Unlocked by io.net

    This research report has been funded by io.net. By providing this disclosure, we aim to ensure that the research reported in this document is conducted with objectivity and transparency. Blockworks Research makes the following disclosures: 1) Research Funding: The research reported in this document has been funded by io.net. The sponsor may have input on the content of the report, but Blockworks Research maintains editorial control over the final report to retain data accuracy and objectivity. All published reports by Blockworks Research are reviewed by internal independent parties to prevent bias. 2) Researchers submit financial conflict of interest (FCOI) disclosures on a monthly basis that are reviewed by appropriate internal parties. Readers are advised to conduct their own independent research and seek advice of qualified financial advisor before making investment decisions.

    Decentralized Compute Networks: Scaling Global Infrastructure

    Nick Carpinito

    Key Takeaways

    • Rapid generative AI adoption through frontier reasoning models like OpenAI's o3 and DeepSeek R1 is exponentially increasing compute demand.
    • Inference and real-time reasoning workloads have created compute bottlenecks, significantly exceeding the training costs and driving demand for decentralized compute.
    • Decentralized compute networks offer cost advantages and flexibility, but must overcome latency, trust, and integration barriers to attract enterprise workloads.
    • Geopolitical risks and hardware shortages, highlighted by tensions around semiconductor tariffs, further underscore the strategic value of decentralized compute infrastructures as an essential resilience mechanism.

    Subscribe to 0xResearch Newsletter

    The Compute Crisis: Generative AI Adoption Soars into 2025

    2025 has facilitated unprecedented growth in generative AI adoption. By late 2024, over 71% of companies were using generative AI in some form, a massive jump from 33% the year prior. Tools like ChatGPT, Perplexity, and open-source models have proliferated, with potentially billions now leveraging AI chat bots. In Q1 2025, OpenAI’s release of the o3 & o4 models marked another leap, touted as a multimodal “reasoning” model capable of breaking responses into smaller components, testing different approaches, and validating solutions until reaching a final output. Meanwhile, open-source communities pushed forward with models like DeepSeek-R1, a 671B-parameter LLM that achieves comparable performance while maximizing efficiency. Jensen Huang, Nvidia’s CEO, noted that AI adoption is driving a 10x increase in data center requirements over the coming years, as companies rush to offer AI-enabled products and services. These advancements have made AI more reliable and accessible but present inherent tradeoffs due to rapidly increasing compute requirements.

    image (1).png

    Flagship reasoning models from OpenAI and DeepSeek perform similarly on standard benchmarks. Source: DeepSeek

    Crucially, as AI capabilities expand, usage grows exponentially rather than linearly. Each new model unlocks novel applications, which brings new users and more demand for computational power. This virtuous cycle, where increased use incentivizes the development of even larger models, further accelerating usage, creates an AI-Compute Flywheel. This trend increasingly indicates that those who control GPU capacity and the semiconductor supply chain hold tremendous economic and geopolitical sway. Geopolitical tensions, particularly around heavy reliance on Taiwanese semiconductor production through TSMC, further underscore this vulnerability. Recent tariffs imposed by the Trump administration, particularly 32% on Taiwanese tech imports with combined tariffs now approaching 54% on key AI hardware components, significantly increase costs and lead enterprises to reconsider their supply chains. Although TSMC is investing in U.S.-based fabs, geopolitical tensions could still escalate into broader disruptions, impacting AI growth and infrastructure stability.

    Nick Compute copy.png

    Reasoning Models & Post-Training Compute: A New Bottleneck

    2025 is marking a turning point in how AI models are deployed and scaled through increasingly sophisticated post-training techniques. OpenAI’s o3 and emerging Web2 tools show that models are no longer simply generating text but are instead reasoning by breaking complex problems into substeps, testing different solution paths, and validating responses dynamically. This ability to perform structured reasoning, especially in domains like law, math, and coding, unlocks new markets and places new demands on infrastructure.

    image copy 2.png

    OpenAI o3 usage consumes >2x more tokens than its predecessor, o1. Source: Effective Altruism Forum (Magnus Vinding)

    In contrast to static, one-shot inference, reasoning models require dynamic, multi-stage inference, often running multiple parallel passes, recursive calls, or ensemble methods to arrive at a final output. This can lead to test-time compute costs that rival or exceed training, especially as search-based methods like Best-of-N sampling, Beam Search, or Diverse Verifier Trees (DVTS) become standard. Industry estimates put fully loaded inference bills at up to 15x the original training spend. Because inference keeps the model “always‑on,” the bill scales directly with user interactions, and the latency tolerance of user‑facing apps is merciless. Hugging Face and others are showing that such techniques can significantly improve reliability, but only by embracing parallelism at inference time.

    image copy.png

    Various recursive post-training behaviors dramatically increase computational load for inference tasks leveraging reasoning models. Source: Medium (Isaak Kamau)

    This shift is colliding with a strained GPU ecosystem. In 2024, startups were forced to “audition” for cloud GPU access, with providers demanding equity or business justifications for capacity. H100 GPUs, critical for training and inference, saw 6+ month wait times early in the year, and even as production ramped, demand from hyperscalers like Meta and Google outpaced supply. While Nvidia planned to ship up to 2 million H100s and AMD’s MI-series added hundreds of thousands more, most of this capacity was immediately absorbed.  While supply constraints have comparatively eased, pressures still remain today,  with pricing pressure alienating smaller labs.

    Latency now becomes paramount. Models are delivering interactive performance at human timescales, but achieving low-latency, high-throughput inference across decentralized, heterogeneous nodes remains a major hurdle. When even single-shot inference is latency-sensitive, multi-shot reasoning compounds the problem. Web2 incumbents blunt much of that latency with vast edge networks and CDNs (Content Delivery Networks). A purely decentralized mesh may place nodes in sub‑optimal geographies or behind fickle consumer ISPs. Regional DePIN (Decentralized Physical Infrastructure Network) clusters or true edge‑side inference could close the gap, but as of 2025 fully decentralized real‑time inference at scale remains elusive, so most DePIN volume still sits in training, batch inference, or other latency‑tolerant jobs. Nvidia’s 2025 GTC keynote underscored this with the announcement of the Blackwell Ultra platform, an AI supercomputer blueprint to handle large models with increasingly low latencies leveraging Blackwell GPUs and NVLink switches.

    Screenshot 2025-05-15 at 9.19.17 AM.png

    Reasoning models provide superior response accuracy, but consume significantly more tokens than traditional LLMs. Source: Nvidia

    To stay competitive, decentralized compute networks must:

    • Prioritize low-latency job scheduling, especially for inference workloads.
    • Enable dynamic scaling, allocating more GPUs on demand for complex prompts.
    • Deploy regional clusters and edge nodes to reduce geographic latency.
    • Explore hardware specialization, such as consumer-grade M-series chips.
    • Adopt inference-specific orchestration

    Unlike training, post-training is less reliant on massive clusters and more on dynamic, flexible deployment. By reliably serving smarter, multi-step inference, without needing to audition for GPUs, decentralized networks can fill critical gaps in a strained AI infrastructure.

    As adoption of reasoning models accelerates in Web2, from legal AI copilots to enterprise math assistants, DePINs must evolve from a training-centric paradigm to one that reflects the new bottleneck: post-training compute at scale.

    The DePIN Compute Stack

    Our 2024 decentralized compute report outlined a three-layer model of decentralized compute networks. Over the past year, this model has proven useful for mapping the evolving landscape:

    Bare-Metal Layer networks focus simply on amassing and exposing raw compute resources via APIs or marketplaces with minimal workload-specific optimization. In 2025, we’ve seen bare-metal marketplaces expand significantly. These bare-metal networks often use straightforward token incentives, minting tokens for every hour a device is online with penalties for downtime, to attract providers globally, essentially monetizing idle hardware. While critical to enabling more advanced functionality, the bare-metal layer accrues the least value due to its positioning as a commodity market with thin margins. 

    Orchestration Layer networks add use-case-specific services, taking commodity hardware and making it usable for a target user base. Orchestration-layer DePINs tend to derive value from job fees, while their tokens often carry governance or staking roles to maintain network security and performance. These projects are effectively cloud platforms built on decentralized hardware. They compete by offering similar convenience to cloud at lower cost, or enabling novel collaborative workloads. Render, specializing in media rendering and inference workloads, and inference.net, focused on LLM inference from open-source models, are key examples of networks operating in this layer.

    Aggregation Layer networks are still nascent but crucial. Aggregators aim to unify multiple networks and resource types under a single umbrella. Instead of a narrow focus, an aggregator provides a cloud-like interface that can route workloads to various decentralized providers underneath. In 2024 we noted that this is a high risk/high reward endeavor, but remains the holy grail in decentralized compute. The value of this layer lies in breadth and flexibility, but inherits the technical challenges of each underlying network and the complexity of federation. If successful, an aggregation DePIN could become a comprehensive compute marketplace, abstracting away which specific network or provider runs a job. It could also enforce novel scheduling policies, such as splitting workloads for reliability or combining hardware from different networks in a single pipeline. Notably, io.net operates an aggregation layer network leveraging its underlying bare-metal and orchestration layers to serve inference functionality, cluster deployment, demand aggregation from networks such as Render, Filecoin and Aethir, on-demand rentals and container deployments, creating a robust service offering with forward looking plans to support Kubernetes, Unreal Engine and Unity workloads.

    No single DePIN has yet “won” any layer outright, but many have found footing in particular domains. Market differentiation has emerged primarily around targeted services, community ecosystems, and distinct incentive models, rather than sheer scale or network size alone. The evolving dynamics across layers suggest that decentralized compute networks may coexist competitively in segmented verticals rather than consolidating into monopolistic entities. Ultimately, the future landscape will likely feature interoperable ecosystems rather than a single network claiming comprehensive superiority.

    Matching Workloads to Hardware (and DePINs to Workloads)

    AI & cloud workloads differ vastly, and so do hardware capabilities. Recognizing this, the decentralized compute ecosystem has begun to segment by workload niche, aligning each job type to an optimal network or hardware set. In this context, Hugging Face's exploration of test-time compute scaling provides valuable strategies for optimizing inference workloads. By implementing search-based methods during inference, models can achieve higher accuracy and efficiency, which is particularly beneficial for decentralized networks aiming to maximize performance on heterogeneous hardware.

    Large-scale training jobs require many GPUs working in parallel, frequent communication, and weeks to months of runtime, typically facilitated via A100 or H100 GPUs connected with high-bandwidth interconnects. Decentralized networks can support this, but only if they can provide a cluster with high-bandwidth interconnects or clever parallelism strategies. Prime Intellect’s distributed training framework demonstrated mid-size model training across continents by cleverly partitioning the model. Traditional projects choose to partner with data centers to get close to data-center-grade networking for the duration of the training. Still, the most reliable way to train a cutting-edge model remains via reserved spots on established superclusters. 

    Batch inference & rendering involves rendering CGI frames, running batched ML inference on a dataset, or scientific simulations, where the primary concern is throughput per dollar, instead of inter-GPU latency. Thus, using a mix of consumer and data center GPUs from a DePIN can drastically lower cost while achieving the throughput needed. Render shines for CGI rendering where each frame or job can go to an independent GPU.

    Nick Compute copy 4.png

    Real-Time inference and interactive applications include latency-critical tasks like powering a live chatbot, running an AI agent that interacts with a user in real-time, or cloud gaming. Decentralization in real-time solutions is challenging, but not impossible if cleverly executed.

    • Edge Networks deploy nodes in many locations and connect users with the closest server to their location.
    • Model Compression uses decentralized networks to train or fine-tune large models, but then serve distilled smaller models to edge devices. Web3 can facilitate this by allowing many contributors to help compress and validate the model, then pushing it out to endpoints.
    • Specialized Hardware: FPGAs or ASICs could be introduced in DePINs for inference. If a provider has an edge TPU or a dedicated inferencing ASIC at home, they could serve lots of requests with low latency. As companies produce AI inference chips for on-premise use, those could trickle down into the DePIN supply pool. While centralized clouds currently maintain an edge through access to proprietary or specialized accelerators, such as Google’s TPUs, which can outperform general-purpose GPUs by up to 5x on certain workloads, this advantage may diminish over time.
    • Apple Silicon's Role: Apple's M-series chips utilize high-performance Neural Engines and energy efficiency, and are increasingly being utilized for inference tasks. EXO Labs has developed an open-source framework that enables the distribution of AI workloads across a network of Apple Silicon based devices. This approach allows for the efficient execution of large models on consumer-grade hardware, effectively transforming a collection of everyday devices into a powerful AI cluster.

    Currently, most interactive AI services still run on centralized servers, but a future vision might be a peer-to-peer network where your requests get routed to an optimal node for quick responses. The gap to achieve this involves both network coordination and trust. Efforts in fully-homomorphic encryption or secure enclaves could allow a node to run your request without being able to snoop on the data or model, which would alleviate privacy concerns and make decentralized real-time AI more viable.

    High-Performance Computing (HPC) & specialty workloads are necessary for use cases such as physics simulations, bioinformatics, or video encoding. These often run on CPU clusters and require large memory or storage bandwidth. For these workloads, matching hardware is key. A network might specialize in CPU cycles and offer a supercomputer alternative for tasks that can be split up. In crypto, Render’s parent company, OTOY, has hinted at facilitating protein folding tasks on their network in the future. 

    Our takeaway is that decentralization will not uniformly penetrate every corner of computing at once. Certain workload–hardware combinations are low-hanging fruit, and we see strong traction there, while others require further innovation to handle efficiently. Over time, as networks prove reliability and performance in one niche, they tend to expand outward. For example, Render is now venturing from rendering into AI inference, and Akash, initially CPU-oriented, added GPU support as demand for AI grew. We anticipate a network-agnostic convergence, where aggregators could automatically route a training job to the best candidate. 

    Decentralized vs. Centralized Cloud: Cost and Performance Benchmarks

    The growing interest in decentralized compute networks is driven by the sharp cost differential with centralized cloud providers and a steadily improving performance profile. Evaluating these networks hinges on understanding the tradeoffs in price, performance, trust, and compliance.

    Nick Compute.png

    Cost efficiency is where decentralized systems offer their most compelling advantage. For example, renting an NVIDIA A100 costs about $0.75/hour on decentralized marketplaces, compared to $2.50–$3.00/hour on AWS on-demand, with AWS spot prices rarely dipping below $1/hour in high-demand areas. The gap widens with newer hardware like the NVIDIA H100, which averages $2/hour on decentralized platforms versus around $12/hour on AWS. These savings stem from decentralized pricing models based on electricity and capital costs, rather than artificial scarcity or service premiums. While CPUs and general resources are also cheaper, the greatest disparities are seen in GPU pricing. Storage and memory are available but secondary in focus. For an nascent AI startup running just four A100s for model training 24/7, switching from AWS on-demand to a decentralized alternative could reduce monthly GPU costs from ~$8.6k to ~$2.1k, allowing operators to speed up training by renting more capacity, or reallocate funds elsewhere, making DePIN an attractive option for resource-constrained teams.

    Performance and throughput are less critical for single-machine tasks like training a model on one GPU where performance is identical to hardware capability, while multi-node tasks introduce overhead. Centralized providers like AWS use high-bandwidth, low-latency interconnects for efficient GPU synchronization, while decentralized systems often rely on slower 1 GB/s Ethernet, which can slow training by 2–3x. Mitigations include gradient compression, asynchronous updates, edge acceleration, and job placement optimization (grouping jobs on nearby or directly connected nodes). Over time, these improvements may allow decentralized systems to close the performance gap for multi-GPU training.

    Security and trust are key concerns in decentralized computing. Trusted Execution Environments (TEEs) enable secure enclaves where code runs without exposing data to the host OS. Efforts such as the Nexus Verifiable AI Lab are pioneering verifiable AI pipelines leveraging cryptographic proofs and secure hardware to enhance transparency and compliance. While GPU TEEs are still early-stage, vendors like Nvidia are beginning to support confidential GPU computing. For highly sensitive tasks, decentralized networks may use secure multi-party computation (MPC) methods, adding privacy but reducing performance. Onchain cryptographic proofs, including systems like ZK-SNARKs (Zero-Knowledge Succinct Non-interactive Argument of Knowledge), offer another approach by verifying execution without revealing data. Though still emerging, these tools could significantly reduce trust requirements in decentralized systems.

    Data egress costs and compliance frameworks present another differentiator. AWS is known for high outbound data transfer fees, whereas many decentralized protocols absorb or eliminate such costs. However, compliance remains a critical barrier. Industries like healthcare impose strict data residency and processing rules that decentralized networks currently struggle to meet. Until robust compliance frameworks are implemented, regulated users may prefer centralized providers with well-established certifications and controls. io.net has achieved early traction in this direction through their achievement of SOC 2 compliance in April 2025.

    While not yet a full substitute for centralized clouds in every domain, especially where compliance or high-speed interconnects are critical, the trajectory of the compute DePIN sector mirrors the early days of cloud computing. 

    Web3-Native AI Models: Progress and Gaps

    The convergence of AI and Web3 has given rise to Web3-native open-source AI efforts, with Nous Research and Prime Intellect representing two of the most compelling experiments.

    Nous Research emerged in 2024 as a decentralized AI lab, with a mission to create open-source, large-scale AI models governed by its community. With a hefty $50M Series A warchest, they aim to challenge closed models from OpenAI and Google. Leveraging Solana, Nous coordinates thousands of global contributors who provide computational resources and curate datasets, pioneering a model where AI systems are partially owned by contributors rather than a single corporation.

    Recently, Nous announced plans to pretrain a 40B parameter model named "Consilience" using their newly introduced Psyche framework, an advanced decentralized infrastructure that builds upon their existing DisTrO system. Psyche significantly reduces data transfer and latency by compressing gradient updates via a JPEG-inspired discrete cosine transform (DCT), enhancing scalability and GPU utilization across distributed hardware. Despite ongoing challenges such as VRAM limitations and hardware heterogeneity, Nous's community-centric approach and technical innovations position them to scale efficiently, moving closer to rivaling traditional, centralized AI labs.

    682108892e5b8c350cb40532_benchmarks.png

    INTELLECT-2 provides comparable, and often slightly superior, performance to its open-source competitors. Source: Prime Intellect

    Prime Intellect has advanced decentralized AI with the release of INTELLECT-2, a 32-billion-parameter language model trained via globally distributed reinforcement learning. Building on the success of INTELLECT-1, which demonstrated the viability of decentralized training for large models, INTELLECT-2 leverages a fully asynchronous RL framework called PRIME-RL. This framework decouples rollout generation, model training, and weight broadcasting, enabling efficient training across a heterogeneous network of permissionless compute contributors.

    Key components include a locality-sensitive hashing scheme ensuring verifiable inference across diverse GPU hardware, and a library for efficient distribution of updated model weights via a HTTP-based tree-topology network. INTELLECT-2 is open-sourced, with the model, training code, and datasets available for the community. Developers can interact with the model via a chat interface and contribute compute resources through the protocol testnet.

    InfraFi: A GPU-Powered Financial Product

    DePIN-Fi, the financialization of decentralized resources is emerging via GPU Income Notes (GINs). In this model, operators or DAOs bundle GPUs into a pool, deploying them across decentralized compute networks to maximize utilization. The revenue generated flows into a smart contract, and tokens are issued, representing a claim on the future cash flows of the GPU pool. Each token is backed by the physical GPUs as collateral, providing tangible asset backing. This approach offers potentially high yields, though actual returns depend on utilization rates and market demand. Some projects, such as USD.AI and GAIB, are beginning to structure synthetic dollar tokens that are backed by AI hardware and compute resources, providing an additional financial layer on top of decentralized compute infrastructure. These synthetic assets allow capital to flow more efficiently into hardware ecosystems, offering exposure to markets without direct hardware ownership.

    However, several risks need consideration:

    • Utilization Risk: Yields are contingent on GPU utilization. A drop in demand can reduce utilization rates, leading to lower returns.
    • Hardware Risk: GPUs can fail or become obsolete. Maintenance and upgrades are necessary, and a portion of revenue might need to be allocated for hardware depreciation.
    • Market Risk: The value of USD.ai tokens can fluctuate with crypto market conditions and the actual earnings of the GPU pool.
    • Regulatory Risk: Given their nature, such tokens might attract regulatory scrutiny and could be classified as securities in certain jurisdictions.
    • Operational Risk: Active management is essential to optimize GPU deployment and maintain hardware, requiring efficient DAO governance or automated protocols. Models like USD.AI's may mitigate some operational exposure through centralized ownership or "sale-leaseback" structures, where the platform owns the hardware and leases it to data centers, aligning incentives and insulating retail investors from direct operational complexity.

    To mitigate these risks, strategies include:

    • Diversification: Using a mix of GPU models and geographic distribution can spread risk and capture demand across different markets.
    • Tokenomics Design: Implementing a two-tier token system, with one representing stable earnings and another capturing growth potential, can cater to different investor risk appetites.
    • Insurance: Setting aside reserves or purchasing insurance can cover hardware failures and other unforeseen expenses. Protocols may also explore insurance-backed synthetic assets or reserve pools funded through network fees, adding stability to token value.

    As decentralized compute grows, the concept of Protocol Owned Compute, where protocols directly acquire, deploy, and manage compute infrastructure, may further align incentives between networks and token holders, offering both stability and capital efficiency. DePIN-Fi products could inject liquidity and accelerate growth in decentralized compute.

    Landing Web2 Workloads: Meeting the Enterprise Where They Are

    While crypto-native projects were first to experiment with decentralized compute, the most transformative near-term growth will likely come from Web2 enterprises. Companies deploying legal copilots, customer support agents, or internal productivity tools are under pressure to contain cloud costs while maintaining performance and reliability. For these organizations, decentralized compute offers an economic and strategic hedge.

    Web2 buyers care about:

    • Reliability and SLAs: Traditional cloud providers offer clear SLAs, support, and consistent reliability. For mission-critical AI workloads, startups often prefer established clouds or specialized providers.
    • Tooling and Integration: Developers are accustomed to tools like AWS, Docker, and PyTorch. Although some decentralized platforms offer familiar interfaces, the ecosystem is still less mature compared to AWS's comprehensive documentation and SDKs.
    • Perception and Trust: Established cloud providers benefit from brand trust. Enterprises often question data safety and job reliability when using new networks or unknown servers.
    • Cost Efficiency: User‑facing inference can outstrip training budgets by multiples, so any added latency or under‑utilization hits the bottom line immediately.
    • Network Liquidity: Ensuring that there is always a sufficient pool of resources available to meet workload specifications.

    DePIN networks are beginning to address these requirements in practical ways:

    • Inference endpoints from io.net and inference.net replicate the simplicity of calling OpenAI APIs, enabling drop-in replacement for LLM inference with lower cost and higher configurability.
    • Container orchestration allows enterprises to deploy custom environments with fine-tuned models or proprietary agents, without hardware procurement.
    • Verification layers such as staking, peer scoring and slashing are emerging to ensure job integrity and uptime, reducing trust assumptions in a decentralized network.
    • Edge acceleration and placement strategies aim to bring resources closer to the user, especially vital for apps like AI customer chat or real-time code assistance.
    • Confidentiality techniques, including secure enclaves and encrypted inference, are in early deployment, allowing sensitive workloads to run without revealing data.
    • Supply & demand balancing in decentralized networks often uses token-based incentives, adjusting reward mechanisms based on network usage.
    • Bi-directional participation is a key feature of decentralized compute ecosystems, allowing data centers to monetize idle resources and recoup sunk costs or apply revenue to cloud instances. 

    Already, some enterprise teams such as Paramount Game Studios and Blender are experimenting with batch inference, fine-tuning, and model distillation on decentralized networks as a cost-saving measure. Over time, as performance and compliance tooling improves, real-time workloads like reasoning agents and legal copilots could follow.

    What will accelerate adoption is abstraction and orchestration: Web2 developers want the power of decentralized compute without needing to manage token wallets, node reputation, or network quirks. Aggregators like io.net and orchestration layers like Render are approaching feature parity with existing cloud offerings, optimized for the cost-sensitive, inference-heavy, latency-aware needs of Web2 AI workloads.

    Enterprises are hungry for affordable, flexible compute. The challenge for DePINs is delivering it without requiring a rewrite of the enterprise stack. Those who bridge that UX gap first, while scaling inference with performance guarantees, stand to win a growing share of the AI infrastructure market.

    Risks

    The decentralized compute infrastructure landscape currently faces significant geopolitical, operational, and economic risks. Foremost among these is geopolitical exposure stemming from reliance on global semiconductor supply chains, primarily centered around Taiwan and TSMC.

    These geopolitical tensions are compounded by hardware-specific risks. Rapid GPU innovation cycles mean decentralized compute networks continually face potential obsolescence, necessitating constant reinvestment and complicating financial sustainability. Moreover, decentralized networks rely on globally sourced, heterogeneous hardware, introducing variability in component quality and reliability, potentially increasing failures and performance issues.

    Economically, decentralized compute networks also face significant volatility risks, particularly in relation to their financialized products such as GPU Income Notes (GINs) or synthetic tokens backed by compute resources. Crypto market downturns or reduced utilization rates could sharply diminish returns, reducing incentive structures critical to network participation. Misalignment of compute resource supply with fluctuating enterprise demand could exacerbate underutilization issues, impacting revenue and undermining long-term network viability.

    Operationally, decentralized networks must manage growing complexity, particularly as workloads shift towards dynamic, real-time inference tasks involving sophisticated multi-step reasoning processes. The orchestration of these latency-sensitive tasks across decentralized infrastructure can lead to execution inefficiencies, service degradation, or increased latency spikes. Troubleshooting and fault isolation in distributed networks present additional operational hurdles, complicating rapid remediation and potentially prolonging outages.

    Enterprise adoption of decentralized infrastructure introduces further risks around trust, reputation, and verification. Enterprises remain cautious about adopting decentralized compute for mission-critical workloads due to perceived reliability and security risks compared to established cloud providers. Failures in verification mechanisms such as staking, peer-scoring, or node slashing could severely damage network reputation, undermining adoption and growth.

    Security and confidentiality are ongoing challenges. Decentralized nodes, reliant on consumer-grade and potentially compromised hardware, face risks of supply chain attacks or unauthorized access, potentially exposing sensitive data. Although emerging solutions such as confidential computing and FHE show promise, these technologies remain nascent and unproven at scale, creating regulatory compliance and data sovereignty challenges.

    Environmental concerns also loom large for decentralized compute networks. Rising scrutiny of the significant energy consumption associated with GPUs could attract regulatory intervention, potentially increasing operational costs or imposing stricter sustainability mandates. Additionally, frequent hardware turnover risks amplifying electronic waste concerns, creating reputational risks and possible regulatory burdens related to disposal and recycling.

    Lastly, the rapid proliferation of competing decentralized networks risks market fragmentation, diluting liquidity, interoperability, and overall efficiency. Aggregation-layer solutions, aiming to unify these fragmented resources, must overcome considerable technical complexity to effectively manage heterogeneous underlying infrastructures, creating additional points of failure and operational challenges.

    The Road Ahead

    The decentralized compute landscape is rich with projects, each with different philosophies and technical approaches. In closing, it’s worth summarizing where things stand and what to watch:

    • Adoption is Steadily Growing: Real usage metrics show upward trends. Render’s outputs tripled year-over-year while Akash’s deployed apps and GPU lease counts are rising monthly. These are still small compared to AWS, but the growth rates are impressive.
    • Community and Ecosystem: The success of these networks will depend on building an ecosystem of tools, users, and skilled contributors. Hackathons are popping up for dApps that utilize decentralized compute. As more such end-user applications succeed, they drive demand back to the networks.
    • Competition from Incumbents: We should expect Big Cloud to respond to distributed offerings with lower prices or programs to co-opt decentralization principles. Multi-billion-dollar giants face the innovator’s dilemma, with decentralization undermining their centralized control and high margins.
    • Network Specialization: DePINs should be prioritizing task specialization and delivering a comparable or better experience compared to their web3 counterparts.
    • Synthetic Data Accelerates: Synthetic Data as a Service will become a powerful application for DePINs, enabling rapid creation of specialized datasets through cost advantages and flexibility. 

    DePIN’s compute subsector stands at the intersection of an AI revolution and the decentralization movement, two of the most powerful trends of our time. Our continued coverage showcases the narrative progressing from vision to early reality. The road ahead will have challenges in technology, economics, and governance, but the momentum suggests that these networks could become a prevalent force in the AI infrastructure landscape if suppliers, orchestrators and aggregators can deliver a familiar experience while preserving the benefits of distributed networks.

    This research report has been funded by the io.net. By providing this disclosure, we aim to ensure that the research reported in this document is conducted with objectivity and transparency. Blockworks Research makes the following disclosures: 1) Research Funding: The research reported in this document has been funded by the io.net Foundation. The sponsor may have input on the content of the report, but Blockworks Research maintains editorial control over the final report to retain data accuracy and objectivity. All published reports by Blockworks Research are reviewed by internal independent parties to prevent bias. 2) Researchers submit financial conflict of interest (FCOI) disclosures on a monthly basis that are reviewed by appropriate internal parties. Readers are advised to conduct their own independent research and seek the advice of a qualified financial advisor before making any investment decisions.