LIVE EVENT
GCN Investor Conference in Newport Beach, CA
OCT 15 · NEWPORT BEACH, CA
LIVE EVENT
By 2025 the question about artificial intelligence had stopped being whether it works. It had become which layer of the stack captures the economics — and capital was being deployed across four incompatible answers at once.
By 2025, the debate about whether artificial intelligence would prove commercially significant had largely resolved. Adoption was measurable, revenue was real, and the technology was in production use across many categories.
The question that replaced it is harder and is the substance of this report: where in the value chain do the economics actually accrue?
This is a different kind of question, and it requires a different kind of underwriting. The first question is technology risk — will this work? It is assessed by technical judgement and by evidence of capability. The second is competitive-position risk — given that it works, who captures the value? It is assessed by analysing market structure, switching costs, bargaining power and defensibility.
Many investors who were well equipped for the first question are less well equipped for the second, and the two are frequently conflated. A conviction that the technology matters is not a thesis about which company captures its economics.
Four answers were live in 2025, and capital was deployed against all four simultaneously:
The uncomfortable observation is that these are partly claims on the same pool of economics. If the compute layer captures most of the value, less remains for models. If models remain differentiated and capture pricing power, applications built on them face a higher input cost. They cannot all be right at the valuations being paid, because they are dividing one surplus.
The shift from "does it work" to "who captures it" is worth dwelling on because it changes what evidence matters.
Technology risk is resolved by demonstration. A capability either works or does not, and the evidence is direct. It resolves relatively quickly and relatively unambiguously.
Competitive-position risk is resolved by market structure, and it takes years. The relevant questions are:
Two historical patterns are frequently invoked here and they point in opposite directions, which is why the analogy debate is unproductive:
Both analogies are argued confidently. The honest position is that the analogy does not determine the answer — the outcome depends on specifics of this technology's economics that were still resolving in 2025, particularly whether frontier capability requires continuously escalating capital or plateaus.
The case: models require enormous computation to train and to serve. Whoever supplies it captures revenue regardless of which model or application wins — a picks-and-shovels position.
The evidence for: revenue at this layer was the most visible and the most measurable in 2025. It appears in the financial statements of listed companies rather than being inferred.
The risks: the capital requirement is enormous and front-loaded, as the 2024 and 2026 global reports describe. Depreciation arrives on a schedule regardless of utilisation. High returns attract competitive entry, from new suppliers and from customers building their own. And the demand is derived — it depends on the layers above generating enough revenue to justify continued purchasing.
Where evidence appears: quarterly filings — capex, depreciation, useful-life assumptions, utilisation commentary. This is the most transparent layer and the first to produce hard evidence, which is why the 2026 reports point there.
The case: the model is the core technology. If frontier capability remains meaningfully ahead of alternatives, the organisations producing it capture pricing power.
The evidence for: frontier models did command premium pricing, and the capital and talent required to produce them is a genuine barrier.
The risks: this thesis depends entirely on capability remaining differentiated. Evidence on both sides was live in 2025 — capable open-weight models narrowed the gap for many uses, while frontier capability continued advancing. If capability converges, models become a commodity input and pricing power moves to whoever controls distribution or data instead.
Where evidence appears: pricing behaviour, and whether customers switch providers on price. Falling prices per unit of capability, with customers switching readily, is convergence.
The case: value accrues where a specific problem is solved for a specific customer. Applications with workflow integration, proprietary data and distribution can build defensibility, and pay a declining input cost as models commoditise.
The evidence for: this is where historical patterns most often placed the durable value, and where switching costs are most constructible.
The risks: many applications in 2025 were thin layers over a model, with limited defensibility. The distinction between a genuine application business and a wrapper was not always visible in early revenue growth — both grow fast initially.
Where evidence appears: net revenue retention and gross margin, both disclosed by listed comparables. High retention indicates the product is embedded; high and stable gross margin indicates the input cost is not eating the value. These take several periods to become meaningful, which is why this layer's evidence arrives last.
The case: an existing business with customers, distribution and proprietary data can embed the technology and capture the benefit without any platform economics. The customer does not switch; the product simply gets better.
The evidence for: distribution is the hardest thing to build, and incumbents have it. Many enterprise buyers preferred to receive AI capability through existing vendors rather than adopt new ones.
The risks: incumbents capture the benefit as improved margin or retention rather than as new revenue, which makes it hard to observe and hard to value separately. And organisational capacity to adopt varies enormously — the technology being available does not mean an incumbent can deploy it.
Where evidence appears: margin improvement and retention at incumbents, which is genuinely difficult to attribute.
The arithmetic point is worth making directly, because it is the report's central claim.
The total value created by a technology is finite in any period. It is divided among the parties in the chain according to their bargaining power. If the compute layer captures a large share, less remains for models. If models capture pricing power, applications face a higher input cost and thinner margins.
Valuations across all four layers in 2025 implied substantial value capture at each. Summed, the implied capture exceeded plausible estimates of the total value being created — not by a small margin.
Two things can be true simultaneously and both are:
Four theses about who captures a surplus cannot all be underwritten at full value. Owning all four is a rational hedge; pricing all four as winners is not the same thing.
The practical implication for an allocator is that exposure across the chain should be sized as a hedge against uncertainty rather than as four independent high-conviction positions. The distinction sounds semantic and is not: it determines whether the position is sized to the probability of being right or to the consequence of being wrong.
Rather than assert an answer, it is more useful to specify the evidence — and to note that it arrives in a particular order.
Compute layer evidence arrives first, quarterly, in filings: capex, depreciation, useful-life assumptions, utilisation. Extending useful-life assumptions is a signal worth watching, because it raises reported earnings without changing economics and it is disclosed.
Model layer evidence arrives next, in pricing: price per unit of capability, and whether customers switch on price. Rapid price declines with easy switching indicates convergence and a commodity outcome.
Application layer evidence arrives last, in retention and margin, and requires several periods to become meaningful.
Incumbent evidence is the hardest to isolate, appearing as margin and retention improvement that is difficult to attribute to any single cause.
An investor wanting to update early should watch the filings rather than the funding announcements. Funding rounds tell you what investors believe. Filings tell you what customers paid.
The value-chain question reduces, ultimately, to bargaining power. Setting out where it sits at each interface clarifies the four theses better than assessing them individually, because bargaining power is what determines how a surplus is divided.
Between compute suppliers and model developers. Power sits with whoever is scarcer. While advanced compute is capacity-constrained, the supplier has it — a model developer who cannot obtain capacity has no alternative at any price. If capacity becomes abundant, the power inverts, because a supplier with idle capacity competes on price. This is the single most consequential interface and it turns on a capacity cycle rather than on anything about the technology.
Between model developers and applications. Power sits with the model developer while capability is differentiated and switching is costly. It moves to the application if models converge — an application that can substitute one model for another is buying a commodity and will pay commodity prices. The observable evidence is whether customers switch on price, which is behaviour rather than opinion.
Between applications and their customers. Power sits with the application to the extent it is embedded in workflow, holds proprietary data, or owns the customer relationship. It sits with the customer where the product is easily replaced. Net revenue retention is the direct measure.
Between incumbents and new applications. Power sits with whoever owns distribution. An incumbent with an existing customer relationship can add a capability and retain the customer without the customer ever evaluating an alternative. Distribution is the hardest thing to build and the most durable thing to own, which is the strongest form of the incumbent thesis.
The chain's implication: if compute stays scarce, the compute layer captures disproportionately and everything downstream is squeezed. If compute becomes abundant and models converge, value moves toward distribution — which favours applications with embedded positions and incumbents with existing customers. These are the two coherent end states, and the four theses map onto them rather than being independent.
The four theses are not four independent bets. They are two possible resolutions of the same bargaining question, and an investor who understands which resolution they are underwriting is holding a clearer position than one who owns all four.
Watch capacity, not enthusiasm, at the compute layer. Pricing power at this layer depends on scarcity. Equipment billings, announced capacity additions and utilisation commentary are the relevant series, and SEMI publishes equipment billings free and monthly.
Track model pricing per unit of capability over time. Provider pricing pages are primary and free, and historical snapshots are available via web archives. Rapid price decline with easy customer switching is the signature of commoditisation, and it is observable well before it appears in anyone's revenue.
Use net revenue retention and gross margin at the application layer. Retention distinguishes an embedded product from a trialled one. Gross margin shows whether the input cost is consuming the value. Both are disclosed by listed comparables and both take several periods to become meaningful, which is why this layer's evidence arrives last.
Watch useful-life assumptions in infrastructure filings. Extending assumed asset life reduces annual depreciation and raises reported earnings without any change in economics. It is disclosed, and it is one of the clearer signals of pressure on a return case.
Size chain-spanning exposure as a hedge, not as four convictions. Owning across the chain is a reasonable response to genuine uncertainty. Pricing all four layers as winners is not the same thing, and the distinction determines whether positions are sized to the probability of being right or to the consequence of being wrong.
Watch the filings, not the funding rounds. Funding announcements tell you what investors believe. Filings tell you what customers paid. Only the second updates a view.
A structural retrospective on AI investment in 2025, focused on the value-capture question and on what evidence would resolve it.
Where figures appear they carry a numbered source. Mechanisms — technology risk versus competitive-position risk, surplus division across a value chain, the order in which evidence appears by layer — are analysis with reasoning shown.
The report deliberately declines to advocate a layer. Its purpose is to specify the question precisely and to identify where the evidence will appear, which is more durable than a position that will date within a year.
Private Credit Report 2024 — Systemic Scale, Untested Claims precedes this report in the asset class sequence.
Private Credit Outlook 2026 — The Cycle Arrives follows this report in the asset class sequence.
Global Investment Outlook 2025 — The Bifurcated Market covers the same year at global multi-asset level.
US Venture Capital Report 2025 — Concentration covers the same year in North American private markets.
Accredited investors receive our market reports, private event invitations and curated deal flow.
.webp){kind=logo}
.png)
.png)