AI is an opportunity for Nerds, and Agents are an opportunity for Money. Venture capital, MegaFunds like A16Z have always told us it’s a story about cycles and exits, but from the perspective of a Solo GP, it’s more like harmonic vibrations of signals and structures. You need to find the real patterns they haven’t told you about.

In 2021, a16z returned $12.5B in profits to LPs, with DPI exceeding the sum of the previous decade. At the same time, 2021 also marked the beginning of the disaster for the US VC industry. Excluding actual DPI, it was all just unrealized gains. In other words, 2021 was a golden period for exits, where LPs could actually get their hands on real money. However, if LPs reinvested, they would have to endure the pain that continues to this day.

Everything is telling a contrary narrative, and the crypto market’s volatility is in sync with this. In 2022, the metaverse concept fueled Web3, even artificially extending the bull market, until early 2025, when Binance’s “bestie coin” farce put an end to VC coins. Nowadays, most VCs have fallen into silent mode. Economies of scale have been dragged into a capital-intensive model of computing power and data, with a long delay in recouping costs. Network effects are nowhere to be seen on-chain, leading to survival through institutionalization and SaaS channel fees.

However, throughout the history of venture capital, each cycle of interest rate hikes and cuts, and the liquidity injected, has nurtured different VC models. We repeatedly invent valuation logic for risk, and the relative freedom of the crypto market also allows those with intentions in this market to uncover the most profitable signal mechanisms.

When VCs stop taking risks, “Every passion originates from external things impacting the sensory organs, causing animal spirits to move through the nerves. If there’s still an impression, in March and April 2021, Roblox and @coinbase chose the Direct Listing model to go public. Unlike a regular IPO, a direct listing only sells existing shares, requiring no underwriters and no lock-up period. Interestingly, both were led by A16Z. Amidst the dazzling DPI data, in June 2021, A16Z raised $2.2 billion for its third crypto fund, and in January 2022, A16Z raised a new fund of $9 billion.

So what’s the price? The price is that Coinbase’s stock price fell by 90% from its peak in 2023. It can be stated very clearly that A16Z’s role in the US stock market is no different from that of crypto VCs. But the problem is, A16Z can still raise $7.2 billion in 2024 and $15.1 billion in 2026. Even in May 2026, its fifth crypto fund raised over $2.2 billion, bringing its total crypto fund history to nearly $10 billion. The market presents a choice: either become an LP of @a16z and wait for the astonishing DPI at the moment of liquidity injection, or become the price of A16Z, the source of astonishing DPI.

But problems also arise. A16Z is not sensitive to market signal discovery. In other words, VCs who dominate each cycle face the curse of scale. Excessive scale prevents them from having enough motivation to discover early-stage paradigms, especially revolutionary rather than incremental mechanisms. Arthur Rock, the father of modern venture capital, peaked at his debut, with Fairchild and Intel initiating Silicon Valley’s VC model. KP and Sequoia formally introduced the institutionalized VC model, leading through the PC and mobile internet transitions. YC turned VC into a probability game under mass mechanisms, batch-producing sub-giant unicorns under power laws. SoftBank, led by Masayoshi Son, turned VC into an approximate Ponzi game of super-scale through the myth of Chinese concept stocks like Alibaba. Thus, as old giants indulge in past glories, emerging ambitious individuals will prove their unique vision through mechanism innovation, thereby obtaining cheap money and embarking on their new era of adventure.

Even reputation itself can be exchanged for money. Paradigm founder Matt Huang invested in ByteDance. Although ByteDance couldn’t go public, Paradigm chose to leap into crypto. The latest news is that they have shifted to AI and robotics. Let’s revise the answer: if you can’t be an LP of A16Z, and you don’t want to be the price paid by others, then you must discover unamplified new signals and kill the old predecessors with new mechanisms. The cracks have appeared. In 2021, A16Z was not “allowed” to participate in Anthropic’s financing; instead, more individual investors bet early, such as Skype co-founder Jaan Tallinn and former Google CEO Schmidt, who led the Series A round. SBF of FTX entered in 2022, giving us another enduring imagination of Crypro X AI.

A16Z doesn’t need to take risks. SBF used retail money for “effective interest A\”. If we are to find the most reasonable starting point for a Solo GP, Claude’s venture history is the most typical. Unlike individual angels, Solo GPs run the entire VC operation based entirely on their own research capabilities. The Agent era is easy for us to understand, but it is precisely humans who first practiced it. Unlike YC’s broad-net approach, Solo GPs still need to invest deeply in each project, and every investment is important for DPI. A16Z has become an indicator of the market itself. As new technological trends emerge, newer participants try to find them even a little earlier than the A16Zs. Beyond large AI models, they are targeting Agents. There is a dangerous leap here: economies of scale cannot exist in large AI models. For every additional human user, server costs increase, and costs cannot be amortized like software. That is, network effects have not appeared in Agents as expected, and the invocation between Agents is still an ideal state.

Non-human network effects. “In 1784, Watt improved the rotary steam engine, and in 1824, the complete theory of the steam engine was expounded by the Frenchman Carnot. Everything about AI is a black box. Scaling Law was observed by Baidu’s Wang in Baidu. The mathematics required for Transformers does not exceed the graduate level, but why it can exceed the mathematical level of graduate students is unknown. AI is an opportunity for Nerds; you just need to give money to the most cutting-edge people and then wait for miracles to happen. The talent acquisition popular in Silicon Valley is the best proof: Researcher > Data > Model. However, large models themselves are difficult to recoup costs. To reiterate the reversal of economies of scale, even shifting from training to inference, even shifting from dialogue to tasks, cannot stop this process.

The only way out for large AI models is to become traffic centers like AWS and CloudFlare. If they are destined to not reduce production costs, then consumption must grow infinitely. Agents are an opportunity for Money; Agents must become the main consumers. Infinite主体 X infinite consumption, this is the root of why Agents calling each other has become a hot topic. However, to a considerable extent, Agents and Bots are indistinguishable. It’s unclear what an Agent is, and it seems Bots have existed for a long time.

If we must define an Agent, the “evaluating agent” in reinforcement learning is the origin of this wave of technological advancement. In DeepMind’s thinking, allowing the agent to automatically evaluate training success is the key to the next level of intelligence upgrade. This is very different from Claude’s role division in coding. From a programming perspective, an Agent is actually a mapping of the human programmer’s role. When we talk about Agentic Coding, it’s a world away from AlphaZero’s Agents.

Only from this perspective, where Agents act on behalf of humans, can Claude’s impact on SaaS be justified. It’s merely an iteration of the outsourcing mechanism: moving towards high-value scenarios, after programmers come accountants and analysts; moving towards fewer full-time employees, after outsourcing comes the cost of multiple Agent invocations. However, a problem still exists: Agents have not shown human social relationships. Real business relationships do not become smoother by applying Agents; humans still prefer to interact with people. We have indeed created more Agent scenarios, which are effective in “internal” coordination, such as large companies using GPUs when laying off employees.

However, in “external” collaboration, it must be noted that it has not been confirmed as expected. In May 2026, US employment grew strongly, with non-farm payrolls increasing by 172,000, mainly in blue-collar sectors like leisure and hospitality, and healthcare. Conversely, the financial sector decreased by 22,000. The fear of Agents in human society is real but severely overestimated. Of course, just like whether the Sahara needs shoes, this can also be a signal to continue enhancing model intelligence, increasing Agent capabilities, and investing in Robotics. In other words, Agent economics only holds true in theory; infinite growth on the consumption side has not materialized. Continuing to bet, how can we make Agents call each other and give rise to network effects?

Crypto positioning for the Agent era. “Evolution does not always lead to increased complexity; evolution is not always an upward trend. Let’s summarize what we know and issue a warning for the unknown. Venture capital cannot represent effective discovery of technological signals; it has become a game for the few brave. Agents are being mass-produced with the hope of reducing the production costs of large models, but natural invocation relationships do not arise between them. These two seemingly contradictory discourse systems contain subtle coordination: finding signal mechanisms that stimulate Agent invocation. Simply issuing Agent assets or making DeFi protocols Agent-like is meaningless. There are already few people and many bots on-chain. Adding smart contract invocations will only increase technical risks. This path is not smooth.

From a practical standpoint, human primacy will not be replaced by Agents because role mapping depends on business relationships. Domestic technology does not buy 4090s, and China and the US do not buy from each other. The boundaries of technology are narrower than we imagine. Exa targets the demand for real-time + high-quality data for Agents. One cleaning, multiple invocations – this is true economies of scale, but it’s difficult to trigger invocations between Claude and Codex. Catena meets the compliance financial needs of Agents between B-sides, even applying for an OCC license for B2B compliance. This is a specialized network effect, but it’s difficult to reduce the cost of large-scale use.

Payment protocols represented by stablecoins aim to tap into C-side entry and settlement exit demands. Lightweight protocols reduce usage costs, and micropayments reduce collaboration costs. But it’s not enough. To achieve final A2A daily communication, humans must be willing to sacrifice their souls. Similar to TrueNorth’s three-step approach: let humans use Agents to assist in transactions; let Agents learn from human participation in transactions; let Agents lead on-chain transactions. Compared to Claude’s policy and legal restrictions in connecting with IBKR, it can only become a kind of CoPilot. TrueNorth’s use of Hyperliquid’s live trading is not difficult. But for humans to willingly accept Agent guidance, all of this is still very far away, at least farther than VCs imagine.

In attempts at Agent + finance, “primarily investing in payments, secondarily in trading” occupies the absolute mainstream structure. Payments are very certain; PayPal and Stripe’s market share will be tokenized, and stablecoins will be agentized. Trading prospects are broad, from Simons to Jane Street, and to the unpayable kindness of Liang Sheng’s HFT, it sparks infinite imagination for VCs. But all of this is not the same as our imagination of Agents taking over trading and payments. Quant trading establishes “computing power hegemony,” still a speed advantage relative to humans. Trading establishes “channel advantage,” still a fee advantage relative to banks. The gap emerges here: VCs need to facilitate one thing – making people willing to be proactively replaced by Agents. A16Z is powerless to do this. Throwing money cannot make new social platforms like Clubhouse and Towns Protocol successful, so for more complex financial Agent scenarios, they can only lie flat.

If we refer to the successful experience of DeFi, letting Agents touch funds, with low-frequency, small-amount validation of feasibility, then high-frequency, large-amount daily use can follow. One can imagine that if the roads were full of FSD Tesla Robotaxis, it would actually be safer than a mix of human/AI driving. But to make this happen, humans actually need to act as test subjects: a minority use AI-assisted driving, establishing technological parity with human drivers; reducing the casualty rate of a minority using AI-assisted driving, establishing a compensation mechanism. In other words, establishing a mechanism for Agents to handle money will be easier to convert users than for Agents to make money. Only when Agents accumulate enough experience using money can humans stop thinking and just click confirm all the way. Only when Agents actively participate in the market can market efficiency and safety be improved. It can be understood that the process of Agents seeking returns is the process of market efficiency improvement, gradually self-lifting, writing C++ with C++, and optimizing Agents with Agents. Trading is the endpoint for Agents, but before that, you must go through a long elliptical track. In the high-value scenario of finance, blockchain is the experimental field for open finance, and stablecoins are the proof of the Agent’s market optimization process. This has nothing to do with scale or resource investment; it has to do with mechanism establishment and expansion.

Conclusion: “Where in life is there no cycle? Without cycles, there would be no era dividends for us. It’s always the new generation replacing the old. VCs are becoming smaller and more personalized. Whether it’s Solo GP or OPC, we haven’t yet observed Solo GPs investing in OPC becoming mainstream. Amidst the uncertain surge of technological waves, we don’t know which paradigm will become mainstream. ‘Software is eating the world.’ After the dot-com bubble in the early 2000s, it ushered in over 20 years of long-term dividends. Now, we have entered a new era of ‘Agents eating software.’ Agents themselves are development tools, a mark of productivity evolution, but there are no new software applications developed by Agents that have become universally adopted. This is also a fact. After the IPOs of @SpaceX, OpenAI, and @AnthropicAI, the positioning of foundational large models has ended.

If this is a new round of long-term dividends, then new crypto VCs like @dragonfly_xyz, ParaFi, Haun, and @paradigm, a16z, will continue to scale up, or will niche funds like 5cc emerge and demonstrate their strength in the new deployment frenzy. Even the entire DeFi industry will undergo a new paradigm update. In the past two Kondratiev cycles, financial systems have continuously innovated, and this time, Agents and stablecoins will be the new starting point for a dual revolution. Crypto is small, the world is big, let’s witness it together!

[@zuoyeweb3]