Can the Market Absorb Three Trillion-Dollar IPOs?

A memo on the upcoming mega IPOs

Last week, SpaceX’s S-1 was released to the public. A number of people pinged me about a private memo I had previously circulated and asked me to publish it so they could share it with others to discuss. I’ve reprinted the memo below. The numbers are accurate as of its April 25 version. Figures and market conditions have since moved, but the principles and framework laid out don’t depend on them. If anything, the updates reinforce the memo.

A few notable updates:

• Anthropic agreed to terms raising $30bn at a $900bn pre-money valuation (after initially having signaled a $50bn raise). They also signed a $1.25bn/month deal with SpaceX for compute through May 2029 (~$45bn in aggregate) and are now in talks with Microsoft to use its Maia 200 chips. The company recorded ~$4.8bn of revenue in 1Q2026 and is reportedly on track to hit $10.9bn and record its first operating profit in 2Q2026.

• OpenAI is preparing to file its S-1 and is targeting a September listing, although CEO Sam Altman reportedly told staff at a company all-hands that filing for an IPO is different from being ready to be a public company, and they would wait until they are. The company recorded $5.7bn of revenue in 1Q2026.

The central question, then and now: can the market absorb three $1tn+ IPOs in a single window?

If this is something you’ve been thinking about, I’d love to hear from you. I’ll be in San Francisco, New York, and London over the next 2 months (email; twitter—DMs open).

---

Memo

Date: April 25, 2026
Re: Scenario Analysis – Financing Paths for the Frontier AI Labs

Introduction – Executive Summary, Orientation, Framing

I’m a long-time tech bull whose largest position over the past five years has been either Nvidia or Google, and I use AI tools daily. But I’m also a student of markets, and what I see right now is unprecedented: three cash-burning companies are planning to IPO at $1tn+ valuations in a single year. In the long run, I’m long AI. But in the short run, I wonder whether the market can absorb them.

A note on velocity: the industry is moving at a weekly cadence. Figures are current as of the date above; structural arguments do not depend on them.

Executive Summary

Over the next 12 months, SpaceX, OpenAI, and Anthropic are each set to go public at $1tn+ valuations. Whether the market can absorb them comes down to one question: Will the reference class of public company history apply, or is AI a truly transformational paradigm shift?

The technology is genuinely remarkable, and what looks unjustifiable against historical baselines may actually be appropriately priced – or even underpriced. But if the reference class holds, the math is brutal. The frontier labs are losing tens of billions of dollars a year, with OpenAI projected to lose $85bn in 2028 alone. They need to raise more capital, but existing capital pools are tapped out. Endowments, pensions, and sovereign funds are at concentration limits for the AI labs. Strategic investors are already heavily invested via circular arrangements and have signaled that further resources are contingent on a public market listing.

But as a result of circular dealmaking, AI lab financials are unclear: ARR is extrapolated from four-week windows and gross margins appear to be negative. Combine that with the labs’ burn trajectories and multi-year compute commitments, and the IPO window is a few quarters, not years. Further complicating matters is SpaceX’s planned IPO. It has already driven index inclusion rule changes that the AI labs can also benefit from, so if SpaceX prices cleanly and the aftermarket holds, Anthropic and OpenAI will follow within 6-12 months. If it stumbles, the labs may have to delay their IPOs – which they may not be able to afford.

But even an IPO doesn’t magically fix things. OpenAI needs an estimated $115-218bn more in capital through 2029 – no single IPO clears that, and traditional corporate debt markets are mostly closed to companies this unprofitable. The path will be a series of equity raises plus structured project finance for years, with the market getting a vote each time.

However, if AI is the paradigm shift its leaders are championing, none of this may matter. In fact, today’s marks would be cheap, not expensive. The capital stack arithmetic that looks impossible against historical baselines becomes mechanical against a TAM that has grown to absorb it.

This memo treats both branches seriously, although the page count may suggest otherwise. Parts I-III are longer because the financing path conditional on the reference class holding is more mechanical and analyzable in detail. Part IV addresses the paradigm shift case directly and argues it is systematically underpriced, in part because humans are structurally bad at exponentials. And the bull case doesn’t even require believing in AGI or ASI – just that its current ability is good enough to take a material share of the $3-4tn in global white-collar services.

Orientation

This memo is organized into four parts:

1. The Fundamental Fork: Stay Private or Go Public

2. Post-IPO Trajectories: Acceleration, Backstop, Acquisition, Nationalization

3. Parallel Capital Channels: Structured Compute Financing

4. Meta Questions: Open-Source and Jevons Paradox

Framing: Five Premises

1. Circularity on both sides of the capital stack

AI lab “raises” are not straightforward equity rounds. They are contingent on milestones (e.g., AGI, IPO) and heavily skewed toward compute-for-equity swaps that, once aggregated, make the dominant capital flow visible: hyperscaler/chipmaker invests in lab -> lab commits multi-year spend on chips/cloud -> counterparty books revenue and narrative support -> equity multiple expands -> enables further investment.

Circularity is a stabilizer on the way up and a risk factor on the way down. The same channels that reinforce valuations when things are going well amplify drawdowns when they aren’t.

Many of the capital flows labeled “investment” are structurally round-trips. The net new cash in the system is smaller than the sum of the reported line items. The same structure shows up consistently: Nvidia-OpenAI, Anthropic–AWS, Oracle–OpenAI, Stargate. None of these are accounting improprieties on their own, but aggregating them produces a capital-stack number that overstates true new capital by a factor analysts will need to estimate individually.

In fact, it should be noted that circularity in business is not inherently problematic – Nvidia generates real cash from real chip sales, then deploys some of that into the ecosystem via investments across the megalabs and neolabs. That’s how successful platforms have always built their user and financial bases, from Microsoft’s developer ecosystem to Apple’s supplier investments to AWS’s startup credits. The specific concern is not that circularity exists, but that the specific magnitude and structure of these arrangements, once aggregated, make the net new external capital materially smaller than the sum of the reported line items suggests.

A parallel dynamic operates on the demand side. Even granting that the labs’ reported revenue is GAAP-legitimate, some material fraction of the demand producing that revenue may be engineered by loss-making pricing rather than reflecting willingness-to-pay at profitable unit economics.

In a functioning market, price is the leveler of supply and demand – genuine excess demand produces higher prices, not sold-out inventory. The labs’ current pricing is below profit-maximizing because the strategic objective is market share, not margin: VC capital (and strategic-partner capital functioning as VC) subsidizes per-token pricing, which engineers additional usage that wouldn’t exist at market-clearing prices.

A corollary worth highlighting: if current pricing is below profit-maximizing as a strategic choice, then a portion of today’s losses is elective rather than structural. The AI labs could raise prices toward unit-economic breakeven (as Anthropic appears to be doing), which could lower demand but would also decrease unit losses. OpenAI’s projected $85bn in 2028 losses embeds the assumption that subsidized pricing continues. This cuts both directions: it makes the loss numbers less terrifying as a fixed constraint, but it also reveals how dependent the demand story is on the subsidy continuing.

Per VC Bill Gurley, the ride-sharing analog is the cleanest historical parallel: Uber’s VC-subsidized “structural demand” reverted when Lyft’s IPO forced partial reconciliation with unit economics. Uber’s peak annual losses topped out around $2bn (as did Amazon’s). OpenAI’s projected loss of $85bn in 2028 alone is an order of magnitude larger than any prior money-losing business that subsequently reconciled to profitable unit economics at scale.

Capital-side circularity inflates the numerator (headline round sizes, reported revenue, balance-sheet asset values); demand-side circularity inflates the denominator (reported usage, ARR). Both produce a version of the AI economy that is observed internally rather than externally: the numbers are generated by the system’s own financial engineering rather than by market-clearing signals.

2. Revenue numbers are contested

Both OpenAI and Anthropic disclose annualized run-rate figures from short reporting windows: OpenAI multiplies a four-week period by 13 and Anthropic combines four-week API revenue x 13 with monthly subscription revenue x 12. Critics such as Ed Zitron argue this systematically overstates revenue, that gross margins as presented are revenue-recognition artifacts, and that inference is unprofitable on a fully loaded basis. The specific numbers are contested, but the direction aligns with what traditional credit investors have been signaling (which is why bonds-for-labs hasn’t worked at scale – see Part III) and with leaked AWS spend figures.

On April 6, the WSJ published a review of confidential financials. OpenAI projected $121bn in 2028 compute spend and $85bn in 2028 losses with break-even not expected until after 2030. Anthropic’s projected training costs reach ~$30bn in the same window with management targeting profitability in 2028-2029. Other analysts have different estimates, with Zitron pinning OpenAI’s cumulative four-year funding-and-revenue need at $852bn through 2030 (~$673bn revenue plus ~$200bn additional funding against ~$218bn cumulative burn through 2029) vs WSJ-implied $115bn. The spread reflects differences in how compute-commitment obligations are amortized and how Stargate-class commitments are allocated.

It’s worth noting, however, that large losses during the land-grab phase of a paradigm shift are a recognized pattern, not an anomaly. The labs are pricing for growth rather than margin, betting that technology and product leadership today will translate into durable economic position tomorrow. This has worked before – most successful platform businesses ran losses for years while building share, and the modern internet giants monetized well after they had won their core categories. The concern is not that the strategic logic is wrong, but that the dollar scale is two orders of magnitude larger than prior cases.

3. Growth forecasts are historically unprecedented

Michael Mauboussin and Dan Callahan at Counterpoint Global applied base-rate analysis to OpenAI’s disclosed forecasts: OpenAI projected $145bn of revenue in 2029 vs $3.7bn in 2024. That’s a 108% 5-year CAGR. Against ~18,900 firm-period observations for US public companies with $2-5bn starting revenue from 1950-2024, no company has ever achieved this growth rate. The OpenAI forecast is ~9.5 sigma against the historical distribution. Rolling forward using OpenAI’s $200bn 2030 forecast yields 72.7% CAGR, which also has zero historical precedent even against an expanded reference class of firms starting above $6.5bn (16,400 firm-periods).

The infrastructure side is equally challenging. Economic geographer Bent Flyvbjerg’s 16,000-project database shows bespoke megaprojects (Olympics, nuclear, large IT) come in on budget, schedule, and benefits only 0.5% of the time; modular projects (solar, wind, transmission) do better. Data center buildouts are closer to modular, but the dollar scale and forward commitments push them toward the harder end.

None of this makes the outcomes impossible, as ChatGPT’s adoption curve is historically fast, and OpenAI’s 2025 growth was ~250% YoY. However, the forecasts embedded in the current valuations are without precedent.

4. SBC is structurally large

OpenAI’s 2025 SBC was ~45% of sales (~$1.5mn per employee), about 7x the next-highest pre-IPO tech issuer, with ~$3bn annual increases planned through 2030. This creates three compounding pressures: persistent dilution (~6-8% per year for public shareholders from grants alone), lockup-window selling concentrated at predictable calendar points, and a flat-stock retention spiral where stagnation kills the equity’s retention power even as past grants keep diluting.

5. Competitive moat is narrower than valuations imply

Open-weight models from five independent frontier families (DeepSeek, Qwen, Kimi, GLM, Mistral) now match or beat proprietary models on many public benchmarks, including specific instances such as GLM-5.1 topping the SWE-Bench Pro leaderboard before Opus 4.7’s release. And pricing compression is already happening: Anthropic cut Opus’ price by 67% in November 2025 and OpenAI launched GPT-5.4 at GPT-4o’s input price. But even after these cuts, self-hosted inference still runs 70–500x cheaper per token than proprietary APIs.

This doesn’t collapse the labs’ revenues overnight though. Closed models retain leads on production coding, safety alignment, and human preference, and the managed API has real value. But the Mauboussin base-rate problem isn’t only about TAM: the labs need to monetize $400B+ in compute commitments at ASPs that are compressing while the cost base is being locked in.

I. The Fundamental Fork: Stay Private or Go Public

Stay Private

On the surface, there seems to be no shortage of AI funding – the sector has accounted for roughly two-thirds of US venture dollars over the past 18 months, and private rounds keep clearing at higher marks. But beneath the surface, the channels available to the labs specifically are more constrained than the aggregate numbers suggest, and the constraints tighten as the dollar amounts grow.

Given profitability (or lack thereof), burn rates, and scale, the AI labs need to continue to raise large amounts of capital. However, the traditional capital pools available to private companies are nearing capacity. Pensions and endowments are already heavily allocated to venture, where capital has concentrated into the AI labs. The biggest pool, sovereign wealth funds, have strict single asset concentration limits for non-state-owned assets that they are coming up against (Exhibit 1). They also, however, have major geopolitical considerations.

Apart from traditional financial investors, the Mag7 companies are and have been large sources of capital and compute for the AI labs. However, recent funding rounds have been either signaled or explicitly tied to going public.

Go Public

Recent regulatory changes driven by SpaceX have laid the groundwork for the AI labs to go public by setting them up for early index inclusion that would drive forced buying. In specific, the Nasdaq removed their 10% float requirement and lowered their seasoning period from 3 months to 15 days.

SpaceX filed their S-1 on April 1, 2026, targeting a $75bn fundraise at a $2tn valuation (3.75% float; 30% earmarked for retail). They will start their roadshow on June 8. OpenAI is targeting a $1tn IPO in Q4 2026. Anthropic is racing to IPO ahead of OpenAI to take advantage of the window and avoid a potential fallout (Exhibit 2). Prediction markets currently give them a ~67-73% chance.

Many public debates treat the IPO as a liquidity event. This misses the bigger picture because no single IPO clears the capital needs. The financing path for both labs will require a series of capital raises plus debt layering. Furthermore, the AI labs effectively have a structural constraint on pricing, as there is no meaningful precedent for a tech company to raise less in an IPO than their previous round of private funding, which would floor OpenAI’s raise at $40bn and Anthropic’s at $30bn.

According to University of Florida Professor Jay Ritter, US IPO proceeds have only exceeded $100bn in one (2021) of the past 46 years (Exhibit 3). SpaceX, OpenAI, and Anthropic proceeds alone would total over $180bn (~2x the largest IPO year on record). This is not just a crowded window – it could be a supply shock with no historical precedent.

Given the recent index rule changes, the three companies will be able to benefit from a select pool of passive capital. However, they will not be analyzed in isolation by the same pool. SpaceX prices first in June, then the labs follow in Q4 2026 and H1 2027. Each consecutive listing absorbs a chunk of the same finite Nasdaq-100 tracking AUM. The math has to be done sequentially.

Under the new float rules, SpaceX will get ~11% weighting (implied Nasdaq 100 weight of 0.7-0.9%) that translates to forced passive buying of $5bn, and active Nasdaq-benchmarked funds (e.g., Fidelity, Wellington, Capital) will move to benchmark weight over 2-4 weeks, adding $10-20bn for a total of $15-25bn of buying. Against a $75bn float, that’s 20-30% of public float absorbed by indexers alone. Incremental buying may come from aerospace/defense, space thematics vehicles, infrastructure mandates, and national-security-oriented sovereign allocations.

Applying the same math to either lab (assume $50bn raised at $1tn valuation for both), the labs will be 15% of full weight, or index weight of 0.75-1 (vs the 4% it would get on market cap alone). Forced passive and benchmarked buying comes out to $10-20bn, which, against a $50bn float comes out to 20-40% of float absorbed by indexers. However, unlike SpaceX, the AI labs will not have similar incremental pools to draw on, meaning flows will have to come from AI/tech/growth.

If SpaceX clears in June with $15-25bn of passive and active-benchmarked demand, that is Nasdaq-100 demand already re-weighted before either lab prices. The passive AUM pool doesn’t reset, meaning reallocations to make room for SpaceX come at the expense of every existing Nasdaq-100 constituent, including whatever weight the labs will eventually claim. By the time the labs price, they are competing for residual demand in the pool that has already absorbed one mega-cap addition, against a peer set of Mag7 constituents that have already been trimmed once.

Combined buying across the three names is on the order of $35-65bn over a 6–9-month period, all of which has to be funded by selling other Nasdaq-100 constituents. For reference, when Tesla joined the S&P 500, it saw a 70% appreciation against the S&P 500’s 2.5% gain in the five weeks between announcement and inclusion. Pre-inclusion estimates put forced S&P 500 passive buying at $72bn, active large-cap rebalancing at $8bn, and total index-tracking buying at $220bn. Actual inclusion-day trading volume was $147bn.

These passive flows will likely come from Mag7 holdings simply because of the dollar amounts involved. And the AI labs offer investors something they don’t already have: a pure play on frontier AI. Once the labs are public, passive and discretionary tech funds may reallocate out of existing Mag7 positions to fund new lab weightings, putting mechanical downward pressure on Mag7 multiples. There is circularity here: the very investors selling Mag7 to buy the IPOs are the same counterparties whose AI story is tied to the labs’ success. The same loop runs in reverse: a Mag7 stock decline pressures the very investors most exposed to the labs.

To be clear, the pure play aspect is a double-edged sword. If ROI disappoints, the Mag7 have other businesses to fall back on while the AI labs have nothing (Exhibit 4). The same feature that justifies a valuation premium in boom conditions becomes the feature that justifies a discount in stress. At loss-making scale, an undiversified business gets punished twice: once in the revenue line, once in the multiple compression. The narrower buyer set compounds this, as the labs are pure play in their business and also pure play in their natural holder base, meaning thinner incremental demand and more sensitivity to sentiment swings.

Of course, delays are possible (and likely), but bounded: 1) recent funding rounds include contingent commitments tied to IPO milestones which may need to be renegotiated, 2) burn rate trajectories require continuous capital raises where private market runway extension is expensive and capped by saturation, 3) compounding SBC dilution effects, and 4) competitive sequencing where one lab may list and disappoint, closing the window for the other.

II. Post-IPO Trajectories: Acceleration, Backstop, Acquisition, Nationalization

Acceleration

Acceleration is the trajectory where the labs outperform after IPO: Jevons demand-side tailwinds meet the forced-buying mechanics of Part I, fundamental revenue growth beats the S-1 baselines, and the stocks compound rather than defend a floor. Mechanically, it would look similar to this:

• Passive flows function as an acceleration mechanism vs downside protection

• Labs raise further capital at higher multiples post-IPO rather than manage sequenced offerings to preserve dilution

• Anthropic reaches profitability threshold for S&P 500 inclusion in 2029/2030 and OpenAI follows 1-2 years later. Inclusion would trigger forced mechanical buying from $13tn+ of passive S&P 500 tracking AUM.

• Acquisition math inverts and the labs become acquirers using their stock to purchase application-layer companies, vertical AI specialists, and adjacent software assets.

• Revenue growth compounds and unit economics improve, making rating agencies revisit lab credit at the edges and opening up additional access to the credit market.

• Pure play options in AI accelerate asset flow from Mag7 to the AI labs. The Mag7’s AI-narrative multiple compresses modestly in relative terms even as absolute level holds.

Circularity is usually flagged as a downside amplifier, but the same loops can run in reverse. The mechanisms below would, in an acceleration scenario, produce increasing returns rather than amplified drawdowns.

• OpenAI stock appreciates -> Microsoft equity stake increase on their balance sheet -> market increases Microsoft’s multiple on AI narrative -> Microsoft commits further capex -> capex supports OpenAI’s compute capacity -> compute capacity supports revenue growth

• Nvidia GPU revenue re-anchors as a core business line rather than an AI-bubble line -> Nvidia’s multiple sustains -> Nvidia continues strategic ecosystem investment -> Nvidia supports the demand side of the compute market

• Amazon’s investment in Anthropic does the same for AWS – the $100bn AWS commitment from Anthropic becomes a validated, growing revenue line -> sustains Amazon’s Cloud-AI multiple -> continued Amazon Trainium investment -> improved Anthropic unit economics

• Google’s latest Anthropic commitment (up to $40bn in new investment + 5 GW capacity) does analogous work for Google Cloud. Anthropic’s incremental compute consumption validates GCP’s AI stack as a viable third pillar alongside Azure and AWS -> upside reflected in Google’s multiple.

Strategic Backstop

If public market demand for AI labs wanes, the Mag7 could backstop them via a combination of equity investment, anchored GPU allocation, aftermarket buying, and narrative support.

The recent Amazon and Google Anthropic deals showcase the hyperscaler backstop mechanism operating in real time, pre-IPO, without distress triggering it. The backstop channel is already both established and active, so the question is not whether hyperscalers would catch a falling lab, but whether the scale of commitments already on the balance sheet leaves sufficient incremental capacity to do so at multi-hundred-billion-dollar drawdowns.

The realistic backstop goal would be “flat,” not “up.” CoreWeave is the illustrative example. Their multiple is being sustained by Nvidia’s support – on its own fundamentals, it would trade materially lower. Applied to the labs: the question is not whether Mag7 backstopping can drive the stocks higher (it cannot plausibly do so at multi-hundred-billion scale), but whether it can prevent a deep drawdown during stress through a combination of equity injection, aftermarket buying during indigestion windows, narrative reinforcement, and customer-commitment acceleration.

The incentives to support the AI labs are strong. Mag7 AI-narrative exposure, revenue commitments, and equity stakes all point the same direction. A sharp repricing of the labs feeds back into the Mag7’s own multiples. However, there are capacity concerns. Mag7 cash and short-term investments run into the hundreds of billions, but 2026 capex commitments are unprecedented.

There may also be a ceiling on backstop effectiveness. CoreWeave works because it’s a small fraction of Nvidia’s revenue and narrative. The labs are a much larger fraction of Microsoft’s and Amazon’s AI stories, and heavy aftermarket buying by a hyperscaler in its own partner-lab, disclosed publicly as it would have to be, could signal distress rather than conviction.

Acquisition

If the stock trades materially below strategic investors’ cost basis (not unlikely given loss scale and aftermarket dynamics), acquirer math changes (assuming no antitrust issues).

A material aftermarket drawdown is the prerequisite for distressed-recap or acquisition-at-a-discount scenarios. But Nasdaq-100 forced-buying (demand floor of $10–25bn per lab over the first month) may, in principle, be enough to prevent a clean crater even if fundamentals disappoint.

So the question is not whether the stocks can fall on fundamentals, but whether fundamentals can push them far enough below fair value to trigger acquisition interest given that indexers are structurally bidding at the same time. The likely outcome is a grinding sideways-to-down pattern – not a crash that opens an obvious entry point, but sustained multiple compression that slowly changes acquirer math without providing a single catalyst moment.

The most likely acquirers are the companies most exposed to the AI narrative through revenue commitments and equity stakes. Acquiring a lab at a discount crystallizes a valuation that goes against the narrative the acquirer has been supporting, translating into multiple compression on the acquirer itself. The most probable “acquisition” path is therefore a distressed recap (rescue financing in exchange for incremental equity/governance concessions) rather than a clean buyout.

But given anti-trust concerns and scale, “acquisitions” may be untraditional. Two potential models:

• The SpaceX-Cursor template: SpaceX agreed to pay Cursor $10bn or acquire them for $60bn. xAI has Colossus-scale compute but weak product-market fit; Cursor has the product and the users but no competitive model. This was a compute-for-product swap with a breakup-fee floor.

• The Microsoft-Inflection template: Microsoft paid Inflection $650mn ($620mn license + $30mn legal waivers) and hired its founders and most of the 70-person team for Microsoft AI. The company remained legally intact to avoid antitrust review.

Nationalization

There are three distinct sub-scenarios for nationalization: 1) Seizure, 2) Equity stake, and 3) Bailout. Seizure would set a destabilizing precedent but has a higher conditional probability under a Democratic administration and for a company like Anthropic likely to resist a forced sale. The equity stake has a precedent in Intel which is politically easier if packed with CHIPS-adjacent support or a compute-security arrangement. Bailout is unlikely now but likelihood rises in a stress scenario.

A federal bailout in today’s environment would be politically far harder now than a comparable bailout would have been in 2008 or 2020 because public sentiment toward AI is negative and worsening, and that has already translated into actions and results in the political sphere.

Recent NBC polling shows roughly 26% of U.S. voters with a positive view of AI versus 46% negative, and the negative feelings have already manifested. In just one week (April 6-10), six city council members across two cities in Missouri were ousted, voters in Wisconsin approved the first-in-the-nation ballot initiative opposing a planned data center campus, a pro-data center city councilor in Indianapolis had his house struck by 13 gunshots and a note reading “No Data Centers” on his doorstep, and a Molotov cocktail was thrown at Sam Altman’s San Francisco house.

Any bailout is more likely to arrive packaged as something else – compute-security funding, strategic-reserve compute procurement, CHIPS-adjacent loan guarantees, Defense Production Act invocation – than as an explicit rescue or unadorned credit backstop. And if a bailout does happen, OpenAI is the more likely candidate – it has already asked once. Anthropic is the more likely seizure candidate given its mission-driven posture around safety, willingness to litigate their federal supply chain risk designation, and Dario’s reported views on foreign-government access to AGI.

Reportedly, when Dario was still at OpenAI, Greg Brockman suggested selling AGI to governments. When Dario asked which ones, Greg responded it would be to the nuclear powers that made up the UN Security Council to not destabilize world order. Dario considered selling AGI to rival powers such as Russia and China tantamount to treason, and considered quitting (which he later did). Furthermore, Dario has been reported to have internally called OpenAI “mendacious,” compared the Altman-Musk legal battle to “Hitler vs. Stalin,” calling a Brockman pro-Trump PAC donation “evil,” and likening OpenAI and peers to tobacco companies.

And retention data supports the cultural claim. SignalFire’s 2025 State of Talent Report: Anthropic two-year retention 80%, DeepMind 78%, OpenAI 67%, Meta 64%. Engineer flow is roughly 8x from OpenAI->Anthropic and 11x from DeepMind->Anthropic versus the reverse. Amodei has stated publicly that Anthropic employees have turned down Meta offers at $100M+ packages and that the company is not matching those packages. Whether or not retention is a predictor of outcomes, it is an observable fact that the cultural story is different – and the strategic implication (harder to pressure into outcomes management opposes) should be weighted accordingly.

III. Parallel Capital Channels: Structured Compute Financing

Structured compute financing is the most important but least-discussed channel in the equity-narrative framing. It is not only a future event, but already the primary way AI capex is financed and the channel through which the equity side supply constraints are being partially worked around.

Many investors seem to believe that credit will “save” AI because the global bond market is enormous and far larger than equity markets. But the framing is a category error. While bond markets are larger, they are heavily segmented, with only $8tn of the $145tn global debt market made up by US IG corporate credit. Furthermore, credit markets are structured around downside protection, which a cash-burning lab doesn’t offer (Exhibit 5). However, the credit channel was opened by a reframe – not in changed credit discipline, but in what the collateral is: compute hardware, power contracts, and hyperscaler offtake instead of lab-level cash flow.

And the scale already seems dizzying. Combined AI-related capex from US hyperscalers is running above $500bn/year, which is larger than the entire US O&G industry’s annual capex ($150-300bn). Amazon alone guided to $200bn for 2026. An April 2026 Morgan Stanley analysis estimates the Mag7 to spend $3tn on AI through 2028 – and generate enough internal cash to cover only about half of that figure, with the balance coming from a variety of credit instruments. The migration is already visible in primary markets: hyperscalers issued ~$121bn in bonds in 2025 alone (more than 4x the 2020–2024 average of ~$28bn).

The capital needs exceed what can plausibly be financed through traditional corporate balance sheets over a multi-year horizon, which is why the financing has migrated to structured channels. But if the hyperscalers are running into walls, the labs will run into even bigger ones. The burn trajectory shows total cash out; the cloud-spend data shows where the cash goes, and reveals that unit economics are worse than aggregate burn alone suggests.

For Anthropic, leaked AWS documents show the company spent $2.66bn on AWS through the first nine months of 2025 on estimated revenue of $2.55bn – 104% of revenue on cloud compute from a single provider, before payroll, R&D, or any other cost. OpenAI is in a similar situation. Leaked documents show inference spend on Azure more than doubled in 9 months, and Microsoft’s quarterly reporting revealed OpenAI lost $12bn in a single quarter. An independent analysis of GPT-5’s August-December 2025 deployment showed ~30% gross margins and -5% operating margins (excluding R&D), well below the 60-80% gross margin typical of mature software – and before accounting for OpenAI’s revenue-share with Microsoft, which would push operating margins further negative. (One nuance: OpenAI’s training spend is largely non-cash (paid via Microsoft’s initial $13bn in credits), but inference is cash. As credits deplete, the cash burn accelerates.)

If the labs are spending 100%+ of revenue on cloud alone, the gross margin that anchors traditional credit is near-zero/negative. This is the deeper reason structured credit, not traditional credit, is the only channel clearing this market (Exhibit 6). SPV-with-offtake lets the hyperscaler balance sheet absorb the unit economics problem rather than requiring lab-level resolution. So existing credit channels are taking on hyperscaler concentration risk, not diversified lab-sector risk.

And the concentration is historically big: per a February Morgan Stanley analysis, the capex cycle is larger in both absolute terms and as a percentage of sales than any prior corporate-led cycle (Exhibit 7). A February Moody’s analysis estimated $662bn in off-balance-sheet future leases across Amazon, Meta, Google, Microsoft, and Oracle, which is ~113% of their adjusted debt. They land sequentially in 3-5 years, meaning the effective capex ratio is even higher than reported.

So credit is already funding the current data center boom, with multiple $10bn+ deals across the US. The structures vary, spanning GPU-backed facilities, real estate and offtake project finance, and hyperscaler strategic commitments. But the common thread is that credit is being underwritten against hyperscaler offtake rather than lab credit. Typical terms: 50–60% LTV, 10–14% coupons, 3–5yr tenors. Insurance general accounts via rated-note feeders (Exhibit 8).

McKinsey estimates $6.7tn in global data center capital through 2030 – several multiples of cumulative global project finance capacity over the same window. Existing channels can likely absorb the first $1-1.5tn over 2-3 years before spreads widen, but the next $5tn+ requires structural expansion of the credit pool or demand destruction.

The current credit risk is being underwritten against hyperscaler offtake counterparty, not AI labs. The lab is the demand anchor, not obligor. The Crusoe-Oracle-OpenAI chain is illustrative: lenders hold paper serviced by Oracle’s IG lease payments; if OpenAI misses, Oracle absorbs/negotiates.

Analysts cite Amazon’s 1999 $1.25bn convertible note as a template, but the current momentum is far more extreme. Labs have neither profits nor cash flow, so credits route through hyperscaler credit, not lab credit (Exhibit 9). This scales because once IG offtake is established, the debt tranche is rated, which opens insurance general accounts, pension private debt, and public credit funds, and a private credit niche transforms into a rated asset class (see CMBS in the 1990s/2000s).

However, there are several binding constraints that have all tightened in the past twelve months:

• Amazon, Google, and neoclouds report full A100 utilization (released 2020), implying GPU depreciation may be overstated, and in fact may even be appreciating assets. However, the asset thesis depends on supply-side pricing power, which has 18-month generational cycles, and CoreWeave already breached covenants on its Blackstone facility.

• With NAIC regulations tightening, the credit channel is narrowing precisely as supply peaks.

• Hyperscaler counterparty concentration is manifesting two risks: 1) Rating action: one-notch downgrade reprices entire stack simultaneously. This is meaningful given the scale (Amazon’s $200bn 2026 capex and Oracle’s $300bn OpenAI commitment), and 2) Capex guidance cuts citing ROI that challenge demand-pull.

• Bank concentration limits binding on Oracle-tenant deals, the market has already rejected at least one Oracle-tenant deal, and while Oracle credit is still BBB, it is trading in secondary markets at junk-equivalent. If Oracle rating comes under formal pressure, every existing Oracle-tenant deal reprices simultaneously.

Take Stargate as an example. The project has ~7 GW in progress with reported total costs of $400bn+. The equity layer is thin, with only $52bn committed vs the $500bn headline. The remaining $448bn is to be raised through debt and only has a portion committed, and Softbank’s CFO said it’s “taking longer than anticipated.” Expected annual revenue upon completion is $75bn (~5x cost-to-revenue, which is consistent with IG infra economics if offtake is durable). If OpenAI can’t pay, the asset base is stranded. Oracle reported -$24.7bn LTM free cash flow as of Q3 2026 with cloud infra as the only growth engine. Stargate softness means Oracle loses both growth story and multiple justification. Generational obsolescence further complicates depreciation schedules.

But beyond corporate concentration, two of the three largest non-Mag7 funders have personal/balance sheet leverage creating forced selling pathways:

• Larry Ellison has pledged 30% of his Oracle shares as collateral for personal indebtedness and is personally guaranteeing his family trust’s equity commitment to Paramount-WBD. If Oracle stock declines materially, Ellison faces three compounding pressures simultaneously: margin calls on existing pledged shares, further margin borrowing to cover Paramount-WBD commitment, and reduced Oracle dividend income to service his personal debt stack. Forced selling of shares to meet margin calls would accelerate any decline.

• SoftBank’s exposure to OpenAI is now routed through debt as much as equity. In addition to its $30bn OpenAI commitment, SoftBank secured a $40bn bridge loan for OpenAI-related investments in March and is now seeking an additional $10bn margin loan specifically collateralized by its OpenAI shares. If OpenAI’s post-IPO valuation declines, SoftBank’s margin loan collateral deteriorates, potentially forcing OpenAI-share sales or early tranche cancellations that would directly affect OpenAI’s capital stack.

The hyperscaler concentration argument above stops at the corporate level, but the Ellison and SoftBank situations show two primary OpenAI-Stargate funders have named, mechanical forced-selling paths if AI-narrative equity falls. This is concentration risk with a personal-finance amplifier; it’s a bearish mirror of “circularity as increasing returns” (Exhibit 10).

The counterweight to the friction above is that the total potential capital is enormous. However, most of it is currently gated by credit ratings, regulatory treatment, and underwriting conventions.

There is ~$10tn in adjacent pools structurally closest to AI paper. If structures clear, the pools most likely to expand AI exposure are: US insurance industry assets of $9tn (with ~$5tn in fixed income, 95% IG, <5% AI exposure today), global private credit AUM of $1.7tn with $500bn+ in dry powder, US pension private debt allocations are rising into the low trillions, and sovereign infrastructure allocations add another $1-2tn distinct from single-name equity concentration limits.

There is also a $145tn global bond market to tap into. The US makes up $50tn of that (excluding MBS and ABS), with $8tn+ in US IG corp. US structured credit holds another ~$14tn. This is the theoretical upper bound of capital that could eventually fund AI if underwriting conventions, rating methodologies, and regulatory treatment all evolved. AI-related structured paper outstanding today is in the low hundreds of billions (<0.5% of this universe).

However, these two pools will most likely stay sidelined for the foreseeable future. The binding constraint is not capital but gating infrastructure (ratings, statutory treatment, underwriting conventions) – and that infrastructure is tightening, not easing. Regulatory direction is actively moving against incremental AI structured paper, underwriting conventions take years to change, and the most plausible unlock paths all require a broader AI financing crisis as the precondition for policy action. Capital is available in principle, but not in practice on the relevant horizon.

IV. Meta Questions: Open-Source and Jevons Paradox

Open Source

The capability gap between frontier models and open source has nearly closed. Five independent open model families (Deepseek, Qwen, Kimi, GLM, Mistral) have reached frontier quality near-simultaneously, making the trend structural rather than a one-off (Exhibit 11). GLM-5.1 even led both Claude Opus 4.6 and GPT-5.4 on SWE-Bench Pro. The release cadence is fast (~3-6 months between meaningful upgrades), and on some benchmarks, open is already ahead.

While the capability gap is closing, the cost gap is not. Self-hosted inference on open models runs 70-500x cheaper per token than equivalent API calls to frontier proprietary models. At sufficient scale, self-hosting becomes economically compelling – and the threshold falls every quarter as open-model quality improves. This has several important implications for the revenue story: 1) it caps pricing power, 2) it compresses gross margins over time, and 3) it fragments the market structurally. On-device inference amplifies the threat.

The case for sustained premium pricing rests on three pillars, each of which is narrower than it sounds: 1) frontier capability lead on hardest tasks, 2) safety alignment and human preference, and 3) managed API convenience.

This doesn’t mean that open source collapses the AI lab revenues, but it does cap growth rates, compress margins, and fragments demand. The Mauboussin base-rate problem described earlier is amplified: achieving the 72% 5-year CAGR baked into current valuations requires expanding share of a total inference market that is simultaneously becoming much larger (Jevons – see below) and commoditizing at the low end. Either can dominate but neither can be assumed away.

Jevons Paradox

The sections above assume finite demand against which $500bn–$1tn+ valuations must justify themselves. The bull case is that this framing is wrong – that AI demand behaves like internet demand did through the 1990s-2020s, expanding the total pie faster than the labs’ share of it contracts. This section treats that possibility with the seriousness the asymmetric stakes warrant.

A specific piece of evidence underlying the bull case: the economic value of frontier tokens appears to be growing faster than the industry’s ability to serve them. Every tier-1 lab is capacity-constrained on inference during peak hours. Tier-2 and tier-3 labs report being sold out of tokens. AWS, Google, and neocloud operators report continued full utilization of A100s released in 2020.

Both OpenAI and Anthropic are trying to secure as much compute as possible (Anthropic’s Amazon and Google deals, OpenAI’s capex ramp). And if compute is genuinely the binding constraint, if willingness-to-pay at frontier quality exceeds the industry’s ability to serve at any price, the Jevons dynamic is already present and visible.

And if Jevons holds, the labs are not merely fairly valued at current marks – they are materially undervalued. The bearish framing is conservative only if the reference class holds. But if the reference class is shifting and AI is a paradigm shift rather than a cyclical overshoot, the conservative read flips: not being long the labs at current marks is the risky position because the upside is structurally much larger than the downside.

While Jevons’ observation was with regards to the analog world, that more efficient steam engines led to more total coal consumption rather than less because efficiency expanded the set of economically viable uses, it has held for many other technologies, including LED lighting, fuel-efficient cars, cheap computing, and now, arguably, inference. This view has been championed by everyone from Jensen Huang and Satya Nadella to Sam Altman and Travis Kalanick.

Jevons paradox can be applied to AI in the following ways:

• Agent-to-agent. The internet is bounded by the number of humans alive. Agent-to-agent has no equivalent ceiling as agents can easily be spun up by any person or agent. Total inference would scale with economically valuable subtasks (exponentially and much larger).

• Numbers today are already enormous. OpenAI last disclosed in Oct 2025 that they were processing ~8.6tn tokens/day. The full commercial LLM API market is estimated to be doing ~50tn tokens/day. OpenRouter alone is doing 1tn+/day. Token volume is growing several-fold annually across every disclosed provider (Azure 5x YoY, OpenRouter ~10x, OpenAI ~20x over two years), against a human labor force growing <1%. Most of this is not priced into anything.

• Compute cost has fallen sharply (inference per token is down 9-900x per year) and AWS, Google, and neoclouds still running A100s at full capacity indicates demand is still meaningfully ahead of supply even on older generations as newer ones come online.

• Public-markets analog. A decade ago, a $1tn company was unimaginable. Today, there are around ten. Internet-as-infrastructure expanded the equity market to hold it. AI at similar/greater scale would make valuation questions mechanical rather than stretched.

Even on conservative assumptions, such as 10-20% of global knowledge work being meaningfully augmented by agents over the next five years, the addressable market for agent inferences runs into the trillions per year, comparable to the entire global cloud market today. Against that backdrop, $1tn valuations for the leading labs look closer to a floor than a ceiling.

A weaker but more rigorous version of the bull case is that you don’t need to believe in superintelligence or transformative AI for the valuations to make sense.

You need to believe that AI at current capability will meaningfully penetrate existing white-collar services markets that people already spend trillions of dollars on. That surface area includes: 1) global legal services ($1tn/year), 2) global management consulting ($1.1tn), 3) global accounting ($600bn/year), 4) global software development ($675bn/year), 5) global customer experience BPO ($100bn/year), and 6) US hospital administrative costs ($687bn/year – roughly 2x what hospitals spend on direct patient care).

All together, this is $3-4tn of immediately addressable TAM that existing AI capabilities can penetrate. Even at 15–20% penetration, that is $450–800bn of annual revenue addressable by AI providers – well above current frontier-lab revenue and easily supportive of $1tn valuations.

This case decouples the bull argument from AGI debates entirely. You don’t have to believe AI will transform civilization. You have to believe it will do to white-collar services what AWS did to enterprise IT: capture a material share of the spend by being structurally cheaper and more scalable than the incumbent approach. That is a much weaker claim than AGI and, on current evidence, closer to observable than speculative.

Skeptics may argue that these scenarios are highly unlikely, but there is a deeper, human reason that Jevons and current-capability cases are underpriced by most investors, including sophisticated ones: humans are structurally bad at exponentials. The cognitive apparatus people use to reason about growth was selected for an environment where almost every quantity of interest varied linearly or was stable. Exponential growth was not a common feature of the ancestral environment, and no cognitive module was selected for estimating it accurately.

When modern humans are asked to estimate exponential quantities, they consistently produce answers biased toward the linear range they can visualize, anchoring to the starting value and extrapolating linearly with some adjustment for growth rate – but not nearly enough.

Think about a stadium that doubles every day until it fills completely at day 30. It would be half-full at day 29, one-eighth full at day 27, and 1/1024th full at day 20. An observer on day 20 looking at a mostly-empty stadium would reasonably conclude not much is happening here – while being just ten days from complete saturation. That observer is applying linear intuitions to an exponential quantity, and the linear intuitions are producing a confident wrong answer at the exact moment the underlying curve is about to deliver most of its effect.

Apply this to AI. Token volume growing 5–10x year-over-year is not “accelerating growth from a large base” – it is exponential-regime behavior that linear-mind readers will systematically under-forecast. OpenAI’s revenue rose from a ~$2bn run rate in late 2023 to ~$25bn in early 2026. Anthropic’s ARR tripled from ~$9bn to ~$30bn in Q1 2026 alone. These trajectories are not consistent with any linear or modestly-exponential model finance has traditionally applied to software companies. When a curve grows at rates exceeding standard model assumptions at every measurement point, the correct conclusion is that the model is wrong, not that the curve is about to revert. Every revision to AI demand forecasts over the past three years has been upward; none should have been surprising. A reader who evaluates the labs using linear intuitions about “how much bigger can this really get” is making the same cognitive error as the day-20 observer.

If the exponential growth continues for even another 3-5 years, investors who dismiss the Jevons case on base-rate grounds will be wrong by orders of magnitude, not percentages. If exponentials break (capacity constraints, S-curve topping, substitution), investors long the labs lose a bounded amount. The asymmetry of being wrong in one direction vs the other, and the systematic human tendency to under-appreciate exponentials, means the prior probability assigned to the Jevons case should be higher than it typically is in institutional framings.

So if you believe AI is real, that AI is truly transformational, your prior on the Jevons case should be higher than your base-rate training suggests.

---

If you’ve made it this far, thank you for reading. If the subject of this memo is something you’ve been thinking about, I’d love to hear from you (email; twitter).

• And if you you enjoyed this memo on AI’s potential impact on markets, you may enjoy this memo on AI’s potential impact on society

Below the paywall: exhibits, sources, and titles of other private memos I may share. Paid supporters can reply or email me with any of the titles they’d like to read and I’ll prioritize accordingly, whether that means sending a draft, deciding to publish it, or discussing it over email/chat.