Số lượng nhà phát triển giảm một nửa: Crypto chưa chết, chỉ là nhường nhân tài cho AI

Original author: Xinyang, Ethan, IOSG Ventures

In 2026, the GitHub activity curve of the Crypto open-source community completed an astonishing "bottoming out". After falling back from a peak of 45K monthly active developers in 2022 to about 23K, this apparent halving on paper sparked discussions about "narrative exhaustion" on social media. However, when we dissect the cross-section of this curve, what we see is not a shrinking industry, but a profound "deleveraging of talent".

▲ Data source: Electric Capital Developer Report, based on Crypto Ecosystems Github

Who Left? Who Stayed?

Those who left were mainly newcomers. In February 2024, the number of new developers reached 5,462 in a single month, followed by a sharp decline, with a churn rate of 52% for those with less than a year in the industry. Most of these people flooded in during the bull market, working on NFT minting contracts, forking DeFi protocols, or building frontends for new L2s.

These roles are highly dependent on market hype. Once the hype faded, projects ceased operations, and the positions disappeared. Data shows that newcomers' code contributions never exceeded 25% of the total. From the start, this group was not part of the industry's core circle.

▲ Newcomers surged in with the bull market and left during the bear market; Established devs (2+ years of experience) hit an all-time high in the same period

Data Source: Electric Capital Developer Report

On the other hand, developers with more than two years of experience not only didn't decrease during the same period but increased, hitting an all-time high and contributing about 70% of the code. Maria Shen, GP at Electric Capital, put it directly: "When we look at the established developers group, it's growing and looks very healthy."

They stay not because they have no other choices.

Technologically, the core work in crypto now is infrastructure development that generally requires years of accumulation to understand: protocol layer development, security auditing, cross-chain architecture. These tasks demand years of experience to truly master and cannot be easily phased out by market sentiment.

Economically, many veterans hold tokens that haven't fully vested, governance powers within protocols, and equity relationships. Their accumulated stake in this industry has formed real barriers and returns.

Looking at the ecosystem distribution, they are voting with their feet: Bitcoin developers grew by 64.3% over two years, Solana by +11.1%, while Cosmos decreased by 51.1% and Polkadot by 46.9%. Veterans are migrating towards ecosystems with real users and revenue, leaving behind projects sustained by narratives.

▲ Source: Coincub Web3 Jobs Report 2025

Data source: Web3.Career

The changing structure of job roles confirms the same thing. Among new Web3 positions in 2025, the largest share wasn't developers, but Project & Programme Management, accounting for over 27%.

For an industry renowned for being tech-driven, this is counterintuitive. But the logic behind it isn't complex: the industry is moving from a construction phase to an execution phase. Over 100 chains need integration. When institutional clients enter, compliance and security requirements become completely different. DAO governance needs to find a balance among stakeholders with diverse interests.

This isn't traditional project management; it's about coordination and judgment in an environment where rules are still being formed.

While the industry appears to be shrinking on the surface, its core density is actually increasing. The 2018-2019 bear market also saw a massive loss of developers, yet it spawned phenomenal projects like Uniswap, Aave, and OpenSea, which defined the 2020-2021 bull run. The builders who stayed in this cycle have more mature infrastructure, and the AI era provides them with a much larger stage than the last one.

What Capabilities Do Those Who Stay Have?

What specific capabilities has the crypto industry cultivated in its builders? To answer this, we need to return to the underlying principles of blockchain. Amidst the alternation of bull and bear cycles, this industry has always operated on the same fundamental rule: code is law, execution is final.

In the 2016 DAO incident, an attacker exploited a recursive call vulnerability to siphon off $36 million. The code had no bug; the logic executed exactly as expected, but the boundaries were not anticipated by the designer. In 2021, the Poly Network cross-chain bridge was attacked, and $610 million was transferred within hours.

No platform could stop it, no institution could reverse it, no legal clause could provide recourse. This is the structural characteristic that distinguishes crypto from almost all other industries: zero margin for error, and post-hoc intervention is practically non-existent.

The capability forced out by this environment is one rarely needed in other industries: the ability to build a functional system from scratch, under conditions of absent rules and absent trust, that allows strangers to willingly participate.

This capability has two dimensions. First, building trust from zero, relying not on any external authority but solely on code and mechanisms to make strangers willing to deposit real assets. Second, making judgments under dual technological and economic uncertainty, designing a working system without a regulatory framework, historical data, or industry standards to reference.

Both dimensions have specific validation in crypto. Uniswap, without a company guarantee, KYC, or customer service, relies solely on trust in a few hundred lines of code and an economic mechanism, achieving hundreds of billions of dollars in daily trading volume.

MakerDAO maintains DAI's stability purely through on-chain governance and collateral mechanisms, without central bank backing or deposit insurance.

The DeFi Summer was even more extreme. Without a regulatory framework, audit standards, or any historical data to reference, builders designed AMMs, lending protocols, and liquidity mining, going from concept to tens of billions in TVL in just a few months. This capability manifests differently in builders at the protocol, application, and governance layers, but the underlying principle is the same.

The AI era is creating a structurally similar problem. Model decision-making processes are opaque, and outputs cannot be independently verified. AI agents autonomously execute trades and allocate funds, but the supporting rules and constraints don't yet exist.

Large model companies control both the models and the evaluation standards, leaving users without effective verification methods. Computational power is highly concentrated in a few major tech firms, creating monopolistic pricing when demand surges. These issues all point to the same core problem: the trust issue in autonomous systems, replaying at the larger scale of AI.

Crypto builders have been dealing with this type of problem for years in environments devoid of external authoritative rules. The previous setting was on-chain protocols; now it's AI. And a cohort of people has already directly transferred the capabilities accumulated in crypto into AI, yielding results.

How Are These Capabilities Being Repriced in the AI Era?

Cases of transitioning from crypto to AI have become common in recent years, but breaking them down reveals they bring different things.

The most intuitive path is the direct transfer of hardware and experience. The three co-founders of CoreWeave, Michael Intrator, Brian Venturo, and Brannin McBee, started mining Ethereum with GPUs in 2017, expanding from one machine to thousands. They shut down their mining operations in 2022. Two months later, ChatGPT was released, and their GPUs directly became AI computing supply. They went public on Nasdaq in March 2025 with an IPO valuation of ~$23 billion, and their market cap subsequently peaked near $70 billion.

OpenSea co-founder Alex Atallah dealt with the aggregation and routing of highly heterogeneous assets in the NFT marketplace. He applied the same experience to AI model routing, founding OpenRouter, which served over 5 million developers within two years and was valued at $500 million.

Another type of migration is more noteworthy. NEAR founder Illia Polosukhin is a co-author of the Transformer paper. After leaving Google, he initially wanted to build AI applications using natural language. However, during development, he encountered a real-world problem: cross-border payments for data labelers worldwide, many of whom lacked bank accounts. Blockchain technology became the optimal solution for this payment challenge.

Now, NEAR is pivoting into an AI infrastructure platform, with core directions being user-owned AI and decentralized confidential machine learning (DCML), allowing users to use AI services without exposing their data. The experience in decentralized architecture accumulated at NEAR has become the hardest-to-replicate starting point in this direction.

Circle co-founder Sean Neville left to found Catena Labs, positioning it as an AI-native bank. He directly transferred his understanding of stablecoin infrastructure to the AI agent financial scenario, with a16z crypto leading an $18 million seed round.

Nader Dabit, a senior developer on Aave and Lens Protocol, moved to Cognition, bringing his experience in building developer ecosystems across multiple crypto protocols into the AI agent tools space.

What this cohort takes with them isn't just GPU hardware or user networks, but an intuition for mechanism design, experience in building developer ecosystems, and the judgment to construct trustworthy systems from scratch when rules are absent. These capabilities precisely correspond to three structural gaps encountered during AI scaling.

Aggregation and Optimization of Computational Power

Computational power is the most immediate bottleneck for AI scaling. Training and inference require vast amounts of GPUs. Demand is volatile, cloud providers are expensive with queues, and enterprises are reluctant to stockpile hardware themselves. This problem has two facets: how to aggregate and allocate computing power, and how to use the aggregated power more efficiently. Crypto builders have directly transferable experience in both areas.

Hyperbolic addresses the allocation and trust problem. Founder Jasper Zhang brought decentralized mechanism design into the AI computing track: tokens encourage scattered GPU holders to contribute idle computing power. However, the more core issue is trust.

Why should you trust that the computation result from an unknown node is correct? The core innovation, PoSP (Proof of Sampling and Protocol), uses random sampling and game theory to make honesty the dominant strategy for nodes. It requires no full verification, has low overhead, is scalable, and yields reliable results. This mechanism is directly migrated from crypto's logic for verifying the behavior of unknown nodes.

MoonMath addresses the efficiency problem. Its predecessor, Ingonyama, focused on ZK hardware acceleration, multiplying the speed of ZK proof generation under extreme computational constraints.

Its direction has now shifted to a Physical AI performance layer, developing sparse attention acceleration for video diffusion models (LiteAttention), low-rank decomposition of FFN layers (LiteLinear), and training backpropagation acceleration (BackLite). From ZK acceleration to AI inference acceleration, the underlying capability is the same: making math run faster under extreme computational constraints. The track has changed, but the accumulated experience hasn't been wasted.

AI Governance and Incentive Mechanism Design

When multiple AI agents collaborate on tasks, how do you ensure they don't harm the overall system while pursuing their individual goals? Each participant optimizes its own objective function, and no one guarantees the system as a whole functions correctly. Furthermore, agent execution speeds far exceed human intervention windows.

This is the exact type of problem crypto builders have repeatedly addressed in DAO governance and tokenomics design: aligning participants with completely different interests to operate according to the system's preset direction without a central authority. Crypto's answer is economic mechanisms. Malicious behavior incurs real economic costs. Rules are written in code and executed automatically.

EigenLayer has directly transferred this mechanism to the AI context. Through its restaking mechanism, nodes must stake assets before participating in collaboration. Failure to perform or malicious behavior triggers automatic penalties. The rules aren't suggestions; they are rigid boundaries with real economic consequences.

EigenCloud extends this logic to verifiable computation and collaborative governance for AI agents, ensuring agents stay within preset bounds while pursuing their goals. Constraining agents with economic mechanisms is far more reliable than constraining them with ethical guidelines.

AI Agent Autonomous Payments

There is another fundamental problem: how do agents pay? Traditional payment systems are designed for humans. Credit cards require accounts. Bank transfers require authorization. Every step assumes the operator is a human with an identity who will wait. An agent won't wait. It might initiate a massive number of requests per second, each possibly involving micro-payments. Traditional payment pipelines fail in this context.

Stablecoins and on-chain rules are the infrastructure already built by crypto builders. They natively support programmability, permissionlessness, and 24/7 operation. These three characteristics are exactly the hard requirements for the agent payment scenario. What's missing is simply a protocol layer connecting stablecoins to the agent workflow.

x402, launched by Coinbase in May 2025, activates the HTTP 402 status code, embedding stablecoin payments directly into HTTP requests. An agent initiates a request and completes the payment simultaneously. No account needed, settlement in about two seconds.

As of April 2026, the x402 protocol has processed over 165 million transactions with a cumulative volume of ~$50 million, involving 69,000 active agents (source: x402 Foundation). Cloudflare, AWS, Stripe, and Anthropic MCP have all integrated. Agent payments are already a track with genuine traffic.

These three directions correspond to three structural gaps encountered by AI scaling: aggregation and efficiency of computational power, incentive alignment for multi-agent collaboration, and infrastructure for autonomous payments. These problems have no ready-made answers in traditional software architectures, but each has corresponding processing experience in the crypto industry. The capabilities haven't disappeared; they've simply found new hosting scenarios.

The Builder's New Role: From Writing Contracts to Setting Rules for AI

The scaling of AI is creating a functional gap that didn't exist before. It's not a gap in technical talent, but a gap in people who can design trust mechanisms within autonomous systems. As the served entity shifts from humans to AI, the role of the crypto builder is also being redefined.

The table below compares the dimensional changes in the specific functional paradigm:

The core difference between the two paradigms isn't the tech stack, but the method of establishing trust and the logic of rule execution. In the Pre-AI era, crypto builders dealt with human participants. Rules were written into contracts. The margin for error was zero, but the system's boundaries were relatively clear.

In the AI-Native era, when the interacting entity becomes an autonomously running AI agent, the problem to solve becomes: agent behavior is unpredictable, execution speed far exceeds the human intervention window, and the system's boundaries themselves need redefinition under greater uncertainty.

The crypto builder's functional positioning is shifting from "writing secure contracts" to "designing trustworthy mechanisms for AI autonomous systems".

Recruitment at leading institutions already reflects this change:

▲ Core AI/Data positions actively opened by leading trading platforms in Q1 2026

Source: Gate Research Institute

Recruitment by top trading platforms and institutions in 2026 clearly reflects this trend: They are no longer simply hiring AI engineers or crypto developers, but seeking individuals who can bridge the two sides – understanding on-chain incentive distortions and governance game theory while being able to deeply embed AI tools into crypto workflows and design mechanisms that keep agents aligned with regulators and users over the long term.

Capital allocation trends already reflect this judgment. Paradigm is raising a new fund of up to $1.5 billion, expanding its investment scope from crypto to AI and robotics.

Haun Ventures completed a $1 billion Fund II, primarily focusing on financial infrastructure at the intersection of crypto and AI, particularly payment, stablecoin, and agent-to-agent economic systems supporting autonomous AI agent trading and coordination.

a16z crypto completed its $2.2 billion fifth fund (Crypto Fund V), explicitly stating that 100% of the fund will be invested in the crypto space. Facing the complexity and opacity of the AI era, they will focus on application directions leveraging crypto's transparency, verifiability, and decentralization.

Furthermore, according to PitchBook data, in 2025, about 40% of US VC capital flowing into the crypto space went to companies also involved in AI businesses, a significant increase from 2024.

While both involve crypto builders moving into AI, the chosen paths differ significantly depending on the market environment.

In the US, as the regulatory environment becomes relatively clearer, protocol-level innovation has genuine room to survive. The capital network is dense, the path from idea to funding is short, and the margin for error is relatively large.

Projects like Hyperbolic, EigenCloud, Gensyn, and Ritual share a common characteristic: designing new mechanisms from scratch, rather than simply integrating applications on existing systems. Top VCs have clear investment theses for directions like "verifiable computation, agent coordination, and decentralized ML" and are willing to provide ample