A technology can keep improving while the companies financing it get into trouble. The economy needs useful work at an acceptable cost. A borrower needs cash on the dates its contracts specify. Those are different requirements.
In The Slope Is the Cycle, the case against calling an early cycle top rested on three moving parts: capability, compute supply, and demand breadth. Rapid capability gains kept opening possibilities before anyone could confidently place a ceiling on demand. The risk was that committed supply would arrive before those possibilities became paying customers.
Alignment introduces another way for that loop to weaken. The best model a lab can build may advance faster than the best model it can release. While the internal frontier moves ahead, competitors may approach the capability customers can actually buy.
That would pace the commercial advantage of the frontier. It could improve the economics of adopting AI while squeezing some of the businesses supplying it.
This is a scenario analysis. Financial disclosures are current through September 13, 2026, with each reporting period identified below. The capability schematic and compute explorer are hypothetical; neither specifies how long a release bottleneck would last.
The frontier a customer can buy
In We Must Pace the Frontier, Dario Amodei advocates slowing capability advancement and strengthening safeguards. Anthropic commits to embedded external evaluators; broader pacing depends on coordination. His proposal also calls for tighter restrictions on unauthorized distillation and chip access.
The narrower possibility explored here is a persistent gap between internal progress and broad release. That is one potential outcome, not a description of an announced industry pause. Its duration is uncertain.
01 / Hypothetical mechanism
Two gaps can move in opposite directions
The private frontier advances; the capability customers can buy converges.
Sources: Dario Amodei — We Must Pace the Frontier. Retrieved 2026-09-13.
Schematic by the author. Line positions and slopes are illustrative, with no measured capability units or calendar horizon. Convergence is a scenario, not a finding from current benchmark data. Amodei also proposes slowing capability advancement itself.
Three slopes matter. Internal capability determines what researchers can do. Available capability determines what customers can access. Economic usefulness determines what customers can accomplish after paying for integration, supervision, and mistakes.
The third slope can rise while the second flattens. A stable model can become more useful through better tools, workflow design, and reliability. Alignment work can itself unlock deployment by reducing supervision. A release bottleneck therefore need not produce an adoption bottleneck.
It can also create two opposite gaps: a widening lead inside the best labs, and a narrowing difference among the models on the market. The first preserves research value. The second puts pressure on pricing.
Catching up does not require winning every benchmark
An enterprise rarely needs the world’s best model for every task. It needs enough quality for a particular job, at a total cost that makes the job worth automating. The point at which a cheaper model crosses that threshold matters more than its position on a general leaderboard.
Distillation supplies one route to that threshold. A useful historical example is DeepSeek’s original R1 release: its distilled 32B model exceeded o1-mini on the three selected evaluations below. These are vendor-reported results from the original 2025 model family, not evidence of where the September 2026 frontier stands. DeepSeek’s evaluation table
02 / Reported historical evaluation
Distillation can make a smaller model competitive
Original January 2025 releases · selected benchmarks · pass@1, higher is better
AIME 2024
MATH-500
GPQA Diamond
Sources: DeepSeek-R1 — original model and distilled-model evaluations. Retrieved 2026-09-13.
Selected benchmarks from the same vendor table. Historical evidence of transfer, not an independent evaluation or a September 2026 frontier comparison. No benchmark average is calculated.
The economic hypothesis does not require open models to equal the strongest closed model on every task. It requires more workloads to become substitutable. A business might route classification, extraction, and first drafts to a cheaper model while retaining a premium model for difficult reasoning or consequential actions.
Here, “open” refers to models whose weights are available for deployment, subject to their licenses. They are not automatically cheaper to operate or equally suitable for every use. Hosting, security, evaluation, and human review remain real costs.
Nor does the hypothesis require China to stop innovating. Compute constraints can limit progress without fixing it at zero. Chinese-developed weights can be deployed on infrastructure elsewhere; research capacity and worldwide serving capacity are distinct constraints. Conversely, tighter access controls can impede distillation, and unavailable teacher models are harder to copy. Convergence remains a condition to test.
The distribution route already exists. In February, AWS added six open-weight models to its managed Bedrock offering. Customers can adopt open models without managing their own servers or leaving their cloud provider. AWS announcement
Cheaper intelligence can use more hardware
Losing a model premium does not mean losing the need for computation. Broader adoption adds inference. More accessible weights invite fine-tuning, evaluation, and experimentation. Deployment inside individual organizations can also require spare capacity, duplicated model copies, and smaller batches of requests.
Those choices can increase the equipment required for the same useful workload. All else equal, a system operating at 20% utilization needs three times the installed capacity of one operating at 60%. That is arithmetic, not an estimate of enterprise utilization. Batching and serving configuration affect how much work a device can deliver. NVIDIA’s inference optimization discussion
Smaller models, quantization, and more efficient serving pull in the other direction. The aggregate result depends on how quickly useful workloads expand relative to those efficiency gains.
03 / Interactive sensitivity — not a forecast
Flat frontier training can coexist with growing compute demand
Change the assumptions. The starting capacity index is 100.
1× means no growth
More tasks or more work per task
More useful work per active device-hour
0.6× means 40% lower utilization
Includes alignment and other post-training
Required capacity index
Required capacity 121.7, up 21.7 percent from baseline.
Starting mix · 100
Your scenario · 121.7
- Frontier training
- 40.0
- Inference
- 66.7
- Other R&D
- 15.0
Starting shares are assumed: 40 training / 50 inference / 10 other R&D. Bars use the same scale. This arithmetic cannot price chips or predict infrastructure profits.
Calculation and boundaries
Capacity = 40 × training multiplier + 50 × inference multiplier ÷ (throughput gain × relative utilization) + 10 × other R&D multiplier. Only inference is adjusted for throughput and utilization. All multipliers are relative to the starting point. The model assumes capacity can be expressed in comparable units; actual hardware, locations, networking and power constraints are not interchangeable.
Sources: NVIDIA — Mastering LLM Techniques: Inference Optimization. Retrieved 2026-09-13.
NVIDIA supports the serving-efficiency mechanism; every numerical assumption above is the author’s.
All starting shares and multipliers are illustrative author assumptions, not estimates of current industry demand. Capacity index, not dollars, energy, chip counts or a forecast. Training and other R&D categories are disjoint; inference efficiency and utilization adjustments affect inference only. No time horizon is assigned.
In the default illustration, frontier training stays flat, inference work doubles, and other research grows by half. A 50% inference-throughput improvement limits the increase in required capacity to roughly 22%. Improve throughput further and capacity falls below the starting point. Reduce utilization and capacity rises much more. None of those starting shares is a measured industry mix.
This is why “compute costs” needs a denominator. The price of a model API, the cost of completing a task, the rental rate for a GPU, and total hardware spending can move in different directions. Adopters can pay less per useful task while the industry buys more equipment.
There is another distinction: slowing releases does not necessarily flatten training. Labs might keep pursuing internal capabilities, and alignment and post-training consume compute too. Flat frontier training is an additional assumption available in the explorer, not a conclusion implied by the title of this post.
Debt keeps its own schedule
The weakest financing structure is the one whose fixed payments outrun the cash it can reliably collect. Large debt alone is an incomplete signal. Interest rates, maturities, contractual protection, available cash, and the need to replace equipment determine how much room a borrower has.
That is where a capability bottleneck could become a credit event. A customer unable to monetize its model may struggle to pay its infrastructure provider. A provider that replaces that customer at a lower rate can keep its GPUs busy while earning less than its financing requires.
The technology does not have to fail. The retained margin can simply be too small.
04 / Reported balance sheets
Cash cushions differ more than the AI label suggests
Unrestricted cash and equivalents as a share of reported debt · each full track represents that company’s debt
Definitions and excluded restricted cash
- CoreWeave: Current and non-current recourse and non-recourse debt. Leases excluded. Unrestricted cash and equivalents only. Restricted cash excluded: $1.38B.
- Applied Digital: Reported current debt plus long-term debt net of financing costs. Predates June 16 refinancing. Leases excluded. Restricted cash excluded: $2.56B.
- Oracle: Current and non-current notes payable and other borrowings. Excludes leases and $708M marketable securities. Restricted cash not quantified in this chart.
- IREN: Current and non-current reported debt. Finance and operating leases excluded; subsequent financing announcements are not added to this snapshot. Restricted cash excluded: $1.72B.
Sources: CoreWeave — Q2 2026 results · Applied Digital — FY2026 Form 10-K · Oracle — Q1 FY2027 results · IREN — FY2026 results and business update. Retrieved 2026-09-13.
A dated balance-sheet comparison, not a solvency ranking. Cash may already be committed to construction. Debt uses reported carrying amounts and excludes leases; restricted cash and marketable securities are excluded from cash. Reporting dates differ. Applied Digital’s snapshot precedes its June refinancing.
CoreWeave is a priority research case. Its June balance sheet reported about $35.1 billion of debt, including $7.5 billion classified as current, against $5.5 billion of unrestricted cash. Second-quarter net interest expense was $640 million, roughly one-quarter of revenue. Its $104 billion backlog supports the other side of the argument: contractual coverage matters. The open questions are delivery, customer credit, cash collection, and future bookings—not whether demand is imaginary. CoreWeave Q2 results
Oracle illustrates why the largest debt balance is not necessarily the most fragile. At August 31 it reported about $125.3 billion in borrowings, but quarterly operating income covered interest approximately 4.7 times. Its existing business provides support even as expansion strains cash flow. Oracle Q1 FY2027 results
Restricted cash needs particular care. Construction reserves or cash pledged to a financing vehicle cannot be treated as freely available corporate runway. Equally, newly raised cash is not surplus if it is already needed to finish a contracted facility. The chart is a starting point for examining those obligations.
Funding can stop before capability does
Andy Constan adds a nearer-term question to the debate over eventual AI returns: can capital markets finance the buildout before those returns arrive? An interruption in issuance could delay projects even if their long-term prospects remain attractive. Constan’s funding discussion
That qualifies the earlier cycle argument. Rapid capability growth keeps opening applications, but it cannot rule out a financial cycle top. Announced spending, contracted spending, and spending with all necessary financing secured are different stages.
05 / Reported cash flows
The buildout needs cash beyond operations
Cash capex versus operating cash flow · latest reported three-month periods
CoreWeave
Q2 2026 · three months ended June 30
Gap $5.74B
Net financing cash inflow that quarter: $10.07B
Oracle
Q1 FY2027 · three months ended August 31
Gap $5.40B
Net financing cash inflow that quarter: $13.11B
What this gap measures
Both companies raised financing in these quarters. A forward funding requirement also depends on existing usable cash, future operating receipts, committed financing that can actually be drawn, debt payments, and remaining project costs. The orange bars do not measure that future requirement.
- CoreWeave: Cash equipment purchases include capitalized internal-use software. Operating cash flow includes working-capital movements; net financing is after debt repayments and other financing outflows.
- Oracle: Operating cash flow includes $11,363M of customer prepayments with a significant financing component. These are real cash receipts that bring forward funding for future service. Do not subtract them from capex while also counting them in operating cash flow.
Sources: CoreWeave — Q2 2026 results · Oracle — Q1 FY2027 results. Retrieved 2026-09-13.
Historical gap = cash capital expenditures minus GAAP operating cash flow. These expenditures were made and financing was raised; this is not an estimate of unfunded commitments. Financing cash flow is shown separately, not subtracted from the gap. Total company spending, not AI-only spending. Excludes non-cash asset additions and is not annualized.
The chart measures a historical cash gap. Both companies raised substantial financing during the displayed quarters; the orange bars are not unpaid bills. Customer advances matter too: they provide cash today against services still to be delivered.
The forward screen is the cash needed to finish projects and meet payments, less usable cash, expected operating receipts, and financing that can actually be drawn. Equity funding reduces the burden of fixed payments, but access to new equity can still become expensive or disappear. Installed models and hardware can keep doing useful work while new construction and equipment orders slow.
The same AI boom borrows at different prices
Financing terms expose differences that an “AI infrastructure” label hides. A project with a strong contractual payer can borrow on different terms from one whose customer’s business still needs to prove itself. Collateral, tenor, guarantees, and execution risk matter as well.
06 / Reported financing terms
Similar hardware. Very different financing costs.
Selected USD notes and GPU facilities · stated annual cash coupon or fixed rate
CoreWeave$1.25B
USD senior notes · due 2032
Applied Digital$2.35B
ComputeCo · due 2030
Applied Digital$2.15B
ComputeCo 2 · due 2031
Applied Digital$1.59B
ComputeCo 3 · due 2031
IREN$3.60B
Microsoft contract · GPU facility
IREN$2.40B
Mackenzie · GPU facility
All bars share a 0–12% scale. Amounts identify each issue or facility, not each company’s total debt.
Transaction dates and qualifications
- CoreWeave, USD senior notes · due 2032: 2026-06-11. Pricing announcement. USD tranche only. Senior unsecured; coupon, not market yield.
- Applied Digital, ComputeCo · due 2030: 2025-11-20. Closed note issue. Issued at 97% of par; coupon understates effective financing cost. Principal amortization begins December 2027. Parent completion guarantee.
- Applied Digital, ComputeCo 2 · due 2031: 2026-03-10. Closed note issue. Issued at 98% of par. Senior secured subsidiary debt.
- Applied Digital, ComputeCo 3 · due 2031: 2026-06-16. Closed refinancing. Issued at par; replaces the bridge facility. Do not add both to debt.
- IREN, Microsoft contract · GPU facility: 2026-08-27. Facility reported in business update. Investment-grade GPU financing. Facility amount is not necessarily drawn debt.
- IREN, Mackenzie · GPU facility: 2026-08-27. Facility reported in business update. Non-investment-grade customer deployment; 9% fixed. Included within $2.8B newly announced GPU financings.
Sources: CoreWeave — June 2026 note pricing · Applied Digital — FY2026 Form 10-K · IREN — FY2026 results and business update. Retrieved 2026-09-13.
Selected transactions, not weighted-average company borrowing costs or current market yields. Terms, collateral and durations differ. Facility size is not necessarily drawn debt. Note discounts and fees can raise effective costs above the displayed coupon; the three Applied Digital issues are subsidiary obligations.
IREN makes the comparison unusually concrete: it reports 6% GPU financing for its Microsoft contract and a separate 9% facility supporting a non-investment-grade customer deployment. The rates reflect different packages of risk; they do not isolate customer credit as the only cause. IREN business update
Applied Digital’s three displayed subsidiary issues total $6.09 billion of principal, implying about $474 million in annual coupons before amortization. Its May balance sheet predates the June issue, which refinanced a bridge loan. Those figures should not be added together. Project financing can limit parent recourse, although completion guarantees can still require additional capital. Applied Digital annual filing
Leverage can also sit above the model developer. SoftBank arranged a $40 billion bridge facility primarily for additional OpenAI investment. On September 9 it announced plans to repay the $25.9 billion outstanding on September 15. At this post’s cutoff, repayment is scheduled, not completed; replacement funding must be examined before calling it deleveraging. Facility announcement, repayment announcement
An equity-funded lab can therefore have a leveraged investor, a leveraged compute provider, and a leveraged landlord. The research task is to follow the payment obligations and identify who ultimately bears a shortfall. A fund manager arranging credit is not necessarily the entity holding all of that credit risk.
These disclosures establish exposure to an ambitious revenue ramp. They do not establish that debt repayment requires any particular rate of model improvement. That connection has to be demonstrated through the economics of the customers signing the contracts.
The guarantee can lead back to the chipmaker
A hardware supplier can also absorb customer credit risk. NVIDIA’s disclosed arrangement with SB Energy and an OpenAI tenant affiliate shows why the obligation map matters alongside the debt table.
07 / Documented payment obligations
A customer default can reach the hardware supplier
PORTS Technology Campus · Pike County, Ohio · agreements dated August 17, 2026
OpenAI → SB Energy
Lease payments
Payments as agreed under the leases.
NVIDIA → SB Energy
Conditional shortfall payment
After specified tenant defaults; recoveries reduce the shortfall.
$105B initial cap
OpenAI → NVIDIA
Reimbursement and indemnity
For amounts NVIDIA actually pays to the lessor.
Initial project scope
4.25 GW
Guarantees take effect with leases and readiness conditions; first readiness expected in 2028.
Recovery matters
A new tenant can reduce the shortfall
The filing allows replacement leasing, sale, or lease assumption among NVIDIA’s remedies.
Read the obligations and their limits
- OpenAI → SB Energy: Lease payments. Payments as agreed under the leases.
- NVIDIA → SB Energy: Conditional shortfall payment. After specified tenant defaults; recoveries reduce the shortfall.
- OpenAI → NVIDIA: Reimbursement and indemnity. For amounts NVIDIA actually pays to the lessor.
OpenAI is a tenant affiliate; SB Energy denotes the lessor and affiliates. The guarantee generally covers the shortfall between the guaranteed minimum lease value and recoveries. Termination events include specified lease termination, the lease’s twentieth anniversary, and OpenAI attaining a satisfactory credit rating. The diagram omits other corporate and project relationships.
Sources: NVIDIA — August 17 Form 8-K. Retrieved 2026-09-13.
Selected obligations in one disclosed arrangement. Arrows show who owes whom, not cash already paid; line widths do not encode amounts. The $105B initial cap is conditional exposure, not drawn debt, a loss estimate, or the total project cost. Additional optional support is excluded.
The economic question is what a replacement customer would pay. Broader open-model adoption could support alternative uses for infrastructure, but high utilization alone does not guarantee enough rental income to protect its original financing. OpenAI’s reimbursement obligation also leaves NVIDIA exposed to the same tenant’s ability to pay.
This project is distinct from NVIDIA’s platforms targeting more than $500 billion of third-party capital over time. That figure is a financing ambition, not a guarantee of the entire amount; NVIDIA describes potential residual-value support on selected opportunities. NVIDIA’s platform explanation
The adopter can keep the savings
The natural beneficiaries are businesses that can change how work gets done. A small company can move quickly because it has fewer processes to unwind. A larger company can bring better data, engineering resources, and distribution. Foresight and execution matter more than size alone.
There are already deployment foundations to study. Walmart has described AI tools for associate assistance and task management. JPMorgan describes enterprise deployment of generative AI. Neither disclosure proves savings from switching to open weights; both show where cheaper capable models could enter existing operations. Walmart, JPMorganChase
Workflow owners may retain part of the benefit too. ServiceNow describes evaluating a portfolio of open and frontier models against specific enterprise tasks. Its customer can buy the surrounding workflow while the underlying model changes. That is a plausible source of bargaining power as models become interchangeable. ServiceNow’s model strategy
Savings are not guaranteed to become profits. Competitors may pass them through as lower prices. Customers may demand more output for the same fee. Open-model developers may gain adoption while a host, application vendor, or end customer captures most of the value.
Where margins could be squeezed
Routine outsourcing, contact-center work, advertising production, and conventional IT services are natural places to investigate. When a customer can obtain an acceptable result with fewer purchased labor hours, the supplier’s pricing model comes under pressure.
Existing providers can respond. They can automate their own delivery, own the integration work, and charge for outcomes. The important question is how much productivity they retain after customers renegotiate.
The early evidence does not justify a blanket verdict.
08 / Reported operating margins
A margin-squeeze thesis still has to meet the evidence
Year-over-year changes within each company · percentage points
Quarter ended May 31 · Non-GAAP operating margin
Half-year ended June 30 · Headline operating margin
Sources: Concentrix — Q2 FY2026 results · WPP — H1 2026 results. Retrieved 2026-09-13.
Both companies use adjusted measures. Concentrix’s denominator is total revenue; WPP’s is revenue less pass-through costs. Compare each company with itself. These results do not identify AI’s causal contribution, and WPP’s margin improved.
Concentrix combines a potential pricing challenge with financial obligations: at May 31 it reported about $4.58 billion of balance-sheet debt against $256 million of cash. Its adjusted operating margin fell, but it continued generating positive operating cash flow. WPP’s adjusted margin improved. Neither result isolates AI as the cause. Concentrix results, WPP interim results
These are two different research screens. One examines companies borrowing to build AI capacity. The other examines businesses whose existing revenue may be repriced by cheaper intelligence. The second can suffer even if compute investment remains strong.
That makes a broad equity-index short an imprecise expression of this thesis. It bundles potential adopters, distributors, hardware suppliers, and vulnerable intermediaries into one trade. A company-specific case still needs evidence on margins, valuation, and the timing of any catalyst.
A hedge within a broadly long portfolio
My portfolio implication starts with AI’s potential to create value across the economy. I would want to remain broadly long equities to participate in that value, including the gains retained by capable adopters. A modest bearish position in selected infrastructure companies could hedge part of that exposure. The intended role would be to dampen portfolio swings when financing or execution disappoints, rather than serve as a standalone bet against AI.
CoreWeave and IREN are the two crossover candidates I would investigate. Both deploy expensive compute infrastructure before collecting all the revenue expected from it. There are two possible routes to pressure. Construction can fall behind schedule because each site encounters its own power, permitting, cooling, or equipment problem. Alternatively, fewer large frontier training runs and more competitive inference hosting can reduce the returns available on new capacity. Debt obligations can persist in either case.
CoreWeave (CRWV) has the clearest connection between delivery, customer commitments, and financing. Its debt facilities include provisions tied to delayed site delivery. Its filing also warns that a shift toward consumption-based purchasing could weaken cash-flow predictability and margins. Yet committed, take-or-pay contracts supplied 98% of second-quarter revenue. Lower market prices would initially threaten new bookings and renewals more directly than revenue already protected by those contracts. The case depends on cash arriving later or future capacity earning less, with the existing contracts providing a meaningful buffer. CoreWeave quarterly filing
IREN (IREN) combines data-center development with ownership of the GPUs used to generate cloud revenue. Delays can postpone customer acceptance and receipts; some Microsoft-related financing guarantees step down as capacity is accepted. Its cloud contracts range from month-to-month to five years, exposing different portions of the business to pricing pressure at different times. Longer contracts generally fix pricing, and its $5.90 billion of June cash provides substantial flexibility. The hedge case needs weaker returns on the buildout, not merely a lower spot rental price. IREN annual filing
The counterargument applies to both companies. Open-model competition has to reach infrastructure returns for this hedge thesis to work. An inference boom could instead improve utilization, extend the useful life of existing equipment, and reward operators with scarce power and efficient serving. The same companies could be beneficiaries. Lower margin percentages alone would not establish weaker profits or returns on capital.
09 / Interactive portfolio illustration — assumed returns
A small hedge changes the portfolio’s payoff
100% long equities, plus a modest bearish overlay split equally between CRWV and IREN.
2.50% of starting capital short each company · 95% net stock exposure. The long equity position stays at 100%.
All stock returns below are invented stress cases. They are inputs, not predictions derived from the companies’ filings.
AI boom rewards the builders
Hedge adds -2.0 pp
Assumed returns: equities +15.0% · CRWV +50.0% · IREN +30.0%
Construction delays hurt the builders
Hedge adds +1.5 pp
Assumed returns: equities -10.0% · CRWV -40.0% · IREN -20.0%
Adopters gain; compute margins shrink
Hedge adds +1.0 pp
Assumed returns: equities +8.0% · CRWV -30.0% · IREN -10.0%
Hedge fails: builders rise as equities fall
Hedge adds -1.0 pp
Assumed returns: equities -10.0% · CRWV +30.0% · IREN +10.0%
How the portfolio arithmetic works
Portfolio return = equity return − (total short notional ÷ 100) × average of CRWV and IREN returns. At the initial 5% setting, $100 of capital holds $100 of long equities, shorts $2.50 of each company, and retains the $5 proceeds as cash. Gross stock exposure is 105%; net stock exposure is 95%.
The final case shows why a hedge can add losses. Higher short exposure magnifies both the offset and the risk. Direct shorts can lose more than their initial notional; the plotted cases are not worst-case bounds. Option premiums, expiries, and nonlinear payoffs are outside this illustration.
Every return and position size is an author assumption, not historical data, a forecast, or a recommended allocation. The long equity position equals 100% of starting portfolio capital. Short notional is additional, split equally between CRWV and IREN, with proceeds held as cash. No rebalancing, borrowing costs, dividends, cash interest, taxes, or trading costs. Linear short-stock payoffs only; options would behave differently. The cases have no assigned probabilities or time horizon and do not establish lower volatility.
The illustration keeps the portfolio’s long equity investment intact and adds small bearish exposures. In the assumed delay case, they offset part of the equity loss. In the assumed adoption case, equities rise while the hedge also contributes. When builders outperform, the hedge costs money; it can add losses even when the broader portfolio falls.
That makes volatility reduction an objective to test, not a property established by these scenarios. Position size, valuation, borrowing or option costs, and the timing of contract repricing determine whether the hedge earns its place. Defined-risk options can limit the loss on the bearish position, but their premiums and expiries change the economics; the chart’s simple short-stock arithmetic does not model them. The company analysis supports a focused hedge research list, not an unconditional short recommendation.
Four ways the scenario resolves
| Path | What changes | Likely economic beneficiaries | Where pressure concentrates |
|---|---|---|---|
| Open models become adequate for more work | Available-model premiums shrink; adoption spreads | Capable adopters and differentiated workflow owners | Replaceable model services and routine labor providers |
| Inference absorbs the buildout | More workloads keep equipment productive | Efficient operators and hardware suppliers | Operators whose debt costs exceed the margins they retain |
| Funding or demand arrives too slowly | Financing access, contract conversion, pricing, or utilization disappoints | Buyers of discounted capacity; strong balance sheets | Borrowers needing refinancing or more capital before cash generation |
| Verified safety or privileged access remains scarce | Customers pay for controlled access and reliability | Labs and distributors that prove a durable advantage | Unqualified substitutes and customers without access |
The last path is the strongest counterargument. Safety may become an economically valuable distinction. A small difference in failure rates can justify a large model-price premium on consequential work. Restricted access could also let a lab capture value through its own products or selected partners. An internal lead does not automatically become commercially useless.
My working expectation is a split market: cheaper interchangeable models for more routine tasks, and premiums for difficult work where reliability and access still matter. That is a judgment to test, not a probability estimate.
Follow the cash that survives competition
The next evidence should come from production substitution and completed-task costs, not model downloads alone. Watch what customers move to cheaper models, how much review remains, and whether usage expands outside coding. Then follow the money: renewal prices, margins after replacement investment, unrestricted liquidity, remaining funding needs, guarantees, and the next debt payment.
The earlier cycle framework also needs a measurement adjustment. Frontier-lab revenue becomes a less complete indicator of AI’s economic value if that value moves into open-model hosting, application margins, and adopters’ savings. A weaker lab-revenue-to-capex ratio can coexist with improving returns elsewhere.
The available frontier can flatten while adoption accelerates. Hardware demand can rise while model premiums fall. The fault line is where a business has financed fixed obligations with cash flows that competition or restricted access makes harder to earn.
Pacing the advantage would give capable adopters more time to catch up. Debt would continue on its original schedule.
Source and data notes
The figures use hand-curated source observations frozen as of September 13, 2026. Every reported chart links to its primary source. Financial observations span May 31, June 30, and August 31; they are not a synchronized quarterly sample. Debt refers to reported balance-sheet carrying amounts unless explicitly identified as note principal or a facility commitment. Lease obligations are excluded from the cash/debt chart, not assumed to be economically irrelevant. Cash excludes restricted balances and marketable securities.
Cash/debt percentages are cash divided by the sum of reported current and non-current debt components. Oracle’s interest coverage is quarterly GAAP operating income divided by quarterly interest expense; it is not a debt-service coverage ratio. Applied Digital’s annual coupon calculation is principal times stated rate for each of the three displayed issues, before amortization and fees. Later issues and refinanced facilities must not be double-counted.
The funding chart compares three-month cash-flow statements ending in different months. Its gap is cash capex minus operating cash flow, without annualization or an AI-only allocation. Customer prepayments included in operating cash flow must not also be deducted from capex. Financing inflows are reported separately; neither the gap nor the prepayment sensitivity is a forward funding forecast. The obligation map represents selected contractual relationships in NVIDIA’s August 17 filing. Its guarantee cap must not be added to the earlier debt balances or treated as an expected loss.
The first figure is a conceptual scenario with no measured axes. The second is a historical vendor evaluation, not a current leaderboard. The compute explorer uses assumed starting shares and transparent arithmetic, without a calibrated time horizon. Margin comparisons are year-over-year within each company and use different company-defined adjusted measures. No chart estimates the causal effect of AI on company results.
The hedge explorer also uses author assumptions throughout. It holds the long equity position at 100% of initial capital and adds equal short notionals in CRWV and IREN, retaining the short-sale proceeds as cash. Its returns exclude costs and cash interest, assume no rebalancing, and do not model options. The scenarios have no assigned probabilities or horizon; they are neither a backtest nor evidence that the overlay reduces portfolio volatility. The selected sizes and stock returns are not recommendations or estimates.
- Dario Amodei — We Must Pace the Frontier — published 2026-09; Why Pace?; Pacing Within Democracies.
- DeepSeek-R1 — original model and distilled-model evaluations — published 2025-01; Distilled Model Evaluation; o1-mini and DeepSeek-R1-Distill-Qwen-32B rows.
- AWS — six fully managed open-weight models — published 2026-02-10; Model availability announcement.
- NVIDIA — Mastering LLM Techniques: Inference Optimization — published 2023-11; Model optimization and in-flight batching.
- CoreWeave — Q2 2026 results — published 2026-08-11; June 30 balance sheet; Q2 statements of operations and cash flows (three-month columns); revenue backlog.
- CoreWeave — Q2 2026 Form 10-Q — published 2026-08; Note 2 Revenue; Note 7 Debt; risk factors on take-or-pay contracts, pricing, and data-center delivery.
- CoreWeave — June 2026 note pricing — published 2026-06-11; USD 1.25B 9.625% notes due July 15, 2032; euro tranche excluded.
- Applied Digital — FY2026 Form 10-K — published 2026-08; May 31 consolidated balance sheet; Note 8 Debt; subsequent events for June 16 refinancing.
- Oracle — Q1 FY2027 results — published 2026-09-10; August 31 balance sheet; statements of operations and cash flows.
- IREN — FY2026 results and business update — published 2026-08-27; June 30 consolidated balance sheet; Capital-efficient GPU financing.
- IREN — FY2026 Form 10-K — published 2026-08; Risk factors: construction, contract pricing and matching commitments; GPU financing guarantees and Microsoft acceptance; June 30 cash balance.
- SoftBank — OpenAI investment bridge facility — published 2026-03-27; USD 40B facility; maturity March 25, 2027.
- SoftBank — planned early repayment — published 2026-09-09; USD 25.9B outstanding; repayment scheduled September 15, 2026.
- Concentrix — Q2 FY2026 results — published 2026-06-29; Quarter ended May 31; non-GAAP operating margin and balance sheet.
- WPP — H1 2026 results — published 2026-08-06; Headline operating margin; adjusted net debt; revenue less pass-through costs.
- ServiceNow — Flexible by Design — published 2026-08-25; Model portfolio and workflow-specific evaluation.
- JPMorganChase — Jennifer Piepszak shareholder letter — published 2026-04; Investing in technology, data and AI.
- Walmart — AI tools for associates — published 2025-06-24; AI-driven task management; early results are manager estimates.
- Andy Constan — AI funding and issuance — published 2026-08-20; Interview discussion of issuance and capital-market interruptions; conceptual attribution only.
- NVIDIA — August 17 Form 8-K — published 2026-08-17; Item 1.01: Residual Value Guaranties; initial $105B cap, trigger events, remedies, and OpenAI reimbursement.
- NVIDIA — compute financing platforms — published 2026-08-11; Bringing Capital to AI Factories; Why would NVIDIA support financing? $500B mobilization target is distinct from project support.