In May 2026, Anthropic closed a US$65 billion Series H funding round at a post money valuation of US$965 billion - almost 2.5 times the US$380 billion valuation it commanded just three months earlier. On 1 June 2026, the company confidentially filed for a US Initial Public Offering (IPO). OpenAI followed a week later, on 8 June 2026, with its own confidential filing with market commentators suggesting a potential IPO valuation approaching US$1 trillion. And on 12 June 2026, SpaceX (having acquired xAI earlier in the year) listed on the NASDAQ at US$135 per share implying a valuation of US$1.77 trillion and raising US$75 billion in what is now the largest IPO in history. Meanwhile, at the other end of the spectrum, the Paris-based start-up AMI Labs founded by former Meta chief AI scientist Yann LeCun raised more than US$1 billion for his new start-up at a US$3.5 billion pre-money valuation - with 12 employees, no released product, and a candid acknowledgement that real-world applications were at least a year away.
These are not stories about irrational markets. They are stories about risk: how it is priced, the assumptions that underpin the numbers and how dramatically reasonable people can disagree about its magnitude. This matters because AI company valuations are increasingly arising as issues in dispute - whether as the subject of disgorgement remedies, the basis for damages calculations in intellectual property and competition disputes, or as part of securities and shareholder claims - and the range of defensible expert conclusions on any given AI business can be startlingly wide.
Valuation methodology requires experts to grapple with two fundamental uncertainties: the reliability of forecasts about a business’s future performance, and the risk that those forecasts will not materialise. The most widely used tool for addressing this is the risk-adjusted discount rate. Under this approach, a business’s expected future cash flows are discounted to present value at a rate that reflects both the time value of money and the specific risks of the investment. The riskier the business, the higher the discount rate applied and, consequently, the lower the present value assigned to its future cash flows.
Where the discount rate alone does not fully capture every dimension of risk, valuers will make further adjustments. These can take the form of refinements to the projected cash flows based on defensible commercial assumptions, using probability-weighted scenarios where the range of outcomes is wide and discrete enough to model meaningfully, or by applying post-valuation discounts to reflect risks that are difficult to quantify within the discount rate itself (such as the illiquidity of a private company’s shares). Conversely, premiums may be applied when the valuation does not fully reflect identifiable upside, such as synergies in an acquisition.
Whichever approach is taken, the objective is to estimate cash flows that reflect the expected outcome of the business - an unbiased, best estimate of what is anticipated. The principal danger in layering multiple adjustments is that the same risk can inadvertently be counted more than once, and a rigorous valuation will therefore be explicit about where each risk is reflected and ensure that no single risk factor is double counted across the analysis.
Before examining the specific risks facing AI companies, it is worth understanding how the sector is structured, because where a business sits in the AI value chain can have a direct bearing on its economics and where its main source of risk is coming from.
The AI sector can be broken down into a series of interconnected layers, each with distinct economics and competitive dynamics. At the base sit the providers of physical infrastructure: chip designers and manufacturers (NVIDIA, AMD, Arm), data centre operators, and the providers of power and cooling. Above them are the cloud and compute layer - the so-called hyperscalers (Microsoft, Google, Amazon) that aggregate infrastructure capacity into scalable services. The foundation model layer (OpenAI, Anthropic, Google DeepMind, Meta AI) sits above this, developing the large-scale models that underpin most AI applications that are increasingly becoming household names. Above that sit AI platforms and tooling providers (Hugging Face, Cohere, Scale AI), and finally, enterprise applications: sector-specific products built on top of foundation models.
A business operating primarily at the foundation model layer faces very different risks from one building an enterprise application on top of an existing model. The former requires enormous upfront capital, competes in a ‘winner-takes-most’ market, and generates value primarily through strategic positioning and future optionality. The latter may generate more predictable near-term revenues but faces the risk that its underlying model provider changes pricing, restricts access, or releases a competing product.
A further interesting feature from a risk perspective is how the relationship between the hyperscalers (and major cloud providers) and the AI foundation developers has developed and become financially intertwined. The clearest example of this is the relationship between Microsoft and OpenAI.
How it works is as follows:
Three features of this arrangement are illustrative of broader patterns emerging across the sector.
First, a large portion of Microsoft’s economic support for OpenAI has reportedly taken the form of compute infrastructure, cloud credits, hosting arrangements, and long-term Azure capacity commitments rather than cash. To the extent that OpenAI utilises that infrastructure capacity, the associated consumption would also contribute to Microsoft’s Azure revenue.
Second, OpenAI’s ongoing demand for compute capacity supports the growth metrics of Microsoft’s AI and cloud businesses, reinforcing the strategic and financial interdependence between the parties.
Third, the implied value of Microsoft’s equity interest in OpenAI is influenced by subsequent private funding round valuations, potentially generating substantial unrealised gains even in the absence of a liquidity event.
Similar dynamics have emerged elsewhere in the sector. In April 2026, Amazon expanded its Anthropic investment commitment while Anthropic committed to spending more than US$100 billion on Amazon’s AWS infrastructure over the following decade. Amazon also disclosed substantial gains associated with the revaluation of its Anthropic investment.
Such arrangements raise an important question for the expert valuer: to what extent are the resulting revenue, margin and balance-sheet metrics representative of arm’s-length, standalone market conditions, and how would the economics of either business appear absent the strategic relationship?
A further dimension to these significant AI investments is that companies such as Amazon, Alphabet, Microsoft and Meta, which have historically operated as relatively asset-light, cash generators are now some of the world’s biggest investors in physical infrastructure, with forecast spend reaching approximately US$725 billion in 2026 alone. The combined free cash flow of these companies is at its lowest level in more than a decade. What’s more, these companies are increasing their debt and reducing share buy-backs to help fund these investments which in turn could impact the risk profile of these companies.
While the above risks should be considered carefully in any valuation, the main risk when valuing AI companies is in forecasting the future. So what risks should we be conscious of?
Those of us who have valued technology companies are all too familiar with the risk of obsolescence and commoditisation, but even the most experienced of us have to acknowledge that the pace of development in AI is without precedent in recent technology history. A model that represented the frontier six months ago may be materially less competitive today. Open-source model releases - freely available, increasingly capable, and rapidly improving - periodically reset assumptions about the commercial value of proprietary model development.
Pricing competition is reinforcing this dynamic. Chinese model developers including MiniMax, DeepSeek and Moonshot are pricing their models at roughly US$2 to US$3 per million output tokens, compared with around US$15 for comparable models from leading US developers - a near-sixfold gap. This price differential is particularly significant in the era of AI agents, which consume orders of magnitude more tokens than chatbot interactions.
For the expert, this creates a specific challenge: the long-run cash flow projections that underpin a DCF analysis, and the competitive moats assumed in a market-based multiple, may both be materially affected by technological and pricing developments that are genuinely difficult to predict. Experts should consider whether terminal value assumptions built on the premise of sustained competitive advantage are more optimistic than a scenario in which open-source alternatives or lower-cost providers commoditise the model layer.
The beta in the discount rate will capture some element of systematic technology risk - the sector-wide volatility that all comparable companies share. But it does not adequately capture company-specific commoditisation risk or the possibility that a particular business is rendered commercially obsolete by an event that does not affect the sector as a whole. An expert should consider whether company-specific commoditisation risk is best reflected in the cash flow forecasts or through probability-weighted scenarios rather than in the discount rate, particularly because the limited pool of pure-play AI comparables makes the observed sector beta itself a less reliable measure than it would be in a more mature industry.
Reported financial metrics in the AI sector require careful analysis before being used as the basis for valuation. Three issues require particular attention: the composition of reported revenue, the distinction between speculative and demand-driven capital deployment, and the gap between AI deployment and AI maturity.
Annual Recurring Revenue (ARR) is frequently cited as the primary valuation metric for AI businesses. The label, however, conceals considerable variation. Revenue classified as ‘recurring’ may include pilot contracts with no committed renewal, milestone-based payments, or revenue generated through strategic relationships with investors or infrastructure providers rather than arm’s-length third-party customers.
Howard Marks of Oaktree Capital Management has drawn an important distinction in this context between training capex - the speculative investment undertaken to build AI models and infrastructure capacity - and inference capex, which is deployed in response to actual demand for AI capacity from end users. The latter is more directly supported by genuine commercial revenue; the former remains a bet on future demand. As Marks observed, some AI revenue is currently ‘circular’ in nature, derived from AI companies buying from each other, and the chain of revenue must ultimately rest on end users paying for real economic value.
Equally relevant is the gap between AI adoption and AI maturity in the customer base. A 2025 McKinsey survey found that while 88% of firms reported using AI in at least one business function, only 1% considered their AI strategies mature. For valuers, this suggests that headline adoption metrics may be a poor proxy for durable, scaled enterprise revenue. The customer concentration risk inherent in many AI businesses - where a small number of large enterprise customers may account for a disproportionate share of revenue - compounds this concern.
Training and running large AI models requires substantial and ongoing access to specialised computing hardware, primarily high-performance GPUs. The cost of this compute is a major determinant of AI company economics - and its trajectory is far from certain.
As discussed above, some AI businesses benefit from preferential compute access through strategic arrangements with hyperscalers.The valuation question is whether such arrangements are sustainable on a standalone basis, and what the company’s economics would look like if it were required to access compute at market rates without the benefit of a strategic relationship. A company whose reported margins depend on subsidised infrastructure may be materially less valuable on a standalone basis than headline figures suggest.
Even with strategic relationships, the amounts being committed are extraordinary.
Dario Amodei, CEO of Anthropic, has cautioned that because data centres take one to two years to build, AI companies must forecast their compute needs years in advance. With annual compute spend in the billions, he has warned that if revenue growth turns out to be fivefold rather than tenfold a year, even leading AI companies could face bankruptcy.
Indeed OpenAI was recently in the news for the gap between its ambitions and its finances, having missed key internal user and revenue targets while having locked in approximately US$600 billion in future data centre spending commitments.
The sector is also exhibiting characteristics typical of a capital cycle. With Big Tech free cash flow at a decade low and tens of billions of dollars in new debt being issued to fund infrastructure, some analysts have drawn parallels with prior capital cycles in capital-intensive industries such as telecoms and chemicals, where over-investment has historically led to overcapacity, depressed margins and weak returns. Similar concerns have been raised in connection with European AI infrastructure financing, where commentators have warned of high risk given uncertain returns and the potential for oversupply of AI capacity given the pace of construction.
The competitive dynamics of the AI sector are still being established. Market shares are fluid, pricing power is untested at scale, and the extent to which AI represents a winner-takes-most or winner-takes-some market remains genuinely uncertain across different application areas.
Marks has drawn a useful distinction in this respect between the established hyperscalers - highly profitable diversified businesses for which AI is one important segment among many - and the early-stage AI start-ups to which multi-billion-dollar valuations are being assigned. The former may be over or under-valued at any given moment, but it is unlikely that today’s prices for businesses such as Microsoft, Amazon and Google will prove ruinously excessive. The latter, by contrast, function in many cases more like lottery tickets: most participants will end up with worthless tickets, but the few winners may generate exceptional returns.
The history of technology markets suggests that the companies that ultimately capture the majority of value are frequently not the ones that led the early technology wave. The companies that dominated the internet era were not the first to operate in their respective markets. In AI, the question of which businesses will prove to have the most durable competitive advantages is not yet answered, and valuations that assume current leaders will maintain their position indefinitely warrant careful scrutiny.
The regulatory environment for AI is developing rapidly and unevenly across jurisdictions. The EU AI Act, which has been in force since 2024 and is being implemented on a phased basis, introduces a risk-based compliance framework with material obligations for providers of high-risk AI systems and general-purpose AI models classified as presenting systematic risk. Compliance costs, restrictions on deployment in certain use-cases, and civil liability exposure are among the direct economic consequences.
More recently, regulatory attention has expanded to encompass systemic risks posed by frontier AI capabilities. In May 2026, the International Monetary Fund (“IMF”) warned that the latest AI models elevate cyber risk to a potential macro-financial shock, citing the capacity of advanced models to identify and exploit vulnerabilities in widely used systems with unprecedented speed and scale. The IMF characterised the risk of correlated failures across financial institutions as systemic in nature and emphasised that cyber defences should be assumed to fail periodically, requiring greater resilience and coordinated regulatory preparation.
For the expert valuer, this development creates a new dimension of regulatory risk. Frontier AI capabilities that drive revenue and competitive positioning may simultaneously create contingent liability exposure, attract accelerated regulatory intervention, and trigger restrictions on deployment or use that may reduce practical addressable market opportunities. The valuation implications include potential effects on addressable market size, increased compliance costs affecting long-run margin assumptions, and liability exposure associated with model outputs.
Whilst all companies in this sector face some regulatory risk, the degree of exposure is highly company specific. A business whose product sits within a high-risk regulatory category, or whose capabilities have attracted systemic-risk classification, faces materially different exposure than one operating in a lower-risk application area. In contentious matters, regulatory risk should therefore be assessed by reference to the subject company’s actual exposure rather than applied as a generic discount rate adjustment.
The value of many AI businesses depends significantly on access to large volumes of training data and the model weights produced from it. The legal framework governing the ownership and licensing of that data is actively contested.
Copyright disputes relating to the use of third-party content in model training - including litigation brought by major media organisations, book publishers, and individual creators - have the potential to affect the defensibility of claimed intellectual property, the cost base associated with obtaining compliant training data, and the ability to continue operating current models without modification or replacement. The recent settlement of Bartz v Anthropic PBC (3.24-cv-05417) illustrates the point: the court found that the training of AI models on copyrighted books was itself defensible as fair use, but that Anthropic’s acquisition and storage of pirated books was not. This distinction resulted in a US$1.5 billion settlement. This case illustrates that legal distinctions relating to data sourcing and acquisition may materially alter the expected cost structure, liability profile and scalability assumptions underpinning AI valuations.
For the expert, the relevant question is not how these disputes will ultimately be resolved - that is for others to determine - but how the uncertainty they create should be reflected in the valuation. This may include adjustments to projected cash flows to reflect potential licensing costs, contingent liability provisions, or explicit downside scenarios in which adverse outcomes constrain the business model.
Corporate governance is a category of risk that warrants particular attention in AI valuations because the governance structures in this sector are unusual and, in several cases, materially affect what minority investors can realistically expect to realise.
The leading AI businesses have adopted governance structures that depart from the conventional shareholder-primacy model. OpenAI operates through a non-profit foundation controlling a public benefit corporation, with the foundation retaining special voting and governance rights. Anthropic is incorporated as a Delaware Public Benefit Corporation governed in part by a Long-Term Benefit Trust whose governance arrangements are designed to balance investor returns against broader public-interest objectives. Several leading AI businesses are controlled by concentrated founder and CEO control, achieved through a variety of mechanisms including dual-class share structures in some diversified technology groups and bespoke trust and foundation arrangements in others. The legal vehicle differs, but the economic effect is similar: a small number of individuals retain disproportionate influence over corporate strategy. The recent SpaceX IPO of June 2026 illustrates this structure in practice: Elon Musk holds approximately 41% of the economic interest in the company but retains 82.4% of the voting power post-listing.
The valuation consequences of these structures are not hypothetical. In May 2026, the corporate structure of OpenAI was subject to high profile US litigation in Musk v. Altman, the claimant (Musk) sought disgorgement of up to US$134 billion and unwinding of the October 2025 recapitalisation through which the Foundation came to hold its 26% stake and Microsoft its approximately 27% stake. The case was dismissed at trial on statute-of-limitations grounds rather than on the underlying merits, and the claimant has indicated an intention to appeal. Whatever the eventual outcome, the case illustrates that unusual governance structures in this sector can themselves become the subject of substantial contentious litigation, with valuation consequences flowing directly from the legal result.
Where minority investors cannot expect ordinary shareholder rights - dividends, control over major decisions, protection from self-dealing, exit liquidity - the value of their interest is correspondingly less than a pro rata share of enterprise value would suggest. The expert valuer should consider whether the governance structure warrants a specific minority adjustment, whether there are contingent risks of value extraction by controlling parties that should be modelled through expected values, and whether the basis of value being applied is consistent with the legal context of the engagement.
The fact that AI valuations involve substantial uncertainty does not mean that expert evidence is unreliable or that valuation is arbitrary. It means that the methodological approach must be designed to address uncertainty explicitly, rather than seeking to eliminate it by forcing artificial precision onto inherently uncertain inputs.
A discounted cash flow (DCF) analysis remains valuable in contentious AI valuations, not because it eliminates subjectivity but because it requires the explicit articulation of every significant assumption. Revenue growth, margin trajectory, capital requirements, terminal value, and discount rate are all stated explicitly, making them available for scrutiny and cross-examination.
In the income approach, risk can be reflected in two places: the cash flow forecasts (by adjusting expected values or by modelling probability-weighted scenarios) or the discount rate (by raising the rate to compensate for higher risk). In principle, the two approaches can produce the same valuation conclusion. In practice, the choice matters - and a well-constructed valuation reflects each risk in the place where it is most transparent and most capable of being tested.
The risks that drive valuation outcomes fall broadly into three categories, and each is appropriately reflected in a different place in the analysis:
The critical discipline is that each risk should be captured in one place only. A risk modelled through a downside scenario in the cash flows must not also be loaded into the discount rate, and vice versa. Double counting risk is one of the most common errors in contentious valuations.
Market-based approaches rely on the availability of comparable companies whose trading or transaction multiples provide evidence of value. In the AI sector, this is challenging for the reasons described earlier: few pure-play public comparables, highly diversified reference companies, and materially different business models across the sector.
Comparable selection in AI valuations will likely be a central area of disagreement between experts. The expert must document explicitly the basis on which companies have been included or excluded from the comparable set, and should assess whether the selected comparables share the relevant economic characteristics of the subject business - not merely its sector label. Where comparables are drawn from transaction evidence, the expert should consider whether those transactions were completed in market conditions comparable to the valuation date, and whether the pricing reflected strategic premiums, specific liquidity preferences, or ecosystem economics that are not present in the subject transaction.
Given the breadth of genuine uncertainty in AI valuations, probability-weighted scenario analysis could be the most intellectually honest approach to communicating the range of reasonable conclusions. The expert develops multiple genuinely distinct scenarios - reflecting materially different assumptions about technology evolution, competitive dynamics, regulatory outcomes, and the durability of strategic relationships - and assigns probabilities to each. Albeit that these probabilities themselves may attract considerable debate amongst experts.
The scenarios should be grounded in identifiable factual hypotheses, not merely mechanical adjustments to a single base case. A well-constructed scenario analysis assists the Judge or tribunal by making explicit which assumptions drive the most significant variation in value, and by demonstrating that the range of reasonable conclusions reflects genuine uncertainty rather than analytical weakness.
One of the most important and frequently underappreciated questions in any contentious valuation (not just AI) is which basis of value (or put simply, the value to whom) applies. The answer is determined by the legal and contractual context, and different bases of value can produce materially different results.
For example, market value (often favoured in the absence of instructions to the contrary) provides a neutral valuation based on what a hypothetical willing buyer and seller would agree which can differ significantly from equitable value, where the purchaser is identifiable and, as a consequence, may exclude minority discounts or take account of strategic synergies.
The expert’s report should identify clearly which basis of value has been applied and why.
In contentious proceedings, the expert is frequently required to value a business as at a past date - the date of a disputed transaction, the date of an alleged breach, or the date at which a relevant event occurred. This creates the problem of hindsight: the expert may know how events subsequently unfolded, but the valuation must be conducted from the perspective of a market participant at the historical date, using only information that was known or knowable at that time.
In the AI sector, where the competitive landscape can shift materially within months, the distinction between contemporaneous knowledge and hindsight can be particularly significant. A model that was considered state-of-the-art at the valuation date may have been superseded by the time the report is written; regulatory proposals that were uncertain at the valuation date may have since been enacted; companies that appeared dominant may have been disrupted by lower-cost competitors. The expert must document clearly the information set used and the basis on which post-date information has been considered or excluded.
Transparency about the key assumptions driving the valuation conclusion is important in any valuation, and particularly so in AI valuation cases, given the breadth of judgment involved. The expert’s report should identify the assumptions to which the valuation is most sensitive, present a range of outcomes across reasonable variations in those assumptions, and explain the basis for the central estimates adopted.
If sensitivity analysis is used, it should focus on the variables that drive the greatest variation in value - typically revenue growth rates, the treatment of strategic relationship revenues, long-run margin assumptions, terminal growth rates, the systematic components of the discount rate, and the probabilities assigned to truncation and other company-specific scenarios.
Valuation disputes involving AI companies are unlikely to become simpler in the near term. The sources of uncertainty described in this paper - technology and pricing pressure, revenue quality, infrastructure economics, competitive dynamics, regulatory evolution, intellectual property, governance complexity, and reporting opacity - are likely to be structural features of the sector rather than temporary characteristics of an immature market. Indeed, AI companies remain subject to the same fundamental disciplines as other businesses: capital flows to companies that demonstrate growth, competitive pressure constrains discretion on safety and quality, and management priorities are shaped by the demands of investors.
The methodological gap between price and intrinsic value is no longer theoretical. SpaceX listed in June 2026 at an implied valuation of US$1.77 trillion which is more that double Morningstar’s published fair value of approximately US$780 billion. Two credible analytical perspectives on the same business, differing by more than a factor of two. That is the territory in which AI valuation disputes are likely to play out for the future, and the methodological question this paper addresses – how to analyse risk rigorously, transparently, and defensibly in the face of such uncertainty is unlikely to become less relevant.
Early engagement with valuation methodology matters. The choice of the basis of value, the scope of the expert’s instructions, and the period for which financial information is available can all significantly affect the quality of the expert evidence. Revisiting these decisions once the expert process is underway will increase the cost of the expert phase.
A wide range of expert conclusions is not, of itself, evidence of inadequate analysis. In AI valuations, reasonable experts applying orthodox methodologies to the same facts can reach materially different conclusions. The width of that range is itself evidential: it reflects the genuine uncertainty that characterises the sector and should be presented to the judge or tribunal as such, rather than treated as a problem to be explained away.
The most defensible expert evidence is not the most precise. An expert who acknowledges the limits of what can be reliably estimated, discloses key judgments transparently, and communicates a well-reasoned range of conclusions is likely to be more useful to the judge or tribunal - and more resilient under cross-examination - than one who presents a single-point conclusion built on a chain of optimistic assumptions.
The expert’s task is not to eliminate uncertainty. It is to analyse it rigorously, communicate it transparently, and help the Judge or tribunal understand the range of conclusions that are reasonably available on the evidence.
Kate Lilleyman is a highly experienced testifying expert specialising in contentious valuations in disputes across a number of sectors and jurisdictions. For more information, please contact Kate.
Five reasons why system implementations fail with actionable recommendations that empower charities to navigate system implementation more effectively.
In this piece, Simon Davidson and Mark O'Sullivan examine the growing tension between data and instinct in CFO decision making and explain why data influenced leadership is becoming essential as organisations accelerate towards AI.
The Amendments to FRS 102 bring in significant changes to the reporting standard for both revenue recognition and lease accounting (amongst other incremental changes), which will broadly align these accounting requirements to those of IFRS 15 ‘Revenue from contracts with customers’ and IFRS 16 ‘Leases’.
Nested relationships in correspondent banking can expose firms to hidden financial crime risks. Learn how to improve oversight, transparency and monitoring.
A practical overview of the FCA’s 2026 wholesale markets priorities, highlighting key risks in financial crime, market abuse and conflicts of interest, and what firms should do to strengthen controls and governance.
Practical guidance on preparing for an FCA sanctions compliance review, covering regulatory expectations, programme readiness and how to respond effectively.
