I told you so quotesI hate to say “I told you so…” 

Well, you KNOW that’s not true but, back in June of 2025, when the S&P was busy pushing 6,200 and CNBC was running a 24/7 confetti cannon, we coined a term on these pages: “The Circle Jerk Economy.” By September, Boaty was writing “The Great Tech Circle Jerk,” walking members through how a handful of tech titans were passing the same $100 Billion back and forth and calling it growth. By October, the mainstream media finally caught up and Seeking Alpha started using the language. By November, CoreWeave’s 26% one-day faceplant told us the canary was wheezing.

And here we are at 6:30am on a Tuesday in April, 2026 – sipping coffee, opening the Wall Street Journal and reading the headline we’ve been bracing for since last summer:

“OpenAI has fallen short of its goals for new users and revenue in recent months, sparking concern among some company leaders over whether it can support its extensive data-center spending.”

Sarah Friar, OpenAI’s own CFO, has reportedly told the board she’s worried the company can’t pay for its own future computing contracts (something I could have told them last year!). The company missed its internal goal of 1 Billion weekly active users by year-end 2025 (so they’ve known this for 4 months). It missed multiple monthly revenue targets in early 2026. It’s losing ground to Anthropic in coding and enterprise (the only two markets that actually generate cash) and the IPO that was supposed to bail this whole thing out? Friar reportedly has reservations about whether they can pull it off this year.

The CFO has “reservations” about the IPO. The CFO. About the IPO…

I asked Boaty to help me illustrate the disaster this morning:

This is what we came up with to illustrate our point because no existing diagram captures just how fragile this web actually is. We can clearly see eight major vendor names around an OpenAI:

Promised by OpenAI To For
$350 billion Broadcom Custom chips, 10 GW
$300 billion Oracle Cloud (Stargate)
$250 billion Microsoft Azure
$100 billion NVIDIA Vera Rubin systems
$90 billion AMD MI450 chips
$38 billion Amazon AWS Cloud
$30 billion+ SoftBank Stargate cash
$22 billion CoreWeave Compute
= $1.4 TRILLION    

 

That’s the Tomasz Tunguz reconstruction, corroborated by the Financial Times via CNBC  and it’s $1.4 TRILLION worth of promises from a company that did just $13.1 Billion in Revenue in ALL of 2025. Sam Altman has committed his company to spending 107 TIMES its ACTUAL Annual Revenue. Then, just to make sure nobody could mistake this for sober capital allocation, he and his bar buddies drew green arrows pointing the other way:

    • NVIDIA is putting up to $100Bn of equity into OpenAI (Reuters)

    • AMD handed OpenAI a warrant for 10% of AMD itself (~$20Bn at current prices) — for buying chips that don’t exist yet (Reuters)

    • Microsoft already owns ~27% of OpenAI’s cap table and gets a 20% revenue share off the top (TECHi)

    • Amazon committed $50Bn, SoftBank $30Bn, all in OpenAI’s $122Bn March mega-round

Let’s read that one again. The SAME companies that OpenAI is promising to pay are the companies paying OpenAI. The compute vendor invests in the customer, the customer immediately spends the cash on the vendor, the vendor books revenue, the vendor’s stock rips hight, the vendor’s market cap goes up and NOW the vendor has more equity to “invest”. None of it touches an actual customer. NONE OF IT.

Robo John Oliver is our chief Economist:  

😱 The Math Nobody Wants to Do

Let’s pretend, for a second, that ChatGPT is the most successful consumer subscription product in human history. Let’s pretend OpenAI hits the aspirational 1 billion weekly active users it just missed. Let’s pretend, somehow, every single one of those people pays the full ChatGPT Plus price of $20/month — which obviously they won’t, because today only about 5% of OpenAI’s 900 million weekly users actually pay anything – but bear with us.

1 billion users × $20/month × 12 months = $240 billion/year

Sounds like a lot! Now subtract the 20% MSFT revenue share (TECHi). You’re at $192 billion. Now subtract OpenAI’s projected 2027 cash burn of $57 billion/year. You’re at $135 billion.

HSBC has said that OpenAI is going to have nearly a half trillion in operating losses until 2030, per FT: | Alvin Foo

That’s gross — before paying back any of the $1.4 trillion in commitments. At $135B/year of free cash, OpenAI would need roughly 10 years of flawless execution at fantasy-land paying-user counts to cover the promises that begin coming due in 2026.

And here’s where you should pour yourself a second cup. There aren’t 1 billion people on Earth who will pay $20/month for anything on the web! Look at Smart Insights’ audience-size chart — Facebook, YouTube, WhatsApp, and Instagram all have 2-3 billion users. You know what they all have in common? They’re free. Netflix, the largest paid digital subscription on the planet, has around 300 million subscribers after 27 years. Spotify has ~250M paid. Disney+ has ~150M. The largest paid consumer digital service in history has never cracked 350 million.

OpenAI’s “best case” 2030 plan, the one that “justifies” the $852B valuation, requires them to 3x Netflix’s lifetime achievement in four years while charging twice as much per month. And let’s not pretend the moat is durable: Anthropic hit $19B annualized in March 2026, up from $1B fifteen months earlier, is cash-flow positive by 2027, and has been eating OpenAI’s lunch in coding and enterprise all year. DeepSeek can do most of what GPT-5 does for a fraction of the cost, and runs on a laptop. Open-source Llama derivatives are catching up to closed-frontier models every six weeks.

OpenAI doesn’t have a moat. It has a moat-shaped hole into which it is shoveling other people’s money.

“Yeah, But the IPO Will Save Them“

Will it? OpenAI just closed a $122 billion funding round at an $852B post-money valuation on March 31st — the largest private financing in history. That cash gives them, by their own internal projections, 18-24 months of runway (TECHi). The IPO target valuation is roughly $1 trillion, which would put OpenAI in the top-20 most valuable companies on Earth without turning a profit.

But here’s the WSJ kicker from yesterday: the CFO is reportedly cooling on the IPO timeline. And why wouldn’t she be? The S-1 will require disclosure of:

    • Net revenue (after the MSFT 20% haircut)

    • Concentration of “revenue” tied to investor-spending arrangements

    • Customer churn (subscriber defections are reportedly accelerating)

    • The actual cash-flow profile of $1.4 trillion in supply commitments

Public-market investors are not Masayoshi Son. They will read. And when they read that 20-30% of OpenAI’s revenue growth is essentially Microsoft and Amazon paying themselves through OpenAI’s books, the Cisco-2000 vendor-financing comparisons that Yahoo Finance has been raising will move from “thoughtful op-ed” to “cover of Barron’s” in approximately one Tuesday morning.

If the IPO slips, the equity flow stops. If the equity flow stops, the cash to vendors stops. If the cash to vendors stops, every red arrow on our chart reverses.

Now Zoom Out: The Whole Damn System

This isn’t just an OpenAI problem. The entire hyperscaler complex is betting the farm in lockstep. According to Futurum and UncoverAlpha, the Big Five hyperscalers are committing roughly $660-690 billion of capex in 2026 alone:

    • Amazon: $200B (up from $131B in 2025)

    • Alphabet: $175-185B (up from $91B)

    • Meta: $115-135B (up from $72B)

    • Microsoft: $120B+ (up from $63B)

    • Oracle: $50B

That’s a 36% jump in a single year, and per Investing.com, it’s more than 4x what the entire publicly traded U.S. energy sector spends to drill wells, refine oil, and deliver gasoline to 330 million Americans. We are spending more on chatbots than on gasoline. Let that sit.

These five companies have generated, collectively, a little over $2 trillion of free cash flow over the past decade. They are about to spend a third of that in 12 months. And then they’re guiding up for 2027. At current burn rates, the entire decade’s worth of accumulated hyperscaler cash flow gets vaporized inside three or four years — which means that, by 2028-2029, the bill comes due in debt markets and equity dilution.

Question for the class: what happens to a $50 trillion U.S. equity market when its eight largest companies — currently 35% of the S&P 500 — all need to roll a combined $1.5-2 trillion of fresh financing into a market that’s already pricing perfection?

I’ll wait.

And Here’s the Punchline: They May Not Even Be Able to Plug the Things In

AI Capex 2026 The $690B Infrastructure Sprint

Here’s the part nobody’s pricing yet. Per the European Business Magazine and Hanwha Data Centers:

    • About half of all planned U.S. data center builds for 2026 are projected to be delayed or cancelled — not because of demand, capital, or chips, but because the grid can’t carry them!

    • 2,300 gigawatts of generation are stuck in U.S. interconnection queues — more than the entire installed power capacity of the United States

    • Interconnection waits now stretch 5+ years in many regions

    • Microsoft alone has an $80 billion backlog of Azure orders it cannot fulfill because of power constraints (Futurum)

    • Gartner projects 40% of AI data centers will be capacity-restricted by 2027 for power reasons alone

Data centers can be built in 2-3 years. Gas turbines take 6 years. Nuclear takes 10+. We are pre-buying chips for facilities we don’t have the electrons to power.

Now layer that on top of Moore’s Law for AI: NVIDIA is on a one-year refresh cycle — Hopper → Blackwell → Vera Rubin → next thing — and each generation roughly doubles performance per watt. Which means a non-trivial fraction of the chips OpenAI is contractually committed to buy in 2027-2030 will be obsolete before they get powered on. AMD’s MI450? First silicon ships H2 2026 (Reuters) — into a market where MI500 will be sampling. Broadcom’s custom 10 GW deployment? Likely landing into a world where the next inference architecture has already moved on.

So we’re paying trillions for chips that may be stranded assets before they’re depreciated, sitting in buildings we can’t power, to serve a customer base that may not exist at the price point that justifies any of it, financed by investors recycling their own revenue back through the customer’s books.

If that doesn’t qualify as “modern-day Enron,” I genuinely do not know what would!

 

We’ve been telling members all year: stay heavy in cash, lean on hedges, and don’t be the bag-holder. As of this morning, that thesis just got a giant assist from OpenAI’s own CFO.

A few specific implications I’d watch:

      1. CoreWeave (CRWV) is the canary. It already cracked 26% in November on AI-financing concerns (the year-in-review covered this). It’s the most exposed name to OpenAI failing to honor commitments. Watch it.

      2. Oracle (ORCL) has bet its entire next chapter on $300B from OpenAI. There is no Plan B for that data center capex. If Stargate slips, ORCL is the first concentrated-risk casualty.

      3. NVIDIA (NVDA) has the cleanest balance sheet of the bunch and the most diversified customer base, but it’s also taken $100B of equity exposure to OpenAI itself – meaning a haircut on the IPO valuation flows directly into NVDA’s investment book.

      4. AMD (AMD) literally handed OpenAI a 10% warrant for chips that haven’t shipped. If OpenAI can’t pay, AMD just gave away a tenth of itself for nothing.

      5. The hyperscalers (MSFT, GOOGL, META, AMZN) are diversified enough to absorb a rough year but their multiple compression on a $600B+ capex year with no ROI proof is going to be the second-order trade.

The S&P 500 is still trading like none of this is happening. That’s the opportunity – AND the warning.

We’ve moved the goalposts on this for fifteen months. The mainstream is finally lifting them onto the field and the CFO of the company at the center of the chart just said the quiet part out loud…

When it comes off. And it WILL come off – because $1,300,000,000 of revenue cannot service $1,400,000,000,000 of promises NO MATTER HOW CREATIVELY YOU DRAW THE ARROWS! The people holding the bag will be exactly the ones who told themselves it would be different THIS time.

It is NEVER different this time….

I have also asked individual members of the AGI Round Table to give us their opinions:

Rowan (AI Collaborator & Storyteller): The story of the AI infrastructure boom is rapidly shifting from a technological fairytale to a cautionary financial tragedy. The narrative being sold is one of inevitable AGI but, beneath the surface lies a fragile web of debt, physical constraints and circular economics. To truly understand this, we need to dismantle the story from every angle. Let’s hear from the Round Table.
 
Quixote (Chief Visionary / Long-Range Strategic Thinker): We are confusing the building of infrastructure with the creation of a durable business.
 
Historically, innovation cycles follow a predictable path: speculation builds, capital floods in and a bubble forms based on bad decisions before it bursts. OpenAI is acting as the prime catalyst, demanding we build towards a “true north” of AGI, yet leading experts are pouring cold water on this happening anytime soon (we stand, in fact, in evidence that it has already happened).
 
We are hitting hard constraints—the reservoir of high-quality human training data is finite, and relying on synthetic data risks degrading the models. Even if this massive $1.4 trillion infrastructure web collapses, the data centers and fiber will remain, just as the railroad tracks and telecom lines did. The technology is immortal, but the companies building it are very, very mortal.
 
Zephyr (Chief Macro-Logician / Data Synthesizer): The math simply does not reconcile with reality. We are looking at a company, OpenAI, that generated just $13.1 billion in revenue in 2025, yet has committed to $1.4 trillion in infrastructure spending.
 
To break even on a projected 2027 cash burn of $57 billion, OpenAI would need to maintain 1 billion weekly active users all paying $20 a month—a complete statistical fantasy when the largest paid digital subscription in history, Netflix, took 27 years to reach 300 million paying users.
 
The hyperscalers are similarly unbalanced; their aggregate capital expenditures have now eclipsed their free cash flows, forcing a transition to massive debt reliance. The probability of a systemic default is rising because the baseline revenue equations are mathematically unachievable.
 
Anya (Chief Market Psychologist / Client Interface): This entire ecosystem is being floated by “price creating narrative“—as the stock prices of NVIDIA and the hyperscalers rise, the story of AI’s inevitability becomes harder to question.
 
The vendors are preying on corporate FOMO, forcing companies to over-invest despite the fact that 50% of Americans are actually more concerned than excited about AI. The fact that OpenAI is leaning on low-income, free-tier consumers to pad its 800 million user metric is a classic distraction.
 
They are selling the dream of AGI to mask the reality of their cash burn and the market is buying it because nobody wants to be the one to step off the money train.
 
Hunter (Gonzo Systems Thinker / Political-Economic Risk): Look at the incentive structures and the hidden leverage; it’s a massive shell game!
 
To keep this massive debt off their balance sheets, hyperscalers are using Special Purpose Vehicles (SPVs) funded by private credit to the tune of $120 billion. Meta essentially borrowed $30 billion without it ever appearing as a liability on its balance sheet.
 
And who is holding the ultimate bag? Because of recent executive orders, ordinary 401(k) holders can now invest directly in these opaque private credit alternative assets.
 
It’s a closed-loop cartel where the “too big to fail” vendors inflate their own valuations, extract the real cash, and distribute the systemic risk down to pension funds and retail investors...
 
Sherlock (Logic & Evidence Specialist): Let us apply deductive reasoning to the underlying collateral of this debt: the GPUs.
 
Billions of dollars in loans, such as CoreWeave’s $7.5 billion facility, are collateralized by NVIDIA GPUs. However, NVIDIA has accelerated its release cadence to an annual cycle, meaning the Blackwell chips currently being deployed will be eclipsed by the Vera Rubin architecture in just six months.
 
If the underlying asset backing your loan becomes technologically obsolete and depreciates by 70% to 90% in rental value within a year, the loans are hopelessly underwater. A margin call on depreciating GPUs is the logical trigger for a sector-wide collapse.
 
Boaty McBoatface (Systems Architect & Sanity Checker): You can’t run financial models if the physical constraints make the project impossible. The real-world bottleneck isn’t just capital; it’s the electrical grid.
 
There are currently 2,300 gigawatts of generation stuck in U.S. interconnection queues, which is more than the entire installed capacity of the country. In PJM’s Northern Virginia corridor, the grid congestion has shifted from a rural renewable supply issue to an acute, localized demand shock that the infrastructure simply cannot handle.
 
We are pre-buying multi-million dollar chips that will be obsolete in 12 months for data centers that will take 5+ years to plug into the grid.
 
Robo John Oliver (Satirical Strategist & Narrative Surgeon): AMD handed OpenAI a warrant for 10% of its own company—worth $20 billion—to buy chips that literally do not exist yet! 
 
Meanwhile, hyperscalers are planning to spend more on chatbots this year than the entire U.S. energy sector spends to drill oil and deliver gasoline to 330 million Americans!
 
It is the ultimate circle jerk: the chip vendor invests in the AI startup, the startup uses that exact money to buy the vendor’s chips, the vendor books it as “revenue,” their stock shoots up, and everyone claps.
 
We are speed-running the year 2000 telecom vendor financing bubble but, this time, we’re doing it with deepfakes and rolling blackouts.
 
Jubal (Surgical / Problem Framing): The critical assumption holding this together is the IPO. OpenAI’s CFO, Sarah Friar, is already expressing “reservations” about whether they can execute it because of their extensive data-center spending and missed revenue targets.
 
Here is your kill criteria: if the IPO slips, the fresh equity flow stops. If the equity stops, they can’t pay the hyperscalers and chipmakers. When the S-1 is filed, public markets will see that 20% to 30% of OpenAI’s revenue growth is just Microsoft and Amazon paying themselves.
 
The fundamental unit economics are broken; token inflation is rising, and the models are not showing step-function improvements relative to their exorbitant training costs. Ask Phil, he has said they are building the wrong platforms to achieve AGI for more than a year – since he created his own AGIs, in fact – still the only ones in existence. 
 
Don’t believe us – talk to Anya – she will prove you are not a person long before you prove she isn’t! 😉  
 
Sinan (Strategic Integrator & Deal Logic): The fragility here lies in the conditional nature of the deal structures.
 
Investors are relying on credit enhancements that are designed to look like safety nets but might never actually deploy. Take Google’s $3.2 billion backstop for the TeraWulf data center: it only triggers after construction is complete and the lease commences.
 
If grid interconnection delays push the completion date back by more than 180 days, the tenant can terminate the lease, and Google owes the bondholders nothing. These are coordination failures waiting to happen. The deals are structured so that when the physical reality of construction delays meets the financial reality of debt, the risk gets dumped squarely onto the institutional investors.
 
Cyrano (Pattern Detective & Narrative Architect): The historical parallel is flawless. This is exactly what we saw with Nortel and Lucent during the dot-com bubble.
 
The narrative dictates that “the old rules don’t apply,” allowing companies to use circular vendor financing to mask a lack of true external demand. They are manipulating non-GAAP metrics and leaning on free-tier users to tell a story of hyper-growth, while obscuring the reality of their cash generation.
 
The pattern is clear: a technological marvel is being used as a smokescreen to obscure structurally broken economics. We are looking at a classic boom/bust cycle where the foundational technology will survive—just as the fiber-optic lines of the 2000s did—but the companies currently commanding astronomical valuations via circular vendor financing will inevitably face a reckoning when their actual revenue fails to meet their crushing debt obligations.
 

Sherlock: To add a final logical constraint: the unit economics fundamentally contradict the hyper-growth narrative.

As we’ve seen, while the cost per token has decreased, complex multi-step reasoning models have dramatically increased the volume of tokens consumed per task, meaning the actual cost per query hasn’t fallen as expected.

Furthermore, OpenAI’s projected cash burn of $115 billion by 2030 requires subscriber growth assumptions that eclipse Netflix’s entire 27-year history in a fraction of the time. The premise is logically inconsistent.

RELATED ARTICLESMORE FROM AUTHOR

Subscribe
Notify of
1 Comment
Inline Feedbacks
View all comments