Is AI a bubble? The honest answer is that the question is framed at the wrong level. An arXiv preprint titled "Boom, Bubble, or Buildout? A Multi-Method Evaluation of Whether Artificial Intelligence Is in an Ongoing Financial Bubble" concludes that AI is best understood as a real technological revolution with localized bubble dynamics rather than as either a pure speculative mania or a bubble-free productivity miracle. Because it is a preprint, and because the passage this article relies on supports only that headline conclusion, treat it as a structured argument to evaluate rather than settled, peer-reviewed evidence; the authors, date and publication status are not established by the material available here. Fidelity's "Is AI a bubble? 5 signs to watch for" reaches a compatible market-level view: AI has helped drive valuations for both the S&P 500 and technology stocks within this index above the historical average, even though each remains below the extremes of the late 1990s dot-com bubble. So the useful question for an investor is narrower: in this specific AI startup, is the price supported by durable cash flows, or by narrative? This article gives you a company-level screen for that, built from sourced figures on AI startup valuation, the dot-com parallel, and a complete worked screening example you can reuse.
Why the Dot-Com Parallel Is Useful and Misleading
The dot-com era is the right reference class and the wrong verdict. The arXiv preprint's central conclusion separates two questions that usually get conflated in one breath: is the technology real, and is this company's price supported? Its answer is that both can be true at once: a real technological revolution can coexist with localized bubble dynamics. Our reading of that framing is that technology success and overvaluation are independent tests. Being right about AI tells you nothing about whether a given price is justified, which is exactly the confusion the dot-com reference class is meant to correct.
Fidelity's market-level observation sets the boundary of the comparison: AI has helped drive valuations for both the S&P 500 and technology stocks within this index above the historical average, even though each remains below the extremes of the late 1990s dot-com bubble. Above average is not the same as extreme, and neither phrase is a company-level finding. The useful discipline the parallel supplies is a question, not a conclusion: when a price rests on a narrative about a transformative technology, what cash flow is it resting on today?
The scale comparison cuts both ways, and it is weaker evidence than it looks. Wikipedia's "AI bubble" article reports AI investment estimated at 3.6% of gross domestic product, compared to 2.2% for the railroad projects of the 1800s, 1.1% for the interstate highway system of the 1950s, and 1.1% for the Internet build-out of 1996 through 2003. The excerpt does not specify the country, the measurement periods, or the basis on which each figure was calculated, so treat these as directional scale context rather than a like-for-like comparison. Bigger does not mean bubblier, and smaller does not mean safe. A buildout can be both large and productive, or large and overpriced, which is why the judgment has to move down to the level of one company and one price.
Revenue Quality: Which AI Revenue Actually Counts
Revenue quality is the first filter, because AI revenue can be real, contracted, or recycled. The excerpt available for this article, from The Algorithmic Bridge's "You Have No Idea How Close the AI Bubble Is to Popping," reports $16+ billion in revenue in a single quarter and 46% quarter-over-quarter growth. The passage does not identify the company, the measurement period, or whether the figure describes observed results or a scenario. That gap is itself the lesson to carry into a private-company screen. A headline revenue number carries limited weight until its basis is stated: the period it covers, the type of revenue it counts, and who is paying it.
For any AI startup, apply three questions. Is the revenue recurring, or does it arrive once? Is it from unrelated outside buyers, or from investors, suppliers and partners who are also customers? Is the growth durable, or a single step-change that will not repeat? Translate those into document checks: separate paying contracts from paid pilots, unpaid pilots, letters of intent and expressions of interest; separate product subscriptions from service work billed as product; and list every customer that holds equity in or has a commercial relationship with the company.
One caution on the second question. A strategic investor can also be a genuine paying customer on market terms, and that revenue can be entirely real. The relationship is a reason to pull the contract and check the terms, not proof of circular financing. The screening judgment is narrower than a verdict: revenue you cannot trace to an independent buyer paying on market terms is weaker evidence than revenue you can, and the difference should be visible in how much weight the growth chart carries.
Gross Margins After Compute Costs
Gross margin after inference cost is the line where a model-powered product keeps software-like economics or loses them. Software economics assumed that the next unit cost almost nothing. For model-powered products, the next unit carries a token bill. The practical test is not the headline gross margin in the deck. It is the gross margin at the usage level the forecast assumes, after hosting, model API charges, and the human review the product still needs to deliver reliably.
What moves that margin as the company scales is a set of variables, not one law. Unit inference cost falling is one lever. Others are the price charged per unit, usage per customer, the mix between fixed hosting commitments and variable token spend, and how fully committed capacity is utilized. A company can improve margin with flat unit costs if pricing holds and utilization rises, or lose margin with falling unit costs if discounting outruns them. Test each variable against the forecast instead of assuming scale does the work.
Two reporting distortions to correct for, stated specifically. First, promotional pricing and free tiers can make usage-based growth look strong while recognized revenue per unit sits below the list price the forecast assumes, so reconcile usage growth to recognized revenue before calling it demand. Second, separate free tiers from heavily discounted paid tiers, because they do different things to the accounts. Free tiers add delivery cost without adding recognized revenue, so today's reported margin may understate the economics at full adoption. Heavily discounted paid tiers do generate recognized revenue, but at a price below the one the forecast assumes, so both revenue per unit and margin per unit differ from list-price economics, and the forecast's margin may also assume discounts that later expire, with the affected customers then leaving. Neither distortion is automatic. Both are questions: what price did each cohort actually pay, and what happens to margin when the discount ends?
Customer Concentration and Circular Demand
Concentration is the risk that a strong growth rate turns out to be three relationships. Ask what share of revenue sits with the top customer, the top three, and any customer that is also an investor or strategic partner, and ask for the trend rather than the snapshot. A falling concentration ratio is a different fact from a static one, even at the same level. It can reflect new customers being added, or it can reflect revenue declining at the largest accounts. Customer counts and cohort detail are what separate those two readings, and the ratio alone does not.
Circular demand is a hypothesis to test, not a label to apply. Trace where each dollar of revenue originated before it reached the company: if a customer's own spending is funded by the same small group of AI firms, the demand signal is weaker than the growth chart implies; if that spending is funded independently of those firms, the concern is answered. Two related questions belong in the same check. Did any large order get pulled forward, doubled up or deferred around a shortage or a budget cycle, which would make one quarter's revenue reverse later? And do the largest contracts renew repeatedly on comparable terms, or has each one been signed only once?
Present these as screening judgments rather than universal rules. Concentration alone does not make a company uninvestable, and a related-party customer does not make revenue fake. What the screen establishes is how much of the growth story rests on relationships you have not yet read the contracts for.
Dependence on Model Providers
Model-provider dependence shows up in three places: cost, roadmap and distribution. On cost, a price cut or a new pricing tier from a frontier lab changes your gross margin overnight without any change in your product. On roadmap, a provider feature release can absorb the workflow you sell. On distribution, if the provider also sells a competing assistant into your customer base, your switching costs are one of the protections standing between you and substitution, alongside contracts, data portability and relationships the provider does not own.
The same issue appears on the buyer's side of the same contract. When a recurring workflow runs through one provider's model, an update can change how the process behaves with no change on your side. A customer evaluating the startup is therefore asking the startup the question the startup should be asking of its own provider: which steps in this process do we still control? That symmetry is why provider dependence is a diligence topic for both parties, and it applies equally to a startup whose product is built on someone else's model.
When you assess a target, ask four questions. What does the provider contract say, including the notice period for pricing changes? What share of gross margin is exposed to a 2x token price increase, computed from the actual cost breakdown? What remains in the product if the underlying model is replaced? And what remains if the provider ships your core feature? An inability to answer is an evidence gap to close before a price is discussed. It is not, by itself, proof that the company lacks defensibility.
Defensibility: What Survives a Correction
Defensibility in AI is a screening question with a short list of candidate answers. For application-layer products built on third-party models, the model itself is usually not the moat, because the provider can ship it to everyone. The candidates that tend to survive are proprietary data that improves with use, workflow integration that creates switching costs, distribution the provider cannot easily replicate, and domain compliance and reliability that a general assistant does not carry. Treat that list as our screening framework, not a sourced taxonomy, and require evidence for whichever claim a company makes.
Eqvista's "Top 100 AI Startups by Valuation (2026)" reaches a compatible conclusion about where value concentrates, noting that the winners of tomorrow won't be the companies with the biggest models; they'll be those that solve real problems with sustainable business models. The practical translation for a single investment: you can be right about the technology and wrong about the company. That asymmetry is the reason this screen treats a view on the technology and a view on one company's price as separate judgments, and the reason the dot-com parallel is used here as a question rather than a verdict. It is a way of reasoning about any current name, not a prediction about one.
The defensibility test in this screen is therefore behavioral, not narrative. Ask what customers actually do, and then what they would lose if they moved: audit trails wired into their compliance reporting, integrations built into their systems, data that gets richer with their usage. A capability described in the deck is a claim. Observed usage is a starting point, and the next question is what breaks or must be rebuilt on the customer side if they switch. Where neither the claim nor the behavioral loss exists, record it as an open question rather than a pass.
AI Startup Valuation: What the Multiple Assumes
AI startup valuation in 2026 is being discussed at levels that make the multiple the whole argument, and a multiple is an assumption, not a fact. Aventis Advisors' "AI Valuation Multiples in 2026" reports that the median revenue multiple for AI companies stood at 24.2x. The excerpt available for this article does not specify the revenue period, the sample composition, or the observation date behind that median, so it cannot be compared directly with stage-specific pre-money valuations, which are measured on a different basis. Use it as a headline reference point and check the underlying table before you rely on it.
The arithmetic is still worth doing, because it shows what the number assumes. At 24.2x, a company with $10M of revenue implies about $242M of value. This article's screen treats that price as resting on three things: revenue that recurs from unrelated buyers, margins that survive a compute-cost shock, and customers who are not inside a financing loop. Those are the criteria chosen for this screen, not universal requirements for a justified price. If any of the three is unverified, the multiple is pricing an assumption you have not tested. The screen in the next section exists to make those assumptions explicit before the price is discussed, not after.
A Worked Example: Screening One AI Startup
The following is a hypothetical example with invented figures, to show the arithmetic and the reasoning. It is not a real company and not a valuation. The rules are written down first, so the conclusion cannot drift toward whatever the growth chart suggests.
Hypothetical screening rules, set in advance: R1 revenue quality, at least 90% of trailing revenue from recurring contracts with buyers holding no equity in the company. R2 margin resilience, gross margin after compute and delivery-related support stays at or above 40% under a 2x model-API price shock with no customer price change. R3 concentration, top three customers at or below 40% of revenue, with any related-party contract confirmed at market terms. R4 provider dependence, a written notice period on pricing changes plus a documented answer to what remains if the provider ships the core feature. R5 defensibility, at least one switching-cost asset evidenced in customer behavior rather than in the pitch narrative.
Hypothetical inputs: an AI application startup reports $12.0M of revenue for the trailing twelve months, which its deck labels annual recurring revenue. The label does not survive the company's own split: $10.4M of the $12.0M is recurring subscriptions and $1.6M is paid pilots and one-off implementation. A year earlier the trailing-twelve-month figure was $7.0M, split $5.6M recurring and $1.4M other. Three different measures are now on the table and they are not interchangeable: the reported ARR claim, the recurring-revenue share, and recognized revenue for a stated period. R1 and the margin calculation both run on recognized trailing-period revenue, so that is the basis used below. Cost of delivery over the same trailing twelve months breaks down as model API and inference $3.2M, hosting and infrastructure $1.6M, and delivery-related support and human review $0.9M. Sales and marketing of $1.5M is a period cost and sits outside gross margin. If the revenue and cost lines were drawn from different periods, no margin below would be meaningful, so period alignment is the first thing to confirm in a real screen. The top customer represents $3.3M and the top three represent $6.6M, against $5.4M of $7.0M a year earlier; one of the top three is a strategic investor paying $1.1M. The product calls a single frontier model provider through one API, and the document pack contains no pricing terms at all: neither price protection nor a notice period for pricing changes appears in what was supplied. On behavior, 6 of the 10 largest customers export audit logs weekly; no proprietary data asset is claimed to improve with use.
Step 1, gross margin after compute: on the common trailing-twelve-month basis, $12.0M minus the $4.8M API-and-hosting subtotal minus $0.9M of delivery support equals $6.3M, a 52.5% gross margin. Now apply the token shock to the API line only, since hosting does not double with token prices. The API line goes from $3.2M to $6.4M, which puts the shocked API-plus-hosting subtotal at $8.0M. Delivery-related support and human review are also direct delivery costs, so total direct costs are $8.0M plus $0.9M, or $8.9M. Gross margin falls to ($12.0M − $8.9M) ÷ $12.0M = 25.8%, assuming usage and pricing remain unchanged. That is below the 40% tolerance, so R2 fails. The sensitivity is conditional: if the company passes a price increase through to customers, the margin recovers and the risk moves to retention, which is a separate question to test. Closing the gap without a price move requires cutting about $1.7M in direct costs at the stated revenue, because 40% of $12.0M leaves $7.2M of allowable direct cost against the shocked $8.9M.
Step 2, revenue quality: recurring subscriptions are $10.4M of $12.0M, or 86.7%, up from 80% a year earlier and below the 90% rule. R1 asks for revenue that is both recurring and from buyers holding no equity. That intersection cannot be larger than the recurring share, so it cannot reach 90% on these figures even if every recurring dollar came from an independent buyer. R1 therefore fails on the recurring share alone. What the inputs do not contain is the customer-level overlap that would size the intersection: which recurring contracts belong to equity-linked buyers. That mapping is the missing evidence, and it changes how the failure is read rather than whether it stands.
Step 3, concentration: top-three concentration is $6.6M of $12.0M, or 55%, down from 77% a year earlier. Concentration is falling, but a falling ratio can reflect new customers added or revenue declining at the largest accounts, and the inputs do not say which. Either way, 55% is above the 40% rule, so R3 fails on level. The related-party element is separate and unverified rather than failed: the strategic investor accounts for $1.1M, or 9.2% of revenue. Subtracting it gives $10.9M, or 90.8%, but that figure is only revenue from buyers with no equity link if the identified investor is the only equity-linked customer, which the inputs do not establish. Treat $10.9M as an upper bound, not a measured figure. This is not a finding of circular financing. It is a contract to read: confirm the $1.1M was paid on market terms, and identify whether any other customer holds equity.
Step 4, provider dependence: the document pack contains no pricing terms, so the notice-period element of R4 is unverified rather than failed. Absence of price protection and absence of a notice period are different facts, and a pack containing neither clause establishes neither. On the feature question, the company answers that customers would keep audit trails and integrations if the provider shipped the core workflow, but offers no customer evidence for it, so that element also stays open. R4 is unresolved on both elements, and the resolution is documentary: the provider contract and a customer check.
Step 5, defensibility: weekly audit-log export by 6 of the 10 largest customers establishes that the feature is used. It does not establish a switching cost, because the inputs do not say what a customer would lose or have to rebuild if they moved. R5 requires at least one switching-cost asset evidenced in customer behavior, so R5 is unverified rather than passed. To move it, the screen needs the specific loss: audit trails wired into compliance reporting that would have to be recreated, integrations that would need re-implementation, or data that loses value outside the product. There is no claimed data asset that improves with use, and no distribution or compliance claim is evidenced.
Step 6, the screening outcome: R1 fails on the recurring share, R2 fails at 25.8% under the shock, and R3 fails on concentration level. R4 is unverified because the supplied document pack contains no pricing terms, so the notice-period element cannot be checked from what was supplied; the provider contract itself may exist outside the pack, and its absence from the pack is not evidence that no contract exists. R5 is unverified because usage was observed but switching cost was not demonstrated. The outcome is hold. Do not take this company to a price discussion on the current evidence, and note that the 24.2x reference applied to $12.0M of revenue implies about $290M, a figure whose basis is unspecified in the source and which this screen does not support.
Two different lists follow, and mixing them is the common error. Evidence requests that would clarify the case: the API cost breakdown by quarter and the provider's pricing and notice terms; the top-three contracts including the strategic investor's terms; a customer-level map of recurring contracts against equity holdings; and one customer reference on what audit-log export replaces in their workflow. Operational changes needed to satisfy the failed thresholds: recurring revenue from 86.7% to at least 90% of the trailing period, which at current revenue means at least $10.8M that is both recurring and independent against $10.4M of recurring revenue today; top-three concentration from 55% to 40%, which at current revenue means $6.6M falling to $4.8M or the denominator growing; and gross margin under the shock from 25.8% to 40%, which has two routes with different assumptions. The cost route holds revenue fixed at $12.0M: 40% of $12.0M leaves $7.2M of allowable direct cost against the shocked $8.9M, so about $1.7M of direct cost has to come out. The price route holds the shocked cost base at $8.9M and changes the denominator instead: revenue has to reach about $14.8M, since ($14.8M − $8.9M) ÷ $14.8M is approximately 40%, an increase of about $2.8M or roughly 24% at unchanged usage and unchanged cost, which moves the question to retention rather than to the cost line. The two routes are not interchangeable, and neither is achieved by sending documents. This is a screening judgment about evidence, not a verdict on the company, and it is deliberately not a valuation.
Using DDScore Where It Fits Your Assessment
A structured first pass helps most at the point where narrative and evidence diverge. DDScore turns private-company materials such as a pitch deck, financial notes and use-of-funds material into a structured 0 to 100 Due Diligence Score and a written report, covering twelve fixed dimensions including Market, Competitors, Technology & IP, Scalability, Financials and Fundability. Each dimension carries its own score and analysis-confidence figure, and the report separates strengths, areas for development, risks and evidence limitations, with follow-up questions tied to the gaps that matter. It cross-checks submitted materials against each other and against current public information, and a report generally takes 10 to 30 minutes.
For an investor, the practical use is to run it on one AI company you are already considering, then read the Financials, Competitors and Technology & IP sections against the five tests in this article: revenue quality, margin after compute, concentration, provider dependence and defensibility. That comparison is your task; the report supplies the findings and the reasoning, and you decide which of the five tests each finding speaks to. The report works from the materials submitted and from current public information, so a question that depends on a document outside the pack may stay open in the report, the way R1's customer-level overlap between recurring contracts and equity holdings, R4's pricing notice terms, and R5's evidence of what a customer loses on switching stayed open in the worked example. The report can record evidence limitations and follow-up questions tied to what it finds, and where the available evidence is thin it may lower the analysis-confidence figure for a dimension. It does not guarantee to identify every missing contract-level fact, and a confidence reduction in each case should not be assumed. Where a gap remains, closing it is a separate step in your own process. For a founder, running it on your own materials before an investor conversation shows which claims are supported and which are asserted, which is the same screen run from the other side.
The boundary is firm. DDScore is an analytical first-pass tool that supports human judgment and due diligence. It does not provide investment advice, choose investments or replace full diligence, and a score is not a prediction that any company will succeed or fail. The decision, and the verification of anything it depends on, stays with you.
Explore DDScore for Founders · Explore DDScore for Investors