A reasonable accredited investor could have spent the last 18 months building what feels like a diversified pre-IPO portfolio — an AI foundation model company, an AI search company, an AI coding tool, and an AI infrastructure play — and still be holding what is functionally a single concentrated bet on the continued expansion of large-language-model compute budgets.
This is not a hypothetical. Among the 28 issuers available on secondary markets today, a meaningful cluster sits in generative AI or AI-adjacent infrastructure. OpenAI, Anthropic, Perplexity, Cohere, Mistral AI, and xAI are all, at their core, foundation-model or model-deployment businesses. Glean and Notion have substantial AI-layer bets embedded in their growth stories. Even Databricks, primarily a data lakehouse, derives a large share of its current multiple from AI/ML workload projections.
What concentration actually means in private markets
In public markets, concentration is easy to measure: weight by market cap, check sector classifications, look at beta correlations. In private markets, none of those signals exist in real time. You have no daily price, no published beta, and often no audited revenue breakdown to anchor a sector analysis.
What you do have is the secondary mark — the price at which interests last traded on a platform — and the primary round valuation, which is typically the 409A anchor the company's board approved for option-grant purposes. Neither tells you much about how correlated your holdings are to each other. That correlation lives in the business model, not the price.
The useful question is not "how many names do I own?" but "how many independent outcomes do I own?" Two companies that both collapse if enterprise AI budgets contract in a downturn are one outcome, regardless of how different their cap tables look.
A practical framework for auditing AI overlap
Before placing an indication on any new name, work through the following four lenses. They do not require financial models — they require honest answers.
1. Revenue source
Where does the company's revenue come from today, not where does management say it will come from in three years? A company billing enterprises on API consumption and a company billing consumers on subscriptions have different recession sensitivities, even if both are called AI companies. Map each holding to its primary customer type (consumer, SMB, mid-market enterprise, hyperscaler) and primary revenue model (subscription, usage, seat-based, transactional).
2. Compute dependency
Which of your holdings require ongoing access to very large GPU clusters at declining per-token cost in order to maintain their margins? Foundation-model companies — those training and serving their own models — are acutely sensitive to compute costs and to any shift in the availability or pricing of H100-class hardware. Companies that sit on top of models (wrappers, applications, workflow tools) have a different cost structure, though they carry their own risks around commoditization.
3. Hyperscaler relationships
Several of the largest private AI companies have deep capital and distribution relationships with Microsoft, Google, or Amazon. Those relationships are strategic assets in good times and potential constraints in bad times. If two of your holdings are both deeply tied to the same hyperscaler for distribution or cloud credits, a shift in that hyperscaler's AI strategy is a shared risk factor across both positions.
4. Liquidity path
AI companies at private valuations above $20 billion have a relatively narrow set of realistic liquidity events: a public offering, an acquisition by a strategic buyer, or a long-term hold. If your portfolio has four names all clustered above that threshold, their liquidity paths are correlated to the same IPO window conditions. A sustained period of risk-off sentiment or rising interest rates narrows all four simultaneously.
Where genuine diversification exists in the 28-issuer universe
Not every private name on active secondary markets is an AI story. Stripe is a payments infrastructure business whose revenue correlates to global e-commerce and SMB card volume. Klarna and Chime have consumer fintech exposures tied to credit cycles and interchange economics. Anduril operates in defense contracting, with revenues shaped by U.S. and allied government budgets. Rocket Lab is a launch and space systems business with its own supply chain and government contract dynamics. Discord and Epic Games carry consumer entertainment exposure. Revolut competes in consumer banking across multiple jurisdictions.
These are not risk-free names. But their primary risk factors are genuinely different from the AI compute cluster. Holding Anduril alongside Anthropic is a more diversified two-name portfolio than holding Anthropic alongside Cohere.
Position sizing when you identify overlap
Once you have mapped your overlapping exposures, the question is whether to trim, hold, or deliberately add more. There is no universally correct answer. Some buyers deliberately run concentrated AI books because they have high conviction in the sector and are comfortable with that concentration. The problem arises when concentration is accidental — when a buyer believes they are diversified and discovers they are not only after a macro shift re-prices the entire cluster.
A few sizing principles that secondary market participants have found useful: keep any single issuer below 20% of your total private portfolio at cost; keep any single thematic cluster (all AI names combined) below 40% of total private exposure; and treat the illiquidity premium of private shares as a reason to demand higher expected return, not as a reason to accept lower conviction.
Secondary markets offer one thing that primary rounds do not: the ability to add or reduce exposure after the fact, without waiting for a company-initiated event. If you identify concentration and want to reduce, a secondary sale is the mechanism. If you want to add to a less-represented theme — defense, fintech, consumer — secondary markets give you that access today.
Using secondary pricing data as a sanity check
Secondary marks, while imperfect, do carry information. When two companies in the same AI cluster trade at very different discounts to their last primary round valuation, that divergence is worth interrogating. It may reflect genuine differences in near-term revenue visibility, in the perceived strength of their investor syndicate, or in how much overhang exists from employee selling pressure. It may also reflect nothing more than thin volume and stale marks. Knowing which is which requires more than looking at the price alone.
We publish hourly-refreshed pricing across all 28 issuers on the marketplace. Comparing how AI-cluster names are moving relative to each other — and relative to fintech or defense names — is a quick way to pressure-test your own portfolio thesis.
Browse current pricing and available supply across all 28 names at /marketplace. If you are already holding multiple AI positions and want to think about a partial sale to rebalance, the /sell page walks through what information you will need to get started.