Dolphin Network

Founded 2024-2025 by Eric Hartford (creator of the Dolphin LLM series) under the Cognitive Computations / QuixiAI brand. Peer-to-pool decentralized AI inference network on Base that lets gamers/GPU owners serve open-weight models during idle time and earn POD tokens. Beta live since October 13, 2025.

Compared to peers Morpheus, OpenServ, and Venice, Dolphin Network has the strongest founder AI credibility but the least-proven decentralized network layer.

What Makes It Architecturally Distinct

• Peer-to-pool, not session-based, not direct P2P matching: operators join/leave on demand without lease commitments. Inference requests go into a pool; tasks randomly assigned to nodes in the relevant model pool. Random sampling enables verification (log-probs, model integrity checks, fingerprinting against model substitution / quantization spoofing). Avoids direct buyer-seller links. Distinguishes from Akash Network/io-net/Render Network (Wave 1 session-based GPU rental) and from pure P2P projects.
• First-party model brand: unlike commodity compute marketplaces, Dolphin Network owns its demand side via the Dolphin LLM lineage (~4-5M monthly Hugging Face downloads per project's own dphnAI org, pre-token, independent of crypto).
• Verification stack: validators sample requests, fingerprint outputs via log-probabilities to detect model substitution/quantization spoofing. Node operators post a cryptoeconomic bond slashable for ~2 months of income on confirmed cheating. Bond also functions as a reward multiplier — higher bonds yield up to ~2x rewards via Curve-like mechanics.
• Multi-asset payment: API accepts POD, ETH, BTC, USDC, XMR, ZEC — broad crypto rails, not just native-token-only.
• Adjacent products: chat.dphn.ai (chat UI), datagen.dphn.ai (synthetic data pipeline). Roadmap claims dynamic shifting between inference/RL/training, exploratory distributed pre-training.

Operational Metrics (May 2026)

Source caveat The traction figures below are project-disclosed and aggregated from secondary sources. They have not yet been independently verified against a third-party on-chain dashboard. Confidence: medium pending corroboration.

• 3.2B+ inference tokens generated by the pool cumulatively
• ~9,400 tokens/second sustained bandwidth across nodes
• $POD price surge ~14x in past month (MC trajectory from ~$12M to peaks implying $80-190M range)
• FDV ~$163-280M depending on source / supply assumption (500M-1B supply)
• Active node count not disclosed

Founder Credibility (the load-bearing part)

Eric Hartford is the single strongest founder profile across these four projects. Verifiable ML practitioner, not crypto-OG-with-AI-claim:

• Currently Chief Scientist at Lazarus AI
• Prior: Principal Applied AI Researcher at TensorWave; Senior Applied AI Researcher at Abacus.AI; Senior Applied AI/ML Researcher at Convai
• Pre-AI: Senior SWE at Microsoft, Staff at eBay, SDE at Amazon
• Creator of the Dolphin LLM series (built on Microsoft's Orca paper) and the Samantha series
• 30+ fine-tuned model releases (Llama 2/3, Mistral, Mixtral 8x7B, Qwen2 7B/72B, Yi 1.5 34B, Phi-2, Phi-3 Medium) on Hugging Face under dphn/, cognitivecomputations/, ehartford/
• Dolphin-R1 reasoning dataset (Apache 2.0, 800k traces) used to train Dolphin 3.0-R1-Mistral-24B
• TheBloke quantized nearly every Dolphin release before retiring — community legitimacy marker

POD Token Mechanics

• ERC-20 on Base. Contract 0xed664536023d8e4b1640c394777d34abaff1df8f (post-DPHN→POD migration)
• Primary venue Uniswap V4 Base (POD/ETH); CEX: LBank, MEXC
• Max supply 1B (per CoinGecko; some sources report 500M — supply opacity remains a yellow flag); FDV ~$163-280M depending on source; ATH $0.3751; ATL $0.00007232
• Circulating supply not reported on CoinGecko (yellow flag for transparency)
• No detailed vesting / allocation breakdown public (gap)

Buyback Flywheel and xPOD Vault

The new operational detail surfaced in the May 14 Grok thread substantially upgrades how the POD utility model should be read:

• 100% of protocol revenue auto-buys back POD on-market, directly offsetting emissions/inflation
• Revenue split example (Qwen 3.6 35B inference @ $0.70/M tokens): ~$0.50 to nodes + ~$0.20 net for buybacks
• Pricing undercuts OpenRouter on equivalent models — competitive pressure on commodity inference
• Holders stake into xPOD vault for: (a) auto-compounding dividends from on-market buybacks, (b) daily free inference credits, (c) premium/subscription perks
• Node operators must bond/stake POD (slashable for misconduct; higher bonds multiply rewards up to ~2x via Curve-like mechanics)
• Flywheel: real usage → buybacks → staking yields/credits → more supply/demand

This puts Dolphin's token utility model meaningfully ahead of typical crypto-AI "stake to access" mechanics: the buyback-into-vault loop converts inference revenue into measurable token-holder yield rather than relying on emissions-driven appreciation.

Traction — Models vs. Network Split

The cleanest analytical move on Dolphin is to separate two questions:

Do the Dolphin models have real demand outside crypto? Yes:

• 5M+ monthly Hugging Face downloads predates the token
• Distribution via Ollama, OpenRouter, Featherless, Puter, Pluely
• Venice selected Dolphin-Mistral-24B-Venice-Edition as their flagship uncensored offering (Feb 2025); OpenRouter benchmarks the Venice Edition at 2.20% censorship refusal rate, lowest in field

Does the Dolphin Network (the decentralized layer) have real demand outside crypto? Not yet demonstrated:

• No disclosed enterprise customers for the network
• No public node count, request volume, or revenue figures
• Operator economics undisclosed
• ~$8M/day trading volume reads more retail-speculative than usage-derived

Critiques

• Cryptoeconomic bonding is the weakest of three verification approaches. Per Dragonfly Research's overview (source), bonding is "the simplest, easiest, cheapest, but least sexy and in principle least secure" vs. ZK proofs or optimistic fraud proofs. Inference.net's LOGIC paper achieves more rigorous log-prob verification at scale (10k+ heterogeneous GPUs, 18 months in production).
• Uncensored-model niche caps enterprise TAM. Useful for Venice-style consumer products; hard sell to enterprise where alignment/safety are buying criteria.
• Low-float speculative dynamics. ATL → ATH spread of ~5,000x suggests typical thin-float pump pattern. Team/insider allocation not publicly disclosed.

Roadmap (May 2026)

Pragmatic staged rollout per project communications:

• Stage 1 (live): Distributed inference beta + synthetic data pipeline for own models
• Stage 2 (1-3 months out): Multimodal nodes — image/video/audio/transcription (e.g. Whisper)
• Later stages: Synthetic data suite; auto-balancing (inference ↔ training/RL); sharded distributed inference (split large models across GPUs); distributed LoRA / SFT / RL / full fine-tuning; user-facing model creation suite; exploratory large-scale distributed pre-training
• Future model releases: Dolphin X1 235B and 405B planned
• Inference SDK targeting games/NPCs, custom bots, worker sandboxes
• Stated edge: ~10x perf/$ for smaller models on consumer GPUs

The sharded inference and distributed-training items align Dolphin with the Prime-Intellect / Nous-Psyche training-stack ambition — currently aspirational vs. the inference beta.

Competitive Slot

• Wave 2 P2P mesh inference for consumer GPUs, leaning uncensored-model niche
• Closest peers: Hyperspace-Pods, MeshLLM, Darkbloom, Grove, QVAC
• Most direct technical competitor: Inference.net (more mature: 10k+ GPU network, 18 months operating, peer-reviewed log-prob verification)
• Differentiated from Akash Network/io-net/Render Network (Wave 1 session-based)
• Differentiated from Bittensor (subnet/Yuma consensus)
• Differentiated from Prime-Intellect/Nous-Psyche/Gensyn (training-focused)

Unique angle: one of very few decentralized-inference projects with a first-party model brand. Most networks are commodity compute marketplaces. Dolphin owns its demand side, which is a real moat versus pure compute plays.

Real-AI-Company Viability

Compared to peers:

• The models survive crypto's disappearance — they're already used in production via Hugging Face/Ollama/OpenRouter
• The network's standalone viability is unproven — too early to assess whether decentralized inference demand exists at meaningful scale
• Hartford's day job is Chief Scientist at Lazarus AI — even if Dolphin Network fails, the model lineage continues