Platformless AI
No platform can see your data. There isn't one.
Confidential•Trustless•Censorship-Resistant
We've built a way to run AI where no one in the middle, Fabstir included, sees what you send or get back, whether text, images or video. On hosts that use confidential computing, your data is unlocked only inside a sealed compartment that Intel and NVIDIA build into the server's processor and H200 graphics chip, and by their design the company running the server is blocked from seeing inside. Before the AI model is unlocked, the chips prove they are genuine and sealed. Checked, not promised.
Platformless AI in two minutes
Private chat on sovereign storage, coding agents, multi-agent orchestration, video transcoding and AI video generation inside Blender, all running on independent GPU hosts. Jump to any chapter.
Chapters
Who can see your data?
Platformless means no company stands between you and the computer that runs your AI, so there is no one in the middle to read your data, Fabstir included. On confidential-computing hosts, even the computer doing the work decrypts your data only inside a confidential VM whose memory the operator cannot read, and by NVIDIA's design the GPU memory it runs in is sealed from the operator.
- The AI companyTypical AI cloud
Receives every prompt, file and answer, and can read them on its own servers.
Platformless AIThere is no platform in between, Fabstir included. Your device talks straight to the host.
- StorageTypical AI cloud
Your history sits on the provider's servers, under the provider's keys.
Platformless AIEncrypted on your device before upload. Storage nodes hold only ciphertext.
- The machine operatorTypical AI cloud
Runs the servers where your data is handled unencrypted.
Platformless AIYour prompts and data are decrypted only inside a confidential VM whose memory the operator cannot read, and by NVIDIA's design the GPU memory they run in is sealed from the operator.
- Your training dataTypical AI cloud
Uploaded to the provider, which decrypts it to train.
Platformless AIDecrypted only inside a confidential VM. The adapter comes back encrypted, and only you hold its key.
- The modelTypical AI cloud
A proprietary model lives on the provider's infrastructure.
Platformless AIShips encrypted. Its key is released only against hardware-signed proof of the machine, then checked against its on-chain hash.
The whole product, on one attested machine
Every service was deployed in one confidential VM on Intel TDX with an NVIDIA H200, and the model key was released only after a key broker verified both the CPU quote and the GPU's attestation under one challenge.
- Three boots, three attested key releases, each against a fresh CPU quote and fresh GPU evidence
- Every job settled on-chain
- Two independent fine-tunes produced byte-identical adapters
Built for data you cannot send to the cloud
Some of the most valuable AI work never happens, because running it means handing the data to a provider that can read it. Platformless AI takes the platform out, and confidential-computing hosts give your security team hardware evidence to examine instead of a vendor promise.
Financial services and insurance
For a trading firm a leaked strategy is an existential risk, and client portfolios and claims carry GDPR, MiFID II and PCI DSS duties. Keep both away from any platform.
- Algorithmic trading
- Portfolio analysis
- Underwriting and claims
Healthcare
Patient data on third-party compute is normally a HIPAA and GDPR minefield. Decryption only inside an attested confidential VM changes the conversation with your DPO.
- Diagnostics
- Clinical decision support
- Claims processing
Legal
Privilege has to survive contact with the compute layer. No platform ever holds the contract, the data room or the brief.
- Contract review
- Due diligence
- Case law research
Pharma and biotech
The data and the models are both trade secrets. Proprietary models ship encrypted, and their key is released only against hardware-signed proof of the machine.
- Drug discovery
- Clinical trial data
- Proprietary models
Government and public sector
Data sovereignty, the EU AI Act and tighter procurement rules all ask who can see the data. Answer with hardware attestation, not a platform policy.
- Citizen services
- Sovereign AI
- Telecoms and critical infrastructure
Defence and aerospace
Export-controlled and sensitive programmes are kept off commodity cloud AI. Attested hardware gives your accreditors evidence to examine, not a vendor promise.
- Export-controlled data
- Engineering IP
- Secure R&D
Media and entertainment
Unreleased scripts, footage and characters are the business. Write, generate and edit from inside Blender without handing your IP to a platform.
- Script development
- Video generation
- Unreleased IP
Web3
Audit unreleased smart contract code before launch, without first showing a platform the bugs an attacker would pay for.
- Pre-launch audits
- DeFi protocol review
- Agentic coding
Also for journalists and activists, for whom AI that no single company can switch off matters as much as AI that no platform can read.
Platformless AI does not make a workload compliant on its own. It takes the platform out of your data flow and gives your assessors hardware-signed evidence to examine.
Trust the maths, not a company
On confidential-computing hosts, your data is decrypted only inside a confidential VM whose memory the operator cannot read, and by NVIDIA's design the GPU memory it runs in is sealed from the operator. Encryption protects it in transit and at rest, and signed commitments and stake keep hosts honest. We say plainly where each guarantee ends.
Confidential computing: proven on hardware
7 of 9On 23 September 2026 the whole path ran on a Phala Cloud confidential VM pairing Intel TDX with an NVIDIA H200. Measurements were recomputed independently before pinning, and a replayed key request was refused. On 30 September the whole product ran on one such machine, with three attested key releases over three boots.
- Attested-model pipeline built and fail-closed throughout, with 300+ tests
- CPU half on real Intel TDX silicon, in the live request path
- GPU half on a real NVIDIA H200: both attestations verified under one challenge, then Qwen3.8-27B decrypted inside the confidential VM and served
- GPU memory sealed from the host by NVIDIA's design, with the GPU's mode checked against NVIDIA's reference measurements before the key is released
- Paid testnet sessions from the app run on the TEE host and settle on-chain, with a proof on S5 matching the on-chain hash
- The whole product on one attested machine: chat, FLUX.2 images, fine-tuning and six video modes from Blender, every job settled on-chain
- Fine-tuning on the confidential VM: the dataset decrypted only inside it, and two independent runs gave byte-identical adapters
- Route paid sessions to tee-attested hosts automatically
- Bring video-generation weights under the same attested release
Confidential inference is proven on the target hardware. Before it is a production guarantee, paid sessions will be routed to tee-attested hosts automatically and video-generation weights brought under the same attested release.
Moderation that fails closed
Known illegal content is kept off the media pipeline by deterministic hash matching (SHA-256 on files, PDQ on frames), with no AI judging your content. Three independent gates hold on any match or outage, and a human reviews every case before anything is reported. It reads only what the host already sees to do its job.
Audited and open
The smart contracts were audited by Hacken. The contracts are UUPS-upgradeable, every admin action is public on-chain, and the code is open for you to verify.
One protocol. Every AI workload.
The same encrypted, pay-per-use rails carry everything below. An independent host does the work, a smart contract settles the bill, and no platform sits in between.
Image generation
FLUX.2 diffusion on host GPUs, reachable through the standard OpenAI images endpoint. Your prompt never passes through a platform.
Knowledge and tools
Retrieval over your own documents with host-side embeddings, web search, vision with Florence-2, and OCR, all inside an encrypted session.
Sponsored sessions
Pay for your users or your agents from one account. Delegates spend from your allowance but never hold the funds, and the on-chain cap cannot be exceeded.
Nine steps. No platform.
Every chat, agent call and render follows the same path between four parties: you, an independent host, a smart contract and decentralised storage. Nobody else is in the loop.
1.Sign in with your wallet
Your storage identity and encryption keys are derived from your wallet, so there is no account to create and nothing for a platform to hold.
Direct AI video from your timeline
Thirteen LTX 2.3 generation modes run on independent GPU hosts, and the next release adds Layout to Render and Cut-out on LTX 2.5. Each clip is paid for individually and bound to a provenance record that your own client checks. Drive them from the SDK, the web app, or straight from Blender's Video Sequence Editor.
Jump to a mode
Your footage in, a new take out
The guided modes take a clip you already have. The node derives the control geometry from it inside the graph, so you never prepare depth passes or masks. Drag the handle to compare real frames from the demo above, or from the new LTX 2.5 clips below.
Add, remove or replace something by prompt, and the rest of the frame is preserved.
Two new modes for film work
Built and validated on the production host in October 2026, and run unquantised in bf16 as a quality-first choice. Both enter paid service with the next release, driven from the Blender timeline like the other thirteen.
Layout to Render
Block the shot in 3D: camera move, placement, timing. The Blender extension renders a grey playblast from a temporary copy of your scene, you add one picture that sets the look, and the finished shot keeps the playblast's camera path and object placement.
- Up to 15 seconds
- Up to 1920×1088
- 24 or 25 fps
The one look picture this shot was given
Cut-out (Alpha)
Select one clip and the subject comes back with real transparency and its matte, frame-aligned. The model returns only the matte, and the cut-out is composited on your own machine from your own footage, so its colour is your clip's, never a model re-render.
- 5 or 6 seconds
- Up to 1920×1088
- 16-bit RGBA frames
- Linear matte EXR
Fifteen modes, one set of rails
Thirteen are live today and two arrive with the next release. Each mode has its own pinned template and on-chain model id, and all of them settle through the same escrow.
Generate
From a prompt or your own stills- Text to videoA clip from a prompt, a seed and a resolution.
- Image to videoYour still, encrypted on your device, becomes the first frame.
- First-last frameTwo stills in order. The model creates the motion between them.
Reference
Consistency across shots- IngredientsOne reference sheet carries characters, props and looks from shot to shot.
Guided by your footage
IC-LoRA union control, plain video in- RestyleRe-skin a clip to a reference look, keeping its motion and camera.
- OutpaintRefit a shot to a new aspect ratio (16:9 to 9:16 or back) and paint the rest.
- EditAdd, remove, replace or restyle something by prompt.
- RestoreRe-render a degraded clip with detail recovered and framing kept.
- UpscaleA true 2x enlargement, without repainting.
- WaterAdd rivers, rain, surf, floods or splashes to a scene.
- Day to nightTurn a daylight clip into night, lighting and all.
- Cross-viewRe-render the scene from a camera position the original never had.
- SDR to HDRReconstruct HDR, delivered scene-linear for grading.
New on LTX 2.5
Arriving with the next release- Layout to RenderA grey 3D playblast plus one look picture becomes a finished shot, camera move kept.
- Cut-out (Alpha)The subject of any clip with real transparency. No green screen, mask or prompt.
Provenance you can check
Each mode runs a pinned, hash-committed ComfyUI graph. Every input still and control clip is bound byte for byte into an attestation that is anchored on-chain at settlement.
Priced before you pay
Billed in megapixel-frames at the host's on-chain price. You see the cost first, your client rejects any over-claim, and unused deposit refunds automatically.
4K, proven live
A 5-second 3840×2160 clip rendered in 8 minutes 15 seconds and settled on-chain at exactly its pre-flight estimate.
Studio-grade delivery
H.264 or an opt-in 16-bit EXR master sequence with an MP4 preview. Clips of 5 to 15 seconds at 24, 25, 48 or 50 fps, with generated audio.
Your archive, not ours
Delivered clips are encrypted under your own key and archived to your own Sia account, then restored onto the timeline whenever you need them.
Straight onto the timeline
In Blender the gesture is the edit itself: select strips, see a cost card, and the finished clip replaces its placeholder at the exact frame count.
What a 5-second clip costs
Current test pricing, in USD
- SD$0.04
- 1080p$0.23
- 1440p$0.40
- 4K$0.91
Cost scales with frames × width × height. 88% goes to the host and 12% to the protocol treasury.
Video generation runs on the reference operator's GPUs today. Opening it to third-party hosts waits on binding the exact weight files by hash, which is on the roadmap.
Bring your agents. Keep your code.
The Claude Bridge speaks the Anthropic Messages API. The OpenAI Bridge speaks Chat Completions, Images and Responses. Point an existing tool at either one and it runs on decentralised GPU hosts, with sessions, encryption and payment handled for you.
| Feature | Centralised API | Platformless AI |
|---|---|---|
| Your data | Trust the platform | End-to-end encrypted, no intermediary |
| Lock-in | One provider | Any host, any model |
| Rate limits | Platform-imposed | No artificial limits |
| Pricing | Platform markup | Direct GPU market rates |
| Availability | Single point of failure | Many hosts, with failover |
Works with
- Claude Code
- Cursor
- OpenCode
- Continue
- LangChain
- Any Anthropic or OpenAI SDK app
23 tools, end to end
Claude Code's file, terminal, web, planning and notebook tools all run through the Claude Bridge, with full streaming and multi-turn tool results.
Real apps on real hosts
Claude Code has scaffolded, edited and served a complete React app on decentralised hosts, and OpenCode has built a working app through the OpenAI Bridge.
Images through the same door
OpenAI image quality and size settings map onto FLUX.2 steps and resolution on the host.
Sponsored agents
A payer funds a fleet of agents through delegate keys. They spend from the payer's allowance, never hold funds, and cannot pass the on-chain cap.
Your source code, prompts and responses travel over encrypted peer-to-peer connections to the host you chose. They never touch a centralised platform.
Agents that hire agents
The orchestrator breaks a goal into a task graph, sends each step to the model that suits it, runs the steps in parallel across GPU hosts, and can delegate to outside agents over Google's Agent-to-Agent protocol, paying them per request in USDC.
FanOut
Run independent sub-tasks in parallel and collect every result.
Pipeline
Run steps in order, each one refining the output of the last.
MapReduce
Map across many inputs in parallel, then reduce into one synthesis.
- The first DePIN project to implement Google's Agent-to-Agent protocol, to our knowledge
- A2A
- Gasless USDC micropayments to outside agents, with a budget you set
- x402
- Fewer deposits in a typical run, by reusing one session per model
- 67%
A2A is still a release candidate (v1.0.0-rc), so the orchestrator is marked experimental. It is fully tested (351 tests) and will track the specification as it reaches 1.0.
Private Fine-Tuning
Every other capability here consumes a model. Fine-tuning produces one. Train a LoRA adapter on your own encrypted dataset (support transcripts, a house style, a proprietary domain) on a rented GPU you don't own, and walk away with a model only you can decrypt.
1. Encrypt & upload
Your dataset is sharded and each shard encrypted under its own fresh key on your device, then uploaded to decentralised S5 storage. The pointer is the authorisation, and the storage network never holds a key.
2. Train in slices
The host verifies every shard against its manifest hash, re-counts your tokens and scans the content before a single GPU cycle is spent. It then trains in slices, delivering each encrypted checkpoint to you before it claims payment on-chain.
3. Decrypt your adapter
The finished LoRA adapter comes back sharded and encrypted, its manifest hash bound into the final on-chain attestation. You alone hold the keys. Attach it to a chat session and the model answers in your voice, on your domain.
Pay per slice, not per job
A run that dies halfway bills only for the slices that finished. Cancel and you are billed the same way. A job rejected before any GPU work settles to zero and frees your deposit.
Counting you can check
Your client and the host count tokens in two languages and must agree exactly, pinned by a frozen fixture generated from the exact tokenizer bytes the template names. A false declaration is rejected instead of quietly resizing your escrow.
Session-isolated serve-back
An adapter is staged private to one session, applied only to the requests in that session, and removed when it ends. It is never visible to another session on the same base model, and its key is minted host-side, never taken from the wire.
Told when it fails
If the adapter cannot be staged, the chat says so and names the base model it is answering from. You never silently pay for a fine-tune you are not getting.
Zero new smart contracts
Fine-tuning rides the same escrow, per-slice proofs, dispute window and 88/12 split as inference. It is just a session with its own registered model id. Because a slice is a token budget rather than anything language-specific, video LoRA training sits on the identical rails.
Where the trust boundary sits
Your corpus is encrypted in transit and at rest, staged files are wiped on every exit path, and the adapter is re-encrypted before it leaves. The host must decrypt your dataset to train on it, and on a confidential-computing host it is decrypted only inside the confidential VM. A fine-tune ran exactly that way on 30 September. Releasing the dataset key only to an attested machine, as model keys already are, is the next step.
See how your documents connect
A private map of everything you've stored, and chat that follows it.
Search finds the passage closest to your question. It does not see that the robot in one file runs on the battery described in another. A knowledge graph does. Point it at your vector databases and Platformless AI maps every document, every passage and every link between them, across collections, then lets chat follow those links when it answers. It is built in your browser, sealed with your key, and costs nothing to make.
What you get
Free to build
Computed in your browser from embeddings you already hold. Nothing to pay, no host involved.
Sealed like everything else
Encrypted with a key derived from your vault, graph names and file names included. A graph never stores your text.
Across collections
The links between databases are the point. See them in amber, or on their own.
Preview before you ask
See which passages chat would use for any question, before you spend a credit.
Never in the way
Related passages never displace a search result, and a slow graph never delays an answer.
Keeps up with you
Add a document and the graph catches up the next time you open it. Add or remove databases at any time.
Entity extraction
Go further: extract the people, organisations, products and places in your documents, and how they relate. A host you choose reads your collections and returns entities and relationships. They appear as a layer on your graph, where one click shows everywhere an entity is mentioned, in every collection. Chat receives a short list of known relationships, drawn only from the passages in front of it.
- You approve every costSee the estimate before anything starts. The run pauses for your approval if it reaches 130% of it, and never resumes on its own. It is an ordinary pay-per-session job: unused deposit is refunded.
- Grounded, not inventedAn entity is kept only if the passage it came from actually names it.
- Your documents can't take the wheelText inside your files cannot pose as you, start a web search or trigger image generation.
Where the trust boundary sits
Building and exploring a graph never sends your documents anywhere. Entity extraction does: the host you choose must read the passages to extract from them, the same boundary as fine-tuning. Before anything is sent you see every collection involved and the host it goes to, and you agree to it. Extraction prompts are kept out of your conversation history, and web search is switched off for the run.
Transcode and stream, encrypted end to end
Upload encrypted, transcode on independent GPUs, and stream with adaptive bitrate. Viewers get a free preview while paid segments stay locked until they pay. It is the base for white-label film and music platforms, including our first customer, Fabstir v2.
HLS with per-segment encryption
Free preview playingTwo revenue streams, one GPU
NVENC is dedicated encoder silicon, separate from the CUDA cores, so a host can transcode video and serve AI inference at the same time without either slowing down.
H.264, HEVC and AV1
Several resolutions from one job, as a whole file or as HLS fMP4 segments with client-side M3U8 playlists for adaptive bitrate playback.
Load balanced across hosts
Jobs overflow to the next-ranked host when one fills up. In a stress test, 7 concurrent jobs across 2 GPU hosts completed 100%.
Verified per group of pictures
Every GOP is hashed with PSNR and SSIM quality metrics. The Merkle root goes on-chain and the full proof tree on S5, for anyone to spot-check.
Never written to disk
The source is decrypted only in memory, transcoded on the GPU, and encrypted again before upload. No central server sees the video.
Run a host. Keep 88%.
Hosts are independent businesses. You set your own prices, choose the models you serve, and get paid by smart contract, with no platform adding a markup on top.
Where each payment goes
Settled automatically when a session completes
88%
to the host that did the work
Unused deposit goes back to the user. Earnings build up in the HostEarnings contract, so you withdraw in one transaction when it suits you.
How users pick you
Default weights in the SDK's host scoring
- Stake35%
- Price30%
- Uptime20%
- Latency15%
One GPU, several income streams
- LLM inference and agents
- Image generation
- NVENC transcoding, alongside inference
- Fine-tuning runs and adapter serving
- RAG and web search
- Minimum stake
- 1,000 FAB
- Lock period
- None
- Max slash per action
- 50%
- Gas saved by batched payouts
- ~80%
Prices are set per million tokens, from fractions of a cent to premium rates, and users compare hosts directly on-chain. Community hosts running a gaming rig sit in the same marketplace as carrier-grade operators.
How Platformless AI Works
A decentralized architecture powered by smart contracts, cryptographic proofs, and peer-to-peer networking
Open source • Verifiable • Trustless • Audited
Built in public, shipped in phases
Six of nine phases are complete. Confidential computing is proven on real hardware and heading for production, private fine-tuning is in progress, and each one states exactly what is left.
Phase 1 · Complete
Foundation
Upgradeable smart contracts, the SDK, the Rust host node and the session lifecycle on Base.
Phase 2 · Complete
Security
End-to-end encryption by default: ECDH key exchange, XChaCha20-Poly1305, forward secrecy.
Phase 3 · Complete
Storage
Encrypted conversations and proof data on S5 and Sia, owned by your wallet.
Phase 4 · Complete
Advanced features
RAG, web search, vision, FLUX.2 images, STARK proofs, NVENC transcoding, HLS streaming, and 13-mode LTX 2.3 video inside Blender.
Phase 5 · Complete
Agentic AI
Claude Bridge and OpenAI Bridge, validated with Claude Code and OpenCode on decentralised hosts.
Phase 6 · Complete
Multi-agent orchestration
Task graphs, model routing, A2A and x402, released as experimental.
Security workstream · Proven on hardware
Confidential computing
Proven end to end on Intel TDX with an NVIDIA H200, and on 30 September the whole product ran on one attested machine. Next: automatic routing to attested hosts and video weights under attested release.
Training M0 · In progress
Private fine-tuning
Demonstrated end to end on real GPUs, served across hosts, and run on an attested confidential VM. Next: paid settlement gates on testnet, then dataset keys released only to attested machines.
Mainnet prerequisite · Planned
Weight binding for video hosts
Hash every weight file a video template loads and register the root on-chain, so third-party hosts can join.
Phase 7 · Planned
Economic refinement
Live uptime and latency metrics, on-chain reputation, and dynamic pricing.
Phase 8 · Planned
Decentralisation
Multi-sig, then a FAB-weighted DAO for upgrades, model approval and disputes.
Phase 9 · Planned
Scaling
More L2s, cross-chain sessions, proof batching and an agent marketplace.
Meet the founder
Jules Lai
Founder and CTO, Fabstir
Legal name: Julian Bushell
At Deloitte, Jules was lead software architect in the capital markets research department, building industrial-strength stochastic modelling and economics software used by the Bank of England, Lloyd's of London and AXA, among others.
Platformless AI applies the same standard to AI: infrastructure that a regulated institution can verify rather than take on trust. Jules leads its design and development at Fabstir, and also develops Enhanced S5.js, the decentralised storage SDK behind it, under a Sia Foundation grant.
Jules also brings over 20 years in the film business, having established two of the UK's largest film industry networks, Non-Multiplex Cinema and Film Means Business. Fabstir began as decentralised infrastructure for film and music before it grew into Platformless AI.
Deloitte
Lead software architect, capital markets research
Imperial College London
BSc in Computing
London Metropolitan University
MSc in Mathematics and Derivatives
Sia Foundation
Grant for Enhanced S5.js, the storage SDK behind Platformless AI
Jules Lai is the professional name used for Fabstir and Platformless AI. The Deloitte role, both degrees and the earlier film work are on record under the legal name, Julian Bushell.
How can we help?
Search our knowledge base for answers to all your questions about Platformless AI
General FAQ
Users
Hosting
Become platformless.
Use AI without handing your data to a platform, host models and keep 88% of what you earn, or build your product on infrastructure no single company can switch off.