There’s a graph circulating online. It shows the Azure logo — Microsoft’s blue “A” — and a single number: 45%. That’s the percentage of Microsoft’s $625 billion in commercial remaining performance obligations tied to a single customer. One company. One client responsible for nearly half the contracted future revenue of one of the largest cloud platforms on Earth.

That customer is OpenAI. And it can’t pay.

This is not speculation. It’s not a bear case built on pessimistic assumptions. It’s the math, laid bare in SEC filings, earnings calls, and leaked internal projections. What follows is a reconstruction of how Sam Altman built the most valuable private company in history on a foundation of circular financing, strategic misdirection, and a personal credibility that — as of this week — may have finally collapsed.

The Roadmap of Decisions

To understand where OpenAI is today, you have to trace the sequence of decisions that brought it here. Not the product launches or the benchmark scores. The financial and structural moves that, taken together, reveal a pattern.

Late 2025: The Corporate Restructuring. OpenAI converted from a capped-profit entity into a for-profit public benefit corporation. The timing wasn’t philosophical — it was transactional. Amazon’s $35 billion conditional investment required a clear corporate structure. Microsoft’s right of first refusal on cloud contracts was eliminated in exchange for a 27% equity stake and a $250 billion Azure purchase commitment. The rules of the game were rewritten to unlock the next round of capital. First you fix the storefront, then you open for business.

The Memory Play. Altman publicly declared that OpenAI would consume roughly 40% of global HBM (High Bandwidth Memory) production. The announcement served two purposes: it inflated the narrative of scale — “we’re so big we’re cornering global supply” — and simultaneously spiked HBM prices for every competitor. It was supply chain weaponization dressed up as a capacity announcement. The problem: it also raised OpenAI’s own costs.

The Multi-Cloud Scramble. OpenAI signed cloud commitments with everyone — $250 billion with Azure, $138 billion with AWS, $300 billion with Oracle, plus deals with Google Cloud and CoreWeave. These are simultaneous, overlapping obligations totaling well over $800 billion. A company projecting $14 billion in losses for 2026 promised to pay almost a trillion dollars to its infrastructure providers. Microsoft reportedly considered legal action over the AWS deal, arguing it breached exclusivity terms. Three hyperscalers, all owed astronomical sums, all watching each other nervously.

The Model Treadmill. GPT-4o gave way to GPT-5 (August 2025), then 5.1 (November 12), then 5.2 (December 11), then 5.3, then 5.4 — all within roughly a year. GPT-5.1 survived exactly 29 days as the flagship model before being replaced by 5.2. Not a quarter. Not a month. Four weeks. Each model cost hundreds of millions (or billions) in training compute. Each one had a shorter lifespan to generate revenue before being replaced. The trigger for the 5.2 rush was Google’s Gemini launch, which threatened to overtake ChatGPT across the full ecosystem — Search, Android, Workspace, Cloud. OpenAI’s response was to compress whatever testing and evaluation 5.2 required into a timeline that raises serious questions about whether standard safety protocols were followed. For a company founded on the premise of AI safety, the irony is difficult to overstate.

The $122 Billion That Wasn’t. In February 2026, OpenAI announced the largest private funding round in history. The headline number: $122 billion. The reality: approximately $37 billion in actual cash reached the company’s accounts. Nvidia’s $30 billion was compute credits, not money. $35 billion of Amazon’s $50 billion was conditional on OpenAI achieving AGI or completing an IPO by year-end. SoftBank’s $30 billion arrived in quarterly tranches, each contingent on progress reviews. The rest came from institutional investors — some of whom are also OpenAI’s own suppliers. The circularity is structural: Nvidia invests, OpenAI buys Nvidia GPUs, the money returns to Nvidia. Amazon invests, OpenAI spends on AWS, the money returns to Amazon. Investment and procurement become indistinguishable.

At the reported burn rate of $14–17 billion per year, $37 billion in real cash buys roughly two years of runway. The IPO isn’t an aspiration. It’s an oxygen line.

The Man Behind the Curtain

On April 6, 2026 — less than a week ago — The New Yorker published a 15,000-word investigation by Ronan Farrow and Andrew Marantz. The piece, titled “Sam Altman May Control Our Future — Can He Be Trusted?”, was built on interviews with more than a hundred sources and hundreds of pages of internal records.

The portrait it paints is not of a visionary who occasionally bends the truth. It’s of a pattern — consistent, documented, spanning decades.

At Loopt, his first startup, employees reportedly requested his removal twice. At Y Combinator, partners alleged he made personal side investments into the best companies while blocking outside investors. YC co-founder Paul Graham reportedly told colleagues that “Sam had been lying to us all the time.”

At OpenAI, the pattern escalated. Board members compiled memos about Altman’s misrepresentation of facts and safety protocols. “Lying” topped one list of behavioral concerns. One board member described him as “unconstrained by truth” — possessing “a strong desire to please people” combined with “almost a sociopathic lack of concern for the consequences of deceiving someone.”

The timing of the exposé is devastating. OpenAI needs to go on an IPO roadshow — a process that is fundamentally an exercise in trust. Institutional investors need to believe the CEO’s projections, his strategy, his numbers. Ronan Farrow — the journalist who brought down Harvey Weinstein — just published a meticulously sourced case that Sam Altman has a lifelong pattern of saying whatever the person in front of him needs to hear.

OpenAI’s response was revealing. Hours after the piece went live, the company announced a “safety fellowship” program. Internal emails from 2024, obtained through discovery in the Musk v. Altman lawsuit, show this is the standard playbook: one high-impact news moment, then “turn the page” with product announcements. When reporters asked to speak with OpenAI researchers working on existential safety, a company representative seemed confused. “What do you mean by ‘existential safety’?” he replied. “That’s not, like, a thing.”

The Numbers Don’t Close

Let’s be direct about OpenAI’s financial position as of April 2026:

Revenue: Approximately $25 billion annualized, with $2 billion per month in recent run rate. Roughly 80% comes from consumer subscriptions. Of 900 million weekly active users, fewer than 50 million pay. That means 850 million users generate cost — every query burns compute — without generating revenue. They are not an asset. They are a liability disguised as a growth metric.

Margins: Gross margin sits at approximately 33%, down from 40% in 2024. Inference costs quadrupled in 2025. Each new, more capable model is more expensive to run, not less. Operating margin is negative. The company burns roughly $2 for every $1 earned on inference before R&D, marketing, or overhead.

Losses: Projected at $14 billion for 2026. Cumulative losses projected to exceed $200 billion before reaching positive cash flow, which isn’t expected until 2029 or 2030.

Obligations: Over $800 billion in cloud infrastructure commitments across multiple providers. Training costs alone are projected at $32 billion in 2026 and $65 billion in 2027. Plus a 20% revenue share paid to Microsoft through 2032.

Assets: 110 patents (42 granted). No owned infrastructure. No datacenters. IP rights encumbered by Microsoft’s exclusive license through 2032. Two acquisitions (Windsurf at ~$3B, io Products at ~$6.5B). A brand that is increasingly associated with controversy.

On the secondary market, institutional investors holding roughly $600 million in OpenAI shares have been unable to find a single buyer. Read that again. Before the IPO even happens, there are shareholders who cannot exit at any price.

The Three Doors

OpenAI is targeting Q4 2026 for its IPO, with a valuation near $1 trillion. Goldman Sachs and Morgan Stanley are retained. The CEO wants to proceed. His own CFO, Sarah Friar, has told colleagues internally that the company is not ready — and Altman has reportedly excluded her from meetings with investors and key financial decisions in response.

Here is what happens behind each door.

Door A: The IPO launches, and the stock falls.

This is the catastrophic scenario, and it’s more likely than the market wants to admit. At a $1 trillion valuation on $25 billion in revenue, OpenAI would trade at 40x sales — for a company with negative operating margins, no clear path to profitability before 2030, and a CEO whose credibility was just publicly demolished. For context, Nvidia trades at 36x earnings while generating $121 billion in actual profit.

If the stock falls materially after listing, the dominoes begin. Amazon’s $35 billion conditional tranche — tied to the IPO — either triggers at a disappointing valuation or Amazon seeks an exit clause. SoftBank’s remaining tranches come under review. Employees holding stock options watch their compensation evaporate and start taking calls from competitors. Enterprise clients begin hedging their OpenAI dependency. Microsoft accelerates its Anthropic integration in Copilot. The narrative of inevitable dominance — the only thing holding the valuation together — breaks.

This is the Uber scenario, but worse. Uber was overvalued and unprofitable, but it owned its marketplace. OpenAI doesn’t own its infrastructure, its IP is encumbered, and its competitive moat is eroding quarterly.

Door B: The IPO succeeds — temporarily.

This is the scenario Wall Street prefers to model. Strong brand recognition, 900 million users, AI hype still warm enough to carry a premium. The stock holds or rises on day one. OpenAI raises $50–100 billion and buys another 18–24 months of runway.

But the fundamentals haven’t changed. The burn rate continues. The model treadmill keeps spinning. Quarterly earnings calls force public disclosure of metrics that were previously hidden — real margins, real customer acquisition costs, real churn rates. Analysts start asking why inference costs keep rising, why enterprise adoption isn’t scaling as projected, why the gap with more efficient competitors keeps narrowing. The stock enters a slow decline as the market reprices from “AI revolution” to “unprofitable API business with structural cost problems.”

This is the path of a thousand cuts. It doesn’t end in a single dramatic crash. It ends in a company that’s public, scrutinized, losing money, and unable to raise additional capital without massive dilution. It’s WeWork in slow motion — except the underlying product actually works. The question is whether “the product works” is enough when the business model doesn’t.

Door C: The IPO never happens.

This is the door nobody in Silicon Valley wants to discuss, but it has precedent. WeWork’s IPO was pulled in 2019 after the roadshow revealed that investor appetite was far below what the company and its bankers expected. The valuation collapsed from $47 billion to a fraction of that. The founder was removed. The company eventually went public via SPAC at a tiny valuation, then filed for bankruptcy in 2023.

The mechanism is straightforward. During the IPO roadshow, institutional investors examine the S-1 filing — the first time OpenAI’s complete financials would be public. They see the circular financing, the conditional tranches, the $800 billion in obligations, the 33% gross margins, the $200 billion projected cumulative losses. They read the Farrow profile. They note that the CFO reportedly disagrees with the timing. They compare with Anthropic — which is also preparing to go public, projects profitability two years earlier, and grows faster on a fraction of the spending.

The book doesn’t fill. The underwriters — Goldman and Morgan Stanley — advise postponement. The postponement leaks. The narrative collapses. Without the IPO, the $35 billion conditional from Amazon dies. The runway shortens dramatically. Emergency funding rounds happen at steep discounts. The $852 billion valuation becomes a memory.

In this scenario, Microsoft’s 27% stake and exclusive IP rights through 2032 become the most valuable assets in the wreckage. Satya Nadella — who has been quietly building the multi-model Copilot strategy all along — picks up the pieces at a fraction of the cost.

The Common Thread

Across all three doors, one dynamic remains constant: Microsoft wins. If the IPO succeeds, Microsoft holds 27% of a trillion-dollar company and collects on $250 billion in Azure commitments. If the IPO fails, Microsoft’s IP rights and Anthropic partnership become even more valuable as OpenAI weakens. If OpenAI collapses entirely, the infrastructure, the trained talent pool, and the market it created don’t disappear — they just get repriced and redistributed, largely to Microsoft’s benefit.

Microsoft built the casino. OpenAI is at the table, betting with borrowed chips. The house always wins.

What This Means

We are not writing this because we want OpenAI to fail. ChatGPT brought AI to the world. It changed how hundreds of millions of people work, learn, and create. That matters.

But the gap between what OpenAI represents as a product and what it represents as a business has become an abyss. A product that works is not the same as a company that can survive. And a CEO whose own board members, former partners, and colleagues describe as having a pattern of deception is asking public market investors to trust him with their retirement funds.

The AI industry will survive whatever happens to OpenAI. The models don’t disappear. The research doesn’t vanish. The demand is real. But the structure built around this particular company — the circular financing, the impossible commitments, the narrative-over-numbers culture — that structure is fragile in a way that should concern everyone who depends on it.

The three doors are open. Sam Altman is about to walk through one of them. The question isn’t which door he chooses. It’s whether any of them lead where he says they do.