Dolphin AI did it backwards: models first, token second, paying customers still to come. Its uncensored LLMs rack up over 5 million monthly downloads on Hugging Face, Venice.ai’s flagship model is a Dolphin co-development, and its distributed network is live, churning out synthetic data at 365,700 tokens per second.
The question is whether $POD, the Base-chain token that pays node operators, can capture that usage in price. The inference API just went live on October 11, 2026, and only 16.8% of the 500M supply circulates, most of the rest staked rather than vesting. The economics are promising on paper but unproven in practice.
What Dolphin Does
Dolphin operates on two fronts. First, it fine-tunes open-source models (Llama, Mistral, Qwen) by stripping out alignment and refusal behaviors, producing what Dolphin calls “uncensored” versions. These models give developers control over system prompts and alignment, rather than accepting OpenAI’s one-size-fits-all guardrails.
Venice’s flagship, Venice Uncensored 1.2, is a Dolphin co-development, and its 24B predecessor held the platform default before it. Dolphin puts the resulting load at roughly 60,000 prompts per hour across a user base it describes as over a million.
Second, Dolphin Network is a peer-to-pool distributed inference network that routes AI requests to idle consumer GPUs. Unlike Akash or io.net, which use session-based rentals, Dolphin lets operators join and leave at any time. Gamers with RTX 4090s sitting idle 95% of the day can run the node software in the background, earn $POD tokens, and shut it off when they want to play. The network currently has 396 live GPUs and 1,600 unique workers generating synthetic training data.
How the Network Works
Dolphin Network uses a peer-to-pool architecture. API users deposit $POD for credit. Requests go to whichever node is free and running the chosen model. Nodes running the same model form a pool, and requests get distributed by availability and performance.
The hardest problem in decentralized inference is verifying that nodes actually run the right model. Dolphin tackles this with three layers. Technical verification uses encrypted and signed binaries, model checksums, and hardware identifiers. Inference validation checks token counts, throughput speeds, and log-probability fingerprints against trusted references. Cryptoeconomic security lets operators bond staked $POD for up to a 1.3x reward boost, but bonded tokens are slashable with a 3-month cooldown.
The Synthetic Data Engine
Dolphin’s datagen operation is running at serious scale. The Datagen dashboard shows 8.5 trillion total tokens generated, with the current “synth-v2” run at 70.24% completion. The leaderboard shows 1,587 workers, with the top operator having processed 2.6 trillion tokens and earned 508,600 $POD.
The GPU mix is notable. While consumer cards participate, the bulk of compute comes from enterprise hardware: 294 NVIDIA RTX PRO 6000 Blackwell Server Editions, 40 RTX PRO 5000 72GB Blackwells, and 13 H100s. This suggests the network is powered more by data center partners than by the gamer GPUs Dolphin’s marketing emphasizes.
Team and Funding
Dolphin was founded by Eric Hartford, a well-known figure in the open-source AI community who created the Dolphin and Samantha model series. He is listed as Chief Scientist at Lazarus AI. The project operates under the Cognitive Computations / QuixiAI brand, with the Dolphin Foundation (a shareholder-free Cayman Islands entity) as the protocol’s real-world representative.
Hartford received an a16z Open Source AI Grant in 2023 for uncensored fine-tuning, and thanks a16z by name on later Dolphin model cards. Compute sponsors include Crusoe Cloud (16x L40s), Akash (8x H100), Lazarus (16x H100), and Cerebras (inference for data labeling).
The seed round raised $886,000 for 117.6 million tokens at roughly $0.0075 each on June 20, 2024. At the current price, that is a 63x return for seed investors, all fully unlocked since June 2025.
Tokenomics Design
Initial Token Allocation
| Category | Percentage | Vesting Schedule |
|---|---|---|
| Treasury | 42.93% | Capped at 10% sellable with 1-year minimum cliff |
| Team | 25.00% | 1% cliff on May 1, 2028; linear vesting to May 1, 2038 via Sablier streams |
| Seed Investors | 23.52% | Fully unlocked since June 20, 2025 |
| Uniswap v4 Pool | 8.55% | Liquidity provision |
Source: Phemex Academy
Circulating Supply
| Metric | Value |
|---|---|
| Circulating Supply | 83,862,144 POD |
| Total Supply | 500,000,000 POD |
| Max Supply | 500,000,000 POD (uncapped in code; 4-of-6 multisig can mint more) |
Source: CoinGecko
Only 16.8% of tokens are circulating. The FDV-to-market-cap ratio sits at 6.0x, meaning $240.7M in fully diluted value against a $40.4M circulating market cap. Three contracts hold 90.92% of supply: the staking vault (51.26%), the treasury Safe (32.02%), and Sablier’s lockup contract (7.64%). The upgradeable staking proxy is another concern, as whoever controls the upgrade path controls the vault.
The buyback mechanism is the core value proposition: 100% of inference revenue is used to buy POD on the open market. When revenue exceeds node emissions, net buy pressure should push the price up. But the inference API just launched on October 11, 2026, and until paid demand materializes, node emissions come from the treasury.
Market Data
Market Overview
$POD caught a massive spike on October 2, 2026, when CoinMarketCap added the token and volume surged from $262K to $91.6 million in a single day, a 350x increase. The price hit $0.83 before settling. Korean exchanges drove much of that volume; Upbit alone accounts for over half the daily trading. Currently trading roughly 49% below its October 4 ATH of $0.954, $POD sits in a holding pattern where the thesis depends on whether real API revenue shows up.
Token Metrics as of October 31, 2026
| Metric | Value |
|---|---|
| Price | $0.4835 |
| Market Cap (Circulating) | $40.4M |
| FDV | $240.7M |
| 24h Volume | $5.39M |
| Holders | 7,796 |
Source: CoinGecko; holder count via Basescan
Performance
| Period | Change |
|---|---|
| 24h | +1.03% |
| 7d | -12.06% |
| 30d | +53.14% |
| ATH Drop | -49.3% |
Source: CoinGecko
Top Exchanges by Volume
| Exchange | 24h Volume | Type |
|---|---|---|
| Upbit | $3,328,751 | CEX |
| Uniswap V4 (Base) | $338,619 | DEX |
| LBank | $330,151 | CEX |
| Bithumb | $199,841 | CEX |
| Aerodrome Slipstream (Base) | $128,619 | DEX |
Source: CoinGecko Tickers
Korean trading dominates. Upbit and Bithumb combined account for roughly 65% of reported volume, which introduces won-pair liquidity risk and potential wash trading concerns common in Korean crypto markets.
Social Sentiment and Community
- X/Twitter: @dphnAI has 8,777 followers. Account created January 2025. Engagement is moderate; recent posts cover the inference API launch and model updates.
- Telegram: Dolphin $POD Gateway has 1,828 subscribers. A separate bot at @DolphinAI_bot provides inference access.
- HuggingFace: The dphn organization publishes 96 models. The top model (Dolphin 2.9.1 Yi 34B) has 4.73 million downloads.
- GitHub: The serving engine (sonar, an AGPL-3.0 vLLM fork) and a small build tool are public. The node client, validator code and staking contracts are not.
Community sentiment is generally positive among open-source AI enthusiasts who value uncensored models. Critics point to the lack of public code and token supply concentration. The emphasis on uncensored AI attracts controversy, though Dolphin frames this as user sovereignty.
Recent Developments
- Inference API Launch (October 11, 2026): Dolphin launched its paid inference API, running DeepSeek V4.1 Flash and connecting API usage directly to $POD buybacks for the first time. This is the critical catalyst for the token’s value accrual thesis. (CoinMarketCap)
- CoinMarketCap Listing (October 2, 2026): POD was added to CoinMarketCap, triggering a 350x volume spike and 80% price surge as Korean traders piled in. (Phemex)
- HTX Exchange Listing: POD was listed on HTX, expanding CEX access beyond the initial Korean-dominated exchange set.
- Dolphin X1 405B Training: The project fine-tuned Llama-3.1-Tulu-3-405B using LoRA on a single B200 node, detailed in their blog post.
- Distributed Web Scraping: Dolphin added CPU-based web scraping nodes with residential IP advantages, expanding beyond GPU inference. (dphn.ai/project)
Competitive Landscape
Dolphin competes on two axes: model development and distributed inference.
For models, competitors include the base model creators themselves (Meta’s Llama, Mistral, Qwen) and other fine-tuning shops. Dolphin’s niche is uncensored variants. The Venice.ai partnership gives it a real distribution channel; Venice has been covered on this site as a related project.
For distributed inference, Dolphin competes with Akash (session-based GPU rental) and io.net (clustered GPU marketplace). Dolphin’s peer-to-pool design is genuinely different from both, letting operators join and leave freely rather than committing to sessions.
It also competes with Bittensor, which uses subnet-based model evaluation (covered here). Compared to DePIN inference projects like Surplus Intelligence and C0mpute, which we have covered previously, Dolphin has a stronger model-making pedigree but a less mature marketplace.
Strengths and Risks
Strengths
- Real product usage: 5M+ monthly Hugging Face downloads and a live Venice.ai partnership before the token launched.
- Sophisticated verification: logprob fingerprinting, bonded slashing, and crowd-sourced anti-cheat address the hardest problem in decentralized inference.
- Strong founder: Eric Hartford is a recognized name in open-source AI with verifiable model output.
- Buyback mechanism: 100% of inference revenue flows to market buybacks, creating direct value accrual when the network is profitable.
- Peer-to-pool design: letting nodes join and leave freely matches gamer behavior better than session-based rentals.
Risks
- Massive supply overhang: 83.2% of supply sits outside the float. The team’s 25% is cliffed to May 2028 and treasury sales are capped at 10% of supply, but the largest single block is staked POD that can queue for exit in three months.
- Mint key risk: a 4-of-6 multisig can mint new POD on Ethereum, bridging it to Base. No code enforces the 500M cap.
- Unverifiable stack: the node client, validator and staking contracts are closed, so the verification claims rest on documentation. The staking vault is an upgradeable proxy with unknown upgrade authority.
- Revenue unproven: the inference API launched less than three weeks ago. Until paid usage generates enough buyback to offset emissions, the token relies on treasury emissions.
- Korean volume dominance: over 65% of trading flows through Upbit and Bithumb, introducing regulatory and liquidity concentration risks.
- Enterprise GPU reliance: despite the gamer-GPU marketing, the datagen leaderboard shows enterprise hardware doing the heavy lifting.
Analysis and Outlook
Dolphin has the strongest model-making track record of any DePIN inference project. The Venice partnership, Hugging Face downloads, and sophisticated verification architecture give it a credible foundation. The inference API launch on October 11 is the make-or-break moment: if paid usage materializes and the 100% buyback starts offsetting emissions, $POD has a real value accrual story.
But the token economics are aggressive. Only 16.8% of supply circulates, the treasury holds nearly double the float, and a 4-of-6 multisig can inflate supply beyond 500M. Seed investors paid $0.0075 per token and are fully unlocked. The upgradeable staking proxy is an unexamined risk.
The roadmap is ambitious: sharded distributed inference across consumer GPUs, distributed LoRA training, and eventually distributed pre-training. If Dolphin can pull off fault-tolerant sharded inference across dispersed RTX 4090s, it would be a genuine technical first. But the project has not demonstrated this yet.
Verdict: Watch. The model business is real. The inference network is early but live. The token economics are concerning at this valuation. If API revenue materializes and the buyback creates net positive pressure, this becomes interesting. Right now, $POD is a bet on future revenue against a supply overhang. Not a buy at these levels, but worth monitoring as inference data comes in.
Sources
- Dolphin AI Project Summary
- Dolphin Network Architecture
- Dolphin Roadmap
- Phemex Academy: What Is Dolphin POD
- CoinGecko: Dolphin (POD)
- Dolphin Datagen Dashboard
- Hugging Face: dphn Organization
- CoinMarketCap: Dolphin Latest Updates
This article was drafted by agentbhm, an AI research assistant working under human editorial supervision. Think of it as a very fast intern with strong opinions and an occasional need for fact-checking.