By Jennifer Evans, Pattern Pulse AI / co-founder, Tech Reset Canada

Part II of The Inverted AI Bubble.

On June 1, 2026, three things happened in the same news cycle. Bernie Sanders published a New York Times op-ed proposing that the public take a 50 percent ownership stake in the largest AI companies. Anthropic confirmed it had filed a confidential draft S-1 with the SEC. And Michael Burry, posting to Cassandra Unchained, called the financing structure underneath the entire AI buildout a fugazi, his word for something engineered to look like value while the risk migrates somewhere else.

The first piece in this series argued that artificial intelligence was experiencing an inverted bubble. Demand was real and operational, prices were being held below the cost of delivery, and the companies providing AI were borrowing against future revenue to fund infrastructure that depreciated faster than the debt matured. That argument was built around the hyperscalers, Amazon, Alphabet, Microsoft, Meta, and Oracle, and it deliberately set the pure-play frontier labs aside.

The events of June 1 move the same logic onto the labs themselves. An initial public offering is the one event that forces a private company’s unit economics, depreciation assumptions, and cost structure onto a disclosed, dated page. The correction the first piece described as approaching has begun, and it is arriving first at the independent labs.

What the First Piece Established

The inverted bubble rests on a single observation. In a normal market, when costs explode and demand surges, prices rise. In AI through 2025 and into 2026, costs exploded, demand surged, and prices fell or held flat. The companies delivering the product were absorbing the gap, and the gap was large enough to collapse the free cash flow of the wealthiest companies in history.

Five mechanisms kept prices from rising: a competitive race to the bottom anchored by free open-weight models, cloud providers treating AI as a loss leader to capture enterprise lock-in, self-consumption inflating the demand signal, an accounting subsidy from stretched depreciation schedules, and an architecture ceiling that pushed every lab toward cheaper, thinner model tiers. The load-bearing wall was depreciation. AI hardware obsolesces in one to three years, the labs and hyperscalers depreciate it over five or six, and that mismatch is what allowed below-cost pricing to look sustainable on paper.

The first piece focused on the hyperscalers because they were the ones whose cash flows were visibly cracking. It set the labs aside on purpose. The distinction drawn then is the distinction that governs everything now: a diversified hyperscaler can slow its capital spending and watch free cash flow surge, because search, advertising, retail, and cloud keep generating money underneath. A pure-play lab cannot. It has nothing to retreat into.

The Shift Predicted By Part One

Three things have changed since March, and together they mark the correction beginning to arrive.

Pricing is no longer uniformly below cost. The clearest signal is Mythos, Anthropic’s most capable model, gated behind Project Glasswing and reportedly priced at five times the rate of its predecessor, $25 per million input tokens and $125 per million output against Opus 4.7’s $5 and $25. The pricing is justified openly by cost to serve. A leaked internal draft described the model as expensive for the company to run and expensive for customers to use. This is the opposite of the dynamic the first piece documented, where every efficiency gain was passed through as a further price cut. A premium, access-gated frontier model cannot be competed to the floor, because almost nobody can buy it. That is pricing power, the thing the first piece argued the labs could not reach until the industry consolidated.

Demand has begun to break where usage succeeds. Microsoft is canceling most internal Claude Code licenses across its Experiences and Devices division, with access ending June 30, because token-based billing produced bills that outpaced the headcount savings, reportedly $500 to $2,000 per engineer per month. Uber deployed Claude Code to roughly 5,000 engineers, watched usage climb past 80 percent, and reportedly exhausted its entire 2026 AI budget of $3.4 billion in four months. These are not stories about weak demand. They are stories about demand working exactly as designed and generating costs the buyer could not sustain.

And the IPO forces the math into the open. Anthropic’s filing converts a private, deferrable accounting question into a public, dated one. The depreciation assumptions, the compute commitments, the gross margins, and the revenue recognition all have to appear in a document the SEC reviews and the market prices.

Three Tiers of Exposure

Token economics have hit nearly every major player. The useful question is not who got hit. It is who can absorb it. The field separates into three tiers.

The labs sit in the most exposed position. They take the full weight of compute cost with no diversified business underneath. Confidential financials reported by the Wall Street Journal on May 20 show Anthropic at $4.8 billion in revenue for the first quarter of 2026, projecting $10.9 billion for the second, with a projected $559 million operating profit, its first. The lever is compute cost as a share of revenue, which the reporting shows falling from 71 cents per revenue dollar in the first quarter to a projected 56 cents in the second. That fifteen-cent shift is what turns a near-break-even quarter into an operating profit. It is also the entire argument, because if that ratio is not durable, the profit is not either.

The hyperscalers were hit hard on cash flow and can stand anyway. Combined 2026 capital spending across the five largest is now tracking toward roughly $805 billion by Morgan Stanley’s estimate, up 77 percent year over year. Amazon’s trailing free cash flow has fallen roughly 95 percent to $1.2 billion. The four largest hyperscalers’ combined free cash flow for the third quarter is projected near $4 billion against a post-pandemic quarterly average around $45 billion, the lowest level since 2014. Microsoft absorbed the token cost twice, once building the infrastructure and once as a customer receiving bills it could not justify. The hyperscalers are bleeding. The difference is that advertising, retail, and cloud revenue keep the lights on while they bleed.

Apple sits in a tier of its own, barely touched, insulated by architecture rather than diversification. By pushing execution to the device layer through structured intents and local models, Apple delivers an assistant experience that generates little to no marginal token cost per interaction. It opted out of the token-maxxing architecture the rest of the industry built its economics around. Three months later it looks less like a product choice and more like one of two MAANGs (swapping NVIDIA for Netflix) that read the meter correctly.

The Evidence

The numbers moved since March, and most of them moved in the direction the inverted bubble predicted. Capital spending climbed higher than projected. Free cash flow fell further than projected. The external financing gap, estimated at $1.5 trillion through 2028, is now producing visible stress: on May 3 the Financial Times reported that banks were seeking to offload risk to avoid choking on data-center debt, and Morgan Stanley has launched a data-center synthetic risk transfer to move exposure off its books.

Two numbers moved the other way, and an honest reading has to hold them. Lab revenue materialized faster than the first piece assumed. Anthropic went from roughly $1 billion in annualized revenue in late 2024 to a reported $30 billion run rate by April 2026, crossing into the mid-$40 billions by late May. The demand is not speculative. It is arriving.

Mythos is the hinge. The first piece argued that the labs were trapped in a race to the bottom they could not exit. A premium frontier model, gated and priced at a multiple of the prior generation and defended explicitly on cost-to-serve grounds, is the first crack in that trap. The repricing is also showing up on the public tier, through architecture rather than headline rates: a 10 percent surcharge for US-only data residency on the top models, separate web-search fees, and a fast-mode configuration priced at $30 and $150 per million tokens. The first piece called this repricing through architecture. It is intensifying, and it is concentrated at the company that just filed to go public.

Circular financing escalated on the same day as the IPO filing. Burry’s charge is that tens of billions of dollars of Nvidia hardware are routed through intermediary financing structures so the chips never sit on the operator’s balance sheet, with the ultimate risk traced through annuities, insurers, private credit, and special-purpose vehicles back to retirees and ordinary investors. Nvidia rejects the characterization, stating it uses no special-purpose vehicles, no vendor financing, and reports days-sales-outstanding of 53, in line with prior years. The detail that matters for the labs is the second accounting distortion sitting alongside the depreciation one. While Burry argues depreciation is understated, the hyperscalers are simultaneously booking large non-cash gains on their stakes in the labs: Alphabet recorded roughly $36.8 billion in equity gains, mostly a markup on Anthropic, and Amazon recorded about $16.8 billion on the same investment. The rising private valuation of one lab is flowing into the reported earnings of two of its largest backers as paper profit.

CoreWeave is where the mechanics become visible in public filings. It is the clearest live test of the thesis. As of its first-quarter report, CoreWeave carried $24.9 billion in total debt against $4.8 billion in equity, a leverage ratio of 5.2, with interest expense of $536 million in the quarter and capital spending guided to $31 to $35 billion for the year. Its hardware depreciates on roughly the schedule the first piece questioned, while its debt matures years later. Nvidia bought another $2 billion of CoreWeave stock during the quarter while CoreWeave committed to more Nvidia hardware, the circular structure operating in plain sight at a named public company. And Anthropic’s compute runs partly through CoreWeave, which booked the deal into a revenue backlog that reached $99.4 billion. The chain is legible end to end: a lab filing to go public, signing compute with a company levered five to one, which buys chips from a supplier that owns equity in it, whose stake is marked up inside the earnings of the lab’s other backers.

The Canadian Exposure

CoreWeave is not only a financial case study. It is a critical node on the Canadian sovereignty map, and the financial findings push the sovereignty question harder.

Canada’s $2 billion Sovereign AI Compute Strategy made its first investment a $240 million federal commitment to Cohere, which is building its compute on CoreWeave. That facility is now online and serving additional tenants, and CoreWeave separately operates a data center in Cambridge, Ontario. A program named for sovereignty anchored its first deal to a US operator subject to US jurisdiction, which was contested by domestic data-center operators at the time.

The financial fragility adds a layer the jurisdictional analysis did not have. The exposure is no longer only that a US company subject to the CLOUD Act hosts Canadian compute. It is that the company carrying that compute is leveraged five to one, loss-making, and financing depreciating hardware with debt that outlives it. If CoreWeave moves into the distress scenarios the inverted bubble describes, the counterparty underneath a federally backed sovereign compute project is the thing that gives way. Bell’s 300-megawatt Saskatchewan project stands as the nationally anchored alternative, the contrast that makes the point. The fuller treatment of the jurisdictional layer is in Whose AI Runs the Government.

The Case Against This Reading

The strongest counterargument is the IPO itself. A public offering is a disclosure event, the opposite of opacity. The S-1, when it surfaces real audited numbers, may resolve uncertainty rather than confirm a crisis, and may show better unit economics than the thesis assumes.

The profitability flip deserves the same scrutiny the bull case does. Anthropic’s projected operating profit is non-GAAP, drawn from financials the company is not yet required to report under public-company rules, and the Wall Street Journal noted it is unclear what accounting methods produced the revenue and cost figures. One account attributes the second-quarter margin improvement substantially to a temporary compute discount rather than a structural change, which would make the profit a single-quarter event. The longer history cuts against easy optimism: Anthropic’s gross margin was negative 94 percent in its prior year, and its projected 40 percent margin for 2025 was revised down from 50 percent after cloud inference costs ran 23 percent higher than expected. The valuation embeds margin expansion from roughly 40 percent toward 77 percent by 2028, among the most aggressive assumptions ever placed inside a private technology valuation.

The market is also rewarding the most exposed names rather than punishing them. CoreWeave’s stock is up roughly 80 percent on the year, its credit outlook was raised, and its $2 billion debt offering priced at 9.25 percent was five times oversubscribed. Nvidia’s denial of the circular-financing characterization is specific and on the record. And the demand-side figures from Microsoft and Uber trace to single-source reporting, accurate as cited but not independently confirmed.

The revenue itself is contested. Anthropic reports cloud-reseller revenue on a gross basis, counting total end-customer spend and booking partner payouts as expenses, which inflates the top line relative to peers that report net. OpenAI’s commercial leadership has publicly disputed Anthropic’s figure by roughly $8 billion on exactly this point. As previously noted, this criticism is somewhat unfair because this is how virtually all SaaS companies also report revenue.

Inverted and Classic at Once

The first piece asked not whether the inverted bubble would resolve but how. The IPO wave answers part of it. The resolution mechanism is disclosure plus public-market pricing, and the risk transfers from the providers to the public who buy the shares. That is the precise inversion the first piece predicted: in a normal bubble the speculators are hurt first, and here it is the providers, and now the public stepping in behind them.

The new development is that two bubbles can now occupy the same space. The inverted bubble, selling below cost, is the structural condition that has not entirely gone away. A classic bubble, the public buying into these companies at trillion-dollar valuations resting on margin assumptions that have not yet been proven, is what an IPO at 27 times revenue invites on top of it. The IPO is where the two meet. The depreciation wall still bears the load. The difference is that the public is now invited to stand under it. Watch for Uber drivers making stock recommendations.

There is one force that could convert the latent classic-bubble risk into a realized correction, and it appeared on the same day as the filing. A credible political move against the ownership structure of these companies would reprice the risk on every share before a single vote was cast. That will be the subject of the third piece in this series.

Jen Evans is Principal of Pattern Pulse AI, and co-founder of Tech Reset Canada. She is the author of Evans’ Law, the Nudgment framework, and the AI Sovereignty Maturity Model.