In 2026, the GitHub activity curve of the Crypto open-source community completed an amazing “bottoming out.” From 45K monthly active developers at its peak in 2022, it fell back to about 23K. This halving in paper data sparked discussions on social media about “narrative exhaustion.” However, when we dissect the cross-section of this curve, we see not a shrinking industry, but a profound “deleveraging of talent.”

I. Who left? Who stayed? Most of those who left were newcomers. In February 2024, the number of new developers reached 5,462, then fell sharply, with a turnover rate of 52% for those with less than one year of experience. Most of these people flooded in during the bull market, making NFT minting contracts, forking DeFi protocols, and doing front-end work for new Layer 2s. These positions were highly dependent on market popularity. Once the popularity faded, the projects stopped operating, and the positions disappeared with them. Data shows that the code contribution of newcomers never exceeded 25% of the total, and this group was never in the core circle of the industry from the beginning.

On the other hand, developers with more than two years of experience rose instead of falling during the same period, hitting a record high, contributing about 70% of the code. Electric Capital’s GP Maria Shen’s judgment is very direct: “When we look at the established developers group, it is growing and looks very healthy.” They stay not because they have no other choice. Technically, the core work of crypto now generally requires years of accumulated infrastructure development work to understand: protocol layer development, security audits, cross-chain architecture. These jobs require years of accumulation to truly get started, and cannot be eliminated by the market when the popularity fades. Economically, many veterans have unvested tokens, governance power in the protocol, and equity relationships. Their accumulation in this industry has formed real barriers and returns.

From the perspective of ecological distribution, they are voting with their feet: Bitcoin developers increased by 64.3% in two years, Solana +11.1%, while Cosmos decreased by 51.1%, and Polkadot decreased by 46.9%. Veterans are concentrating on ecosystems with real users and income, leaving those projects that are still maintained by narratives. Changes in the job structure also confirm the same thing. Among the new Web3 positions in 2025, the highest proportion is not developers, but Project & Programme Management, exceeding 27%. This is counterintuitive for an industry known for being technology-driven, but the logic behind it is not complicated: the industry has entered the execution phase from the construction phase, more than 100 chains need to be integrated, institutional customers have completely different requirements for compliance and security, and DAO governance needs to find a balance between stakeholders with different interests. This is not project management in the traditional sense, but coordination and judgment in an environment where rules are still being formed.

The industry appears to be shrinking, but the core density is rising instead. The bear market of 2018-2019 was also accompanied by a large loss of developers, but afterwards, phenomenal projects such as Uniswap, Aave, and OpenSea emerged, defining the bull market of 2020-2021. The builders who remain in this round have more mature infrastructure, and the AI era has given them a bigger stage than the previous round.

II. What abilities do the people who stay have? What special abilities has the crypto industry cultivated in builders? To answer this question, we need to go back to the underlying principles of blockchain. Between the bull and bear cycles, this industry has always operated on the same underlying rule: code is law, and execution is the end. In 2016, The DAO event saw attackers transfer $36.00M using a recursive call vulnerability. The code had no bugs, and the logic was executed as expected, but the boundaries were not anticipated by the designers. In 2021, the Poly Network cross-chain bridge was attacked, and $610.00M was transferred in a few hours. No platform can call a halt, no institution can revoke, and no legal terms can be recovered. This is a structural feature of crypto that distinguishes it from almost all other industries: the margin for error is zero, and ex-post intervention is almost non-existent.

What this environment forces out is a set of capabilities that are rarely needed in other industries: building a workable system from scratch that strangers are willing to participate in under conditions of missing rules and missing trust. This capability includes two levels. The first is to build trust from scratch, without relying on any external authority, but only relying on code and mechanisms to make strangers willing to put in real assets. The second is to make judgments under dual technical and economic uncertainties. Without regulatory frameworks, historical data, or industry standards to refer to, it is still possible to design a system that can operate. Both levels have specific verification in crypto. Uniswap has no company guarantee, no KYC, and no customer service. Anyone who puts funds into the liquidity pool relies only on trust in hundreds of lines of code and an economic mechanism, and achieves tens of billions of dollars in daily trading volume. MakerDAO has no central bank endorsement and no deposit insurance, and relies purely on on-chain governance and collateral mechanisms to maintain the stability of DAI. During DeFi Summer, it was even more extreme. Without regulatory frameworks, audit standards, or any historical data to refer to, builders designed AMMs, lending protocols, and liquidity mining, and it only took a few months from concept to billions of dollars in TVL.

The AI era is creating a structurally highly similar problem. The model decision-making process is opaque, and the output results cannot be independently verified. AI agents are starting to autonomously execute transactions and allocate funds, but the supporting rules and constraints do not yet exist. Large model companies control both the model and the evaluation criteria, and users lack effective verification methods. Computing power is highly concentrated in a few top large factories, forming a monopoly pricing when demand explodes. These problems point to the same core: the trust problem of autonomous systems, which is being replayed in the larger scale process of AI. Crypto builders have been dealing with this type of problem for many years in an environment without external authoritative rules, but the previous scenario was on-chain protocols, and now it has been replaced by AI. And a group of people have already brought the capabilities accumulated in crypto directly into AI, and have achieved results.

III. How are these capabilities being repriced in the AI era? Cases of switching from crypto to AI have become increasingly common in recent years, but when broken down, they take away different things. The most direct path is the direct transfer of hardware and experience. CoreWeave’s three founders, Michael Intrator, Brian Venturo, and Brannin McBee, started mining Ethereum with GPUs in 2017, expanding from one to thousands. They closed their mining business in 2022. Two months later, ChatGPT was released, and the GPUs in their hands directly became AI computing power supply. In March 2025, they were listed on Nasdaq, with an IPO valuation of about $23.00B, and the market value once approached $70.00B. OpenSea co-founder Alex Atallah handled the aggregation and routing of extremely heterogeneous assets in the NFT market, and moved the same set of experiences to AI model routing, founding OpenRouter, which served more than 5.00M developers in two years, with a valuation of $500.00M.

Another type of migration is more worthy of attention. NEAR founder Ilia Polosukhin is one of the co-authors of the Transformer paper. After leaving Google, he initially wanted to build AI applications with natural language, but encountered a real problem in the development process: he needed to make cross-border payments to data labeling workers around the world, most of whom did not have bank accounts, and blockchain technology became the best solution to this payment problem. Now NEAR is transforming into an AI infrastructure platform, and the core direction is user-owned AI and decentralized confidential machine learning (DCML), allowing users to use AI services without exposing data. The decentralized architecture experience accumulated in NEAR has become the most difficult starting point to replicate in this direction. Circle co-founder Sean Neville left and founded Catena Labs, positioning itself as an AI-native bank, directly migrating the understanding of stablecoin infrastructure to AI agent financial scenarios, with a16z crypto leading an $18.00M seed round. Aave and Lens Protocol’s senior developer Nader Dabit turned to Cognition, bringing the developer ecosystem construction experience accumulated in multiple crypto protocols into the AI agent tool field. What this group of people took away is not only GPU hardware or user networks, but also the intuition of mechanism design, the construction experience of the developer ecosystem, and the judgment to build a credible system from scratch when rules are missing. These capabilities happen to correspond to the three structural gaps encountered by AI scaling.

Computing power aggregation and optimization Computing power is the most direct bottleneck for AI scaling. Training and inference require a large number of GPUs, demand fluctuates greatly, cloud vendors are expensive and queuing is required, and companies do not want to hoard hardware themselves. This problem has two levels: how to aggregate and allocate computing power, and how to use the aggregated computing power more efficiently. Crypto builders have directly transferable accumulation in both levels. Hyperbolic solves the allocation and trust problem. Founder Jasper Zhang brought decentralized mechanism design into the AI computing power track: tokens allow distributed GPU holders to contribute idle computing power, but the more core problem is trust. Why believe that the calculation results given by a strange node are correct? The core innovation PoSP uses random sampling plus game theory to make honesty the dominant strategy of the node, without full verification, low cost, scalable, and reliable results. This set of mechanisms is directly migrated from the logic of crypto verifying the behavior of strange nodes. MoonMath solves the efficiency problem. Its predecessor Ingonyama focused on ZK hardware acceleration, and increased the ZK proof generation speed by several times under extreme computing constraints. Now the direction is turning to the Physical AI performance layer, doing sparse attention acceleration (LiteAttention) of video diffusion models, low-rank decomposition (LiteLinear) of FFN layers, and training backpropagation acceleration (BackLite). From ZK acceleration to AI inference acceleration, the underlying layer is the same set of capabilities: making mathematics run faster under extreme computing constraints. The track has changed, but the accumulation has not been wasted.

AI governance and incentive mechanism design When multiple AI agents start to collaborate to perform tasks, how to ensure that they do not damage the overall system in the process of pursuing their respective goals. Each participant is pursuing their own objective function, and no one guarantees that the system can still operate normally after they are added together, and the execution speed of the agent far exceeds the window of human intervention. This is the type of problem that crypto builders have repeatedly dealt with in DAO governance and tokenomics design: allowing participants with completely different interests to operate in the direction preset by the system without a central authority. The answer given by crypto is economic mechanisms, illegal operations will produce real economic costs, and the rules are written in the code and executed automatically. EigenLayer directly migrated this set of mechanisms to the AI scenario. Through the restaking mechanism, nodes need to stake assets before participating in collaboration, and non-performance or illegal operations will trigger automatic penalties. The rules are not suggestions, but rigid boundaries with real economic costs. EigenCloud extends this logic to the verifiable calculation and collaborative governance of AI agents, allowing agents to fall within the preset range when pursuing their own goals. It is much more reliable to constrain agents with economic mechanisms than to constrain agents with ethical guidelines.

AI Agent autonomous payment There is also a more basic problem: how do agents pay. Traditional payment systems are designed for people. Credit cards require account opening, and bank transfers require authorization. Each step assumes that the operator is human, has an identity, and will wait. The agent will not wait. It may initiate a large number of requests per second, and each request may involve micro-payments. Traditional payment channels directly fail in this scenario. Stablecoins and on-chain rules are the infrastructure that crypto builders have already built, natively supporting programmability, no authorization, and all-weather operation. These three characteristics happen to be the rigid requirements of the agent payment scenario, and what is missing is only a layer of protocol that connects stablecoins to the agent workflow. x402 was launched by Coinbase in May 2025, activating the HTTP 402 status code, embedding stablecoin payments directly into HTTP requests, and completing the payment at the same time as the agent initiates the request, without an account, and settling in about two seconds. As of April 2026, the x402 protocol has processed more than 165.00M transactions, with a cumulative transaction volume of about $50.00M, and the number of active agents has reached 69,000 (data source: x402 Foundation). Cloudflare, AWS, Stripe, and Anthropic MCP have all been connected. Agent payment is already a track with real traffic.

IV. The New Positioning of Builders: From People Who Write Contracts to People Who Set Rules for AI The scaling of AI is creating a previously non-existent functional gap. It is not a gap in technical talent, but a gap in people who can design trust mechanisms in autonomous systems. When the service object changes from people to AI, the role of crypto builders is also being redefined. The core difference between the two paradigms lies not in the technology stack, but in the way trust is established and the logic of rule execution. In the Pre-AI era, crypto builders faced human participants, the rules were written into contracts, the margin for error was zero, but the boundaries of the system were relatively clear. In the AI-Native era, when the interaction object becomes an autonomously running AI agent, the problem that needs to be solved is: the behavior of the agent is unpredictable, the execution speed far exceeds the human intervention window, and the boundaries of the system itself need to be redefined under greater uncertainty. The functional positioning of crypto builders is changing from “writing secure contracts” to “designing credible mechanisms for AI autonomous systems.”

The recruitment of leading exchanges and institutions in 2026 clearly reflects this trend: they are no longer simply recruiting AI engineers or crypto developers, but are looking for people who can connect the two sides, who understand on-chain incentive distortions and governance games, and can deeply embed AI tools into crypto workflows, and design mechanisms that allow agents to align with regulators and users in the long term. The capital allocation direction has also reflected this judgment. Paradigm is raising a new fund with a maximum size of $1.50B, and the investment scope has been expanded from crypto to AI and robotics. Haun Ventures completed a $1.00B Fund II, focusing on the financial infrastructure that integrates crypto and AI, especially supporting AI agent autonomous trading and the payment, stablecoin, and agent-to-agent economic system for coordination. a16z crypto completed a $2.20B fifth fund (Crypto Fund V), clearly stating that the fund will invest 100% in the crypto field. Faced with the complexity and opacity of the AI era, they will focus on the transparency, verifiability, and decentralization characteristics of crypto application directions. And according to PitchBook data, in 2025, about 40% of VC investment in the US crypto field flowed to companies that also involved AI business, a significant increase compared to 2024.

Similarly, the paths chosen by crypto builders to switch to AI in different market environments show obvious differences. In the United States, with the relative clarification of the regulatory environment, protocol layer innovation has gained real living space. The capital network density is high, the path from idea to financing is short, and the margin for error is large. The common characteristics of a group of projects such as Hyperbolic, EigenCloud, Gensyn, and Ritual are to design new mechanisms from scratch, rather than simply integrate applications on existing systems. Top VCs have clear investment theses on directions such as “verifiable computing, Agent coordination, and decentralized ML”, and are willing to provide sufficient fault tolerance for early technology exploration. The situation in Asia is different. Singapore and Hong Kong assume more of the role of compliance implementation and institutional fund transfer. The regulatory framework is relatively conservative and has a low tolerance for pure protocol layer innovation. When builders with crypto backgrounds switch to AI, they are more likely to choose application layer and industrial integration paths – using the user base, payment capabilities, or data assets accumulated by crypto to quickly access AI products and services. This is not a gap in capabilities, but a difference in path selection caused by different market signals and regulatory environments.

Returning to the GitHub curve at the beginning. The number of monthly active developers has dropped from 45K to 23K, which seems to indicate that the industry is shrinking. However, among the people who remain, the proportion of established dev has hit a record high, and they are flocking to ecosystems with real users, and are being repriced by the AI industry in an unprecedented way. When AI scaling encounters structural bottlenecks such as computing power aggregation, Agent autonomous payment, data and decision verifiability, and privacy coordination, these Builders are at the intersection of Crypto and AI. These long-term accumulated sensitivities to rules, incentives, and authenticity are gradually being transformed into scarce system-level capabilities in the AI era. As an investment institution that has been deeply involved in crypto infrastructure since 2017, IOSG’s judgment on this line is not only at the observation level. We participated in the investment when EigenLayer’s restaking mechanism was not widely recognized by the market, led Ingonyama’s (now MoonMath) seed round to bet on the migration of ZK hardware acceleration to the AI performance layer, and invested in Hyperbolic in 2024, optimistic about its path of using crypto-native verification mechanisms to solve the trust problem of decentralized computing power. The common logic behind these layouts is that the trust, coordination, and verification problems encountered by AI scaling will ultimately need the mechanism design capabilities accumulated by the crypto industry to solve. We believe that the intersection of crypto and AI is not a narrative, but a structural opportunity that is happening.

[IOSG]