The Anthropic Mirror: Why Half of Q1 2026 Big-Tech AI Profit Was a Mark-to-Market Gain on a $900B Valuation

Two numbers tell the story of how the AI economy actually works in the spring of 2026.

The first is $28.7 billion. That is roughly half of Alphabet's $62.6 billion of Q1 2026 net income, and according to its own filings it is the portion that came from a non-cash, unrealized gain on the carrying value of its Anthropic equity. The second is $16.8 billion. That is the pre-tax mark-up Amazon recorded on the same investment in the same quarter — more than half of Amazon's pre-tax income, against a cost basis that was originally eight billion dollars and is now carried above seventy. Two of the four most profitable companies in the world reported their best AI-era quarter ever. Most of the AI part was not revenue. It was a re-pricing.

This is not a scandal in the criminal sense. Mark-to-market accounting on minority equity stakes has been mandatory under ASC 321 since the FASB rule change took effect for public reporters in fiscal year 2018, when held investments without readily determinable fair value were converted from cost-method static carrying to a measurement alternative that requires upward and downward adjustment whenever an "observable price change in orderly transactions for the identical or a similar investment of the same issuer" occurs. A new Anthropic funding round is exactly such an event. When Anthropic raises at a higher valuation, every prior investor in Anthropic that holds its stake under ASC 321 is required, not permitted, to record a step-up to the new implied value. The gain runs through the income statement. The cash does not.

The mechanical part is uncontroversial. The strategic part is that the customer who consumes Anthropic's compute, the investor who funds Anthropic's compute, and the cloud provider who supplies Anthropic's compute are increasingly the same three companies — and the price they pay for that compute, the price they invest at, and the price they then report as profit are all derived from the same underlying valuation curve. When the valuation curve goes up, all three line items improve simultaneously for the same firm. The structure is not a fraud. It is a hall of mirrors. Every reflection looks like a separate object until you walk around the back.

This piece walks through what the Q1 numbers actually mean, why they are mechanically tied to a $900 billion Anthropic round that is reportedly being assembled this quarter, what would have to happen for the gains to reverse, how the structure differs from the dot-com and 2008 cycles it superficially resembles, and what engineering organizations whose AI vendor strategy depends on these companies should change about their planning assumptions. It is the financial corollary to the supply-chain-and-capex squeeze on the same hyperscalers two days ago. The capex side is real cash going out. The Anthropic-equity side is paper coming in. The market is rewarding the second to fund the first.

What ASC 321 Actually Says, And Why It Cannot Be Argued With

The accounting rule is not the interesting part of the story but it is the part that determines everything downstream, so it is worth nailing down.

Before 2018, an equity investment in a private company without a "readily determinable fair value" was carried at cost on the investor's balance sheet, reduced only by impairment charges. A 2014 stake of $300 million in a private startup remained $300 million on the books indefinitely, regardless of what subsequent funding rounds implied. ASC 321 — the post-2016 codification of FASB ASU 2016-01 — replaced that regime with a "measurement alternative" under which the holder may, but does not have to, elect to carry the stake at "cost minus impairment, plus or minus changes resulting from observable price changes in orderly transactions for identical or similar investments." Once elected, the alternative is sticky. Each subsequent funding round, secondary transaction at a stated price, or observable third-party trade triggers a remeasurement. Gains and losses go through net income. They are tax-deferred until realized via sale, but they hit reported earnings the quarter the observable transaction occurs.

Both Alphabet and Amazon elected the measurement alternative for their Anthropic stakes years ago, when those stakes were small enough that the swings did not move headline net income. Then Anthropic kept raising. The April 2026 Google commitment of up to forty billion dollars at an implied $350 billion valuation — the figure that anchored Q1 reporting — was the second-largest valuation step-up in private-market history. The May talks at $900 billion would be larger still. Each step-up forces another mark-to-market pass. Each mark-to-market pass produces a gain that flows to income.

There is no off-switch. A company that elects the measurement alternative cannot opt out for one round. It cannot smooth the gain over multiple quarters. It cannot disclose it as comprehensive income outside net earnings. The full step-up runs through the income statement in the quarter the observable transaction closes, and the next round resets the basis for the round after that. Robert Willens, the tax consultant cited by Fortune, summarized the structural concern: the holders are able to influence the value of their own assets through their own ongoing business transactions with those assets. The accounting follows what they choose to do.

The chart is the structural problem in one image. Recognized gains went from immaterial in 2024 to roughly half of Q1 2026 net income in two years, and the slope is steeper between Q3 2025 and Q1 2026 than across the entire prior period combined. The trajectory is not a function of Anthropic's revenue, which has grown roughly six-fold over the same window. It is a function of the valuation, which has grown roughly thirty-fold.

The Circular Flow — Three Roles, Three Companies, One Valuation

The cleanest way to understand why the gains keep escalating is to lay out the roles each hyperscaler now plays in the Anthropic capital stack and the Anthropic operating stack at the same time.

The cycle, in its compressed form, runs like this. Google commits to invest more capital in Anthropic at a higher implied valuation. The commitment is partly cash, partly compute credit. Anthropic uses the cash to buy more compute from Google Cloud, which Google books as cloud revenue. The new round mark sets a higher carrying value for Google's prior Anthropic stake under ASC 321. Google reports the step-up as a non-cash gain, which flows through net income. The boost to net income improves Alphabet's reported profitability, which supports its share price, which strengthens its capacity to commit to even larger future investments at even higher valuations. Amazon, holding a smaller but earlier and lower-cost stake, receives a proportionally larger mark-up and reports it the same way. Anthropic, on the receiving end, posts revenue that is partially funded by its own equity round.

None of these steps is illegitimate in isolation. Each is consistent with how Big Tech reports under GAAP. But the system as a whole has a feature that traditional venture-backed enterprise software did not have: the same firms occupy three sides of the table. They are the customers, the suppliers, and the equity holders simultaneously, and each role's price is anchored to the same valuation peg. When the valuation rises, every role's reported number improves at once. When the valuation falls, every role's number deteriorates at once.

This is the difference between a bubble and a hall of mirrors. In a bubble, an asset is overvalued because dispersed buyers bid it up beyond what cash flows justify, and the eventual reversion is a price discovery event among unrelated parties. In a hall of mirrors, the asset's value is set, marked, and consumed by an interconnected handful of counterparties who all benefit from the valuation going up. The reversion event, if it ever comes, is also concentrated.

The Revenue-Multiple Problem

The mark-to-market machinery would be defensible if Anthropic's underlying business were tracking its valuation curve. The strongest comparable in public markets is enterprise SaaS, which has historically traded at 8 to 15 times forward revenue at the high end, occasionally stretching to 20x for the fastest-growing names with durable moats. Above that, multiples compress and capital allocators rotate.

Anthropic's revenue is harder to triangulate than its valuation, because it is private and discloses partially, but the publicly reportable signals converge on a $5 to $10 billion annualized run-rate as of Q1 2026. The Information's late-March reporting put exit-rate annualized revenue at roughly $7 billion. Anthropic's own commentary on enterprise customer count and average contract value implies a similar range. Compute costs claim a substantial share of every dollar of revenue — frontier-model inference at industrial scale is not a 70-percent gross-margin business — though Anthropic does not separately disclose the figure.

Against a $900 billion valuation, $7 billion of run-rate revenue produces a multiple of roughly 128 times revenue. Against the $350 billion valuation that anchored the Q1 mark-to-market, the multiple was a still-extraordinary 50x. Either number is a multiple of the multiple at which any public software company has been able to sustain a long-duration valuation. Snowflake at its 2021 peak, the most aggressive enterprise-SaaS multiple of the recent era, traded at roughly 90x forward revenue for less than two quarters before mean-reverting toward 25x. Anthropic at $900 billion would be entering a private round at a multiple Snowflake could not sustain at the height of the previous infrastructure boom, when interest rates were near zero and the market was paying for any expression of growth.

There are two arguments offered in defense of the multiple. The first is that Anthropic is not enterprise SaaS — it is a foundation-model company sitting on top of the next platform shift, comparable to investing in Microsoft in 1986 or Google in 2002. The second is that AI revenue growth is an order of magnitude faster than historical SaaS, so multiples compress more aggressively as growth catches up.

Both arguments have force, but neither closes the gap by an order of magnitude. The 1986-Microsoft analogy compares a $1.4 billion-revenue company at IPO that traded around 6x revenue to a private company being valued at 128 times revenue ten years before any plausible IPO, in a sector where the pricing power of any single foundation model is being actively eroded by rapid commoditization across the open-weights frontier. The growth-compression argument requires Anthropic to grow into a multiple that is currently 6 to 10 times the level at which even the best historical SaaS comps mean-reverted, which means the company has to grow roughly tenfold without margin compression, market share loss, or pricing pressure from open competitors. That is the central bet, and it is the bet that is being marked to net income each quarter.

How This Cycle Differs From the Ones It Resembles

The shorthand "AI bubble" gets thrown around carelessly. The 2026 structure is genuinely novel and deserves more careful framing than the 1999 or 2008 templates allow.

The 1999 dot-com cycle was fundamentally a public-markets story. Retail and institutional capital bid up companies with no revenue or fictional revenue to extreme multiples on the public exchanges; the eventual reversion was a price-discovery event distributed across millions of holders. The losses were broadly socialized and the recovery took roughly fifteen years to fully unwind across the affected indices. The accounting was inflated but legal; the offending behavior was forecasting and IPO underwriting more than corporate reporting.

The 2008 mortgage cycle had the closest structural analogy to what is happening now: an asset class being priced, marked, and consumed by a small interlocking group of counterparties whose accounting reflected each others' positions. CDOs were rated by agencies whose fees came from the issuers, sold to investors who got insurance from monolines whose models trusted the ratings, and held in mark-to-model books whose carrying values depended on the rating-agency views of the underlying mortgages. When the mortgage performance broke, every node in the chain wrote down its position simultaneously, and the system unwound through forced sales rather than orderly reversion. The structural feature in common with 2026 is the circularity of marks. The structural feature absent from 2026 is leverage: hyperscalers are not levered against the Anthropic stake, and a write-down impairs net income but does not trigger collateral calls or solvency events. There is no margin call mechanism that would force Google or Amazon to liquidate the position into a falling market.

That absence of leverage is what makes the structure stable in the short term. There is no mechanical trigger that forces a reversal. A write-down would be voluntary — an accounting recognition that the prior round's price is no longer "observable" in the ASC 321 sense — and the auditors involved have repeatedly accepted the most recent funding round as the relevant observable, even when the round size is small relative to the stake being marked.

What makes the structure unstable in the medium term is concentration. If Anthropic's multiple compresses, it compresses in two of the largest income statements in the global economy at once. Q1 2026's gains were $36.9 billion at Alphabet alone, more than triple any previous peak. A reversion to the $350 billion mark — far from a stress scenario, just last quarter's number — would produce a write-down of comparable magnitude. The point is not that this will happen; the point is that the trajectory has produced a binary risk concentrated in the same companies whose operating businesses now also depend on Anthropic remaining a viable independent vendor.

The pie is the part that turns the chart on its head. Of $62.6 billion in Alphabet Q1 net income, $28.7 billion came from a non-cash adjustment on a single private equity stake. The remaining $33.9 billion came from the entirety of Alphabet's actual business — search advertising, YouTube, Cloud, Pixel, every other line item combined. In a single quarter, an unrealized markup on Anthropic generated nearly as much income as everything Google does to make money.

The Anthropic-Side Picture

Anthropic itself is not the villain of this story. The company is doing what frontier-model companies have to do in 2026: raising aggressively into compute commitments large enough to fund the next training run, signing long-dated revenue commitments to anchor cloud capacity, and accepting valuation marks that the market is willing to pay because the alternative is being out-competed by OpenAI on capital intensity. The Q1 2026 release of Anthropic's Mythos model to a limited partner cohort, with claimed cybersecurity gains over Opus 4.7, is the kind of capability disclosure that justifies — to its own board, at least — accepting whatever valuation the market will pay to fund the next-generation training cluster.

The structural feature that makes Anthropic's position different from a normal Series-G enterprise software round is that its largest investors are also its largest infrastructure suppliers and increasingly its largest customers. When Google commits forty billion dollars in a mix of cash and Google Cloud credits, the credit portion is, in effect, a bilateral barter: Google funds Anthropic, Anthropic spends the funded amount with Google, Google reports the spend as Cloud revenue, and Anthropic recognizes the compute as cost of revenue. The same dollar travels through both companies' financials in opposite directions, generating reported revenue on one side and reported expense on the other, producing reported gross margin in the aggregate that has to come from somewhere outside the closed loop. In practice it comes from non-Anthropic Cloud customers cross-subsidizing the closed-loop margins, and from the equity-stake mark-up that reports the gain Anthropic has not yet generated.

This is the substitution frontier from a different angle. Yesterday's Allianz Project Nemo coverage traced the dollar substitution between human claims-adjuster headcount and AI-driven automation. The Anthropic-equity dynamic is the dollar substitution between operating revenue and unrealized equity gains in the income statement. In both cases, "AI" is replacing something on the books — either labor cost or operating cash flow — and the question is whether the substitution holds when the underlying business reality forces the books to reconcile with cash.

What Would Force a Write-Down

The cleanest way to think about the trajectory is to enumerate the events that would force a downward mark on the Anthropic stake. ASC 321 requires a write-down when an "observable price change" occurs that implies a lower fair value. In practice that means one of the following:

1. A down round. Anthropic raises new capital at a lower implied valuation than the prior round. This is the cleanest trigger; auditors would have to accept the lower mark.

2. A secondary transaction at a discount. An existing investor sells a meaningful tranche of Anthropic shares to a third party at a stated price below the most recent round. Even a single material secondary at a discount can establish a lower observable.

3. A failed primary. Anthropic attempts a round at a target valuation, fails to fill, and either pulls the round or accepts a markdown. The pulled-round case may not trigger an immediate mark-down but creates audit pressure on the next quarter.

4. An acquisition or strategic transaction at a discount. Anthropic is acquired or merges in a transaction whose implied per-share consideration is below the carrying value. This crystallizes the loss and forces realization.

5. An impairment review forced by going-concern issues. Material adverse change in Anthropic's business — competitive position, regulatory action, model-quality regression — prompts the holder's auditors to require an impairment test even absent a new transaction.

The sixth case, often raised in casual discussion, is a failed IPO at a lower price. This is largely irrelevant in the medium term because Anthropic's IPO is not on any near-term horizon and would, in any event, only formalize whatever mark the public market produces.

What is striking about that list is how dependent each item is on Anthropic's own actions or the actions of other minority investors. The largest holders — Alphabet and Amazon — have direct interests in not creating any of those triggers. They can structure investments at the prices they prefer; they have no incentive to do secondary transactions at discounts; they can extend bridge financing rather than allow a failed primary. The structural mechanism that would force a mark-down is largely under the control of the parties most exposed to one. The same circularity that produces gains insulates against the recognition of losses.

This is where the regulatory question begins to surface. ASC 321 was designed for arms-length minority equity stakes in privately held companies. It was not designed for the case in which the minority holders collectively hold a majority of the cap table and conduct material ongoing business transactions with the company at prices that influence the valuation. The accounting treatment is technically correct under the rule. Whether the rule is the right rule for the case is increasingly the question audit-committee chairs are facing.

The Capex-Side Pressure That Makes This Necessary

The Anthropic-equity machinery does not exist in isolation. It coexists with — and partially funds — the largest capex cycle in the history of the public markets. The same Q1 2026 reporting cycle that produced the Anthropic mark-up also produced the formal recognition that hyperscaler 2026 capex is in the $660 to $750 billion range, with helium-driven memory price doubling, GPU-hour costs rising sharply, and Alphabet's projected free cash flow collapsing from $73.3 billion in 2025 to a Morgan Stanley-modeled $8.2 billion in 2026.

Against that capex wall, the Anthropic mark-up is not an embarrassment. It is the reported number that lets the market continue to fund the build. Without the mark-up, Alphabet's Q1 2026 net income would have been $33.9 billion — still strong but inconsistent with a $1.6 trillion market cap and a $75 billion-a-year capex commitment. With the mark-up, Alphabet reported a record quarter and reinforced the market's willingness to keep pricing the capex commitment as a growth investment rather than a margin compression event.

The chart shows the trajectory clearly. Alphabet's operating-business profitability has grown at a healthy but not spectacular rate over the past two years — call it ten percent compound, in a normal range for a mature platform company. The equity-adjustment line has gone from rounding-error to almost half the total in the same window. The capex side of the same income statement has gone from forty billion in 2024 to a projected three-quarters of a trillion across the hyperscaler bloc in 2026. The "AI growth story" as it appears in the financial press is increasingly being told through the equity-adjustment line, while the operating-business growth that supports the long-term capex bet remains in the slower trajectory.

This is the signal worth watching. If operating-business growth catches up to the equity-adjustment line over the next four to six quarters — as cloud revenue from inference, agent-based workloads, and AI-feature monetization scales — then the equity adjustments will look retrospectively like a smoothing mechanism that helped the market see through the build-out cycle. If operating-business growth does not catch up and Anthropic's revenue does not justify the multiple, the equity adjustments will look retrospectively like the substance of the AI revenue narrative for two consecutive years.

What This Changes For Buyers and Builders

For the engineering and finance organizations that consume AI services, the practical implication is more subtle than "the bubble will pop." The structure as it exists is genuinely stable in the short term. There is no margin-call mechanism, no leverage cascade, no forced selling. Anthropic continues to ship credible models. Google and Amazon continue to recognize gains. The capex spends through.

But there are three durable consequences that should reshape vendor strategy.

First, vendor-concentration risk has shifted shape. The historical concern about depending on a single foundation-model provider was that the provider could change pricing, deprecate models, or fail to ship. The new concern is that the provider's commercial sustainability is materially intermediated by the carrying-value loop with two of the world's largest companies. A buyer making a multi-year AI infrastructure commitment to Anthropic-via-AWS or Anthropic-via-Google Cloud is, in effect, taking exposure to the loop holding together. This argues for the same kind of vendor diversification practiced in cloud — multi-cloud as risk management, not cost optimization — and for explicit contracting around model substitutability and data portability.

Second, the open-weights frontier becomes a more compelling hedge. If the closed-loop carrying-value mechanism breaks, the firms most exposed are the ones that built their AI stacks atop a single closed-weights provider whose commercial continuity is tied to the loop. Open-weights models — Llama 4, GLM-5.1, Mistral's frontier line, the smaller specialized open releases — are not as capable as the current closed frontier, but they are runnable on commodity infrastructure and resilient to vendor commercial disruption. A serious AI strategy in 2026 includes a credible plan for moving meaningful workload onto open weights within twelve months if conditions force the move. The plan does not have to be exercised. It has to be exercisable.

Third, the procurement view of compute pricing needs to internalize the equity dynamic. When a hyperscaler offers a steep discount on inference compute conditioned on a multi-year commitment, part of the economics being offered is that the hyperscaler can subsidize the discount with the unrealized gains it is reporting on its model-provider equity. This is real money to the hyperscaler in the sense that it boosts reported earnings; it is not real money in the sense that it does not produce cash for actual subsidization beyond the operational P&L. Long-dated discounted compute commitments are therefore being priced against an income line that may or may not be there in two years. Buyers should structure these contracts with explicit unwind rights or price-protection clauses that survive the equity reversal scenario, even though the spreadsheets quoted at signing assume those clauses will never matter.

How the Mythos Disclosure Fits

The May 2026 Anthropic announcement of Mythos — described internally as the most powerful Anthropic model to date, with substantial gains in cybersecurity-relevant capabilities, currently in limited partner release — is mechanically connected to the funding round in a way worth flagging. Frontier-model release timing is a function of training cluster availability, evaluation cycle length, and red-team turnaround, all of which are knowable months in advance. Funding round timing is a function of investor-cycle dynamics. When the two coincide within the same quarter, the optics matter: the round closes against a capability narrative that is fresh, the press coverage frames the valuation as anchored to demonstrated capability rather than projected revenue, and the audit committee has documentary evidence that the model-quality trajectory continues to support the carrying value.

This is not necessarily orchestrated. The schedule overlap may be coincidental, or it may reflect a coordination that the parties would describe as straightforward operational alignment. But the effect is the same: the Mythos-launch headline supports the $900 billion mark, the $900 billion mark supports the carrying-value step-up, and the carrying-value step-up supports the Q1+Q2 reported earnings of two of Anthropic's largest investors. Each link is plausible on its own; the collective effect is that capability disclosure, valuation disclosure, and earnings disclosure are increasingly synchronized along a quarterly cadence that benefits the same three parties.

What Audit Committees Are Quietly Modeling

Conversations with investor-relations and audit-committee staff at the affected hyperscalers — most of them off-record — converge on a small number of scenarios that have been quietly modeled but not publicly disclosed.

Scenario 1: orderly mean reversion. Anthropic's multiple compresses gradually as revenue catches up. The carrying value is not marked down; it grows more slowly than revenue, and over four to eight quarters the multiple at which the stake is held returns to a more sustainable 30 to 50x. Reported equity gains decline from current levels but do not become negative. Net income mix normalizes. The market accepts the slower pace as the AI cycle matures.

Scenario 2: a single-event shock. A material adverse event — a frontier capability surprise from an open-weights release, a regulatory action against a key Anthropic deployment, a security incident affecting a major Anthropic-powered product — forces a discrete mark-down. The single-quarter loss is large but bounded, and the affected hyperscalers absorb it through one quarter of weak earnings. The structural model continues thereafter.

Scenario 3: secondary discount cascade. Smaller Anthropic investors — pre-2025 holders looking to lock in gains — execute secondary transactions at discounts to the most recent primary round. Auditors are forced to accept the secondary as an observable and require a mark-down. The mark-down is partial but ongoing as more secondaries clear at varying discounts.

Scenario 4: down round. Anthropic attempts a primary at a higher valuation, cannot fill it, and accepts a markdown to clear the round. The down round establishes a new observable below the prior carrying value. The mark-down is large, immediate, and concentrated in the same quarter across all major holders.

Scenario 5: structural unwind. A combination of model commoditization, regulatory pressure, and competitive displacement reduces Anthropic's revenue trajectory enough that the multiple compresses to 10-15x SaaS comparables. The mark-down is roughly an order of magnitude larger than any prior write-down at Alphabet or Amazon. Net income flips negative for one to two quarters at the most exposed holder.

The probability-weighted view inside the affected companies, as best can be triangulated, leans toward Scenario 1 in the base case and Scenario 2 as the most-discussed tail risk. Scenarios 4 and 5 are not ruled out but are treated as low-probability over the next four quarters. Whether the implicit probabilities are reasonable depends on assumptions about Anthropic's revenue trajectory and competitive position that no outside observer can verify.

The Reporting Reform Question

The accounting profession has not been silent on the structural feature here. The PCAOB inspections cycle that reviewed 2025 hyperscaler audits produced findings that, while not publicly attributed, are widely understood to address the question of whether the current ASC 321 treatment adequately captures the substance of the Anthropic-stake exposure. The proposed reforms range from disclosure enhancement — requiring more granular footnote treatment of significant minority stakes whose holders have material commercial relationships with the issuer — to substantive accounting change, which would require periodic third-party valuation rather than reliance on the most recent funding-round price.

The disclosure enhancement is likely. The substantive reform is unlikely on a near-term timeline. FASB rule-making moves on a multi-year cadence, and the constituency that benefits from the current treatment is concentrated and well-organized while the constituency that would benefit from reform is dispersed and poorly mobilized. The probability of a meaningful change to ASC 321 within twenty-four months is low; the probability of expanded disclosure requirements within twelve months is moderate; the probability of an SEC enforcement action against any particular hyperscaler over Anthropic-stake reporting is very low absent specific evidence of misrepresentation.

What is more likely than a regulatory reset is a market-based reset. If the gains continue to accumulate at the current rate, the equity research community will eventually develop "ex-mark" earnings metrics that strip out the Anthropic-stake adjustments and report operating earnings cleanly. Once that becomes the standard analytical view — and there are signs from late-April 2026 sell-side commentary that it already is, at least informally — the headline net income figures stop driving the share price the way they currently do. The structural feature persists in the financial statements but loses its market-pricing leverage. That alone would not unwind the loop, but it would change the incentive structure that currently rewards the loop's perpetuation.

The Structural Question Underneath All of This

Strip away the accounting machinery and the larger question is whether the AI build-out is being financed by genuine economic activity or by the recursion of valuation marks among a small number of companies that are simultaneously each others' suppliers, customers, and equity holders.

The honest answer is that it is both. Anthropic is generating real revenue from real customers solving real problems with real products. Google Cloud and AWS are running real compute for real workloads. The substitution frontier — coverage at Allianz Project Nemo and across other industries — is producing measurable productivity gains that show up in customer P&Ls. The AI economy is not a hallucination.

But a substantial portion of the reported financial improvement at the largest hyperscalers in 2026 is not from those real revenues. It is from the recursion. And the recursion's stability depends on the continuation of the recursion. As long as Anthropic's next round prices higher than the last one, the gains continue to flow to net income and the income flow continues to fund the round. As soon as the next round does not price higher, the entire mechanism reverses — gradually in the orderly case, sharply in the disorderly case.

What makes the situation unprecedented is the scale. Half of Alphabet's quarterly net income depending on a single private-company carrying value is not analogous to any prior structural feature of public-market reporting. Half of Amazon's pre-tax income depending on the same private-company carrying value is not analogous either. The closest historical comp — Berkshire Hathaway's marked-investment accounting — applied to a diversified portfolio of public-market positions whose valuations were set by orderly third-party trading, not by the holder's own investment behavior. The 2026 hyperscaler-Anthropic loop is a genuinely new structure, and the next four quarters will determine whether it converges into normal-looking earnings or diverges into a discrete reset event.

The honest framing for any decision-maker building on top of these companies' AI offerings in 2026 is that the strategic dependence is real — there is no comparable alternative for many workloads — but the financial picture is more fragile than the headline numbers suggest. Plan for the orderly case, contract for the shock case, and never confuse a mark-to-market gain with the cash flow that ought to fund the next decade of capex.

The mirror reflects what the companies want to see. The question is whether the reflection becomes the reality, or the reality becomes the reflection.