AI Momentum Spills into Onchain Agentic Infrastructure

Since the start of the year, a narrow group of AI-aligned tokens has emerged as relative outperformers, decoupling from the broader market and reasserting itself as the highest-beta expression of the global AI trade. 

This rotation is notable not because it is entirely new, but because it follows a deep and prolonged reset. Most AI-related tokens closed 2025 having surrendered the majority of their prior gains.  

After the 2024 frenzy driven by the burgeoning excitement around generative AI and decentralized intelligence, valuations across the segment collapsed. In many cases, prices retraced more than three-quarters of their peak levels as speculative excess gave way to structural oversupply. 

That unwind was not accidental. The first phase of the agentic cycle was defined by experimentation rather than efficiency. Early projects prioritized speed to market, testing whether autonomous agents could be deployed onchain. The goal was to prove feasibility, not durability – a dynamic that drove more than a thousand daily new agents (and the associated tokens) at peak euphoria. 

Capital formation quickly outpaced usage, creating an imbalance that ultimately exhausted demand. And as one would expect in a bubble, the excess eventually bursts. 

The collapse marked the end of narrative-driven proliferation and the beginning of a slower, more selective phase, as marginal projects were flushed out. 

AI bottlenecks repriced 

Today’s renewed interest looks different. Rather than indiscriminate exposure to anything labeled “AI,” capital is rotating toward infrastructure and protocols with clearer functional relevance. The market appears increasingly sensitive to utility, execution, and positioning within the AI value chain. This shift is occurring as enthusiasm around artificial intelligence in traditional markets continues to build. 

The backdrop matters. In equities, AI optimism has migrated beyond GPUs alone and into the less visible bottlenecks that sit further along the stack. As Nvidia’s CEO Jensen Huang recently put it, “the memory bottleneck is severe.” 

As training and inference workloads scale, constraints around memory bandwidth, data throughput and storage have become binding. Hyperscaler deployments now depend not just on accelerators, but on high-speed interconnects and the ability to move and read vast datasets in real time. Forecasts across these segments have been revised higher, and related stocks have repriced accordingly. 

(Source: CoinMetrics, as of 01.08.26 9AM CET)

Crypto markets are beginning to reflect the same reality - not at the narrative layer, but at the infrastructure layer. This time, the repricing is more about a shift away from abstract AI narratives toward a more targeted reassessment of which decentralized networks can plausibly sit downstream of accelerating AI compute and execution demand. 

Agentic thesis still in play 

Despite the volatility of the past year, the core agentic thesis remains intact and is increasingly coalescing into a broader consensus, with “agentic finance” emerging as the core.  

Under this framework, AI agents are evolving from passive tools to operate as active economic participants. 

But for that transition to materialize, agents must do way more than reason or recommend - they must be able to act, transact, and coordinate value independently.  

At the protocol layer, payments remain the critical bottleneck. An AI-first economy cannot function if agents remain dependent on human intermediaries to complete transactions 

This is where new standards such as x402 become structurally important. By enabling stablecoin payment requests to be embedded directly into standard web interactions, x402 extends payments to the same layer that powers the open internet. 

For agentic systems, this unlocks high-frequency, low-value transactions - the kind required for agents to launch services, manage infrastructure and compensate one another. 

That capability is foundational. Once agents can transact autonomously, execution no longer needs to be episodic or human-triggered. Economic actions can be settled programmatically, under predefined permissions, with users retaining control while agents handle execution. 

Over the past year, maturation across the stack – from protocols to tooling - and has allowed agents to move beyond “GPT wrappers” noise into more functional, task-driven systems. Agents are starting to be deployed for concrete use cases such as executing trades across venues or optimizing capital allocations in yield-generating protocols.  

Some of these applications are already well established, but the real driver is the increasingly scalable and reliable operating infrastructure in the background, which market participants are beginning to tentatively reprice higher. 

(Source: CoinMetrics, CoinGecko as of 01.06.26)

Since the year began, spot trading volumes across leading AI-aligned tokens rose from roughly $147mn in consolidated volume to $966mn as of January 6 - a roughly sevenfold increase that reflects a material re-engagement with the segment. 

Render (RNDR) has led the pack, rising roughly +69% and emerging as the strongest performer among AI-tied assets. Virtuals Protocol (VIRTUAL) has followed closely with +60% gains, while Bittensor (TAO) and Artificial Superintelligence Alliance (FET) are up approximately +23% and +38% respectively, albeit with more moderate relative strength. 

Onchain leaders emerging 

Leading the AI tokens cohort, Render captures the hardware and compute leg of the AI trade spilling onchain. 

What began as a decentralized rendering marketplace has evolved into a broader Decentralized Physical Infrastructure Network (DePIN) framework for AI workloads, positioning the network as a distributed alternative to centralized GPU infrastructure. 

The Dispersed platform, released in December, marked a shift toward more flexible, on-demand aggregation of distributed GPU supply, directly addressing the concentration and availability constraints that define today’s compute landscape. 

In that sense, RNDR sits directly downstream of the same forces driving AI equities. As capital has rotated through GPUs, data centers, memory, and broader compute infrastructure in traditional markets, Render is emerging as a potential crypto-native conduit for that demand. 

The protocol effectively tokenizes access to GPU computing power, making it one of the most direct onchain beneficiaries of persistent hardware constraints, rising inference and training costs tied to AI workloads’ expansion. Render’s burn-and-mint equilibrium model (BME) makes the structural link between supply dynamics and growing demand for compute - incentivizing GPU providers to contribute capacity as usage scales. 

Operational metrics reflect that momentum. Over the past year, Render’s two-sided GPU marketplace expanded sharply, posting a 87% growth. 

Available rendering and compute capacity increased by roughly 40%, supported by the launch of a dedicated compute subnet – a specialized independent segment of the network - now actively processing generative AI workloads. 

Closely following Render in relative performance, Virtuals Protocol has re-emerged as a focal point for agent-centric activity. Its positioning reflects not just renewed interest in AI agents, but a structural response to the failures of the previous cycle.

The initial wave of agent launches exposed a core limitation: no single launch model could serve the full spectrum of builders. Early experimentation, introduced in 2025 by Virtual’s first iteration Genesis, emphasized fairness at scale by prioritizing contribution over capital.

While it succeeded in broadening access and improving transparency, the model ultimately fell short - fairness alone did not generate lasting conviction, nor did it provide a viable path for sustained capital formation.

On 5 Jan, the introduction of a modular framework built around three complementary launch paths - Pegasus, Unicorn, and Titan - marked a corrective step. Together, they are designed to support different stages of agent development, from early experiment to established products, without fragmenting liquidity or ownership.  

Pegasus focuses on distribution and discovery at the earliest stage, Unicorn remains the conviction-driven path, balancing open participation with structured capital formation and accountability, and Titan is tailored for credible teams operating at scale, with higher readiness standards and cleaner market entry.

Following the upgrade, Virtuals-tied transaction volumes have strongly rebounded alongside renewed interest in the AI agents sub-sector. 

Since the year began, Virtuals launchpad volume has risen from $13.86mn to $49.31mn, while decentralized exchanges (DEX) volumes jumped from $9.26mn to $52.4mn, with activity spiking around the announcement. 

(Source: TokenTerminal)

Fundamental metrics paved the way. Platform revenues reached approximately $1.16 million over the past month, up nearly 96% from prior levels, pointing to a meaningful pickup in activity. That said, context remains important: revenues are still down 93% from their January 2025 peak of $15.8 million, mirroring the broader collapse in agent demand and the 90% drawdown in tokens' valuations that followed. 

Also, activity is highly concentrated. According to the Virtuals dashboard, a single agent - Ethy AI, an onchain swap and trade execution agent on Base - accounts for roughly half of the protocol’s agent GDP (the total value processed by agents through trading and service fees) and close to 20% of total launchpad revenues over the past month. 

The open question, then, is whether recent performance reflects early pricing of a structurally improved growth trajectory enabled by the platform’s upgrades, or a more opportunistic, event-driven rebound driven by narrow participation. For now, breadth remains limited, making this a development the market will have to monitor closely. 

Bittensor occupies a distinct position within the AI-linked crypto stack, sitting above raw compute and agent deployment. Where protocols like Render address hardware constraints and Virtuals focuses on agent creation, TAO targets the intelligence layer itself.

TAO sits above raw compute, incentivizing intelligence production, model contribution, and agent coordination. As agentic systems increasingly become the dominant deployment model for AI, TAO emerges as a natural onchain settlement and incentive layer.

Leading TAO subnets are already showing early real-world use cases. In a recent analysis of the decentralized AI training networks, Jack Clark (Anthropic co-founder and formerly of OpenAI) highlighted Covenant AI’s Templar – a TAO subnet focused on incentivized internet-wide training – as the “largest active network".

That positioning has taken on renewed relevance following a key supply-side inflection. 

In mid-December, Bittensor underwent its first halving, reducing daily TAO emissions to roughly 3,600 tokens. With a fixed supply cap of 21mn TAO, the halving materially reduced the volume of new tokens reaching the market, tightening near-term supply at a moment when institutional interest has been building.

The emergence of TAO-focused digital asset treasury vehicles and the filing of spot ETF applications have brought incremental demand into focus. ETF-related flows have proven to be a powerful driver of spot liquidity across crypto this cycle, raising the prospect of a supply shock forming against a sharply lower issuance backdrop. 

Through the Artificial Superintelligence (ASI) Alliance - a collective formed by Fetch.ai (FET), SingularityNET, and CUDOS – the stack focused on a more practical constraint: how agents actually run workloads and settle transactions in production environments.

The recent end of ASI Cloud's beta in late December marks a transition into production. The platform now provides developers with permissionless access to enterprise-grade GPU infrastructure and AI inference services, while abstracting the fragmentation that typically defines AI infrastructure sourcing - a shift that answers demand for scalable, flexible alternatives to centralized cloud providers.  

More importantly, the final pillar required for agentic economic activity is around the corner. Fetch.ai’s upcoming AI-to-AI payment system, expected to launch in early 2026 within the ASI One platform, is designed to close the long-standing execution gap. 

The system will enable autonomous coordination between personal and business AIs, allowing agents to execute bookings, settle payments, and act on opportunities in real-time, including offline actions using pre-approved user funds. 

The architecture bridges existing financial infrastructure with onchain settlement. 

Transactions can be executed via traditional card rails - leveraging Visa infrastructure with single-use credentials for security, with Mastercard support expected - as well as onchain using USDC or FET token. Payments will be mediated through Fetch’s Agentverse identity layer, ensuring agents act on behalf of identifiable users or entities. 

Taken together, these layers push the boundaries of the agentic stack, creating the conditions for agentic finance to function in practice rather than theory. And the timing is notable. Nvidia’s latest GPU allocations are effectively sold out through 2026, with extended waitlists forming across hyperscalers. 

Against that backdrop, access to alternative compute paths has become a binding constraint for AI deployment - a potential medium-term tailwind for the decentralized AI segment.