A headline scrolled past my feed twice before I stopped and read it three times โ not for the argument, but for the metadata. Insight Partners' Devin Parekh explains why his $90B firm isn't going all-in on any single AI bet. Eleven words, two errors. First, the name: the Insight Partners managing director who has long covered software, internet, and fintech is Deven Parekh, not "Devin." Second, "his $90B firm" โ Parekh is a managing director. He is not an owner. Insight Partners was founded in 1995 by Jeff Horing and Jerry Murdock; assigning a firm's AUM to a single individual is a category error, the kind that tells you the copy was assembled from a press release rather than a primary document.
In on-chain forensics, when a labeled entity is wrong at the header, you do not trust the body until you have re-derived it. Chain links don't lie โ but press releases do, and the first lie is usually structural, not malicious. So before I analyze what a growth-equity giant means when it says it will not concentrate on any single AI thesis, I have to tell you what the source actually is: a paraphrase of a paraphrase, with no publication date, no interview transcript, and no first-person quotation. That absence is itself the signal.
Context: what the statement is, and what it isn't
The parsed record of this story contains no technical detail. No model architecture. No parameter counts. No benchmark results. No compute-economics. The only extractable facts are commercial and structural: a firm managing roughly ninety billion dollars in disclosed assets has publicly declined to anchor its strategy to one AI direction. The negative object of the phrase "not going all-in" is not a technology โ it is a direction. That matters.
I have spent years doing forensic work on capital that claims to know something. In 2017, at twenty-four, I audited the EVM bytecode of a privacy coin called Project Aether. The whitepaper claimed a fixed supply. The bytecode contained a minting function behind an owner-only modifier. I cross-referenced wallet clusters on Etherscan against the stated distribution and found a twelve-thousand-ETH supply discrepancy. I wrote forty pages around transaction hashes, and three exchanges delisted the token within a fortnight. The lesson I carried out of that audit was not that projects lie. It was that the mechanism always contradicts the marketing eventually โ and the only question is how much runway the contradiction has.
So when a ninety-billion-dollar fund says it will not pick an AI winner, I do not hear a market call. I hear a description of its own balance sheet.
Insight Partners is a software growth-equity and venture platform. Its economics are not those of a concentrated hedge fund. It earns from ARR growth across a portfolio of dozens to hundreds of operating companies, and from exits, layered on top of an operational arm โ Insight Onsite โ that deploys roughly a hundred people to embed inside portfolio companies. That revenue engine is structurally diversified. A firm whose return is the aggregate of three hundred software companies does not have the institutional option of being a single-thesis investor. The statement "we are not all-in on any single AI bet" is therefore not insight. It is arithmetic.
Which is exactly why the crypto reader should pay attention โ because the same arithmetic governs the AI-token complex, and almost nobody applies it.
Core: what the non-bet actually prices
Let me work through this the way I work through a balance sheet: mechanism first, narrative last.
If the smartest, best-connected growth investor in enterprise software cannot converge on a winner in AI infrastructure, then the market is pricing genuine technical uncertainty. The unresolved parameters are known to anyone with an inference bill: whether scaling laws continue to hold at current compute multiples, whether the transformer gives way to state-space or hybrid architectures, and what the shape of the inference-side demand curve looks like once model weights commoditize. None of these have converged. A fund that abstains from direction is, mathematically, short the assumption that a single architecture or layer captures the surplus. Diversification is a position. It is a paid option on "we don't know yet."
That has a direct on-chain analogue. In the summer of 2020 I wrote a Python script to monitor real-time liquidity ratios across Uniswap V2 pools. One protocol, YieldFarm X, was reporting a headline TVL that the pool-level data refused to confirm. The same five hundred ETH of collateral was being recycled across five pools, each venue counting the rehypothecated same coins as its own liquidity. I mapped the wallets and the currency never left the family. The protocol collapsed seventy-two hours after I published the thread. Capital recycles to wherever the narrative sits, and the narrative always reports a bigger number than the ledger.
Hold that against the AI-token complex. The category trades on the same recycling logic. A token announces an "AI strategy," a narrative wallet accumulates, funding rates on perpetuals spike, and the inference revenue โ the only metric that would justify a valuation โ is either zero or undisclosed. The team holds the narrative; the retail holder holds the liquidity. Follow the gas, not the hype. When I trace the fee flow on these tokens, the on-chain footprint of actual usage is a rounding error against the market cap. The market has priced a thesis it cannot verify.
Now, the deeper mechanism inside the Insight non-bet that no headline will print: the firm's diversification is defensive, not opportunistic. Insight's core asset is a software book built on seat-based and subscription pricing. Generative AI compresses exactly that model โ it decouples software capability from headcount, and the per-seat subscription was always a proxy for the labor the software augmented. When a model can perform the work of three analysts, the buyer does not renew three seats. A portfolio of mid-market enterprise SaaS is, mechanically, a short position on the repricing of knowledge labor. So "we diversify our AI bets" is partly a hedge against the revaluation of the firm's own holdings. The opportunity and the threat are the same technology.
There is a commercial observation worth pinning here. The parsed source claims the firm will "reshape the software industry" through its AI strategy. That mechanism is misplaced. A fund does not reshape an industry by spreading capital. Capital reshapes an industry when it is concentrated enough to force a pricing model change across a portfolio. What actually reshapes middle-market enterprise software is a small number of portfolio companies re-architecting their pricing from seat-based to outcome-based โ and the fund's operational arm pushing that template laterally across the book. That is a specific, traceable mechanism with an on-chain parallel: it looks less like a venture bet and more like an index position in the repricing of an entire sector.
Let me put the mechanism into a table, because tables force precision.
| Layer | Who captures surplus | Visibility of the bet | |---|---|---| | Foundation models | Marginal โ weights commoditize, capex is brutal | High narrative, low moat | | Inference / compute | Volatile โ a function of demand curve shape | Priced in real capex, not tokens | | Application / vertical software | Broad โ pricing power returns to the workflow owner | Low narrative, high multiple | | AI tokens (crypto) | Mostly none โ narrative capture | Fully visible, least verifiable |
The table explains the non-bet. Insight is not abstaining from AI. It is positioning at the application layer, where its existing book already sits, and refusing to pay foundation-model valuations for exposure it can get through portfolio companies it already owns. That is not courage. It is discipline โ and it is the correct read for a growth-equity fund whose edge is operating leverage, not architecture bets.
I saw the same discipline fail in 2022. Three days before the Terra announcement, I was monitoring reserve addresses and watched the collateral quality drop roughly forty percent. The on-chain liquidity depths were telling a story the marketing had not yet admitted. I wrote a risk assessment called "The Inevitable Decay" and hedged UST through Curve pools for clients โ an estimated two hundred thousand dollars protected. What I learned was not that I could predict collapses. It was that the data is almost always visible before the narrative updates its language. The reserve addresses knew before the press did.
That is the correct lens for the Insight headline. The interesting data is not the quote. It is the fact that a ninety-billion-dollar manager with the best deal access on earth cannot yet justify a concentrated position โ which means the AI-token tokens trading at institutional-scale valuations are, structurally, doing the opposite of what the professionals do. Retail is concentrating where the smart capital is diversifying. Wallets connect the dots โ and right now the dots trace away from single-thesis exposure.
Contrarian: correlation is not causation, and the position is not a prophecy
Here is where the mainstream reading gets it wrong. The consensus interpretation of this story is "a big fund is cautious on AI, therefore AI is overheated." That is a causal leap the evidence does not support. The fund is not cautious because it has a bearish view. It is diversified because its book requires diversification. The statement is a description of a structural constraint, not a forecast. Confusing the two is how you end up treating a balance-sheet property as market intelligence.
There is a second blind spot. I have watched analyst after analyst read a managing director's interview as if it were the firm's strategy. It isn't. A managing director speaks for a practice area, not a partnership. The statement that this one person is not all-in on a single AI direction tells us nothing about where the fund's aggregate capital sits. It is entirely possible โ and common โ that the firm's overall AI exposure is concentrated in the exact foundation-layer assets the individual says he avoids, because different partners hold different mandates. The headline collapses a person and a partnership into a single voice. That is a labeling error of the same severity as calling a managing director an owner.
And the third blind spot is the most important for anyone building a thesis on this: the source itself has no primary documentation. No date, no transcript, no first-person quote. In on-chain terms, this is a transaction with no verified signature and no block confirmation โ and yet it has been circulated as if it were settled fact. The most useful information gain in this whole story is not the AI opinion. It is that a ninety-billion-dollar firm's strategy is being reported through circular citation, and the market is pricing off it anyway.
That should worry you more than any single allocation decision. Not because the fund is wrong. Because the informational layer you are reading โ the one that shapes the AI-token narrative and the funding rates that follow โ cannot distinguish between a partnership's strategy and a managing director's offhand remark. If the reporting layer has this failure rate, the asset layer it describes is trading on noise, not signal.
Takeaway
Watch the application layer, not the foundation-model headlines. Watch the portfolio companies that quietly re-price from seats to outcomes, because that mechanism โ not any capital allocation โ is what actually reshapes enterprise software and, by extension, the AI-token narratives that orbit it. And watch the reporting layer itself: the next time a ninety-billion-dollar fund's strategy reaches you with no date and no transcript, ask which block that transaction was mined in. Code is the only witness. Everything else โ including this article โ should be verified against the ledger before you risk a single dollar on it.