The AI Capital Absorption Test: When the Compute Boom Must Prove Itself
DN analyzes $700B+ hyperscaler capex, cloud backlogs, GPU demand, falling AI prices, depreciation, power constraints and rates to identify the real AI overbuild risk window.
The AI Capital Absorption Test 2027–2028
The AI boom has reached the point where asking whether artificial intelligence is “real” is almost beside the point. The technology can transform the economy and still produce a destructive capital cycle. The question that matters now is whether the profits arrive before hundreds of billions of dollars of new compute begin aging.
What Matters
The evidence does not currently look like a completed AI overbuild.
The strongest operating indicators still look like scarcity.
Microsoft says demand remains constrained by available capacity and ended fiscal 2026 with $678 billion of commercial remaining performance obligations.
AWS grew 37% year over year in the second quarter.
Google Cloud grew 82%.
Oracle says demand for AI training and inference is growing faster than supply.
Nvidia Data Center revenue grew 117%, AMD Data Center revenue grew 107%, and Broadcom's AI semiconductor revenue grew 221%.
The August ISM Manufacturing PMI stood at 54.6 and the Services PMI at 55.4, both consistent with expansion.
Those are not the statistics we would expect to see if a large inventory of unusable AI capacity had already appeared.
But the opposing evidence is becoming equally important.
Amazon, Microsoft, Alphabet and Meta are now guiding toward roughly $720 billion to $745 billion of 2026 capital investment under their respective reporting definitions.
Free cash flow is already under pressure at several companies.
Depreciation is rising rapidly.
The U.S. 10-year Treasury yield has moved around 5%, sharply increasing the hurdle rate against which long-duration capital projects compete.
And the economic price of AI inference keeps collapsing.
The critical test therefore probably comes after the infrastructure is installed.
DN's base hypothesis is that 2027–2028 becomes the AI Capital Absorption Window: the period in which the extraordinary investment wave must convert from capacity into economically productive workloads fast enough to outrun price deflation, depreciation, obsolescence, power costs and financing costs.
Technology Truth Is Not Capital-Cycle Truth
One of the biggest analytical mistakes in technology investing is assuming that a transformative technology guarantees attractive returns for everyone financing its infrastructure.
It does not.
Railways transformed economies.
Electricity transformed economies.
Fiber-optic networks transformed economies.
The internet transformed economies.
All could still experience periods where the capital installed to capture the opportunity exceeded the amount economically justified at the time.
Federal Reserve research published in July 2026 makes essentially this distinction. It notes that general-purpose technologies can create rational investment booms because participants repeatedly update their expectations upward as the technology exceeds prior assumptions.
Overinvestment does not require mass irrationality.
It can emerge from rational competition under extreme uncertainty.
The same Fed research found that U.S. intellectual-property and equipment investment as a share of GDP had risen sharply by the first quarter of 2026, sitting only slightly below its 2000 peak, with recent acceleration comparable to the 1990s but beginning from a higher base.
The Scale of the 2026 Buildout
| Company | Current 2026 CapEx Indication | Evidence of Demand | Pressure Point |
|---|---|---|---|
| Amazon | Approximately $220B current plan | AWS +37%; AI and chips businesses each above $25B annual run rate; capacity still constrained | TTM free cash flow moved to a $7.6B outflow |
| Microsoft | Approximately $175B after lease-accounting reclassification | $678B commercial RPO; Azure +43%; demand remains capacity constrained | Roughly two-thirds of quarterly CapEx is short-lived CPUs/GPUs |
| Alphabet | $195B–$205B latest guidance | Google Cloud +82%; capacity remains constrained | Depreciation and infrastructure operating costs rising rapidly |
| Meta | $130B–$145B | Core advertising revenue remains strong and AI supports engagement/monetization | Q2 FCF only $784M; long-term debt $83.66B |
The figures above are not perfectly accounting-comparable. Companies differ in their treatment of finance leases, operating leases and property/equipment. The aggregate is therefore an indication of scale rather than a GAAP-comparable accounting metric.
The aggregate is nevertheless remarkable.
Four companies alone are committing roughly three quarters of a trillion dollars in a single year.
Oracle adds another large layer outside that comparison.
Its fiscal Q1 2027 remaining performance obligations reached $664 billion after the company booked more than $30 billion of additional AI cloud contracts.
Oracle simultaneously sold $20 billion of common equity during the quarter as part of its capital investment program.
This Is Not Yet Dark Fiber
The strongest case against an imminent infrastructure bust is straightforward:
providers are still reporting scarcity.
Microsoft says it expects capacity constraints even while adding large amounts of compute.
Alphabet has been using third-party capacity while internal infrastructure catches up.
Oracle says AI cloud demand continues to grow faster than supply.
Amazon has likewise said it cannot satisfy all expected demand despite sharply raising capital investment.
That matters.
Overcapacity and insufficient capacity are opposite operating conditions.
| Demand Signal | Latest Evidence | DN Interpretation |
|---|---|---|
| Microsoft Commercial RPO | $678B, +84% | Strong Demand |
| Google Cloud | $24.8B quarterly revenue, +82% | Accelerating |
| AWS | $42.2B quarterly revenue, +37% | Accelerating |
| Oracle Cloud Infrastructure | $7.4B quarterly revenue, +121% | Supply Constrained |
| Nvidia Data Center | $89.0B quarterly revenue, +117% | Very Strong |
| AMD Data Center | $6.7B quarterly revenue, +107% | Very Strong |
| Broadcom AI Semiconductors | $16.7B quarterly revenue, +221% | Very Strong |
That does not prove the next wave of capacity will be equally well utilized.
It tells us something narrower:
the overbuild has not yet clearly revealed itself in the operating data.
The $700B Question Is Not Demand. It Is Marginal Return on Capital.
Cloud revenue can grow while returns on incremental infrastructure decline.
Imagine capacity doubles.
Workload demand rises 70%.
The industry is still growing rapidly.
But the marginal capacity may be underutilized.
Now add falling inference prices.
A workload that generated $1 of revenue yesterday might generate $0.60 tomorrow even if the number of tokens consumed rises.
AI infrastructure economics therefore depend on three variables simultaneously:
- workload growth;
- price per unit of intelligence;
- cost of supplying that intelligence.
DN Inference Absorption Elasticity
This may become one of the defining variables of the AI infrastructure cycle.
The OECD estimates that the quality-adjusted price index for text models fell almost 80% between January 2024 and April 2026.
Separate research in the Journal of Economic Perspectives finds that the market price of model intelligence has fallen roughly a thousandfold over its broader development.
Normally, that sounds disastrous for infrastructure returns.
But AI has a powerful counterforce.
Cheaper intelligence creates more demand for intelligence.
And agents intensify this dynamic.
A chatbot may use one inference sequence.
A persistent agent may:
- search;
- reason;
- call tools;
- inspect results;
- retry;
- delegate;
- monitor;
- repeat the process continuously.
The OECD explicitly notes that agentic workloads can consume orders of magnitude more tokens per task.
The AI Boom Has a Duration Mismatch That Fiber Did Not
This is where the dot-com analogy becomes incomplete.
Fiber installed in 2000 could remain physically useful many years later.
A GPU cluster has a much shorter competitive life.
Microsoft disclosed that roughly two-thirds of its latest quarterly capital expenditure was for short-lived assets, primarily CPUs and GPUs.
Alphabet has similarly indicated that roughly 60% of its technical infrastructure capital expenditure goes toward machines, with the remainder largely in longer-lived data centers and networking assets.
The distinction matters enormously.
A data-center building may remain useful for decades.
The accelerator inside it can become economically inferior within a much shorter period.
This creates a different failure mode from traditional infrastructure.
Unused fiber could wait for demand.
Underutilized accelerators may not have that luxury.
New chips can deliver:
- more tokens per watt;
- lower cost per token;
- more memory;
- better interconnect;
- greater throughput;
- better performance per dollar.
Old capacity can therefore lose economic value even while remaining technically functional.
The Depreciation Wall Is Already Appearing
Capital expenditure does not hit the income statement all at once.
A large portion first appears on the balance sheet.
Depreciation follows as assets enter service.
That creates a lag.
Alphabet's property-and-equipment depreciation rose from $9.5 billion in the first half of 2025 to $13.6 billion in the first half of 2026.
Alphabet has explicitly warned that its infrastructure buildout will create higher depreciation and data-center operating costs.
Microsoft likewise reports that a large share of spending is in short-lived CPUs and GPUs.
This is one reason 2027–2028 matters more than the headline 2026 spending number.
Many facilities being financed or constructed today have not yet reached full productive service.
The relevant question is what happens when they do.
Accounting Life Is Not Economic Life
Another analytical mistake is to equate accounting depreciation schedules with the real competitive life of technology.
Microsoft recently extended the estimated useful life of its data-center and office buildings from 15 years to 25 years.
That affects the timing of depreciation and lease classification.
It does not imply that a GPU installed inside the building remains frontier-equivalent for 25 years.
An AI capital model therefore has to separate:
- land;
- buildings;
- power infrastructure;
- networking;
- CPUs;
- GPUs and accelerators.
Each asset has a different economic duration.
DN AI Capital Absorption Ratio
This is not a GAAP metric.
It is an economic stress framework.
The purpose is to answer a more useful question than:
“How much revenue is AI generating?”
The better question is:
“How much profit is the next dollar of AI infrastructure generating relative to the economic burden created by installing it?”
Why Revenue Can Mislead
Suppose a new $100 billion capacity block generates $40 billion in annual revenue.
That sounds extraordinary.
But assume:
- gross margin is 55%;
- annual depreciation is $15 billion;
- economic capital charge is $8 billion;
- power and operating overhead is $6 billion.
Gross profit would be $22 billion.
The capital burden would be $29 billion.
The infrastructure could therefore be strategically important and revenue-generating while still failing a simplified capital-absorption test.
DN AI Capital Absorption Stress Engine
AI Capital Absorption Stress Engine
Stress-test whether a hypothetical block of AI infrastructure can generate enough economic gross profit to absorb its depreciation, cost of capital and operating burden before compute ages.
This is a simplified scenario engine, not a company valuation, accounting model, investment recommendation or forecast of actual returns. It intentionally separates short-lived compute from longer-lived infrastructure.
Why the Cost of Capital Suddenly Matters Again
AI infrastructure was easier to justify when capital was cheap.
That environment is changing.
On 16 September 2026, the Federal Reserve raised the federal-funds target range by 25 basis points to 3.75%–4.00%.
The Fed described economic activity as expanding at a solid pace and capital investment as robust, but inflation remained elevated.
Meanwhile, the 10-year Treasury yield moved around 5%, its highest area since 2007.
A 5% risk-free yield changes the mathematics of infrastructure.
Every capital project must compete with a materially higher hurdle rate.
This does not automatically cause an AI bust.
It does mean that weak future cash flows become less forgivable.
Why Hyperscaler Stock Underperformance Is Not Enough
Market price is informative.
It is not an infrastructure-utilization statistic.
A hyperscaler can underperform the broader market because:
- bond yields rise;
- valuation multiples compress;
- investors rotate sectors;
- CapEx reduces near-term free cash flow;
- future returns become more uncertain.
None of those requires unused data centers.
That is why DN does not treat relative equity performance as a sufficient overbuild signal.
We prefer operating evidence.
The Macro Economy Is Not Yet Signaling a Dot-Com Replay
August 2026's Manufacturing PMI was 54.6.
Production stood at 58.3.
New Orders were 53.7.
Services PMI was 55.4, with Services New Orders at 60.9.
That is an expanding economy.
There are weaknesses.
Services employment was below 50 and both manufacturing and services price indices were above 70.
In other words:
growth is currently resilient, but inflationary pressure limits the ability of monetary policy to rescue weak capital economics.
Do Not Turn the Yield Curve Into a Calendar
Yield curves contain useful information.
They are not clocks.
As of 16 September 2026:
- 10-year minus 2-year Treasury spread was approximately +0.27 percentage points;
- 10-year minus 3-month spread was approximately +0.87 percentage points.
Both are currently positive.
Federal Reserve research has also found that there is no single universally best term-spread recession predictor.
Different spreads perform differently over different horizons and samples.
The DN Four-Gate Overbuild Test
DN would not call a systemic AI infrastructure bust until multiple independent economic domains deteriorate together.
Supply constraints disappear, utilization weakens and providers begin reporting excess available compute.
AI/cloud revenue, bookings or backlog decelerate faster than inference economics and efficiency improve.
Depreciation, power, financing and infrastructure costs rise faster than incremental gross profit.
Orders and activity weaken while credit conditions and required returns remain restrictive.
The Current September 2026 Dashboard
| Indicator | Current Evidence | DN Signal |
|---|---|---|
| Real Economy | Manufacturing PMI 54.6; Services PMI 55.4 | Expansion |
| Cloud Demand | Multiple providers still report supply constraints | Scarcity |
| Cloud Backlogs | Microsoft $678B RPO; Oracle $664B RPO | Strong |
| Accelerator Demand | Nvidia +117%; AMD +107%; Broadcom AI +221% | Very Strong |
| Free Cash Flow | Pressure visible at Amazon, Alphabet and Meta | Watch |
| Depreciation | Accelerating as infrastructure enters service | Watch |
| Inference Pricing | Quality-adjusted prices falling extremely quickly | Return Pressure |
| Long-Term Rates | 10-year Treasury around 5% | Higher Hurdle |
| Power Availability | Data-center electricity demand rising much faster than grid demand | Binding Constraint |
The dashboard is not yet consistent with a completed bubble break.
It is increasingly consistent with a boom entering a more demanding proof-of-return phase.
Power Constraints May Actually Delay the Overbuild
Power is usually described as a problem for AI.
It may also function as an accidental capital-discipline mechanism.
The International Energy Agency expects global data-center electricity demand to more than double to roughly 945 TWh by 2030.
In the United States, data centers are projected to account for nearly half of electricity-demand growth through the end of the decade.
Data centers can be built in two or three years.
Transmission, generation and broader energy infrastructure can take much longer.
This mismatch restricts how quickly theoretical compute supply becomes usable compute supply.
The Capital Structure Is Changing
The first wave of AI infrastructure was financed largely by the enormous internal cash flows of the largest technology companies.
The financing mix is broadening.
Alphabet raised $49.6 billion through common and preferred equity in June 2026 and another $20.3 billion through senior unsecured notes.
Meta has materially increased debt while also using outside infrastructure partnerships.
Oracle sold $20 billion of common stock in its fiscal first quarter as part of its capital investment program.
Large AI infrastructure projects increasingly use:
- debt;
- equity issuance;
- finance leases;
- operating leases;
- joint ventures;
- customer commitments;
- customer-supplied hardware;
- prepayments.
This Is Where Systemic Risk Could Eventually Grow
A hyperscaler funding infrastructure from a fortress balance sheet can absorb years of weak returns.
A leveraged data-center operator cannot necessarily do the same.
A private-credit vehicle cannot necessarily do the same.
A project financed against aggressive future utilization assumptions cannot necessarily do the same.
The infrastructure layer therefore becomes more fragile as:
- leverage rises;
- contract concentration rises;
- refinancing dependence rises;
- hardware residual value falls;
- counterparty quality falls.
This is why the eventual weak link may not be Microsoft, Amazon or Google.
It may be capital farther down the financing chain.
AI Adoption Is Enormous. Agent Adoption Is Still Early.
Stanford's 2026 AI Index reports that 88% of surveyed organizations used AI in 2025 and 70% used generative AI in at least one business function.
But deployment of AI agents remained in the single digits across almost every business function.
That gap is strategically important.
There is already widespread AI adoption.
There is not yet widespread autonomous-agent consumption.
If agentic workflows mature, they could represent a second demand wave layered on top of existing AI usage.
General-Purpose Technologies Often Have a J-Curve
AI infrastructure may also look less productive before complementary organizational investments catch up.
Research by Brynjolfsson, Rock and Syverson describes a Productivity J-Curve for general-purpose technologies.
New technologies often require significant complementary investment in:
- workflow redesign;
- software;
- skills;
- organizational processes;
- business models;
- data;
- management systems.
The costs appear early.
The measurable productivity can arrive later.
This is another reason to avoid interpreting a temporary gap between AI CapEx and economy-wide productivity as proof of failure.
But the same lag creates financial risk.
Capital providers must survive long enough for the productivity payoff to arrive.
The 2027–2028 AI Capital Absorption Window
DN therefore divides the infrastructure cycle into three phases.
Phase 1: Installation and Scarcity — 2024 through 2026
Demand runs into available compute, power and supply-chain constraints.
Cloud providers race to install capacity.
Backlogs expand.
Semiconductor suppliers experience extraordinary growth.
Capital efficiency matters less because scarcity makes incremental compute valuable.
Phase 2: Capital Absorption — 2027 through 2028
Large portions of the 2025–2027 build begin operating.
Depreciation expands.
Power and operating costs become more visible.
Inference competition pushes unit prices lower.
Investors begin differentiating between:
- capacity with contracted demand;
- capacity with speculative demand;
- high-utilization infrastructure;
- low-utilization infrastructure;
- assets with durable economics;
- assets being technologically leapfrogged.
This is the phase where the boom has to prove itself economically.
Phase 3: Bifurcation
If AI workloads, agents and enterprise adoption absorb the infrastructure profitably, the buildout can evolve into a durable utility layer.
If not, capital expenditure falls, weaker providers consolidate, financing structures break and capacity is repriced.
Importantly, both outcomes can coexist.
AI can transform the economy while some infrastructure investors suffer severe losses.
DN Base Case: Digestion Before Collapse
The most defensible current forecast is not a straight-line supercycle.
It is also not an imminent dot-com-style collapse.
The operating evidence currently points toward continued near-term demand strength.
But the financial evidence points toward declining tolerance for undisciplined spending.
That suggests the next phase is likely to involve:
- greater differentiation among providers;
- pressure on free cash flow;
- higher scrutiny of depreciation;
- more scrutiny of utilization;
- more emphasis on customer commitments;
- greater importance of power access;
- consolidation among weaker infrastructure operators;
- increasing focus on return on invested capital.
In other words:
the AI trade can become more selective before it becomes universally bearish.
Scenario 1: The Agentic Absorption Supercycle
This is the strongest upside scenario.
Agentic AI creates a new class of workload intensive enough to absorb the buildout.
Persistent agents:
- monitor systems continuously;
- research;
- write software;
- operate browsers;
- make purchases;
- optimize businesses;
- manage workflows;
- interact with other agents.
Inference-volume growth dramatically exceeds price compression.
Cloud backlogs remain strong.
Free cash flow recovers after the investment wave.
Under this scenario, today's CapEx resembles the early buildout of an economic utility.
Scenario 2: Healthy Digestion and Bifurcation
This is the DN base case today.
AI demand remains real and large.
But returns become increasingly uneven.
The strongest providers maintain high utilization while weaker projects struggle.
Inference prices fall.
Efficiency rises.
Margins become more important than revenue growth.
Infrastructure financing migrates toward companies with:
- strong balance sheets;
- committed customers;
- power advantages;
- proprietary silicon;
- high utilization;
- large software ecosystems.
Capital spending decelerates without necessarily collapsing.
Scenario 3: The Overbuild Break
The bearish regime requires more than a high CapEx number.
DN would look for:
- cloud backlog growth slowing materially;
- providers no longer describing capacity as constrained;
- utilization falling;
- AI infrastructure price competition intensifying;
- inference prices falling faster than usage grows;
- depreciation rising faster than AI gross profit;
- FCF deterioration spreading;
- capital-market financing becoming harder;
- PMI/new orders turning persistently contractionary;
- credit stress appearing among neoclouds or data-center financing vehicles.
That combination would look much more like an actual capital-cycle break.
What Would Make DN Turn Bearish Earlier?
The thesis can be falsified.
An earlier overbuild call would become more credible if several of the following appeared together before 2027:
- major cloud providers report available capacity materially exceeding demand;
- RPO or AI backlog growth reverses;
- hyperscalers cut CapEx because of weak customer utilization rather than supply constraints;
- GPU supplier growth collapses while inventory rises;
- large AI customers renegotiate or cancel commitments;
- power-connected facilities struggle to find tenants;
- cloud gross margins deteriorate despite utilization improvements;
- credit losses appear around data-center projects.
What Would Make the Bull Case Stronger?
The risk window would move further into the future if:
- agentic inference becomes a major workload category;
- enterprise AI usage broadens from pilots into core processes;
- cloud backlog growth remains above infrastructure growth;
- utilization remains high after the 2026–2027 capacity wave enters service;
- custom accelerators materially reduce inference cost;
- AI-driven productivity gains increase corporate willingness to pay;
- free cash flow recovers despite continued infrastructure investment.
The Most Important Metric May Be Free Cash Flow After AI
Revenue demonstrates demand.
Gross profit demonstrates some economic value.
Free cash flow tests whether the business can fund the entire machine.
Amazon's trailing-12-month free cash flow moved to a $7.6 billion outflow in Q2 even as operating cash flow increased strongly.
Meta generated just $784 million of quarterly free cash flow while investing heavily.
Alphabet has warned that infrastructure investment will pressure free cash flow.
None of those figures proves overinvestment.
They show that the bill is becoming real.
Why AI May Produce a Bifurcated Bust
A future AI correction may not resemble March 2000 exactly.
The modern infrastructure ecosystem is heterogeneous.
It contains:
- cash-rich hyperscalers;
- high-margin semiconductor suppliers;
- neoclouds;
- data-center REITs;
- power developers;
- private-credit vehicles;
- startups renting compute;
- model companies buying compute;
- enterprises consuming inference.
Their balance sheets are not equivalent.
Their contract quality is not equivalent.
Their residual asset values are not equivalent.
The first real break may therefore occur outside the hyperscalers.
The strongest platforms may acquire distressed capacity rather than become distressed themselves.
The Irony of an AI Infrastructure Bust
If overbuilding eventually occurs, the collapse in infrastructure economics could actually accelerate AI adoption.
Excess capacity lowers prices.
Lower prices make new applications economical.
Those applications create demand.
That demand eventually absorbs the excess capacity.
This is one reason infrastructure bubbles can be socially productive while financially destructive.
The 10 Indicators DN Would Track Every Quarter
| Indicator | Bullish Absorption Signal | Overbuild Warning |
|---|---|---|
| 1. Cloud / AI RPO | Growth stays strong | Backlog contracts or cancellations rise |
| 2. Capacity Commentary | Demand exceeds supply | Excess available capacity appears |
| 3. AI Revenue Growth | Accelerating | Decelerates below infrastructure growth |
| 4. Utilization | High and stable | Falls as new clusters enter service |
| 5. Depreciation Growth | Matched by gross-profit growth | Outruns incremental profit |
| 6. Free Cash Flow | Recovers after build phase | Deteriorates structurally |
| 7. Inference Pricing | Usage growth outruns price decline | Price decline outruns workload growth |
| 8. Semiconductor Demand | Supplier growth remains broad | Inventory and cancellations rise |
| 9. Power / Grid Access | Capacity remains physically constrained | Powered capacity becomes readily available |
| 10. Credit Conditions | Funding available on rational terms | Refinancing stress and defaults rise |
DN Prediction
As of 17 September 2026, the highest-conviction interpretation is:
The AI infrastructure cycle is more likely in a late installation / early absorption phase than in a confirmed overcapacity bust.
Operating demand remains too strong and capacity too constrained to support a high confidence “dark fiber now” thesis.
The more important vulnerability is being pushed forward into 2027–2028, when a much larger installed asset base must compete against falling inference prices, rapidly improving hardware and materially higher capital costs.
The first sign of failure is more likely to be dispersion than a simultaneous collapse: weaker infrastructure projects, leveraged providers and poorly contracted capacity should deteriorate before the most diversified hyperscalers.
A broad AI capital-cycle break becomes substantially more credible only if today's supply constraints turn into excess capacity at the same time that backlog, monetization and macro conditions weaken.
Why This Matters for Crypto
Crypto remains one of the world's most liquidity-sensitive asset classes.
A genuine AI capital-cycle break could affect crypto through:
- equity risk appetite;
- venture funding;
- credit conditions;
- long-duration valuations;
- global dollar liquidity;
- semiconductor and AI-token narratives;
- retail speculative appetite.
But crypto could also benefit from the next AI demand wave.
Agentic finance increasingly requires:
- programmable wallets;
- stablecoin settlement;
- machine payments;
- 24/7 markets;
- API-accessible liquidity;
- agent identity;
- verifiable transactions.
That creates an unusual possibility:
AI infrastructure economics could weaken while agentic-finance adoption accelerates.
The technology cycle and the financial asset cycle do not have to peak together.
Monitor the Capital Absorption Regime
Track hyperscaler equities, semiconductor stocks, Treasury yields, sector relative strength, economic data and cross-asset conditions as the AI infrastructure cycle moves into its proof phase.
Explore TradingView Explore ASCN for Crypto Market Intelligence DN Agentic Finance Frontier Index AI Agent Kill-Switch BenchmarkDN may receive compensation from eligible registrations or purchases through selected partner links. Affiliate relationships do not determine the macroeconomic conclusions in this research.
DN Research Methodology
This research uses a multi-domain framework rather than a single market indicator.
Evidence is divided into six categories:
- Physical capacity: infrastructure deployment and supply constraints.
- Demand: cloud revenue, RPO, backlog and accelerator sales.
- Capital economics: CapEx, depreciation, margins and free cash flow.
- Technology economics: inference pricing, hardware progress and agentic workload intensity.
- Physical constraints: electricity, data-center construction and grid availability.
- Macro/financial conditions: PMI, interest rates, yield curves and capital-market financing.
DN does not infer a market crash merely from high investment.
High investment becomes evidence of overinvestment only when capacity growth persistently exceeds economically productive demand.
The 2027–2028 Capital Absorption Window is a conditional research hypothesis rather than a deterministic forecast.
What Would Prove This Research Wrong?
There are two major ways.
First, the cycle could break earlier.
If current supply constraints suddenly disappear, customer commitments weaken and cloud utilization deteriorates during late 2026, the absorption problem would already be arriving.
Second, the expected absorption problem may never become severe.
Agentic systems, AI-native applications and enterprise deployment could create enough inference demand to absorb the infrastructure faster than anyone currently expects.
AI might then resemble a capacity-constrained utility buildout for much longer.
Both possibilities remain open.
Frequently Asked Questions
Is AI currently in a bubble?
AI infrastructure investment has reached historically unusual levels and therefore creates meaningful overinvestment risk. However, current operating evidence still shows strong cloud growth, large contracted backlogs and capacity constraints. DN therefore distinguishes high bubble risk from a confirmed infrastructure overbuild.
When is the highest-risk period for AI overinvestment?
DN's current hypothesis identifies 2027–2028 as the more important Capital Absorption Window because large amounts of infrastructure being financed and constructed during 2025–2027 will increasingly enter productive service and begin generating depreciation, operating costs and return requirements.
What is the AI Capital Absorption Ratio?
The DN AI Capital Absorption Ratio compares gross profit economically attributable to new AI capacity with the estimated annual burden created by depreciation, capital cost and incremental infrastructure operating expense.
What is Compute Duration Mismatch?
DN Compute Duration Mismatch is the risk that AI hardware requires longer to generate an acceptable economic return than the period during which that hardware remains technologically and economically competitive.
What is the Depreciation Wall?
The DN Depreciation Wall is the delayed phase in which a large capital-investment wave begins flowing through operating expenses as newly constructed infrastructure and compute enter service.
Why could AI agents prevent an infrastructure bust?
Agentic systems can consume substantially more inference than one-shot chatbot interactions because agents plan, use tools, verify outputs, retry and remain active over long periods. If this growth in workload volume exceeds falling inference prices, agents could absorb a large amount of new compute capacity.
What is Inference Absorption Elasticity?
DN Inference Absorption Elasticity measures whether growth in AI workload volume is large enough to offset falling quality-adjusted inference prices. If usage grows more quickly than prices fall, revenue potential can continue increasing despite rapid commoditization.
Does a 5% Treasury yield make an AI crash inevitable?
No. Higher risk-free yields raise the required return on long-duration capital and reduce the present value of future cash flows, but they do not prove overcapacity. Physical utilization, revenue, profitability, financing and demand must be analyzed together.
Can the technology succeed even if the AI investment boom crashes?
Yes. Historically, transformative technologies can produce excessive infrastructure investment. A financial correction can lower infrastructure prices and ultimately make the technology cheaper and more widely accessible even when some investors suffer losses.
Primary Evidence Base
- Federal Reserve: Do Major Technology Advancements Lead to Overinvestment?
- Federal Reserve: September 16, 2026 FOMC Statement
- Federal Reserve: There Is No Single Best Predictor of Recessions
- ISM: August 2026 Manufacturing PMI
- ISM: August 2026 Services PMI
- Microsoft FY2026 Q4 Earnings Call
- Amazon Q2 2026 Results
- Alphabet Q2 2026 Results
- Alphabet Q2 2026 Form 10-Q
- Meta Q2 2026 Form 10-Q
- Oracle Q1 FY2027 Results
- Nvidia FY2027 Q2 Financial Results
- AMD Q2 2026 Form 10-Q
- Broadcom Q3 FY2026 Results
- OECD: Artificial Intelligence Markets
- Journal of Economic Perspectives: The Emerging Market for Intelligence
- Stanford AI Index 2026: Economy
- Stanford AI Index 2026: Research and Development
- IEA: Energy and AI
- NBER: The Productivity J-Curve
Affiliate Disclosure: Decentralised News may receive compensation from eligible registrations or purchases through selected TradingView and ASCN links. Commercial relationships do not determine the economic thesis, source selection or conclusions in this research.
Forecast Disclaimer: The 2027–2028 Capital Absorption Window is a conditional DN research hypothesis. Economic cycles, technological progress, monetary policy, enterprise AI adoption and infrastructure deployment can develop differently from current expectations.
Data Disclaimer: CapEx definitions vary across companies. Finance leases, operating leases, property and equipment and infrastructure commitments are not always reported on identical bases. Aggregate figures should be interpreted as estimates of investment scale rather than accounting-comparable totals.
Investment Disclaimer: Nothing on this page is personalized investment, trading, tax, legal or financial advice. Technology equities, semiconductor companies, cryptocurrencies and other financial assets can experience substantial volatility and loss. 18+.
Related reading:
Agentic Finance Frontier Index 2027: Astra vs Grok Bot vs Claude vs Gemini vs Muse
AI Agent Kill-Switch Reliability Benchmark 2027: Revocation, Limits & Emergency Controls
Best Agentic Checkout Systems 2027: ACP vs UCP vs Visa vs Mastercard
Fastest Crypto Trading APIs 2027: Latency, Rejections & Reliability Ranked
Best Crypto Platforms for AI Agents 2027 | Agentic Finance Rankings
Agentic Wallet Security Index 2027: The Safest Wallets for AI Agents
Best Wallets for AI Agents 2027: Security, Payments & Autonomous Finance
