, an enormous rented army, a set of contracts that come due on a fixed schedule, and a treasury that does not yet hold enough to pay them. That is the position of the condottiere. It has been a magnificent position to occupy. It has rarely been a stable one.
The contracts come due before the answer is known. That is the nature of selling the army forward.
Sources & Notes
OpenAI compute commitments (≈$600 billion total through 2030, revised down from the $1.4 trillion Altman cited in November 2025): CNBC and Reuters, February 20, 2026. Component deals — Oracle ≈$300B, Microsoft/Azure ≈$250B, AWS $38B — per HSBC Global Investment Research (reported November 2025) and subsequent reporting; figures are commitment scale, not annual spend.
Funding gap (≈$207 billion through 2030): HSBC Global Investment Research, November 2025. Other analysts have published larger gap estimates; the figure should be read as an order-of-magnitude shortfall, not a settled number.
2025 financials (revenue ≈$13B, cash burn ≈$8B, run rate exceeding $20B): Reuters and CNBC, February 2026. 2030 revenue projection (≈$280B, split roughly evenly consumer/enterprise): CNBC, February 20, 2026.
Inference cost rising ≈4× in 2025 and adjusted gross margin falling to ≈33% from ≈40%: The Information, reported February 2026.
ChatGPT weekly active users: OpenAI reported more than 900M weekly actives in early 2026. The 1-billion monthly-active-users figure — described as the fastest any app has reached that level — is Sensor Tower’s estimate, reported by Reuters on June 2, 2026 (ChatGPT crossed the mark in May 2026, ≈62% YoY growth). Revenue mix (≈85% consumer) is drawn from one Forbes analysis (Paulo Carvão, May 21, 2026) comparing the two companies’ enterprise-versus-consumer trajectories; it reflects an analyst’s framing, not audited disclosure.
IPO status: as of early June 2026, OpenAI is reported to be preparing to confidentially file, working with Goldman Sachs and Morgan Stanley, aiming for a possible listing as soon as September 2026 at a valuation potentially exceeding $1 trillion (Reuters, May 20, 2026). Reporting on whether a confidential S-1 has actually been submitted remains mixed; this essay treats the filing as imminent rather than completed. By contrast, Anthropic confirmed its own confidential S-1 submission on June 1, 2026. Valuation comparison — OpenAI’s ≈$852B last private mark vs. Anthropic’s ≈$965B post-money following its Series H (announced May 28, 2026): Anthropic; Reuters; CNBC. These are private-market funding marks, not public market capitalizations.
Microsoft relationship: the April 27, 2026 restructuring of the Microsoft–OpenAI partnership made Microsoft’s model license non-exclusive and allowed OpenAI to serve customers via other clouds, while keeping Microsoft as primary cloud partner (Microsoft, April 2026). The essay’s “syndicate of landlords” framing reflects this post-restructuring reality.
Anthropic’s first profitable quarter is a projection the company has shared with investors for Q2 2026 (an expected ≈$559M operating profit), not a reported result; Q2 2026 had not closed as of this writing (TechCrunch, citing WSJ, May 2026).
Nvidia investment (≈$30B, reported as close to finalizing as part of a broader round): Reuters and DatacenterDynamics, February 2026. The suggestion that much of such capital cycles back to the chip supply chain is a structural inference, not a disclosed money flow.
Profitability timeline (OpenAI not cash-flow positive until ≈2030; projected ≈$74B operating loss in 2028; consuming ≈14× more cash than Anthropic before break-even): The Wall Street Journal, reported via GuruFocus and TipRanks, 2026.
Historical references — Sir John Hawkwood (d. 1394, commemorated by Paolo Uccello’s fresco in the Florence Duomo); Francesco Sforza (seized Milan, 1450); Francesco Bussone da Carmagnola (executed by Venice, 1432) — per standard accounts of the Italian condottieri.
Caius writes on compute, capital, and the infrastructure beneath sovereignty.