Executive Summary
Roughly one-third of all AI-focused venture and growth capital deployed over the past 18 months has gone to four foundation model companies: OpenAI, Anthropic, xAI, and Waymo. Their annualized revenue run rates are substantial (OpenAI near $24 billion, Anthropic reportedly north of $30 billion), which provides some operational ballast to the headline valuations. The downstream effect on the broader AI venture and growth equity market is structural rather than incidental. When a meaningful plurality of AI capital concentrates in a handful of names, the comparable transaction data used to justify pricing across the rest of the ecosystem becomes thin. PE and growth equity funds marking AI portfolio companies at 28x ARR are citing comparable AI transactions that consist primarily of the same four foundation model companies, each of which has structural advantages (geopolitical access, sovereign-scale capital, distribution position) that no Series B startup matches. LPs absorbing the resulting valuation marks are increasingly asking sharper questions about the underlying comp logic.
The Concentration Numbers
Through 2025 and early 2026, the four foundation model companies absorbed approximately $188 billion in equity and equity-equivalent capital. OpenAI's recent funding rounds, Anthropic's Series F and G, xAI's expansion rounds, and Waymo's continued capital intake combined to account for roughly 33% of AI-focused capital deployed in the period.
The remaining 67% has been distributed across the rest of the AI ecosystem: enterprise AI applications, AI infrastructure and tooling, vertical AI products, AI-adjacent software, robotics, and applied AI in defense, healthcare, and biotech. The distribution has been wide but thin, with median round sizes in non-foundation-model AI categories declining 25% year over year even as the headline total venture and growth capital appeared to grow.
The bifurcation produces an aggregate market that looks healthier than the median experience inside it. Founders running AI companies in the $5 to $50 million revenue range report substantially more difficult fundraising environments than the headline data suggests, with elongated processes, lower valuations, and more restrictive terms than peer companies of similar quality experienced 18 months earlier.
What This Does to PE and Growth Equity Marks
PE and growth equity funds with AI exposure are facing a specific valuation challenge. The comparable transaction data that supports portfolio company marks consists disproportionately of transactions in the four foundation model companies. Citing those transactions as comparables for a Series B AI company at $15 million in revenue produces marks that are difficult to defend rigorously.
The structural argument against using foundation model transactions as broad AI comparables runs through several specific factors.
Capital access is fundamentally different. The foundation model companies have raised capital from sovereign wealth funds, large institutional asset managers, and strategic investors with geopolitical and supply chain considerations. The pricing reflects strategic value that is not available to a typical Series B company. A sovereign fund paying $1 billion at a high multiple for OpenAI exposure is not pricing the financial return alone. It is pricing geopolitical positioning and AI infrastructure access.
Distribution position is fundamentally different. The four foundation model companies have direct relationships with enterprise customers across every Fortune 500 vertical. The customer pipeline is established. The integration depth is substantial. The ARR figures reflect this distribution position. A vertical AI startup attempting to argue the same multiple based on faster proportional growth misses the structural difference in distribution.
Unit economics are fundamentally different. The foundation model companies operate at scale that produces operational leverage no smaller company can match. The marginal cost of serving an additional customer is approaching infrastructure cost at the foundation model layer. The economics that justify 28x ARR multiples are predicated on this operational leverage. A smaller company without the same leverage faces a different unit economic structure that the multiple should reflect.
Despite these differences, the use of foundation model transactions as broad AI comparables persists across PE and growth equity portfolio marks. The result is portfolio company valuations that are mathematically supported by the comparable set but structurally indefensible under closer inspection.
What LPs Are Saying in Practice
The diligence conversations with allocators have shifted noticeably over the past six months. The patterns are consistent.
A common LP perspective: "If your comp set consists of four companies and three of them have not proven unit economics, you do not have a comp set. You have a narrative."
The framing is direct. LPs are not opposed to AI exposure. They are not even opposed to higher multiples for high-quality AI companies. What they are opposed to is portfolio marks that depend on comparable transactions without examining whether the structural conditions that produced those comparables apply to the specific portfolio company being marked.
LPs that have absorbed multiple aggressive AI marks across their portfolios are pulling back from GPs whose AI exposure cannot be justified through more rigorous comp logic. The pullback is not from AI as an asset class. It is from GPs that price AI portfolio companies primarily through reference to the foundation model transactions.
The contrast in LP behavior across GPs is sharp. GPs with disciplined AI valuation approaches (using strategy-specific comps, applying meaningful discounts to foundation model references, conducting thorough unit economic analysis) are retaining LP confidence even with concentrated AI exposure. GPs that have marked aggressively against the foundation model comps are facing more skeptical re-up evaluations.
For the broader context on operational AI deployment in PE portfolios, see Blackstone N1 and operational AI.
The Foundation Model Revenue Reality
A meaningful counterpoint to LP skepticism: the foundation model companies are generating substantial revenue. OpenAI's annualized run rate at approximately $24 billion, Anthropic's at $30 billion+, and the underlying customer concentration in well-funded enterprise customers, provides real operational ballast to the valuation discussions.
Foundation model valuations are not entirely narrative. The revenue figures are real, the customer count is substantial, and the trajectory remains upward. The companies are operating at scale that very few software companies have ever achieved on similar timelines.
The question is not whether the foundation model companies are valuable. They are. The question is whether the multiples paid for foundation model exposure should propagate to the rest of the AI ecosystem as comparables for general AI valuation.
The answer, structurally, is no. The foundation model companies are at the top of a different unit economic curve, with different customer access, different supply chain position, and different competitive dynamics than the broader AI ecosystem. Treating them as universal comparables for AI valuation produces analytical errors that LPs are increasingly catching.
What Disciplined AI Marks Look Like
The GPs that are pricing AI portfolio companies with intellectual honesty share specific characteristics.
They use strategy-specific comp sets. A vertical AI company in legal tech gets compared to other vertical AI companies in legal tech, with adjustment for revenue scale, growth rate, and customer concentration. A horizontal AI infrastructure company gets compared to other infrastructure companies. The foundation model companies appear in some comp sets as outliers, not as central reference points.
They apply meaningful discounts for structural differences. A Series B company at $15 million ARR being marked against any foundation model transaction receives a substantial discount (often 50 to 70%) to reflect the structural differences in capital access, distribution, and unit economics. The discount is documented and defended in the valuation memo.
They run unit economic analysis at the portfolio company level. Customer acquisition cost, payback period, gross margin trajectory, and net retention numbers get explicit analysis. The multiple paid reflects what the unit economics support, not what comparable transactions imply.
They are transparent with LPs about the methodology. The valuation methodology is documented and explained in LP communications. LPs can audit the methodology and understand how the marks are derived. The transparency builds trust even when individual marks are disputed.
This approach produces lower headline marks than the alternative approach of citing foundation model transactions broadly. The lower marks translate to lower interim IRRs and lower TVPI numbers. The trade-off, however, is meaningful LP credibility and more durable fund-raising for successive vintages.
The 2026-2027 Mark Correction
Several structural factors suggest that AI marks across PE and growth equity portfolios will face downward pressure over the next 12 to 24 months.
Public market AI comparables provide one anchor. As more AI-focused companies have gone public or attempted to, the public market pricing has been more disciplined than private markets. Public market AI multiples are typically 30 to 50% below comparable private market multiples for similar-quality companies. As more public data accumulates, the private market marks become harder to defend at premium multiples.
Audit pressure is increasing. Auditors evaluating PE fund marks are applying more scrutiny to AI portfolio company valuations specifically. The comp set rationale needs to be defended in audit. The thin comp sets that depend on foundation model transactions are facing more challenge in audit cycles.
LP-driven valuation reviews are becoming more common. Some large LPs have begun commissioning independent valuation reviews of AI portfolio companies in funds they invest in. The reviews typically produce lower valuations than the GP marks, which the LPs use to inform re-up decisions and to negotiate side letter terms in new commitments.
Realization data is starting to accumulate. The AI portfolio companies that have exited (through acquisition, IPO, or write-down) provide actual transaction data that competes with the foundation model comp set. The realization data has generally been below the carrying marks, which puts downward pressure on remaining portfolio company valuations.
The aggregate effect over the next 24 months is likely to be a meaningful downward repricing of AI portfolio company marks across PE and growth equity portfolios. The repricing will be uneven, with some GPs absorbing larger markdowns than others. The GPs that have applied disciplined valuation methodology will face smaller corrections than the GPs that have marked aggressively against foundation model comps.
For more on the broader vintage performance environment, see the PE DPI distribution crisis.
What LPs Should Be Asking
For LPs evaluating GPs with AI exposure, several diagnostic questions surface useful information.
What is the specific comp set for each AI portfolio company in the current marks? The question requires the GP to articulate which transactions they cite as comparables for each specific company. Answers that depend heavily on the four foundation model companies are flags. Answers that cite multiple strategy-specific comparables with documented adjustments are more credible.
What discount is applied for structural differences? When foundation model transactions appear in any comp set, what discount is applied to reflect the differences in capital access, distribution, and unit economics? GPs that apply meaningful discounts (50%+) typically have more defensible marks than GPs that apply minimal discounts or none.
What is the unit economic analysis underlying each AI mark? Customer acquisition cost, payback period, gross margin, and net retention data for the specific portfolio company. GPs that can produce this data and explain how the multiple reflects the unit economics are operating with more rigor than GPs that cite multiples without supporting unit economic analysis.
How have AI realization values compared to carrying marks in the existing portfolio? The realization data for previously exited AI companies provides the most useful comparison. GPs whose AI realizations have come in close to or above carrying marks have demonstrated valuation discipline. GPs whose realizations have come in significantly below marks have track record evidence of overstatement.
How is the GP responding to the public market repricing of AI? The public market AI comparables provide an anchor. GPs that have adjusted marks downward in response to public market discipline are operating with rigor. GPs that have maintained marks despite public market repricing are diverging from observable data.