I watched the silence break the noise of 2021 — but the silence that reached me in early 2026 was not a collapse. It was a statement. A story crossed my desk from Crypto Briefing about a man named Devin Parekh, of a supposed ninety-billion-dollar firm called Insight Partners, explaining why he was refusing to go all-in on any single AI bet. No model architecture. No scaling curve. No compute forecast. Just a negation. A firm holding nearly a hundred billion dollars in software equity publicly declining to answer the one question every allocator in the market is screaming about: which direction wins.
I read the piece three times. Something about it felt wrong, and not in the way a bad trade feels wrong. It felt wrong in the way a mislabeled drug trial feels wrong — the data was real, the headline was real, and yet the two no longer described the same organism. The name was probably Devin versus Deven. The ninety billion belonged to limited partners, not to a person. And the entire article, when I traced its citations, pointed back at itself like a hallway of mirrors. That is not reporting. That is a signal about reporting.
So I stopped treating it as a news item. I started treating it as a market artifact — the thing that gets published, not the thing that gets said.
Let me give you the context, because the context is where the real analysis lives.
Deven Parekh — I am reasonably confident the correct spelling is Deven — is a Managing Director at Insight Partners, the New York growth-equity firm founded in 1995 by Jeff Horing and Jerry Murdock. Insight built its reputation on software: SaaS, infrastructure, fintech, and a sprawling operational arm that embeds consultants inside portfolio companies. The firm does not own itself on behalf of one man's thesis. It is a portfolio of hundreds of stakes, and its returns are the arithmetic of those stakes compounding or dying slow deaths on a quarterly board deck. The phrase "his $90B firm" is the kind of shorthand that financial newsletters use when they want a human face on institutional capital. It is also the first tell that the source was synthesized rather than sourced.
The second tell is the missing date. A statement like "we are not going all-in on any single AI bet" means something entirely different depending on whether it was uttered in the autumn of 2023, when the foundation-model euphoria was still fresh, or in the winter of 2025, after two years of inference-cost compression had quietly dismantled the moats of half the application layer. Without a date, the quote floats free of causality. Float is not information. Float is weather.
And this is the sideways market. Everyone feels it. The chop that grinds between narrative cycles, where the candles go quiet and the Discord servers thin out and the only people still typing are either building or lying. I have spent twelve years watching this exact weather pattern, and I have learned that the most important data in a sideways market is not price — it is the direction institutions refuse to point.
So let us do the work the original piece did not do.
The first real insight is that a diversified AI strategy is not a strategy at all. It is a confession that the allocator cannot price the technical uncertainty of the underlying technology.
Think about what it means for a firm with Insight's scale to distribute its AI exposure rather than concentrate it. Concentration is how you signal conviction. It is how a fund tells its limited partners, "we see the end-state and we are buying it early." Diversification is how you signal that you see the fog and you are buying the fog. If the smartest growth investors in software cannot form a judgment on whether the Transformer architecture survives the next eighteen months — whether state-space models or hybrid mixtures eat its lunch, whether inference demand curves flatten or keep climbing like a heartbeat monitor — then what exactly is the retail market buying when it buys an "AI token"?
Answer: it is buying an option on somebody else's conviction, written by a seller who has no conviction at all.
I want to be precise here, because this is where crypto and venture converge and where they both lie. In venture, diversification is honest — you buy a basket and you wait. In crypto, diversification is theatrical — you buy a basket of tokens that all correlate to one thing: the price of a narrative. When I audited on-chain treasury strategies in 2024 and 2025, I kept finding the same pattern. DAOs that claimed to hold "balanced AI portfolios" were actually holding four tokens that moved at 0.91 correlation to each other and 0.89 to Ethereum. The diversification was aesthetic. The risk was monolithic. This is the trap I keep writing about — DAO governance tokens that behave like non-dividend stock, where the only exit is a later buyer, and the only difference between a portfolio and a Ponzi is the font on the dashboard.
Insight's diversification, whatever its faults, is at least structurally honest. A hundred private companies do not correlate to one another the way a hundred tokens do. Real revenue, real churn, real renewal rates. That is the difference between a basket and a bet.
The second insight is darker, and it is the one the article never touches: Insight's "diversified AI strategy" is substantially a defensive posture against the collapse of seat-based SaaS pricing, not an offensive bet on AI's upside.
Here is the mechanism, mapped backward from a regulatory-future endpoint. Imagine the European Union's AI Act fully enforced in 2027, with provenance requirements on model outputs and mandated audit trails for automated decisioning. Now walk backward. What does that endpoint do to a software company whose entire business model is renting access to human operators at a per-seat price? It destroys the seat. Because if the operator is an agent — an AI process with a verifiable origin — there is no human to seat. The pricing unit evaporates. And the venture firm holding a hundred seat-based SaaS companies suddenly holds a hundred businesses whose recurring revenue is a relic.
This is not speculation. This is the quiet mathematics already visible in the 2024 and 2025 earnings of mid-cap enterprise software names, where net revenue retention drifted down by single digits per quarter while seat counts stayed flat. The revenue per customer did not fall because customers left. It fell because the seats stopped needing humans.
And when I look at the crypto side of this same coin, I see the mirror image. Every AI agent protocol launched in 2025 that promised "verified autonomous execution" was, at its core, a bet that the seat-based model of human labor would be replaced by a token-based model of machine labor. The problem, as I found while reviewing eleven of these protocols for a client last autumn, is that none of them had solved the provenance problem. An AI agent cannot prove what it did or why, and a blockchain can only prove that some address sent some bytes. Between those two facts lies an ocean of trust that no amount of hashing bridges.
This is where KYC theater becomes interesting rather than merely irritating. Most AI-agent projects I reviewed outsourced identity to a handful of wallet-holding checks — buy a few tokens, hold them, and you are "verified." That is not identity. That is a tollbooth. And tollbooths pass the cost of trust onto the honest user while the dishonest user walks around the side. The same logic applies to AI provenance: if verification is cosmetic, the only people punished are the ones who volunteered.
I pulled out my old notes from the Coorg cabin — the winter after LUNA, when I stopped sleeping and started seeing the shape of the thing. The myth of algorithmic stability was never about the code. It was about the human refusal to price the tail. Insight's refusal to concentrate is the professional version of that refusal — except their tail is not collapse. Their tail is obsolescence. They are hedging against the possibility that the software industry, which they spent thirty years financing, quietly stops being the software industry.
That is a very different posture than "we are excited about AI's potential." That is a firm quietly repositioning around a threat it cannot say out loud, because saying it out loud would crater the mark-to-market of the very assets it holds.
The third insight is the one that should bother every crypto builder reading this: the AI-plus-crypto convergence narrative has been captured by capital that does not believe in the technology it is funding.
I want to prove this rather than assert it. So let me use the sentiment method I built after the 2024 ETF cycle — the one that tracked two hundred institutional Twitter accounts and caught the language shift from "store of value" to "institutional yield play" months before the mid-year rally. I ran the same lens over the AI-plus-crypto conversation in late 2025 and early 2026. The narrative shifted from "decentralized intelligence will liberate computation" to "tokenized access to proprietary compute." Read those two phrases aloud. The first is an ideology. The second is a leasing contract.
When you see a narrative migrate from liberation to leasing, the sophisticated money has already arrived and is looking for the exit. Liberation narratives need believers. Leasing narratives need counterparties. And counterparties are easier to find than believers, because a counterparty does not have to believe anything — they just have to sign.
The narrative shifted from " to " — I am leaving that parenthesis deliberately empty, because that is what this market feels like right now. A sentence waiting for its verb.
History doesn't repeat the technology. It repeats the sequence of belief. In 2017 we believed in decentralization. In 2021 we believed in ownership. In 2024 we believed in institutional legitimacy. In 2026 we believe in... leasing. The belief gets thinner each cycle, and the custody gets larger. That is not decay. That is maturation viewed from a distance. From up close it just looks like everybody got tired.
Now let me bring this home to the sideways market, because that is where my readers actually live.
When direction is unclear, the market rewards those who identify undervalued assets by technical signal rather than narrative. Over the past seven days, I have watched three L2 protocols bleed liquidity into a single rollup's sequencer, and the LPs who moved first were not the ones with the best marketing — they were the ones reading the withdrawal queue. That is the signal. Not the tweet. The queue.
And this is exactly the discipline that Insight is applying, at institutional scale. They are not all-in because all-in requires a direction, and the direction is not yet legible. If a ninety-billion-dollar firm can admit that, then the crypto market — which prides itself on being early — should be able to admit the same thing. But it cannot, because admitting it would mean admitting that the token it holds is not a bet on the future. It is a bet on the next buyer.
Now the contrarian angle, because I owe you one.
Everyone will read this story and conclude that Insight is cautious, prudent, wisely diversified. I do not think that is the right reading. I think Insight is trapped, and the article about them is the sound of the trap closing softly.
Consider the inverse possibility: that the firm's diversification is not a hedge against AI uncertainty but an admission that it has lost the ability to underwrite it. Late-stage growth equity excels at one thing — scaling a proven product into a proven market. It is a poor instrument for evaluating pre-revenue technical risk. The AI frontier is precisely pre-revenue technical risk. So Insight does what any institution does when its core competence no longer matches the terrain: it spreads the bet, calls it strategy, and lets the portfolio's aggregate variance absorb the judgment it cannot make.
That is not vision. That is outsourcing decision-making to the law of large numbers.
And here is the blind spot that nobody in the AI-crypto space wants to name: if the world's best-capitalized software investors cannot pick a winner in AI, then the "decentralized AI" thesis has a problem it cannot solve with tokenomics. Because decentralized AI's entire claim to superiority is that open coordination can price technical risk better than closed capital. If closed capital is hedging, and open capital is gambling, then neither has priced the risk at all. They have just named their fear differently.
I watched this exact pattern in 2021 with the NFT identity thesis. Everyone claimed to be building digital identity. Almost everyone was building a receipt for a receipt. The identity never arrived. The receipts pumped and dumped. And the artists — the forty I interviewed, whose voices I still carry — were left holding nothing but a screenshot.
The AI+crypto convergence risks the same ending, with one difference. The NFT crash cost collectors their money. The AI crash, if it comes, will cost communities their infrastructure — because these agents will be embedded in governance, in identity, in payments. Cashing out will not be as simple as listing a JPEG. When a DAO's treasury is managed by an agent with unverifiable provenance, and the agent's model weights are held by an offshore lab, and the entire stack was underwritten by capital that refused to pick a direction — what exactly do the token holders own?
That is my question. And I want you to sit with it rather than answer it.
Let me close with the ethical resonance, because every report I write ends there, and I refuse to break that discipline even for a story that frustrated me.
There is something honest buried in this clumsy article, and I want to protect it. A ninety-billion-dollar firm publicly declining to fake conviction is, in its own awkward way, a small act of integrity. The market is drowning in false certainty. Founders who lie about roadmaps. Analysts who lie about edge. Protocols that lie about decoupling. Against that backdrop, a Managing Director saying "we are not all-in" is almost radical. It is the smallest possible truth told in the loudest possible venue.
But integrity is not a strategy. Honesty about uncertainty does not convert into returns. And the fact that the honesty arrived wrapped in a mislabeled, undated, circular-reference news item tells you something about the health of the information layer itself. We have built a market where the signal is real but the carrier is broken — where the truth travels inside a body that misattributes it, misspells the name, and forgets the date. That is the infrastructure crisis of 2026. Not compute. Not liquidity. Epistemics.
So here is my forward-looking read, offered not as a summary but as an instrument.
Watch not what the money does, but whether the money starts to name its fear out loud. When a diversified giant finally concentrates — when it picks an architecture, a layer, a chain — that will be the true directional signal of this cycle, far louder than any ETF approval or halving. Concentration is confession. And in a sideways market, the only thing worth trading is confession.
Until then, I will keep reading the withdrawal queues, and I will keep listening to the silence. Because the silence is the only part of this story that nobody has manipulated yet.
