How To Escape The Five Riders Of The False AI Apocalypse

2026-06-12 | A-0033 | research-brief | 30% AI

Imagine if Microsoft spied on all Excel users, and then made Excel run slower when they detected their competitors using it.

Imagine if Google stopped providing search results for questions about basic biology.

Imagine if Apple discovered you were a cryptographer, and remotely locked you out of your MacBook.

Would you not be outraged at such paternalistic, user-hostile behavior?

Yet that is how Anthropic treats their customers.

Anthropic and the other proprietary frontier AI labs all try to:

• censor what people can read
• dictate what people can create
• spy on behalf of the CIA/Palantir
• strangle their competitors with a thicket of costly regulations

In a grimly amusing turn of events, Amodei’s efforts to scare the public into giving him and his cronies monopoly control over the AI industry have backfired.

On June 10, Amodei declared

“In addition to transparency, I now believe frontier models should face mandatory third-party testing for cyber, bio, and autonomy risks—with the power to block or revoke deployment of models that pose catastrophic risk.”

Amodei got his wish.

On June 12, Anthropic announced:

“The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.

The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.”

Unfortunately, Anthropic won’t be the only AI firm to suffer the ill effects of its behavior. The USG is likely to impose similar bans on other AI firms.

How can we free ourselves of the domination of both self-serving paternalistic AI firms and authoritarian governments?

We have to develop AI protocols that enable the user to be fully sovereign:

• open source
• decentralized
• private
• censorship-free
• spyware-free

Toward that end, I’ve put together a survey of the current companies / projects that are working to advance truly sovereign AI technology. Note that being listed here does not mean that I support everything that they do or that their leaders advocate–only that they are making contributions to a key part of the sovereign AI stack.

Which projects do you think should be on the list, but aren’t? Which projects are you most excited about?

• Tier 1: Full-Stack
• Tier 2: Training Data + Open Models
• Tier 3: Decentralized Compute
• Tier 4: Decentralized Inference
• Tier 5: Permanent Storage
• Tier 6: Data Markets
• Tier 7: Privacy-First Tools
• Tier 8: Privacy Infrastructure

Tier 1: Full-Stack (Data + Training + Inference)

Bittensor (TAO)

• What: Decentralized network where validators and miners compete to produce the best AI outputs. Miners host models; validators assess output quality. Token-incentivized. Multiple subnetworks (subnet 1 is text generation, others for image, compute, etc.).
• Censorship resistance: No central authority to impose alignment. The network’s incentive structure rewards useful outputs regardless of content.
• Stack position: Primarily inference and model hosting. Does not control training data or pre-training (miners bring their own models).
• Key differentiator: Oldest and most established decentralized AI network. Market cap in the billions. 32+ subnetworks for different AI tasks.
• Homepage: https://bittensor.com
• Twitter: @opentensor

Morpheus AI

• What: Decentralized protocol for AI agents, compute markets, and open intelligence services. Runs a live mainnet compute marketplace (Morpheus-Lumerin Compute System). Participants earn MOR tokens as capital providers, code contributors, or compute providers.
• Censorship resistance: No central authority can shut down or filter agents. Token incentives reward open, useful outputs regardless of content.
• Stack position: Full-stack decentralized AI protocol – smart agent platform + compute marketplace + capital providers + code contributors.
• Key differentiator: First live decentralized AI agent network with real token incentives and a functioning mainnet compute marketplace. Morpheus ecosystem partners include Lumerin (routing), Exabits (GPU), AlphaTON (TON/Telegram deployment), and Matrix AI (agent builders).
• Homepage: https://mor.org
• Twitter: @MorpheusAIs

Nous Research

• What: Train and release uncensored/alignment-resistant open-weight models (Hermes family, Hermes Agent). Building Psyche – a decentralized training protocol (peer-to-peer distributed pre-training across heterogeneous GPUs). Also building the Nous Portal for decentralized inference routing.
• Censorship resistance: Explicitly trains models that refuse fewer prompts. Removes RLHF refusal patterns. Open-weights on HuggingFace.
• Stack position: Training data curation + distributed training (Psyche) + uncensored models (Hermes) + agent framework (Hermes Agent) + inference routing (Nous Portal).
• Funding: General Catalyst, others.
• Key differentiator: Only org that operates at every layer – datasets, pre-training, post-training, open weights, agent framework, inference.
• Open source note: Hermes model weights are open on HuggingFace. Psyche training protocol is being developed openly. However, the full training data pipeline and Nous Portal inference routing are not open source. Open weights, not fully open source.
• Homepage: https://nousresearch.com
• Twitter: @NousResearch

Prime Intellect

• What: “The Open Stack for Self-Improving Agents.” Provides decentralized GPU marketplace (Intel Compute), RL post-training framework (Prime-RL), evals, sandboxing, and deployment/inference. Recently released INTELLECT-1 (first decentralized pre-trained >1B parameter model on heterogeneous compute).
• Censorship resistance: Trains on community compute with open configs. Focused on agent self-improvement loops rather than alignment restrictions.
• Stack position: Compute marketplace + RL training + evals + deployment. Training data side is weaker (uses existing open datasets).
• Funding: Founders Fund, Andrej Karpathy, Dylan Patel (SemiAnalysis), Clem Delangue (HuggingFace), Tri Dao.
• Key differentiator: First to successfully decentralize pre-training of a 1B+ parameter model. Verifiable training on heterogeneous hardware.
• Homepage: https://primeintellect.ai
• Twitter: @PrimeIntellect

Tier 2: Training Data + Open Models

Allen Institute for AI (AI2 / OLMo / Dolma)

• What: Produces fully open-source language models (OLMo series), training datasets (Dolma), and evaluation frameworks. Everything – weights, training data, training code, and eval harnesses – is released publicly.
• Censorship resistance: Unlike most labs, AI2 releases the complete training pipeline. Researchers can audit, reproduce, and build on every layer without a gating process.
• Stack position: Training data (Dolma) + pre-training code + open models (OLMo series) + evaluation. No inference product.
• Key differentiator: The only major research organization that releases the complete training pipeline, not just weights. Genuine open science rather than open weights.
• Homepage: https://allenai.org/olmo
• Twitter: @allen_ai

Common Crawl

• What: Nonprofit that publishes a monthly crawl of the public web as petabyte-scale archives in WARC format. The foundational raw material used in The Pile (EleutherAI), C4 (Google), RedPajama, and most major open training datasets.
• Censorship resistance: Operates with a mandate to make web data universally accessible. Not subject to corporate policy decisions about what content to include or exclude.
• Stack position: Raw training data only. Does not process, filter, or model – just crawls and stores.
• Key differentiator: The de facto substrate of internet-scale AI training. Without Common Crawl, most open training datasets would not exist.
• Homepage: https://commoncrawl.org
• Twitter: @CommonCrawl

EleutherAI

• What: Research collective that created The Pile (825GB open training dataset), GPT-Neo/GPT-J/GPT-NeoX (open pre-trained models), Pythia (open research models). Pioneered open pre-training methodology.
• Censorship resistance: Explicitly anti-censorship mission. The Pile contains data that commercial datasets exclude. All models and data released openly.
• Stack position: Training data (The Pile, Pile-CC, etc.) + pre-training code (GPT-NeoX, Megatron-DS) + open models (Pythia family). No inference product.
• Key differentiator: The foundation layer. Almost every open/uncensored model traces its training data or methodology back to EleutherAI’s work. Non-profit research collective.
• Homepage: https://www.eleuther.ai
• Twitter: @ai_eleuther

Hugging Face

• What: Open platform for models, datasets, and spaces. Hosts 1M+ models, 200K+ datasets. Provides Transformers library, TRL (training), PEFT (parameter-efficient fine-tuning), and open inference (HF Inference Endpoints, serverless).
• Censorship resistance: Platform policy is permissive; hosts uncensored models openly. Removed some models only under significant legal pressure. Community norms favor open access. Instrumental in securing the open-source exemption in the EU AI Act.
• Stack position: Platform/marketplace for all layers. Training data (datasets hub), training frameworks (TRL, Accelerate), model hosting, inference. Does not build its own frontier models.
• Key differentiator: The GitHub of AI. Every project on this list publishes models and data there.
• Homepage: https://huggingface.co
• Twitter: @huggingface

LAION

• What: Large-scale AI Open Networks. Created LAION-400M, LAION-5B (open image-text datasets used to train Stable Diffusion, among others). Data transparency advocate.
• Censorship resistance: Datasets are open and unfiltered. Actively fought against takedown demands. Controversially (and correctly for censorship resistance) includes content that commercial operators would filter.
• Stack position: Training data only (image-text pairs). No models, no training infra, no inference.
• Status: 5B dataset was temporarily taken down due to CSAM concerns; later working on filtered versions. Founder Christoph Schuhmann continues the mission through other orgs.
• Key differentiator: The single most important open image-text dataset in AI history.
• Homepage: https://laion.ai
• Twitter: @laion_ai

Open Knowledge Foundation

• What: UK-based nonprofit that defines and promotes the Open Definition, Open Data Standards, and the CKAN data portal software. Foundational to the infrastructure of publicly accessible datasets used in AI training.
• Censorship resistance: Advocates for data freedom as a legal and policy matter, not just a technical one. Has successfully lobbied for open licensing in government and academic datasets globally.
• Stack position: Standards and tooling for open data. Does not produce AI models.
• Key differentiator: The organization that wrote the definition of “open data.” Every dataset that claims to be “open” implicitly depends on their standards work.
• Homepage: https://okfn.org
• Twitter: @OKFN

RedPajama (Together AI)

• What: Open effort by Together AI to reproduce and improve the training datasets used for models like LLaMA. RedPajama v1 reproduced LLaMA’s 1.2T token dataset. RedPajama v2 expanded to 30 trillion tokens from Common Crawl with quality annotations.
• Censorship resistance: All data and quality filters are publicly released, allowing anyone to audit what’s included or excluded.
• Stack position: Training data only. Together AI also operates inference infrastructure, but RedPajama is the open data contribution.
• Key differentiator: The largest publicly available filtered and annotated training dataset, directly enabling open replication of frontier models.
• Homepage: https://www.together.ai/blog/redpajama-data-v2
• Twitter: @togethercompute

Sentient Foundation

• What: Initiative building open-source AI systems owned by their community of contributors rather than a corporation. Focuses on ensuring the economic benefits of AI flow to those who contribute compute, data, and code.
• Censorship resistance: Decentralized governance and ownership prevents any single entity from redirecting the project toward commercial or regulatory constraints.
• Stack position: Protocol design + model development. Early stage.
• Key differentiator: Explicit focus on ownership and economic rights for AI contributors, not just open weights.
• Homepage: https://sentient.foundation
• Twitter: @SentientAGI

Tier 3: Decentralized Compute + Training Infrastructure

Aethir

• What: Distributed GPU cloud built for enterprise AI and gaming workloads. Aggregates underutilized GPU capacity from data centers and validators into a decentralized compute marketplace.
• Censorship resistance: No single provider can refuse a workload. Compute is allocated by smart contract, not by a corporate policy team.
• Stack position: GPU compute marketplace. Does not produce models or data.
• Key differentiator: Enterprise-grade SLAs in a decentralized setting – targeting the same workloads as AWS and Azure but without centralized control.
• Homepage: https://aethir.com
• Twitter: @AethirCloud

Akash Network

• What: Decentralized peer-to-peer cloud marketplace using a reverse-auction mechanism. Open-source deployment of Llama-series models is live. Launching a Confidential Computing layer using Trusted Execution Environments (TEEs) to provide hardware-level privacy guarantees.
• Censorship resistance: Reverse-auction marketplace means any provider can participate and any user can deploy. No provider can discriminate against workloads by content policy.
• Stack position: Cloud compute marketplace. Not a model trainer or inference engine itself.
• Key differentiator: Most mature decentralized cloud marketplace. Reverse auction keeps costs near hardware cost. TEE layer adds hardware-level privacy on top of decentralized availability.
• Homepage: https://akash.network
• Twitter: @akashnet

AlphaTON Capital (ATON)

• What: Nasdaq-listed investment company that announced Morpheus AI as the preferred AI infrastructure for its portfolio across the Telegram and TON ecosystem in 2025.
• Censorship resistance: Brings decentralized AI infrastructure to large-scale deployment across one of the world’s largest messaging ecosystems. Removes the OpenAI/Anthropic dependency from Telegram-adjacent AI products.
• Stack position: Ecosystem partner and capital allocator. Does not build its own AI technology.
• Key differentiator: First publicly traded company (Nasdaq) to formally commit to decentralized AI infrastructure as its primary platform choice.
• Homepage: https://alphatoncapital.com
• Twitter: @AlphaTON

AO (AO The Computer)

• What: Hyper-parallel compute layer built on Arweave. Actor-oriented, message-passing computation model. Processes run in parallel across independent “sub-graphs” in Arweave’s storage. Designed for AI agent computation, autonomous processes, and composable smart contracts.
• Censorship resistance: Computation is eternal and verifiable. AO processes live alongside Arweave data – code and messages are permanent. No single node can censor a running process.
• Stack position: Compute/execution layer on top of Arweave storage. Not a model trainer or inference server – it’s where autonomous AI agents live and compute forever.
• Key differentiator: The only compute environment where AI agents can run perpetually and autonomously with permanent state. If Arweave is the hard drive, AO is the CPU.
• Homepage: https://ao.thepermaweb.ai
• Twitter: @aoTheComputer

Arweave

• What: Permanent decentralized storage network. Data stored on Arweave is immutable – pay once, store forever. The storage layer for the permaweb: models, datasets, code, and smart contracts that can never be taken down.
• Censorship resistance: Maximum for data permanence. Content is replicated across nodes with a Sustainable Profit Margin endowment. No one can delete or modify stored data. The closest thing to “uncensorable storage” that exists.
• Stack position: Storage only (the permanent data layer). Other projects build compute, inference, and applications on top.
• Key differentiator: Only storage network with true permanence guarantees backed by an endowment. Base layer for censorship-resistant AI – if your model weights, training data, or agent code are on Arweave, they can never be removed.
• Homepage: https://arweave.org
• Twitter: @ArweaveEco

Exabits

• What: Decentralized GPU compute provider focused on enterprise AI workloads. Partners with Morpheus AI to ensure AI agents have access to GPU cluster capacity.
• Censorship resistance: Provides the hardware substrate for decentralized agent compute. No single data center operator controls the available capacity.
• Stack position: GPU infrastructure. Does not build models or agents.
• Key differentiator: Focuses on high-density GPU cluster access for agent workloads, in contrast to consumer-GPU marketplaces.
• Homepage: https://exabits.ai
• Twitter: @exabits_ai

FedML (TensorOpera)

• What: Open-source federated and distributed learning platform. Allows organizations to collaboratively train models without centralizing raw data. Rebranded as TensorOpera in 2024 to reflect a broader platform scope.
• Censorship resistance: Training happens at the data source. Data never leaves the organization’s infrastructure. No central server sees the raw data.
• Stack position: Training infrastructure (federated). Provides the platform; organizations bring compute and data.
• Key differentiator: Most complete federated learning platform with cloud, on-premise, and edge device support. Production-grade at enterprise scale.
• Homepage: https://fedml.ai
• Twitter: @FedML_AI

FLock.io

• What: Decentralized federated AI training marketplace. Incentivizes participants to contribute compute and data to collaborative model training without centralizing the underlying data. Integrates with Akash Network for decentralized compute.
• Censorship resistance: Smart contract governance of the training process means no single party can veto contributions or redirect the model.
• Stack position: Federated training marketplace.
• Key differentiator: Adds token incentives to federated learning, making it economically viable to participate even for small contributors.
• Homepage: https://flock.io
• Twitter: @flock_io

Flower / Flwr (Federated Learning)

• What: Open-source framework for federated learning. Train models across decentralized data sources without centralizing data. Used by research labs, hospitals, and enterprises for privacy-preserving training.
• Censorship resistance: No central data aggregator needed. Data stays on-device. Training can happen across any heterogeneous nodes.
• Stack position: Training infrastructure (federated). Provides the framework; you bring the compute and data.
• Key differentiator: Most mature open-source federated learning framework. Production-grade with support for heterogeneous hardware.
• Homepage: https://flower.ai
• Twitter: @flwrlabs

Gensyn

• What: “The network for machine intelligence.” Decentralized compute verification layer. Proves that machine learning computations were executed correctly without trusting the compute provider. Building a full compute marketplace (training + inference) with cryptographic proof of work.
• Censorship resistance: Decentralized compute means no single provider can refuse your training job. Verifiable computation means the results are trustworthy.
• Stack position: Compute verification + marketplace. Does not produce models or data.
• Key differentiator: Solves the verification problem for decentralized AI. If someone runs your training on their GPU, Gensyn proves they actually did the work.
• Open source note: 37 public GitHub repos. Popular tools (rl-swarm, codeassist) are MIT-licensed. However, the core verification component (REE – Reproducible Execution Environment) is dual-licensed: SDK/scripts are MIT, but the compiler and runtime binaries are proprietary. Peripheral tooling is open; the critical verification layer is closed.
• Homepage: https://gensyn.ai
• Twitter: @gensynai

Internet Computer (ICP)

• What: Blockchain platform by the DFINITY Foundation that runs decentralized applications (dApps) including AI services. Uses “canister” smart contracts that can run WebAssembly code at scale on a globally distributed node network.
• Censorship resistance: Applications deployed on ICP are governed by the Network Nervous System (NNS) DAO. No AWS takedown can remove a canister – the platform has no single infrastructure provider to pressure.
• Stack position: General-purpose decentralized computing platform. Can host AI inference canisters and store data on-chain.
• Key differentiator: The only blockchain that can host full-stack web applications (front-end + back-end) entirely on-chain with no cloud dependency.
• Homepage: https://internetcomputer.org
• Twitter: @dfinity

io.net

• What: Decentralized GPU network that aggregates idle GPU capacity from crypto miners, independent data centers, and cloud providers into a single marketplace. Focuses on ML inference and training workloads.
• Censorship resistance: Distributed GPU pool with no single controlling entity. Smart contract-mediated job scheduling.
• Stack position: GPU compute marketplace.
• Key differentiator: Aggregates the largest pool of GPU capacity in the decentralized space by pulling simultaneously from multiple source types.
• Homepage: https://io.net
• Twitter: @ionet

Lumerin

• What: Routing protocol for the Morpheus AI network. Operates the Morpheus-Lumerin Compute System, the live mainnet marketplace where AI inference jobs are routed to compute providers. Handles payment, routing, and job verification.
• Censorship resistance: Smart contract routing ensures no intermediary can selectively block or throttle AI workloads.
• Stack position: Compute routing infrastructure for the Morpheus protocol.
• Key differentiator: The first live on-chain marketplace for AI inference routing. Real paying users and compute providers are operating on mainnet – not a testnet demo.
• Homepage: https://lumerin.io
• Twitter: @LumerinProtocol

Matrix AI Network

• What: AI developer building specialized agents on the Morpheus network. Products include M.A.C. (an AI-powered investment analysis agent) and MANTOR. Emphasizes user data sovereignty: data is processed locally, not uploaded to a central server.
• Censorship resistance: Agents run on decentralized infrastructure with user-controlled data. No central platform can restrict what the agent analyzes or reports.
• Stack position: Agent application layer on top of Morpheus.
• Key differentiator: Demonstrates Morpheus’s practical deployment – live agents with real users rather than a protocol demo.
• Homepage: https://matrix.io
• Twitter: @MatrixAINetwork

Open Gradient

• What: Protocol building decentralized infrastructure for AI compute and inference. Focuses on making AI compute accessible as a permissionless network resource, with an emphasis on model provenance and verifiable computation.
• Censorship resistance: Permissionless design means no gating on who can deploy inference workloads.
• Stack position: Compute and inference infrastructure.
• Key differentiator: Early stage; positioned as a decentralized alternative to cloud AI inference APIs with on-chain provenance tracking.
• Homepage: https://www.opengradient.ai
• Twitter: @OpenGradient

OpenMined / PySyft

• What: Open-source community building the PySyft library for privacy-preserving machine learning. PySyft enables federated learning, differential privacy, and secure multi-party computation within the standard PyTorch/TensorFlow workflow.
• Censorship resistance: Training on sensitive data without centralizing it. Medical, financial, and government datasets can contribute to AI training without exposing the underlying data.
• Stack position: Privacy-preserving training infrastructure (library/framework layer). Not a model producer.
• Key differentiator: The most widely used library for privacy-preserving ML. Lets researchers add differential privacy or federated training to existing PyTorch code with minimal changes.
• Homepage: https://www.openmined.org
• Twitter: @openminedorg

Psyche Network

• What: Decentralized AI coordination network focused on privacy, autonomy, and network resilience. Designing protocols for AI agents to coordinate without centralized orchestration.
• Censorship resistance: No single point of control for agent coordination. Resilient to node failures and censorship attempts.
• Stack position: Agent coordination infrastructure.
• Key differentiator: Focuses on coordination as a primitive – most decentralized AI projects address compute or data; Psyche addresses how agents communicate and coordinate.
• Homepage: https://psyche.network
• Twitter: @psycheoperation

Tier 4: Decentralized Inference

Allora Network

• What: Decentralized AI inference network with a coin ($ALLO). Focus on context-aware inference – models that adapt based on feedback from other models in the network. Weighted ensemble of models compete to produce the best predictions.
• Censorship resistance: Decentralized inference marketplace. No single gatekeeper controls what queries get processed.
• Stack position: Inference only (with feedback loops). No training data or model production.
• Key differentiator: Context-aware inference that improves over time through network effects. Competing models drive quality up without central curation.
• Homepage: https://allora.network
• Twitter: @AlloraNetwork

Anyscale

• What: Company behind Ray, the dominant open-source framework for scaling AI applications across distributed compute. Ray is used by OpenAI, Spotify, and most major AI labs for distributed training and serving.
• Censorship resistance: Open-source means no vendor can remove Ray from your infrastructure. Ray’s distributed design means no single compute provider dependency.
• Stack position: Distributed compute framework. Not a model or data provider.
• Key differentiator: The de facto standard for distributed AI in Python. If you need to scale inference or training beyond a single machine, Ray is almost certainly the right tool.
• Homepage: https://anyscale.com
• Twitter: @anyscalecompute

Exo

• What: Open-source project that enables distributed AI inference across heterogeneous consumer devices – MacBooks, iPhones, Raspberry Pis. Devices cooperate to run models that wouldn’t fit on any single device.
• Censorship resistance: Maximum. Runs on devices you own with no cloud dependency. No API key, no provider, no account.
• Stack position: Inference only (distributed across consumer devices).
• Key differentiator: The only project that successfully chains consumer devices together for inference. A household cluster of phones and laptops can collectively run a model that none could run alone.
• Homepage: https://github.com/exo-explore/exo
• Twitter: @exolabs

Groq

• What: Semiconductor company that developed the Language Processing Unit (LPU), a chip architecture optimized for sequential, token-by-token AI inference. GroqCloud offers dramatically faster inference for open-weight models with a zero data retention policy.
• Censorship resistance: Supports open-weight models (Llama, Mixtral, Gemma) with no training on user prompts.
• Stack position: Inference hardware + managed cloud inference service.
• Key differentiator: Fastest available inference for open-weight models. The speed advantage changes what’s practical for agentic use cases that require many sequential LLM calls.
• Homepage: https://groq.com
• Twitter: @GroqInc

Kuzco Network

• What: Decentralized inference network that distributes LLM inference across a network of GPU contributors. Recently rebranded; follow @inference_net for current developments.
• Censorship resistance: Distributed inference means no single node controls what queries get processed.
• Stack position: Inference only (distributed across contributor GPUs).
• Key differentiator: GPU contributor network model – anyone with a consumer GPU can participate and earn from providing inference capacity.
• Homepage: https://kuzco.xyz
• Twitter: @kuzco_xyz

ollama

• What: Local inference engine. Run any GGUF model on your machine. No cloud, no API, no censorship possible.
• Censorship resistance: Maximum. You control the hardware. Nobody can refuse your prompt.
• Stack position: Inference only. One machine, no distribution.
• Key differentiator: Easiest way to run uncensored AI locally. Handles model download, quantization selection, and serving in a single command.
• Homepage: https://ollama.com
• Twitter: @ollama

Petals

• What: Decentralized inference for large models. Distributes model layers across volunteer nodes (like BitTorrent for neural nets). Allows running models larger than any single participant’s VRAM.
• Censorship resistance: Decentralized. No single operator. However, individual node operators could filter content.
• Stack position: Inference only (distributed across volunteer nodes).
• Status: Active but lower profile since 2024. Works best for models under 70B parameters.
• Homepage: https://petals.dev
• Twitter: @bigscience

vLLM

• What: High-performance open-source inference server for open-weight models. Uses PagedAttention for efficient KV-cache management; the production standard for self-hosted inference.
• Censorship resistance: Self-hosted deployments apply no filters by default.
• Stack position: Inference server software.
• Key differentiator: The de facto standard for production open-weight inference. Handles high-throughput serving with continuous batching.
• Homepage: https://vllm.ai
• Twitter: @vllm_project

SGLang

• What: High-performance inference server optimized for structured output generation and multi-call LLM workflows. Faster than vLLM for agentic use cases that chain many sequential LLM calls.
• Censorship resistance: Self-hosted deployments apply no filters by default.
• Stack position: Inference server software.
• Key differentiator: Outperforms vLLM on structured output benchmarks and multi-step agent pipelines. If your workload involves JSON generation or tool-calling chains, SGLang is faster.
• Homepage: https://sglang.ai
• Twitter: @sglang_project

Tier 5: Permanent Storage & Permaweb

AR.IO

• What: Develops the gateway network and Arweave Name System (ArNS) for the permaweb. ArNS provides human-readable names (like DNS) for Arweave-stored content. The gateway network provides fast HTTP access to the otherwise low-level Arweave protocol.
• Censorship resistance: Gateway operators are independent – no single entity controls access. ArNS names are registered on-chain and cannot be seized or redirected by a registrar.
• Stack position: Web infrastructure layer on top of Arweave storage.
• Key differentiator: Bridges permanent Arweave storage and the ordinary web browser. Without AR.IO, accessing Arweave requires specialized clients; with it, you use a normal URL.
• Homepage: https://ar.io
• Twitter: @ar_io_network

Community Labs

• What: Venture studio that raised $30M to accelerate Arweave ecosystem development. Builds developer tooling, consumer applications, and enterprise products on the permaweb.
• Censorship resistance: Funds and builds products that make censorship-resistant storage usable at scale for ordinary developers and end users.
• Stack position: Ecosystem development and application layer.
• Key differentiator: The primary commercial arm for Arweave ecosystem growth. Ships products to real users rather than staying at the protocol layer.
• Homepage: https://communitylabs.com
• Twitter: @CommunityLabs_

Forward Research

• What: The primary R&D lab for Arweave. Responsible for core protocol development, the ecosystem architecture, and web services that provably respect user rights.
• Censorship resistance: Designs the protocol-level guarantees that make Arweave censorship-resistant by default.
• Stack position: Core protocol R&D.
• Key differentiator: The architectural steward of the permaweb. Without Forward Research, Arweave’s decentralization would drift toward de facto centralization over time.
• Homepage: https://fwd.arweave.net
• Twitter: @fwdresearch

Irys (formerly Bundlr)

• What: High-throughput data upload infrastructure for Arweave. Handles batching (bundling) of transactions to enable the upload speeds required for large-scale AI data storage. Evolving into a programmable datachain supporting data provenance for AI and DePIN workloads.
• Censorship resistance: Handles the operational layer of uploading to Arweave at scale, making permanent censorship-resistant storage practical for large artifacts.
• Stack position: Upload infrastructure on top of Arweave.
• Key differentiator: The practical path for bulk uploads to Arweave. Direct Arweave transactions are too slow for large AI artifacts; Irys handles the throughput gap.
• Homepage: https://irys.xyz
• Twitter: @irys_xyz

Mirror

• What: Decentralized publishing platform where articles are permanently stored on Arweave. Writers own their content; it cannot be deleted, paywalled, or deplatformed by the platform.
• Censorship resistance: Maximum for long-form text. Once published to Mirror, an article is cryptographically immutable and permanently accessible to anyone.
• Stack position: Publishing application layer on top of Arweave.
• Key differentiator: The first practical censorship-resistant publishing platform. Works like Medium but with no centralized control over your content.
• Homepage: https://mirror.xyz
• Twitter: @MirrorXYZ

Tier 6: Decentralized Data Markets

Grass

• What: Distributed network that compensates participants via a browser extension for contributing their internet bandwidth to collect and index public web data. The collected data is used to build AI training datasets.
• Censorship resistance: Distributes data collection across millions of residential IPs. No single corporate crawler can be blocked or regulated as a point of failure.
• Stack position: Training data collection layer.
• Key differentiator: The only large-scale residential web data collection network with direct token compensation. Bypasses Cloudflare blocks and bot detection that stop centralized crawlers.
• Homepage: https://www.getgrass.io
• Twitter: @grass

HiveMapper

• What: Community-driven mapping network where contributors install dashcams that automatically upload street-level imagery to build a real-time, globally distributed map dataset. Compensates contributors with HONEY tokens.
• Censorship resistance: Distributed data collection makes the dataset impossible to suppress or corrupt. No single entity controls the mapping data.
• Stack position: Specialized geospatial training data layer.
• Key differentiator: Real-time map data collected continuously by a distributed contributor network. Coverage updates as contributors drive, in contrast to infrequent Google Street View sweeps.
• Homepage: https://hivemapper.com
• Twitter: @Hivemapper

Masa Network

• What: Decentralized marketplace where users own and monetize their personal data. Data DAOs aggregate user data contributions for AI training while preserving privacy and providing token compensation.
• Censorship resistance: User ownership of data prevents platforms from selling it without consent or corporate entities from monopolizing training data.
• Stack position: Personal data collection and monetization layer.
• Key differentiator: Directly addresses the “data extracted without consent” problem of current AI training. Users set the terms for how their data is used.
• Homepage: https://www.masa.ai
• Twitter: @getmasafi

Ocean Protocol

• What: Decentralized marketplace for sharing and monetizing datasets. Data providers publish datasets with on-chain usage licenses; consumers pay to access them. Compute-to-data allows running algorithms on private data without the data ever leaving the owner’s infrastructure.
• Censorship resistance: Decentralized marketplace means no central platform can block a dataset or dictate licensing terms.
• Stack position: Data marketplace and compute-to-data infrastructure.
• Key differentiator: Compute-to-data is the key innovation – lets AI researchers train on sensitive datasets (medical, financial) without custodial risk. The data never moves; only the model does.
• Homepage: https://oceanprotocol.com
• Twitter: @oceanprotocol

Vana

• What: Ecosystem of user-owned Data DAOs that aggregate personal data contributions for AI training. Users contribute data, verify it, and receive tokens in proportion to their contribution. Targets data that currently flows to Big Tech for free.
• Censorship resistance: User ownership and DAO governance prevents data from being controlled by a single corporation with a terms-of-service change.
• Stack position: Personal data aggregation and governance layer.
• Key differentiator: Focuses specifically on data that current AI labs get for free – social media posts, search history, etc. – and creates a mechanism to compensate contributors economically.
• Homepage: https://vana.org
• Twitter: @vana

Tier 7: Privacy-First AI Clients & Tools

Abacus.ai

• What: Enterprise AI platform providing foundation model APIs and MLOps tooling. CEO Bindu Reddy was one of the most vocal opponents of California’s SB 1047 and federal AI licensing proposals, arguing they are designed by incumbents to entrench competitive advantages.
• Censorship resistance: Actively opposes regulatory frameworks that would require licenses for open-source models. Supports user-borne liability rather than developer liability.
• Stack position: ML platform and API layer on top of existing open models.
• Key differentiator: One of the few commercially successful AI companies that has publicly and aggressively opposed regulatory capture, not just endorsed open source.
• Homepage: https://abacus.ai
• Twitter: @abacusai

Fireworks AI

• What: Production inference engine for open-weights models. Provides fast, enterprise-grade inference for Llama, Mixtral, Gemma, and others with a zero data retention policy for open models by default.
• Censorship resistance: Zero data retention means prompts are not stored or used for training. No surveillance of user queries.
• Stack position: Managed inference for open-weight models. Does not produce models.
• Key differentiator: Best-in-class latency for open-weight inference combined with strong privacy defaults. The enterprise path to open AI without building your own serving infrastructure.
• Homepage: https://fireworks.ai
• Twitter: @fireworks_ai

Jan AI

• What: Open-source desktop application for running AI models entirely locally. Supports GGUF models from HuggingFace Hub with an intuitive GUI. All inference happens on your hardware.
• Censorship resistance: Complete. No network access required after model download. No telemetry by default.
• Stack position: Local inference client (desktop application layer).
• Key differentiator: Best GUI experience for local AI. Designed for non-technical users who want full privacy without needing to understand Ollama’s CLI.
• Homepage: https://jan.ai
• Twitter: @jandotai

LM Studio

• What: Desktop application for downloading, managing, and running open-weight LLMs locally. Provides a clean GUI, model discovery from HuggingFace Hub, and a local API server compatible with the OpenAI API format.
• Censorship resistance: Complete local execution. No prompts, responses, or usage data leave your machine.
• Stack position: Local inference client with an OpenAI-compatible API compatibility layer.
• Key differentiator: OpenAI API compatibility means you can drop it in as a private replacement for OpenAI in any existing application that uses the OpenAI Python SDK – no code changes required.
• Homepage: https://lmstudio.ai
• Twitter: @lmstudio

LocalAI

• What: Open-source, self-hosted drop-in replacement for the OpenAI API. Runs models locally on any hardware with no GPU required (CPU inference supported). Compatible with OpenAI’s REST API format.
• Censorship resistance: Maximum. Self-hosted with no telemetry. Designed specifically to replace cloud AI APIs with a local equivalent.
• Stack position: Local inference API server.
• Key differentiator: If you have an application built on the OpenAI API, LocalAI is the fastest path to making it fully private and self-hosted. Just swap the base URL.
• Homepage: https://localai.io
• Twitter: @LocalAI_API

Mintplex Labs (AnythingLLM)

• What: Builds AnythingLLM, an all-in-one desktop and server application for building private, local AI workspaces from your documents. Connect any LLM, any vector database, any embedding model.
• Censorship resistance: Explicitly designed for privacy. Optional telemetry, offline-capable, does not centralize your documents to any cloud service.
• Stack position: Document AI workspace (application layer).
• Key differentiator: The most complete local RAG application. Lets non-technical users build a private ChatGPT-over-your-documents without touching the command line.
• Homepage: https://anythingllm.com
• Twitter: @mintplexlabs

NanoGPT

• What: Pay-per-prompt AI aggregator providing access to 400+ AI models with no subscription and high anonymity. Payments handled via Nano (XNO) and Monero (XMR). Uses proxy routing to avoid attaching personal identifiers to model queries.
• Censorship resistance: Crypto payments provide financial privacy. Proxy routing breaks the link between your identity and your queries. No account required.
• Stack position: Anonymous AI access aggregator. Does not run its own models.
• Key differentiator: The only multi-model aggregator with serious payment privacy. If you need to make sensitive queries without those queries being attached to your credit card, this is the tool.
• Homepage: https://nano-gpt.com
• Twitter: @nanogpt_com

Nomic AI

• What: Produces GPT4All, the most widely used application for running LLMs on consumer hardware with no internet connection. Also develops Nomic Atlas for AI data visualization and model transparency tools.
• Censorship resistance: GPT4All works entirely offline on CPU. Data never leaves the device.
• Stack position: Local inference application (GPT4All) + data visualization tooling (Atlas).
• Key differentiator: GPT4All is the most accessible entry point for fully private local AI. Runs on a laptop CPU with no GPU required.
• Homepage: https://nomic.ai
• Twitter: @nomic_ai

Open WebUI

• What: Open-source, self-hosted web interface for interacting with local and remote AI models. Works with Ollama, OpenAI-compatible APIs, and direct model downloads. Supports RAG pipelines, multi-modal models, and team collaboration features.
• Censorship resistance: Self-hosted with full control over data. No conversations leave your server.
• Stack position: User interface layer (web client for local or remote AI).
• Key differentiator: The most feature-complete self-hosted AI UI. Combines the polished UX of ChatGPT with complete data sovereignty.
• Homepage: https://openwebui.com
• Twitter: @OpenWebUI

Pin AI

• What: Developer of “Personal Intelligence” – on-device agentic AI that processes personal data locally to bypass the privacy risks of cloud AI. Designed for personal assistant use cases with calendar, contacts, and file access.
• Censorship resistance: On-device architecture means no data upload. No surveillance of personal schedules, communications, or file contents.
• Stack position: On-device AI agent (application layer).
• Key differentiator: Privacy-first personal AI agent. Competes with Siri and Google Assistant but with data that never leaves your device.
• Homepage: https://pinai.io
• Twitter: @pinai_io

Venice.ai

• What: Private, permissionless, uncensored AI interface for open-weights models, founded by Erik Voorhees. Provides API and web access to open models with no logging, no training on your prompts, and no alignment filters restricting responses.
• Censorship resistance: Explicitly designed as an alternative to captured AI. Voorhees has publicly argued that AI safety regulations are regulatory capture by Big Tech incumbents.
• Stack position: Privacy-first AI client and API service. Does not train its own models.
• Key differentiator: The leading commercially available alternative to censored AI APIs. If you need uncensored API access without running your own infrastructure, this is the tool.
• Homepage: https://venice.ai
• Twitter: @venicedotai

Tier 8: Privacy & Financial Infrastructure

Cake Labs

• What: Develops Cake Wallet and Monero.com, the leading mobile wallets for Monero (XMR) and Zano. Also integrates Bitcoin and Zcash. Cake Wallet is the most-used mobile interface for privacy coin payments.
• Censorship resistance: Privacy coin infrastructure is the payment layer for censorship-resistant AI. Paying for compute, data, and models with traceable payment methods creates surveillance risk; Monero payments are unlinkable by default.
• Stack position: Privacy coin payment infrastructure (mobile wallet).
• Key differentiator: The most polished mobile app for Monero, which has no optional privacy – every transaction is private by default.
• Homepage: https://cakelabs.com
• Twitter: @CakeWallet

Electric Coin Co. (ECC)

• What: Primary developer of Zcash, responsible for the core protocol, the Halo 2 zero-knowledge proof system (which eliminated the original trusted setup ceremony), and Zcash Improvement Proposals (ZIPs).
• Censorship resistance: Zcash provides programmable, shielded transactions. Halo 2 removed the trusted setup requirement, eliminating the main attack surface in the original system.
• Stack position: Core protocol developer for Zcash.
• Key differentiator: Halo 2 is one of the most significant advances in practical zero-knowledge cryptography. The recursive proof composition it enables has implications beyond cryptocurrency – including verifiable AI computation.
• Homepage: https://electriccoin.co
• Twitter: @ElectricCoinCo

Monero

• What: The leading privacy cryptocurrency. All transactions are private by default using ring signatures (hides the sender), stealth addresses (hides the receiver), and RingCT (hides the amount). No optional privacy – every transaction is indistinguishable regardless of whether the parties care about privacy.
• Censorship resistance: The de facto payment standard for censorship-resistant AI services. NanoGPT and Venice.ai both accept XMR. Paying for AI compute or model access via traceable payment methods creates a surveillance link between your identity and your queries; Monero severs that link.
• Stack position: Payment infrastructure. Used as the privacy payment layer across decentralized AI tooling.
• Key differentiator: Unlike Zcash (where privacy is optional and most transactions are transparent), Monero has no transparent transactions. Every Monero transaction is private. This makes the anonymity set vastly larger – you cannot identify privacy-sensitive users by the fact that they used the privacy feature.
• Homepage: https://getmonero.org
• Twitter: @monero

Zano

• What: Privacy coin and smart contract platform with confidential assets (CA). Uses a hybrid PoW/PoS consensus mechanism. Confidential assets allow issuing privacy-preserving tokens on top of Zano, enabling confidential DeFi and private token economies.
• Censorship resistance: Privacy-by-default transactions with Mixin and Stealth addresses. Confidential assets make it possible to build markets and applications where transaction amounts and parties are hidden from chain observers.
• Stack position: Privacy payment infrastructure + smart contract layer.
• Key differentiator: Confidential assets extend Monero-style privacy to custom tokens. A marketplace built on Zano can issue its own tokens while inheriting the same transaction privacy as base-layer Zano transfers. Cake Wallet supports Zano natively.
• Homepage: https://zano.org
• Twitter: @ZanoProject

Zcash Foundation

• What: Non-profit organization supporting Zcash governance, security research, and the independent Zebra full node software (written in Rust). Manages the Zcash Community Grants program.
• Censorship resistance: Two independent node implementations (ECC’s zcashd and Zebra) ensure no single entity controls the network.
• Stack position: Protocol governance, second full node implementation, grant funding.
• Key differentiator: The Zebra node provides implementation diversity – a single bug in one codebase doesn’t take down the whole network. Also provides a governance check against ECC’s influence over protocol direction.
• Homepage: https://zfnd.org
• Twitter: @ZcashFoundation