AI Stack

Companies selling AI "agents" that automate workflows — software meant to do multi-step work, not just answer questions. The core risk is the gap between a slick demo and software reliable enough for daily production use: agents that impress in controlled settings often stumble on accuracy, speed, or recovering from their own errors in the field. There is also a squeeze risk from below — if the base models can do the agent work directly, the platform layer's value thins out. Contra tags this as a context classification: it describes the company's position in the AI stack rather than firing as a standalone bearish signal.

Look for evidence of production deployment at scale, not pilots or partnership announcements: named customers running the product daily, usage or seat numbers, revenue attributed to the agent line. Then read the risk factors and customer commentary in the filings — companies increasingly name accuracy problems, slowness, or error-recovery friction in their own disclosures. A company marketing an agent platform hard while disclosing deployment friction is telling you the demo-to-production gap in its own words.

Three levels. The weak tag applies when the company markets an agent platform but shows little real-world deployment at scale — narrative ahead of evidence. The medium tag adds the company's own filings naming customer friction: accuracy problems, slowness, or trouble recovering from errors. The strong tag adds either customer retention disclosed as declining, or the underlying model providers starting to offer the same agent abilities directly — the squeeze from below arriving. The ladder runs from unproven, to friction admitted, to the platform's reason for existing coming under direct attack.

No — it means the layer's economics are unresolved, and the tag keeps that visible while the narrative runs hot. Some agent platforms will build durable workflow lock-in before the base models catch up; others are renting a capability that will be given away beneath them. The tag carries no scoring weight by itself. Its job is to make other firings on the same ticker — margin erosion, retention decline, capex strain — read against the right structural backdrop. The framework flags the position; the outcome is decided in the field.

A software company built mostly on top of someone else's AI models — its product is a layer of workflow and interface renting intelligence anyone can rent. The pressure comes from two sides: as the underlying models become interchangeable, the wrapper's edge fades; and established players in the same industry can bolt AI onto products their customers already use. The tell is in the gross margin. When a company is really just a thin layer over a rented model, the inference cost — the "token tax" — eats the margin a genuine software business would keep.

Because a real software business keeps 80–90% gross margins — the marginal cost of serving another customer is near zero. A company paying a model provider for every query carries a real per-unit cost that scales with usage. If gross margin sits well below the software norm and is falling, the market is watching value transfer from the wrapper to the model layer underneath it. This is the software-margin leg of AI disruption — distinct from AI bypassing search and marketplace front doors, and from AI automating human-service middlemen.

The weak reading fires on an application-layer company built on third-party models whose gross margin has slipped to 75% or below — or is still higher but eroding at least 3 points year over year. Exposure visible, not yet biting. The medium reading fires at 70% or below: the token tax is visibly eating the model, well under the 80–90% a real software business keeps. The strong reading requires margin at 60% or below (or compressed at least 6 points year over year to under 70%) and the moat breaking at the same time — net revenue retention below 100%, meaning the existing customer base is now shrinking. Margin and moat eroding together.

It is a grind, not a crash — margins erode quarter by quarter as inference costs and competitive pressure compound. The two numbers to watch each quarter are gross margin and net revenue retention. Margin falling with retention above 100% means the company is paying more to serve customers who are still expanding — painful but survivable. Both falling together means customers are leaving while unit economics worsen, and there is no obvious internal fix. The framework flags the configuration; whether a company can escape it by moving up the stack is the judgment left to you.

No. The pattern targets companies whose product is primarily a layer over third-party models. Software businesses with proprietary data, deep workflow lock-in, or their own models carry different economics — the token tax is smaller relative to the value they add. The 2025–2026 wave of AI-native SaaS makes the distinction urgent, because thin wrappers and durable platforms currently trade on similar narratives. Free registration shows which application-layer names are firing this pattern now.

Companies whose main business is financing the AI buildout rather than building or running it: lenders against data centers, chip-leasing operations, AI-focused private credit vehicles. They sit one layer removed from the technology, but their fortunes are tied to it completely — their assets are loans and leases to AI-infrastructure operators. Contra tags this as a context classification: financing the buildout is a business description, not a flaw. The tag exists because the risk shape of this layer is distinctive and easy to miss behind a lender's ordinary-looking balance sheet.

Because the borrowers all fail the same way at the same time. A diversified lender's defaults arrive scattered across unrelated industries; an AI-infrastructure lender's book is spread across the same kind of borrower, exposed to the same spending cycle. If the buildout slows, the loans go bad together — the diversification that normally protects a credit book is absent by construction. The risk compounds when the exposure concentrates further, down to a handful of operators or a single named AI lab or cloud giant.

The weak tag applies when the company's main exposure is financing AI infrastructure — the business-model fact itself. The medium tag adds concentration: more than 50% of the loan book in AI-infrastructure operators, meaning a buildout slowdown hits most of the assets at once. The strong tag is single-point concentration: exposure concentrated in one named AI lab or cloud giant, where the lender's book is effectively a credit bet on one institution's spending program. Each level narrows the number of things that must go right for the book to perform.

Watch the health of the borrowers, not just the lender's own reported numbers — a concentrated credit book looks pristine until the cycle turns, because the loans were all written in good times. The useful questions: how concentrated is the book, what happens to collateral values (data centers, chips) if the buildout slows, and does the lender's yield actually compensate for correlated default risk? The context tag keeps the concentration visible; pairing it with the hyperscaler capex-divergence pattern shows whether the spending that services these loans is itself under strain.

It is the set of companies supplying the physical guts of AI: chips, data centers, power generation, networking gear. Right now that layer is riding a wave of demand that outstrips supply. But the boom rests on multi-year spending promises from a small number of buyers, with no independent confirmation that real end-customer revenue justifies the pace. Contra tags this as a context classification, not a bearish signal by itself — being an AI infrastructure supplier is a fact about a company's exposure, not a flaw. The tag exists so that other patterns firing on the same ticker can be read against the right backdrop.

The closest one is the telecom fiber buildout of 2000–02: enormous physical capacity built against projected demand, financed by a compressed group of spenders, ending in a sharp reversal when the spending pace broke before the end-customer revenue arrived. The parallel is not a prophecy — supply-demand conditions differ — but it defines the risk shape: companies whose future growth assumes the spending stays this hot are exposed to the pace, not just the direction, of the buildout.

Three levels. The weak tag applies when the company itself names AI-infrastructure demand as a main growth story. The medium tag adds a balance-sheet fact: forward capital spending exceeds 1× annual revenue — the company is spending more than it takes in to ride the wave. The strong tag adds concentration: more than 30% of forward revenue depends on a small group of AI-lab buyers. The ladder runs from "exposed by narrative" to "exposed by spending" to "exposed by customer concentration" — each step makes the company's fortunes harder to separate from the buildout's pace.

No. Context classifications carry no bearish weight on their own — they describe where a company sits in the AI stack so you can interpret other signals correctly. A supplier with the strong tag and no other firings is a company with concentrated exposure and a clean operational record. The same tag alongside capex-divergence or circular-capital-flow firings is a different picture. The framework's job is to keep the exposure visible so the narrative doesn't hide it; what you do with a clean-but-concentrated name is a judgment the tag alone can't make.

Three at once. They spend far more than they earn on the model business — training frontier models costs multiples of the revenue the models generate. The price of model access keeps falling as the capability becomes a commodity. And enterprise customers can switch providers easily, so retention is structurally fragile. Contra tags this layer as a taxonomy classification with zero scoring weight — it is a description of where a company sits in the AI stack, not a falsifiable risk on its own. The tag matters because the same economics read very differently for a focused lab versus a diversified group.

For a focused lab, the model business is the company — the spending gap, the pricing pressure, and the churn risk apply to its entire revenue base. For a diversified member — a commerce or cloud group that also trains a frontier model, like a Qwen inside a larger group — the tag still applies, but the model-economics pressure reads against that segment, not the whole company. Same mechanism, very different blast radius. That segment-versus-company distinction is the main reason the tag exists.

The weak tag applies when the company builds and serves its own large AI models as a core business line — primary revenue for a focused lab, or a material segment for a conglomerate member. The medium tag adds two facts: capital spending on the model business exceeds 1× the related revenue, and the company is signaling pricing pressure on model access. The strong tag adds a fragility marker: either customer retention is disclosed as declining, or more than 50% of computing capacity comes from a single named partner — a dependency that turns a commercial relationship into a structural one.

Yes — that is its function. A margin-compression or capex-divergence firing on a lab-tier company lands on a business whose core economics are already spending-heavy and pricing-pressured, so the composite reads harsher than the same firing on a diversified software company. Conversely, a strong operational record on a lab-tier name is more informative, because it is being achieved against structural headwinds. Context tags sharpen the read of everything else; they never fire alone. The Live Tape shows which tickers carry the tag today.

Investors apply a discount to the stock — and the discount sticks until the company proves the spending earns a return. The pattern fires when large optional capital spending flows into a business line that has lost money for several quarters straight, and the company discloses no figure connecting the spending to revenue. The market is not punishing ambition; it is pricing the absence of proof. The canonical case is Meta's Reality Labs, which has lost a cumulative $83.5 billion — roughly $4 billion a quarter as of early 2026 — while Meta raised its 2026 capital-spending plan to $125–145 billion with no revenue line showing a return on the AI investment.

Three things to check in the filings. First, is optional capital spending large relative to the business — 15% of revenue or more? Second, has the segment receiving the spending lost money for multiple consecutive quarters? Third — the decisive one — does management disclose any figure that proves the spending converts to revenue: a segment revenue line, a return metric, a backlog tied to the buildout? If the answer to the third is no, the spending is being justified by narrative rather than numbers. The longer that combination persists, the larger the discount the market applies.

The weak reading fires when optional capital spending reaches at least 15% of revenue and the business line has lost money for two to four quarters — early, still ambiguous. The medium reading adds two conditions: losses have run at least four quarters and there is no disclosed figure proving the spending earns money. The strong reading is the full deterioration: spending at 25% of revenue or more, losses running at least eight quarters, free cash flow cut at least in half year over year, and still no proof of return. Each step up is the same mechanism with less room for a benign explanation.

No — it means the market cannot verify the bet, so it discounts the stock until it can. Some large speculative programs eventually produce revenue and the discount lifts. The pattern is a flag on the current state of evidence, not a forecast of the project's outcome. What it tells a retail investor is that holding the stock means underwriting an unproven spending program, and the position should be sized with that in mind. Contra flags the condition; the decision stays with you.

Directly. The 2025–2026 AI data-center buildout is the largest optional-capex cycle in decades, and several of the biggest spenders disclose no revenue figure tied to the AI investment. That is exactly the configuration this pattern measures: heavy spending, sustained segment losses, no verification. The Live Tape shows which tickers are firing this pattern today.

It is when a chip supplier takes a large equity stake in its own biggest customer, and the money circles back: the supplier funds the customer, the customer buys the supplier's chips, and the purchase shows up as the supplier's revenue. It looks like demand, but part of it is the supplier financing its own sales. The textbook case is NVIDIA's roughly $100 billion stake in OpenAI — OpenAI uses the money to buy NVIDIA chips, while a small group of customers drives a large share of NVIDIA's sales. The dot-com era had a name for the same structure: vendor financing.

Two reasons. First, reported revenue overstates independent demand — some of it is the supplier's own capital coming home. Second, the structure stacks credit risk on a single relationship: the supplier is now exposed to the customer as an equity holder and as its revenue base at the same time. The risk is sharpest when a few customers already make up most of the supplier's revenue, because if the loop breaks the unwind is fast and concentrated — the investment loses value at the same moment the revenue disappears.

The weak reading requires a confirmed investment from the supplier into a customer that accounts for at least 5% of the supplier's revenue. The medium reading tightens all three dimensions: an equity stake, the customer at 10% or more of revenue, and industry buildout spending rising faster than revenue for at least two years — the loop plus a stretched backdrop. The strong reading is the acute case: a stake of $5 billion or more, the customer at 25% or more of revenue, and the buildout running at least three years. Bigger stake, tighter concentration, longer cycle — less margin for error at each step.

Contra never issues sell signals — the framework flags the structure and you decide. Circular flows can persist for years while the underlying spending cycle runs, which is why the pattern doesn't predict a date. What it tells you is the shape of the exposure: revenue that partly depends on the supplier's own capital, concentrated in one relationship. Historically, when spending cycles built on this structure turn, the reversal is compressed rather than gradual. The pattern is a position-sizing input, not a timing call. Free registration shows the current firings.

The gap has to close one way or the other — and both directions move stocks. In 2025, big-cloud capital spending grew about 60% while revenue grew 16.5%; 2026 plans imply spending up 80% against revenue up about 15.5%. A divergence that wide is binary: either returns on the spending show up within roughly six to eight quarters, or the spending pace itself forces a pullback. When hyperscalers pull back, the cut cascades through the entire AI supply chain — chips, networking, power, construction — which is why the pattern reads as a warning well beyond the spenders themselves.

Two ratios do most of the work. First, capital spending as a share of revenue: is it rising, and how fast year over year? Second, capital spending against free cash flow: once capex crosses 100% of free cash flow, the company is spending more than the business generates. Then check whether management discloses any return on the AI investment — segment revenue, utilization, contracted backlog. Rising spend ratio, shrinking free cash flow, and no disclosed return together mean the divergence is being financed on faith. Each big spender — Microsoft, Google, Amazon, Meta — can trip this on its own numbers.

The weak reading fires when the capex-to-revenue ratio rises at least 5 percentage points year over year, or capex crosses 100% of free cash flow. The medium reading requires a 10-point rise plus either free cash flow down at least 20% or a spending ramp with no disclosed AI return — Microsoft, Google, and Amazon fit this configuration. The strong reading is the extreme: a 20-point rise, free cash flow down at least 50%, and still no disclosed return. Meta fits, with spending guided up about 80% against revenue up about 15%. The ladder measures how much of the business is being converted into unverified buildout.

The framework's working window is roughly six to eight quarters: either returns become visible in the disclosed numbers within that span, or the spending pace becomes unsustainable and gets cut. That is not a prediction of which quarter — it is the horizon over which the binary resolves. For a retail investor the useful discipline is watching the resolution evidence, not the announcements: does a revenue or return figure appear, does free cash flow stabilize, does guidance hold? The pattern flags the tension; the quarterly numbers resolve it.

Because the hyperscalers' capex is the revenue of everyone downstream. Chip designers, networking vendors, power producers, data-center builders — their growth stories assume the spending pace holds. If the divergence resolves through a pullback rather than through returns, the reset propagates through the whole supply chain at once. The ETF X-Ray decomposes AI-themed funds at the holdings level, showing how much of a fund's exposure sits downstream of the same few spenders.

It is what happens when a supplier stakes a huge share of its future contracted revenue on a single AI-lab customer: its fate gets tied to that one customer's funding and survival. The textbook case is Oracle's $300 billion OpenAI contract — about 57% of Oracle's total contracted-but-not-yet-delivered revenue of $523 billion. The revenue is contracted, not collected: it arrives over years only if the customer can keep paying. When the supplier also borrows heavily to build capacity for that specific customer, the potential loss becomes larger and one-sided — the debt stays even if the customer does not.

Look at remaining performance obligations (RPO) — the contracted-but-undelivered revenue figure companies disclose — and ask what share of it traces to a single customer. Then check two amplifiers: is that customer profitable (an unprofitable customer's ability to pay depends on continuous external funding), and is the supplier's own debt rising to fund the buildout (leverage converts a customer problem into a balance-sheet problem)? Companies often name or strongly imply the customer in filings and earnings calls; the concentration math is usually one division away from disclosed numbers.

Weak (M1): a single AI lab at 30–40% of contracted future revenue, with the concentration present but nothing amplifying it. Medium (M2): a single lab at 40% or more, AND either that customer is unprofitable OR the supplier's debt is rising. Strong (M3): a single lab at 50% or more, AND the customer is unprofitable, AND debt is rising — the full concentration-risk case, as with Oracle. The three dials — concentration share, customer profitability, supplier leverage — measure how much of the downside is structural rather than hypothetical.

No — the framework is explicit that this is a risk-weighting flag, not a call to bet against the stock. A giant contract from a well-funded lab may be exactly the growth the market is paying for, and the customer may keep raising capital for years. What the pattern does is name the dependency so it gets priced consciously: a backlog that is 50%+ one unprofitable customer is a different asset than a diversified backlog of the same size, even though both print the same RPO headline. Users decide what discount that difference deserves.

Because the buildout's financing structure has concentrated enormous contracted sums onto a small number of AI labs, most of which are unprofitable and funded by continuous capital raises. Suppliers across the stack — cloud, data centers, chips, power — are signing multi-year commitments whose fulfillment assumes those labs keep raising money at increasing scale. That is a shared, correlated assumption: the framework tracks a related pattern for suppliers exposed to multiple AI labs, because when funding conditions tighten they tighten for the whole cohort at once. Free registration shows which tickers are firing the concentration patterns today.

It means the supplier's growth rides on a handful of accounts. A specialized supplier selling custom AI chips to a few giant cloud companies has no diversified demand base — if any single customer delays or shrinks its program, the supplier's entire AI growth line moves with it. Broadcom is the clearest live case: it designs custom AI accelerators for a small set of cloud buyers (Google, Meta, and one other), so a slip at any one of them shows up in Broadcom's numbers immediately. Concentration is invisible while everyone spends; it becomes the whole story the day one buyer pauses.

Read the customer-concentration disclosure in the 10-K — companies must disclose customers above 10% of revenue — and read how management describes the AI segment. The questions that matter: what share of total revenue is the AI segment, and what share of that segment do the top few customers represent? A supplier where the AI segment is a third of revenue and three customers are most of that segment has effectively three points of failure. Also look for evidence of new customers ramping; a broadening base dilutes the risk, a static one compounds it.

The weak reading fires when the AI segment is at least 30% of total revenue and concentration is high, but the top three customers are still under 75% of that segment — meaningful exposure, not yet acute. The medium reading fires when the top three reach 75% or more of the segment while the segment is at least 30% of revenue; one program slip would move the whole AI line. Broadcom fits here, with its top three at roughly 85% of the segment and AI around 31% of FY2025 revenue. The strong reading adds the top three at 85%+, the segment at 40%+ of revenue, and no sign of new customers over the next 24 months.

Not necessarily — concentration is a fragility measure, not a verdict on the business. A concentrated supplier with deep design lock-in can compound for years. What the pattern tells you is how the downside arrives if it arrives: not gradually, but in one step, when a single customer's program changes. That asymmetry matters for position sizing and for how much valuation premium you are willing to pay. The framework surfaces the exposure; whether the premium is worth it is your call. The Live Tape shows which suppliers are firing this pattern today.