2025 DeFi On-Chain Revenue $80 Billion Panorama: Half of Lending Demand Is Borrowing from Oneself

marsbitPublicado a 2026-03-26Actualizado a 2026-03-26

Resumen

This analysis provides a comprehensive breakdown of the $8 billion in on-chain yields generated in DeFi in 2025. The five primary sources are: AMM trading fees ($4.2B), lending interest ($1.76B), perpetual funding rates ($300M), real-world assets ($600-900M), and staking/MEV rewards. A key finding is that approximately half of all borrowing demand is recursive, where users borrow to leverage other yield sources. The report highlights that 58% of stablecoin TVL yields less than 3% APY, below U.S. Treasuries. It also notes that significant potential yield sources like on-chain options and insurance remain underdeveloped. A case study of the Sky protocol (formerly MakerDAO) shows its success in offering a 3.75% savings rate by sourcing ~70% of its yield from off-chain, real-world assets.

Author: Vadym

Compilation: Deep Tide TechFlow

Deep Tide Guide: This is the most complete breakdown of DeFi revenue sources to date—where the $80 billion comes from, which protocols it is distributed across, and how much is circular arbitrage.

The most noteworthy conclusion: about half of the lending demand is recursive, with users borrowing money to chase another revenue source; meanwhile, the 30-day average yield for USDC on Aave is only 2%, and 58% of stablecoin TVL has an annualized yield of less than 3%, lower than U.S. Treasury bonds.

This is the most direct data reference for evaluating the sustainability of current DeFi yields.

Full Text Below:

According to a detailed analysis released by researcher Vadym, DeFi generated approximately $80 billion in on-chain yield in 2025. The analysis provides a complete map of the true sources of DeFi returns. It reveals that yield is not scarce in aggregate, but it is extremely unevenly distributed, often circular, and in many cases difficult to package into structured products.

This result comes as yields across DeFi have narrowed significantly. Borrowing rates on major lending platforms have approached the Federal Reserve's policy rate, and the supply yield for "safe" stablecoins currently averages about 3%—lower than U.S. Treasury bonds and the Secured Overnight Financing Rate (SOFR). On Aave, the 30-day average yield for USDC and USDT is about 2%. The report points out that among the over $20 billion stablecoin pools on Ethereum and its L2s, 58% of the TVL has an annualized yield of less than 3%.

Where Does This $80 Billion Come From

The analysis identifies five main sources of yield, each with different risk characteristics and scale constraints.

AMM trading fees are the largest single category, reaching approximately $4.2 billion, with Uniswap, Meteora, and Raydium accounting for 62% of that total. However, the analysis warns that such fees are extremely difficult for structured products to capture. Liquidity providers—especially those using concentrated liquidity—frequently lose money due to toxic order flow, and LP manager pools have also failed to gain substantial market recognition.

Borrowing interest generated a total of about $1.76 billion across various money markets, involving Aave, Morpho, Spark, Maple, and Fluid. Money markets account for over 60% of DeFi's total TVL, making lending the economic backbone of the industry. However, the analysis found that about half of the lending demand is recursive—users borrow and then cycle into other yield sources, such as liquid staking tokens or yield-bearing stablecoins. In Aave's Ethereum deployment, about 39% of the borrowing demand is used to leverage ETH staking yield, with another 11.6% cycling Ethena's sUSDe.

Perpetual contract funding rates, pioneered on-chain primarily by Ethena, contributed approximately $300 million. Ethena's sUSDe derives yield from staking rewards and short funding rates—a mechanism that garnered both praise and caution when it launched in 2024.

Real-world assets generated an estimated yield of about $600 to $900 million, with U.S. Treasury bonds representing the largest share of the RWA market at about 41%, and private credit accounting for 25%.

Network staking rewards and MEV constitute the remainder, with Ethereum issuing approximately 1 million ETH in 2025. The portion of staking yield coming from MEV continues to decline—private order flow routing currently handles about 90% of swap volume, reducing opportunities for front-running.

Untapped Yield Sources

The analysis also points out several categories where yield capture remains minimal. Insurance underwriting generated only $5.5 million in premiums in 2025, primarily through Nexus Mutual. Options—despite having $30 to $50 billion in open interest on centralized exchanges—have about $1.8 billion in on-chain open interest, with no breakthrough structured products emerging. Volatility selling and protocol risk transfer are largely undeveloped, and the analysis sees this as a potential opportunity as risk management competition intensifies.

Sky's Yield Balancing Act

As a case study of how protocols integrate these disparate yield sources, the analysis examined Sky (formerly MakerDAO). Against the backdrop of yield, its 3.75% USDS savings rate has attracted significant capital. Sky's TVL surged 38% in March, making it the fourth-largest DeFi protocol, with the sUSDS savings pool alone absorbing about $6.5 billion in deposits.

The analysis reveals that about 70% of Sky's revenue comes from off-chain origins—primarily USDC earning Coinbase rewards through the Peg Stability Module (PSM), and RWA exposure through products like BlackRock's BUIDL and Janus Henderson funds. The remaining 30% comes from on-chain sources, with Spark, as Sky's primary capital allocator, directing funds into Sparklend, Maple institutional lending, Anchorage, and other yield-bearing opportunities based on current rates.

The analysis suggests the implication of this structure is: even as traditional finance yields increasingly flow through permissioned channels, their redistribution still occurs on-chain, providing a floor for DeFi interest rates and potentially creating conditions for the next generation of yield derivatives—including fixed-rate products, interest rate swaps, and structured tranche products.

Lecturas Relacionadas

Diálogo entre Sequoia y Jensen Huang: Se avecina un cambio de 60 años en los modelos de computación, no serás reemplazado por la IA, pero podrías sufrir un 'ataque multidimensional' de quienes la usan hábilmente

**Fuente: Capital de Sequoia** **Compilación: Yuliya, PANews** En una conversación entre Konstantin Buhler, socio de Sequoia Capital, y Jensen Huang, fundador y CEO de NVIDIA, se explora la transformación histórica en la computación: desde un modelo basado en almacenamiento y recuperación hacia uno generativo en tiempo real. Huang destaca que estamos ante la mayor construcción de infraestructura de la historia, centrada en las "fábricas de IA", que producen tokens de inteligencia personalizados para cada contexto. Según Huang, el cambio fundamental radica en que, en lugar de simplemente recuperar datos preexistentes, la IA ahora genera respuestas únicas y adaptadas, abriendo paso a agentes capaces de razonar, resolver problemas y realizar tareas útiles. Esta evolución representa la tercera gran red global, después de la energía y las telecomunicaciones: una red de inteligencia que envuelve el planeta. Huang desglosa el ecosistema de IA en cinco capas de inversión: 1. **Energía** – La base, con oportunidades masivas en energías sostenibles. 2. **Chips y computación** – Incluye procesadores, redes y tecnologías como la fotónica de silicio. 3. **Infraestructura** – Terrenos, energía, edificios y operación de centros de datos. 4. **Modelos** – Modelos de lenguaje y sistemas que aprenden no solo texto, sino también leyes físicas, biología, etc. 5. **Aplicaciones** – Startups que transforman sectores como finanzas, derecho, logística y más. Respecto al impacto laboral, Huang rechaza las narrativas alarmistas: la IA no eliminará empleos, pero quienes no la adopten quedarán en desventaja frente a quienes sí lo hagan. Explica que las tareas automatizadas (como analizar radiografías o escribir código) no sustituyen el propósito central de una profesión, sino que potencian su valor. Por ejemplo, los radiólogos o ingenieros de software son hoy más demandados porque la IA amplía sus capacidades y eficiencia. La tecnología, lejos de reemplazar, "eleva" las habilidades humanas, permitiendo a profesionales ofrecer servicios más complejos y creativos. En conclusión, Huang insta a abrazar la IA como una herramienta de empoderamiento, subrayando que el futuro pertenecerá a quienes sepan utilizarla para innovar y resolver problemas, no a quienes teman su evolución.

marsbitHace 21 min(s)

Diálogo entre Sequoia y Jensen Huang: Se avecina un cambio de 60 años en los modelos de computación, no serás reemplazado por la IA, pero podrías sufrir un 'ataque multidimensional' de quienes la usan hábilmente

marsbitHace 21 min(s)

"No necesito un modelo mejor": El panorama de la IA bajo un popular post de Reddit

"Claude Fable 5, el nuevo modelo flagship de Anthropic, marcó un 80.3% en el benchmark SWE-Bench Pro, superando ampliamente a modelos anteriores. Sin embargo, una publicación viral en Reddit titulada 'Claude Fable me hizo darme cuenta de que no necesito un modelo mejor' refleja una desconexión entre las métricas técnicas y la percepción de muchos usuarios. Los comentarios más votados expresan 'fatiga' por las mejoras incrementales, argumentando que modelos como Opus 4.8 ya son 'suficientes' para sus flujos de trabajo diarios. La queja principal gira en torno al costo (el doble que Opus) y, sobre todo, a los estrictos 'guardrails' de seguridad de Fable 5. Usuarios reportan que solicitudes relacionadas con ciberseguridad son rechutadas con frecuencia, siendo derivadas a Opus, lo que genera frustración especialmente entre suscriptores de pago. No obstante, usuarios con tareas complejas y de gran escala, como simulaciones de física de miles de líneas de código, defienden el modelo, describiendo una diferencia sustancial y capacidades superiores para contextos largos y análisis profundo. El debate subraya una posible meseta en la percepción pública de la utilidad de la IA: mientras los benchmarks siguen mejorando, para muchos usuarios el 'techo' de necesidades prácticas ya fue alcanzado. La discusión también señala la brecha entre los modelos de acceso público y las versiones más potentes y restringidas, como Mythos 5, disponible solo para gobiernos y empresas críticas. El futuro de estos modelos públicos dependerá de equilibrar capacidad, costo y usabilidad."

marsbitHace 31 min(s)

"No necesito un modelo mejor": El panorama de la IA bajo un popular post de Reddit

marsbitHace 31 min(s)

Trading

Spot
Futuros
活动图片