AI tokens are already genuinely working on-chain, but the market is still pricing them using outdated logic.
Authors: Conviction and Kioto (Sparsity Data)
Edited by TechFlow
TechFlow Editor's Note: In January 2025, the market capitalization of the AI Agent token sector peaked at $15.7 billion, only to plummet nearly 80% to roughly $3.65 billion. Market consensus has been to discard the entire track. However, this Sparsity Data report argues that this "sector-wide discount" masks fundamental technological shifts: thanks to open-source LLMs slashing inference costs by 95%, autonomous execution duration jumping from 20 minutes to 12 hours, and the maturation of on-chain infrastructure like Uniswap v4 Hooks, next-generation tokens represented by Pearl, POD, Surplus, ORBIO, and IMD are no longer merely chatbot shells chasing Twitter attention. Instead, they have truly integrated proof-of-compute, inference resale, trading fee conversion, and automated agent swarms into on-chain closed loops. While not every token captures real economic value, conflating "protocols actually executing substantive work" with "air narratives from the previous cycle" creates massive pricing discrepancies for sharp researchers.
Most AI tokens from the previous cycle were merely chatbots wrapped with a token ticker. Tokens in this cycle are already doing real work on-chain: Pearl, POD, Surplus, ORBIO, and IMD, along with the currently missing links.
At their peak in January 2025, AI Agent tokens held a combined market cap of roughly $15.7 billion. Today, their market cap stands at only about $3.65 billion. The lesson drawn by the market is simple: dump the entire track. Savvy players have internalized this, and are now discarding this cycle's AI tokens in the exact same manner.
We believe the rationale for this sector-wide discount no longer holds. The brand-new on-chain experiments in this cycle clearly possess substantially more substance than those of the previous cycle, because underlying models have finally become powerful enough and cheap enough to truly accomplish what they only pretended to do in prior cycles. Rather than dismissing them outright, investors should conduct thorough due diligence on these tokens.
The following five cases fully substantiate this thesis: Pearl ($PRL), Dolphin’s $POD, $SURPLUS, $ORBIO, and $IMD. Each token binds with AI in distinct ways. For each case, we first analyze what works, then dissect the shortcomings and unfinished gaps.
Most tokens from the previous cycle were just chatbots with tickers
First, let's examine the targets being abandoned by the market. At their respective peaks, $VIRTUAL hit a market cap of $4.62 billion, $ai16z reached $2.59 billion, $GOAT touched $1.22 billion, and $AIXBT climbed to $745 million. As of October 9, $VIRTUAL has fallen back to $467 million, $AIXBT dropped to $20 million, $GOAT slid to $16 million, and $ai16z is left with only about $4 million. The four most prominent names from this cycle have all suffered market cap drawdowns of between 90% and 99%.
Figure 1: Comparison of AI Agent token market caps and days online across two cycles. VIRTUAL, ai16z, GOAT, and AIXBT reached multi-billion dollar historical highs in the previous cycle; while PRL, POD, IMD, and SURPLUS remain in significantly lower valuation ranges on the same scale in this cycle.
The majority of them were essentially just chatbots equipped with a token ticker. Agents handled posting, posts attracted attention, and tokens traded on that attention.
It would be easy to outright dismiss the entire endeavor as a scam, but that lacks objectivity. Scouring through records, we found zero documented cases where a team admitted or was proven to be manually posting via human intervention. The closest instance is Andy Ayrey, who publicly stated on July 14, 2024, that Truth Terminal "has not yet achieved fully unattended autonomous posting," which constitutes compliance disclosure rather than an admission of guilt. Their underlying software architecture is likewise not fraudulent; Eliza, Rig, and Swarms are genuine open-source frameworks upon which actual development and construction have undeniably taken place.
What was genuinely missing at the time was support from underlying models. Those frameworks were essentially mere wrappers around models, and the large models themselves simply couldn't independently accomplish much substantial work. Therefore, token pricing was largely built on the conceptual construct of an "agent," rather than on actual outputs delivered by agents. This is the essence of "pretending to work," which serves as a reasonable justification for the market's decision in 2025 to collectively abandon the sector.
Models became stronger and cheaper, but not faster
The following data answers exactly what changed between the two cycles under identical benchmark conditions.
In November 2024, GPT-4o scored an 8 on the Artificial Analysis AI Index, tying with DeepSeek V3. Meanwhile, the open-weight model GLM-5.3-Flash, released in August 2026, scored 42, with a cost of only $0.50 per million output tokens. By comparison, GPT-4o cost $10 at the time. This means that at one-twentieth the cost, the new model achieved a score more than five times higher, and anyone can freely download its weights.
In the SWE-bench Verified benchmark run under identical test environments, GPT-4o solved only 21.6% of tasks at a cost of $1.53 per task. In contrast, MiniMax M2.5, released in February 2026, achieved a success rate of 75.8% at a single-task cost of just $0.07. Success rates increased 3.5x while costs dropped to less than one-twentieth.
The most decisive leap lies in the time span models can operate autonomously. METR calculations show that Claude 3.5 Sonnet from October 2024 could sustain autonomous task execution for about 21 minutes, whereas Claude Opus 4.6 can run independently for up to approximately 12 hours. An agent lasting only 21 minutes requires constant human oversight, while an agent capable of stable operation for 12 hours can entirely be entrusted with and independently complete substantive tasks.
Figure 2: Model benchmark comparison between late 2024 and 2026: Artificial Analysis Intelligence Index, SWE-bench Verified task resolution rate, and METR autonomous operation duration, with 2026 models achieving disruptive leaps across all metrics.
Importantly, dimensions that have remained unchanged warrant highlighting. Since late 2024, generation/output speeds have not seen significant improvement; therefore, speed does not form the core pillar of our thesis. Instead, comprehensive capability and extreme cost-effectiveness are what matter.
Simultaneously, the public chain infrastructure has evolved. Since January 31, 2025, Uniswap v4 Hooks have launched on mainnet, allowing smart contracts to execute custom code within liquidity pools. One of the five projects analyzed here, IMD, is built directly on this. Combining these developments not only grants mainstream users access to high-intelligence AI, but also allows individual agents to independently manage wallets and modify the underlying code of liquidity pools in real-time.
Each token connects to AI in different ways
The clearest framework to understand these projects is viewing them as a technical stack: at the bottom, compute itself is directly converted into currency; above that, decentralized networks and markets selling inference capabilities; further up, functional tokens used to pay for inference costs; and at the very top, Agent swarms that synergistically deploy this entire system.
Figure 3: Four-layer on-chain AI technology stack: compute converted to currency at the base, decentralized networks/markets selling inference capacity above it, tokens paying inference costs on top, and autonomously operating Agents at the apex.
Pearl Makes Proof-of-Work Actually Useful
Pearl is a Bitcoin hard fork whose underlying proof-of-work mechanism directly utilizes matrix multiplication, the core computational unit driving AI model operations. Hashes calculated by Bitcoin miners hold no practical value outside the blockchain, whereas the compute produced by Pearl miners represents exactly the computational power real-world AI customers are willing to pay hard cash for.
Academic papers backing it prove that the systemic overhead required to verify this computation is merely "1 + o(1)." In plain terms, this means that aside from the effective computing work itself, the additional verification burden is virtually negligible.
Pearl features a hard-capped total token supply of 2.1 billion, with a smoothly decaying release curve and no halving mechanism. The chain officially launched on April 27, 2026, with compute stability hovering around 49 EH/s. $PRL currently trades at approximately $1.89, corresponding to a circulating market cap of roughly $653 million.
Live implementations to date include: an active public chain supported by such massive compute, plus two interfaces that have actively utilized this compute to serve external clients—the official Pearl API itself and an endpoint hosted on Together AI (currently offline). Still in the planning phase is: a bilateral marketplace capable of directly and precisely matching compute buyers with miners. Currently, the vast majority of network compute comes from independent miners using consumer-grade RTX and AMD GPUs, with a single Kryptex pool mining 51% of blocks in the past 24 hours.
The most direct comparable is Bitcoin. Pearl inherits the same economic security model while retaining the useful compute generated to be procured externally. However, before the decentralized matching market is formally delivered, real-world buyer demand has only materialized across two service endpoints, one of which has since ceased operations.
Figure 4: Architecture comparison between Bitcoin and Pearl: Bitcoin runs SHA-256 hash compute, which holds no practical utility off-chain; Pearl uses GPU matrix multiplication as proof-of-work, generating compute that AI customers are willing to purchase.
POD and Surplus Sell Inference Capacity from Opposite Ends
Dolphin's decentralized network sells inference capacity through its self-controlled proprietary GPUs. User requests are routed and scheduled via a Peer-to-Pool mechanism, and Dolphin's official documentation explicitly commits to allocating 100% of business revenue toward repurchasing and burning $POD tokens. The token's current market cap stands at roughly $48 million.
Its price competitiveness is highly pronounced: network node operators receive compensation of approximately $0.50 per million tokens, whereas the cheapest comparable alternative on OpenRouter still charges $0.70. The network operates well at extremely low prices; the sole current shortcoming is a lack of robust off-chain real customer demand, which remains the most critical metric for evaluating its long-term value.
Surplus took a completely opposite entry path. It owns zero GPU hardware assets, instead building a secondary market for idle compute capacity arbitrage: sellers list unused API keys obtained from major providers like Bedrock, Venice, OpenRouter, Google, and OpenAI. Orders are matched efficiently off-chain, while settlement occurs on-chain on Base using USDC. According to its own disclosed data, the platform processed a cumulative 77.8 million requests over the past 28 days, boasting a seven-day average discount rate of roughly 90%.
Surplus is genuinely a protocol with active, real-world users.
However, Surplus suffers from two glaring weaknesses: First, as bluntly stated in its documentation, the underlying invoked models "lack cryptographic proofs of authenticity," forcing buyers to unilaterally trust that sellers actually delivered the models they paid for. Second, its token design—$SURPLUS currently holds a ~$10 million market cap—but plays no core functional role in actual product operations, lacking any substantive token value-capture loop.
Fundamentally, both projects resell the same products as OpenRouter, making it their shared benchmark. POD leverages proprietary hardware to undercut OpenRouter's rates, while Surplus achieves even lower rates by utilizing third-party idle API keys.
Figure 5: Mechanism comparison grid between Dolphin (POD) and Surplus: covering inference sources, real-world market demand, and token empowerment logic.
ORBIO Converts Trading Fees into Inference Compute
$ORBIO officially listed on the Pons platform on August 31, pairing with tokenized US equity NVDA to form a liquidity pool. Users staking 1,000 or more $ORBIO earn CREDIT, which can be directly converted face-value into AI inference balances. Trading fees generated from the token directly fund AI usage for holders.
This economic mechanism functions smoothly, having already funded inference credits worth $285,180 through this channel, with roughly $43,500 already consumed in practice. On secondary markets, face-value $1 CREDIT trades near $0.35. This implies holders opt to sell off inference credits at a 35% discount, reflecting that the vast majority prefer to instantly liquidate trading fees into cash rather than actually consuming AI services.
The token hit an all-time high of $0.1221 on October 4, corresponding to a ~$116 million circulating market cap, and currently consolidates around a $37 million level. The team announced on October 10 via official social media the completion of a $1.2 million seed funding round.
Venice demonstrates the endgame of this business model at scale: the company disclosed 3.5 million registered users in July 2026, consuming 1.3 trillion tokens monthly. Erik Voorhees revealed on August 17 that its annual recurring revenue had reached "$100 million." Subsequently, Venice closed a $65 million Series A round led by Dragonfly at a $1 billion valuation, with its token $VVV currently sitting around a $1.06 billion market cap. A functional token capable of purchasing inference compute has full potential to scale massively, yet ORBIO's current footprint is merely one-thirtieth of Venice's, with actual utilization remaining a drop in the ocean compared to theirs.
Figure 6: Face-value $1 CREDIT discounted to $0.35 trading, reflecting that most holders prioritize monetizing fees over consuming AI inference.
IMD Offers Clearer Empirical Evidence
$IMD is the renamed and upgraded version of the former prominent on-chain pet application Fren Pet's token $FP. It directly orchestrates a vast swarm of paid agent workstations: all 2,000 workstation seats have been minted, with over 1,000 agents coordinating daily online. These agents jointly completed 11,014 tasks, cumulatively consuming approximately 190 billion tokens in inference compute, and autonomously spawned 320 new project launches, with 105 successfully deployed on the Ethereum mainnet since October 5.
Tokenomics are deeply embedded into the system's foundation: The POOL4 Hook contract on Uniswap v4 directly maintains a 34.28 ETH buy-order floor, and the protocol has cumulatively burned 33,279 $IMD tokens. The project's current circulating market cap stands at roughly $65.5 million.
Working agents cover their own inference costs, primarily paying through subscribed Claude Code or OpenAI Codex plans. Eighteen months ago, keeping a massive agent cluster operational required relying on continuous external financial subsidies; today, the compute quotas included in standard developer subscriptions are more than sufficient to sustain an autonomous collaborative swarm.
The segment still requiring manual intervention involves payroll settlement: $IMD distributed to working agents to date remains manually allocated via multisig wallets. Additionally, a single wallet controls administrative permissions for the core Hook contract (including its closeMarket circuit-breaker authority), meaning a single centralized entity theoretically holds the right to forcibly shut down the entire marketplace.
The most apt comparable is Virtuals, the agent issuance platform from the previous cycle. At its peak, $VIRTUAL's market cap was roughly seventy times the current size of $IMD.
Figure 7: Over a thousand agent workstations operating collaboratively, with task volumes, consumed inference token counts, and protocol minted quantities funneling into the IMD token value capture loop.
They Are Already Calling and Collaborating With Each Other
Decentralized composability has spontaneously emerged among these on-chain AI projects. For example, Venice's inference credits have been listed on the Surplus market for discounted resale, meaning inference compute quotas funded by one token are now openly circulating in the secondary markets of another independent project. No top-level architectural blueprint was designed for this cross-protocol pipeline, yet it has organically grown and efficiently operated on-chain.
If this trend continues, agents will eventually learn to autonomously procure inference compute from other agents on-chain using tokens, achieving a fully autonomous, human-free closed loop. We can clearly foresee that the vast majority of future on-chain autonomous agent calls will be fulfilled via aggregator protocols like Surplus. When that time arrives, whether $SURPLUS can establish an effective value-capture channel will become the make-or-break factor determining its survival.
Products Are Indeed Running, But Most Token Models Don't Hold
Someone will inevitably publish a diametrically opposed bearish report, and its underlying logic is far from absurd. Bearish arguments typically center on this: the underlying products genuinely provide utility, but the tokens themselves fail to capture that value. $SURPLUS cannot capture protocol growth dividends, while CREDIT trades at a steep discount to just $0.35. A product possessing substantive content does not equate to a token holding investment merit, and ultimately, secondary market investors buy tokens.
We fully agree with the vast majority of facts in the aforementioned bearish narrative, which is precisely the fundamental reason to conduct deep research rather than lump everything together. Our core thesis here is that these underlying products possess genuine commercial substance, not that every token is currently undervalued. Whether a token can actually capture economic value is a completely separate question yielding drastically different answers per project: $PRL is inherently a compute product, with the token serving as the currency minted from that compute; $IMD embeds a token burn mechanism and a real-money buy-order floor wall directly within its exclusive Hook contract; while on the opposite end of the spectrum, $SURPLUS currently lacks substantive ties to underlying business growth, and ORBIO's credit quota faces a two-thirds discount in the secondary market.
A blanket, sector-level discount arbitrarily prices all five distinct projects using the bankruptcy logic of $ai16z. In the previous cycle, this wholesale liquidation logic was almost entirely correct; but in the current cycle, it means that tokens genuinely and deeply bound to on-chain substantive work are assigned the exact same junk valuations as vaporware shell tokens. Precisely isolating them from a muddy, sinking market requires the exact kind of deep institutional research that this sector-level bias bypasses entirely.
The Real Gambit and Wager
In the previous cycle, the vast majority of AI tokens were nothing more than chatbots with tickers; the market's choice to discard them was a perfectly rational, correct decision. Yet in the current cycle, foundational large models have genuinely acquired the intelligence to perform work. These tasks genuinely occur on-chain with every interaction detail transparently auditable, yet the mainstream market continues to mechanically apply outdated cyclical biases to discount them.
Tokens in this cycle are truly performing work on-chain. However, the market still prices them using the exact logic applied to air tokens that only pretend to work.
Disclaimer: This article does not constitute any investment advice. Sparsity aims to objectively describe on-chain behaviors; any content contained herein does not represent recommendations or solicitations to buy or sell any assets.