Why These Eight AI Pressure Tests—and Why Not More?

Eight is not a claim that the AI investment cycle has only eight risks.

It is a design constraint. BoomRisk needs enough independent evidence to test whether capital, demand, returns, financing and market exposure are still reinforcing one another. It also needs to stop before closely related indicators begin counting the same economic pressure several times.

The decisive distinction is therefore not between a small dashboard and a large one. It is between additional information and repeated information.

At publication, the practical consequence is visible. Open-model Pressure moved down by one point under the fixed rules, from 3/5 to 2/5. With all eight tests available and equally weighted, that change lowered the composite from 3.25 to 3.125, displayed as 3.13. The supportive evidence affected the reading even though Capex & Cash Returns, AI Monetisation, Market Concentration and Data-centre Investment/GDP remained at 4/5.

Eight is a coverage decision, not a magic number

Each test must earn its place through four gates: a clear economic relationship to the AI investment thesis, a sufficiently reliable public source, a repeatable observation method and thresholds fixed before the next result is known.

The eight current channels ask different questions:

1. Capex & Cash Returns: is aggregate hyperscaler free cash flow keeping pace with the burden of the buildout?
2. Credit Stress: are broad investment-grade funding conditions becoming materially less supportive?
3. Semiconductor Demand: is the physical demand signal at the front of the AI infrastructure chain still holding?
4. Open-model Pressure: is lower-cost open-weight capability increasing competitive pressure on model economics?
5. AI Monetisation: is observable cloud revenue growth keeping pace with matched capital-expenditure growth?
6. Market Concentration: how dependent is broad US equity exposure on a small group of companies?
7. Valuation Expectations: how demanding is the forward earnings multiple already embedded in the market?
8. Data-centre Investment/GDP: how large and fast-growing is data-centre construction relative to the wider economy?

Fewer tests would leave the result too dependent on one company, one market or one part of the capital cycle. More tests may eventually improve the framework. They do not improve it automatically.

Where double-counting can hide

Several BoomRisk tests are related by design. The cycle itself is related. The safeguard is to score a different economic question in each channel.

Capex appears beside the Capex & Cash Returns test, but capex growth is context there; the score comes from aggregate free-cash-flow growth. Capex is used again in AI Monetisation, where the scored variable is the gap between matched capex growth and AWS plus Google Cloud revenue growth. Scoring capex growth separately in both places would make one investment surge raise the composite twice before its effect on cash or revenue had been observed.

Market Concentration and Valuation Expectations come from the same State Street fund page, but they do not measure the same exposure. One asks how much index weight sits in ten companies. The other asks how much investors are paying for forward earnings. A shared source does not make two indicators duplicates; a shared economic mechanism might.

The same separation applies across the chain. NVIDIA Data Center revenue tests current accelerator demand. Data-centre construction relative to GDP tests how much fixed investment the wider economy is accumulating. Strong orders can coexist with a large future utilisation burden. Treating them as interchangeable would erase the tension BoomRisk is meant to preserve.

Why equal weight is the least misleading choice today

Equal weight is not a claim that every test has identical economic importance in every market regime. A sudden funding shock could matter more than valuation. A collapse in compute demand could dominate a modest change in concentration.

The strongest objection to the current design is therefore valid: an equal-weight average can understate a genuinely dominant risk and give the same numerical influence to a low-confidence proxy and a high-confidence company filing.

BoomRisk accepts that limitation because the alternative currently asks for more certainty than the evidence can support. Stored live readings began on 22 August 2026. That is not a long enough history to estimate stable unequal weights without overfitting, hindsight or subjective intervention. A transparent one-eighth contribution from every available verified test is easier to audit than a set of calibrated-looking weights whose precision has not been earned.

When eight tests are scored, a one-point move in any test changes the arithmetic mean by 0.125. That visibility is a feature. Readers can identify exactly what moved and how much it contributed instead of reverse-engineering a hidden weighting model.

What happens when evidence is missing

The live implementation normally publishes all eight tests. If a refresh fails, the system can retain the last verified observation, mark it stale and disclose that the latest refresh was unavailable. It does not silently replace the missing value with a neutral score.

The composite requires at least six scored tests. Below that threshold, the framework reports incomplete coverage rather than a number. When six or seven are available, the arithmetic mean uses only those scored observations. That keeps the service usable through a temporary source failure, but it also changes the influence of each remaining test and must be treated as a limitation, not normal operating precision.

Source frequency matters too. A daily system does not imply that every input changes daily. Company filings, quarterly results, construction data, credit markets, model catalogues and fund data update on different schedules. BoomRisk refreshes when a comparable source observation exists; it does not manufacture movement from stale fundamentals.

Why important subjects can remain outside the score

Relevance is not sufficient for scoring.

The AI Circular Financing Watch is material to BoomRisk’s thesis, but it remains unscored. Public disclosures do not provide a reliable denominator for related financing, contractual terms are not comparable, parts of the risk already appear in credit, capex and monetisation, and no tested 1–5 thresholds yet exist. Forcing it into the composite would convert an acknowledged evidence gap into an arbitrary number.

China is also a research lens rather than a ninth test. Evidence about model price-performance can affect Open-model Pressure. Evidence about capex, cloud revenue or financing may inform other tests or remain contextual when company definitions are not comparable. Adding a standalone China score on top could count the same competition and capital-allocation evidence twice while replacing financial questions with geography.

Unscored does not mean unimportant. It means the subject can change how BoomRisk interprets evidence and what it investigates next without pretending that the composite already measures it.

The rules for changing the framework

The framework is not permanent, but it cannot change merely because a different design produces a preferred reading.

• Add a test only when it captures a distinct economic channel, passes the four inclusion gates, adds information after an explicit overlap review and has enough comparable history to set thresholds without using the latest result as the answer.
• Remove a test when its economic relationship weakens, its source is no longer reliable or comparable, its thresholds cease to represent the stated risk, or a better measure makes it persistently redundant.
• Reweight a test only when a sufficiently long live record can support a documented calibration and robustness test. Confidence labels alone are not weights, and editorial conviction is not calibration.
• Publish every structural change through a new methodology version, with the reason, effective date and expected effect on comparability. Historical published readings should not be silently rewritten to fit a new design.

This governance is deliberately conservative. The cost is that the score may remain incomplete or imperfect for longer. The benefit is that new information cannot be smuggled into the result through an attractive but untested model adjustment.

What the composite cannot prove

The number is not a probability of a market decline, a forecast of timing or a valuation of the AI industry. It is the average pressure across a defined set of observable channels.

Several inputs are proxies. Cloud revenue is not AI-only revenue. Broad investment-grade spreads are not AI-issuer credit. NVIDIA revenue does not capture every accelerator supplier, cancellation or lead time. Data-centre construction excludes servers, chips and much of the related power investment. Hosted model prices can change faster than realised customer economics. The valuation test does not yet include a reliable earnings-revision feed.

An average also permits offsetting evidence. Four 4/5 readings can coexist with Semiconductor Demand at 1/5 and Open-model Pressure at 2/5. That is not a flaw to smooth away. It is the central BoomRisk proposition: technological and demand evidence can remain supportive while the financial structure around them stays under pressure.

The conclusion

BoomRisk uses eight pressure tests because the AI investment cycle must be examined across capital, funding, demand, competition, returns and market exposure without allowing one story to dominate the result.

The number should expand only when the information set expands. Adding a near-duplicate indicator, assigning unsupported weights or turning an important unknown into a score would make the framework look richer while making its conclusion harder to trust.

Eight is therefore not the final answer. It is the current boundary between useful breadth and unsupported precision.

BoomRisk is a financial-risk monitoring framework, not an investment recommendation or market forecast. The composite is an arithmetic summary of the available scored tests, not a probability, price target or timing signal. Methodology changes require a new published version and disclosure.