FLOPs: The Decentralized AI Economy and the Collapse of centralized Trust

The centralization meta

In the last few years artificial‑intelligence research has become defined by its gatekeepers. State‑of‑the‑art models such as GPT‑3 were trained on clusters of thousands of NVIDIA V100s—equivalent to 355 years of training on a single device>—and the cost of building models like DALL‑E is measured in hundreds of millions of dollars. These models are trained behind closed doors and then released as APIs; the public pays a toll to query them but cannot examine the models, adjust them or verify the data used. The Gensyn litepaper notes that compute complexity for AI is doubling every three months, vastly outstripping supply^[2]. Without a mechanism to pool resources, only the richest corporations will continue to train frontier models.

centralization does not just hurt innovation; it creates single points of failure and invites censorship. Italy's data‑protection regulator temporarily banned ChatGPT in March 2023, accusing OpenAI of having "no legal basis that justifies the massive collection and storage of personal data". OpenAI complied and disabled ChatGPT in Italy. The chatbot is also unavailable in mainland China, Hong Kong, Iran and Russia ^[3]. When AI services are tied to centrally‑managed clusters, governments can shut them down or demand back‑doors. Export controls on advanced chips restrict who can build or even run cutting‑edge models. The future of AI risks being dictated by those who own datacentres and those who regulate them.

Idle GPUs and the case for decentralization

While a handful of firms hoard clusters, the rest of the world is swimming in unused compute. Nous Research points out that enormous amounts of computing power sit idle or under‑utilised worldwide. The Psyche Network is a response to this imbalance: instead of requiring "massive infrastructure with thousands of accelerators in a single location," Psyche coordinates training across distributed, heterogeneous hardware. The abundance of under‑utilised GPUs provides an opportunity to train and use large models without large capital expenditure^[4].

Gensyn goes further, arguing that "connecting all devices through a single decentralized network provides a level of scalability that is currently impossible to achieve through any existing provider". By turning the world's latent compute into an open marketplace, prices can settle at a fair equilibrium and the compute oligopoly vanishes. Ethereum miners leaving proof‑of‑work stand to earn more by renting out their GPUs to train machine‑learning models; Gensyn estimates that a decentralized marketplace could offer V100‑equivalent compute at roughly $0.40 per hour, 80 % cheaper than AWS on‑demand^[2].

The technical frontier: making global training possible

Decentralizing training is not trivial. In a decentralized setting we gain access to cheap compute, but communication between instances is expensive. Bandwidth across the public internet ranges from 200 Mb/s to 100 Gb/s—orders of magnitude slower than the 800 Gb/s links inside datacentres. Hardware is heterogeneous and nodes can appear and disappear at random^[5]. To overcome these constraints, a wave of research is emerging:

• DisTrO (Distributed Training Over‑the‑Internet) shows that training large models over consumer connections is possible. It introduces optimisers that reduce communication requirements by four to five orders of magnitude and demonstrates that DisTrO‑AdamW matches the convergence of standard AdamW while massively reducing bandwidth during pre‑training^[6].
• Psyche builds on DisTrO by using discrete cosine transforms to compress momentum updates. It overlaps communication and training so that communication latency no longer bottlenecks GPU utilisation and quantises the momentum's sign, reducing bandwidth by another 3×. Psyche coordinates runs via P2P networking on Solana, with on‑chain randomness determining assignments and witnesses.
• Pluralis' Protocol Models apply low‑rank compression to pipeline parallelism. By leveraging rank collapse in transformer projection layers, they compress both activations and activation gradients. The method provides up to a 100× improvement in communication efficiency and enables training billion‑parameter models over consumer‑grade 80 Mb/s connections, matching the convergence of 100 Gb/s datacentre systems. They even train an 8‑billion‑parameter LLaMA model across four geographic regions connected via the internet and achieve datacentre‑level convergence ^[7].
• DiLoCo (Distributed Low‑Communication Optimisation) synchronises gradients only every ~500 steps, spreading communication over minutes of computation. It uses an inner–outer optimisation scheme where each worker performs local updates with AdamW and periodically synchronises pseudo‑gradients via Nesterov momentum. This reduces communication and allows dynamic scaling: compute can be added or removed mid‑training.
• SWARM Parallelism uses a stochastic pipeline where faster devices take on more work and slower devices less, allowing model‑parallel training across unreliable connections. Workers send outputs to any worker in the next stage and tasks are reassigned if a node drops. Experiments show multi-billion‑parameter LLMs can be trained on pre‑emptible instances with bandwidth <200 Mb/s while maintaining high throughput.

These advances suggest a path toward truly decentralized training. Instead of requiring InfiniBand and identical GPUs, future algorithms will adapt to the variability of the public internet, compress communication aggressively and schedule work dynamically. Yet challenges remain: verifying that off‑chain work was done correctly, ensuring privacy and dealing with models whose sheer size exceeds the memory of any single node.

FLOPs: a protocol for decentralized AI and a memetic bet

Enter FLOPs. Its design recognises that the raw unit of AI is not a token but a FLOP of computation. The protocol creates a two‑track economy:

• FLOPs Credits (Track A) are IOUs for compute. Users deposit USDC to buy credits denominated in real‑world FLOPs. The Treasury allocates these credits to high‑conviction decentralized AI miners such as Gensyn, Nous Psyche and Prime Intellect; the resulting mining rewards and airdrops are returned pro‑rata to credit holders. The credits are non‑transferable and separate from the governance token, making the compute spend auditable and regulatory risk minimal.
• FLOPs Token (Track B) is an ERC‑20 governance and treasury token. Tokens are distributed via a long.xyz auction, with the float split roughly 30 % public, 15 % team/operations, 15% for protocol incentives, and 40 % community treasury^[8]. Token‑holders vote on how to allocate the treasury—whether to reinvest into more compute, fund research, run marketing campaigns, perform buy‑backs or seed new decentralized training experiments. There is no predetermined buy‑back or burn schedule; instead, the community can layer in value accrual mechanisms once the protocol is established.

This separation is deliberate. Compute credits give contributors immediate exposure to mining rewards without coupling them to governance decisions; the token captures governance rights and the future revenue of the protocol. FLOPs operates as a community-owned training foundry with flexible governance—the team prioritizes simplicity and transparent operations before layering in complex tokenomics. In a space crowded with poorly‑designed crypto tokens, FLOPs provides passive DeAI exposure through a proven dual-track model.

Why trust FLOPs?

Crypto's recent cycles were dominated by the extraction meta—a frenzy of tokens with predatory unlock schedules and insiders dumping on retail. Many "tier‑1" venture‑funded infrastructure projects commanded valuations in the billions without meaningful users or revenue. Their founders cashed out long before any product traction, leaving communities holding bags.

FLOPs attempts to invert that dynamic. It is built on a simple, auditable contract: credits turn USDC into compute; tokens let holders steer the treasury. The protocol will post regular allocation reports and on‑chain governance votes. A pro‑rata return of mined DeAI tokens directly to credit holders ensures that contributors benefit immediately and transparently. Treasury decisions are subject to token‑holder oversight.

Most importantly, FLOPs aligns itself with an ecosystem that is actually delivering. Gensyn's network provides hyperscale, cost‑efficient compute for deep‑learning models, connecting devices worldwide and offering compute at a fraction of AWS prices. Nous Psyche allows anyone to contribute to LLM training using under‑utilised hardware. Prime Intellect's OpenDiLoCo framework trains 1.1B up to 32B parameter^[9] models across continents while maintaining 90–95 % compute utilisation. Pluralis' Protocol Models enable 100× compression and training of 8B‑parameter models on 80 Mb/s links^[7]. SWARM parallelism and Varuna push the frontiers of model‑parallel training with dynamic pipelines. FLOPs is designed to route credits toward whichever of these networks is most promising at any given time; its treasury grows as the DeAI sector grows.

Narratives and trust in the age of decentralized AI

Robert Shiller's Narrative Economics argues that markets are driven more by stories than by spreadsheets. Bitcoin's value is not purely technical; it is sustained by the myth of Satoshi, the rebellion against central banks, and the promise of digital gold. In decentralized AI, the story is about reclaiming compute from oligopolies and building an open infrastructure for collective intelligence. It is about resisting censorship, as when Italy banned ChatGPT or when AI chip export controls restrict who can participate in the AI race.

FLOPs embraces this narrative. It offers a way for anyone to buy compute, earn exposure to decentralized AI protocols and steer the direction of the treasury. The meme is powerful: "Buy FLOPs, own decentralized training." As more people believe in that story, the treasury grows, more compute is purchased, more DeAI tokens are mined, and the cycle reinforces itself.

Challenges ahead

Decentralized AI is still experimental. Most algorithms assume homogeneous hardware and stable nodes; real‑world networks are messy. Many of the techniques highlighted above are tested on models up to a few billion parameters; nobody has yet trained a hundred‑billion‑parameter model across volunteer nodes. Work verification remains a hard problem. There is regulatory uncertainty: the line between a governance token and a security is blurry, and global privacy laws may restrict where data can flow. FLOPs itself must navigate U.S. geofencing for its auction and ensure that credits do not become de‑facto securities.

Yet the direction is clear. Models are becoming commodities; compute is becoming an asset class. Governments will continue to regulate the use of AI, and if we do nothing the ability to innovate will be concentrated in a few jurisdictions. Decentralization provides an alternative path: one where innovation is permissionless, compute is abundant and no single entity can veto access.

Conclusion: an invitation to build

FLOPs is not a guarantee; it is an invitation. It invites GPU owners, model builders, optimisers, token‑holders and dreamers to co‑create a new AI economy. It recognises that trust is earned by delivering and by aligning incentives, not by marketing alone. It stands on the shoulders of a growing body of research—DisTrO's communication compression^[6], Pluralis' low‑rank model compression^[7], Psyche's P2P architecture and the scheduling innovations of DiLoCo and SWARM—and wraps them into a simple, accessible framework.

If you believe that the future of AI should belong to everyone—not just the few who own datacentres or write the regulations—then FLOPs offers a way to stake that belief. The story is just beginning.