The Bitcoin blueprint—decentralized trust, transparent incentives, scarcity baked into the code—has inspired a lot more than digital money. Some researchers have tried to transplant the blueprint into the realm of artificial intelligence, hoping to build a decentralized marketplace for models, evaluations, and compute. The paper we’re looking at today asks a simple, stubborn question: does this new breed of decentralized AI, embodied by Bittensor, actually live up to the Bitcoin dream, or does it reveal fresh fragilities that crypto design can’t quite hide?
The study, conducted by Elizabeth Lui and Jiahao Sun at FLock.io, digs into on-chain data from all 64 active Bittensor subnets. It’s a rare, granular look under the hood of a living experiment: real people, real wallets, real emissions, all hooked into a multi-role consensus about AI outputs. The authors don’t just measure things; they quantify how power concentrates, how that power translates into rewards, and where the system strains against the ideals of decentralization. They also don’t stop at criticism—they propose concrete tweaks that could realign incentives and shore up security, while keeping the system’s unique marketplace logic intact. In other words, they treat Bittensor as a meaningful testbed for a broader question: can we engineer a trustworthy, scalable deAI platform without letting the whales own the show?
Before we dive in, a quick note on the right-hand side of the story: the authors aren’t a university team, but researchers affiliated with FLock.io. Their work sits at the intersection of economics, distributed systems, and AI governance. And while the paper treats Bittensor as a case study, the questions it raises—how to align incentives, how to prevent cartel-like control, and how to design robust reward schemes—are the same questions that haunt any attempt to crowdsource complex, high-stakes AI work on decentralized networks.
What Bittensor Is Trying to Do
Think of Bittensor as a Bitcoin-inspired arena for AI work, where participants aren’t just holding coins but actively contributing to and evaluating AI tasks. The network vends a capped supply of TAO (the native token) and splits each emission across three roles: miners who run inference or store data, validators who score the miners’ outputs, and subnet creators who define the incentive mechanisms for their AI marketplace. Subnets are the individual AI markets—each a self-contained experiment with its own task, scoring rules, and emission split. The root network of validators sits above the subnets, coordinating the overall emission through a mechanism the authors call the Yuma Consensus (YC). In aggregate, the design tries to fuse the incentives of a blockchain with the practical needs of building usable AI systems in a decentralized, trust-minimized way.
In Bittensor, a block mints TAO and immediately divides it: 41% to miners, 41% to validators, and 18% to the subnet creator. Subnets are where the rubber meets the road: each defines its own on-chain off-chain incentive mechanism (IM) that specifies what counts as good work, how it’s evaluated, and how rewards flow back to participants. Validators don’t just passively approve results; they stake TAO and publish weight vectors that reflect their judgments about which miners produced the best outputs. The YC then clips some of those weights, blends them with stake, and pays out miners and validators from the same pool of emissions. Delegators—TAO holders who back validators—also participate by sharing in the validator’s emissions. It’s a multi-party marketplace, with stakes acting as the scaffolding for governance and paychecks.
What makes the Bittensor design ring true to the Bitcoin idea is the emphasis on scarcity, transparency, and verifiability: a fixed supply, a long-run emission schedule, and an on-chain ledger that records staking, rewards, and performance. But there’s a crucial difference. In Bitcoin, mining power loosely tracks block rewards; in Bittensor, rewards are supposed to reflect a mix of stake, performance, and consensus alignment. The authors emphasize that this is a double-edged sword: stake is a powerful predictor of rewards, but performance (the actual quality of AI contributions) is only weakly rewarded in practice. That disparity is the core puzzle the paper sets out to explore—and, crucially, to fix.
What the Data Reveals
The authors analyze nearly two years of on-chain data, spanning 6.6 million events across 121,000 wallets in 64 subnets. This is, by many standards, a dataset that most experiments would envy: the entire ecosystem’s staking, rewards, and performance traces laid bare for study. The central finding is stark: concentration is extremely high. In most subnets, the top 1% of wallets control a shocking share of total stake—roughly 38% to 100%, with a median around 90%. In other words, a tiny handful of players can muster a majority of voting power simply by owning the right amount of TAO. The implication is sobering: it’s not hard to assemble a coalition that can press for preferential outcomes or push through changes in preference that reflect wealth rather than merit.
Rewards tell a similar story, though a touch less extreme. The richest wallets capture a nontrivial portion of the rewards (anywhere from 12% to 56% of all rewards across subnets, with a median around 24%). While the concentration in rewards is a shade lower than stake, it’s still highly skewed. The Gini coefficient for stake sits near 0.98, and the Herfindahl-Hirschman Index (HHI) hovers around 0.14. Those are the numbers of a system where a few wallets own the lion’s share of the pie. Compare that to Bitcoin’s historical flavor of decentralization, where no single pool has controlled the majority for long—a stark contrast to Bittensor’s current state, at least in several subnets and time windows.
The paper doesn’t stop at broad strokes. It teases apart how size and maturity affect decentralization. Larger, older subnets tend to be less concentrated: as the number of unique wallets grows and the subnet ages, the HHI drifts downward, indicating a diffusion of power. The math behind this is intuitive: more participants mean more diverse voices, and over time those voices start to matter. Conversely, tiny or new subnets remain fragile: a handful of large holders can still dominate the scene, making a 51% attack wiring a lot easier in practical terms.
When the authors look at correlations, the message becomes more nuanced. They compute three correlations per subnet: stake vs. reward, stake vs. performance, and performance vs. reward. For validators, stake strongly predicts reward, and performance tracks reward only moderately. For miners, the story is even starker: while more stake nudges higher rewards, performance scores have little to do with earnings, especially when measured through the YC lens. In short, the system currently pays for wealth more reliably than it pays for proven quality. That’s a powerful signal: if you’re designing a decentralized AI marketplace, you either embed stronger mechanic to reward quality, or you risk producing an economy that values ownership more than merit.
On the security front, the paper translates stake concentration into a concrete, practical risk: how many wallets would have to collude to seize 51% of a subnet’s stake and short-circuit the consensus? The answer, again, is sobering. In many subnets, fewer than 1% of wallets would suffice to reach majority control. Mid-sized subnets typically need 1–2%, while even large subnets can be toppled by a few whales if stake isn’t broadly dispersed. The numbers imply that, without intervention, a new subnets or niche markets could be captured by a small club almost at once, undermining the very premise of decentralization that Bittensor hopes to embody.
Paths to Fix Decentralization
If the data paints a cautionary portrait, the authors also offer a pragmatic playbook. They outline two broad classes of protocol-level interventions: incentive realignment and security hardening.
To realign incentives, the authors experiment with three ideas that could be deployed in parallel or in sequence. The first is a performance-weighted emission split. Instead of giving the validator-heavy bucket a fixed share, the emission could tilt toward high-performing participants. In their tests, this approach nudges the correlation between performance and reward upward for miners, with a modest hit to the stake-to-reward alignment. The second idea is composite scoring, where an effective rank for miners becomes a mix of their baseline, stake-weighted YC rank and their trust or performance. This is a more aggressive adjustment: you could push performance much higher in the payout equation, but the price would be an erosion of the pure stake-to-reward link. The third option is a gentler trust bonus multiplier—a small, bounded boost to rewards that scales with performance. The math shows this is the least disruptive option, delivering a modest improvement in performance-to-reward without badly denting wealth-based incentives.
In their visual summary, the authors show a spectrum of trade-offs. The composite scoring protocol delivers the strongest performance uplift but risks dissolving the stake signal almost entirely. The performance-weighted split lands in the middle, offering a tangible improvement in performance-to-reward correlation with a relatively small drag on stake-to-reward. The trust bonus is the slow burn: it nudges the system toward quality without provoking broad disruption of existing wealth-based incentives. For a community wrestling with the age-old tension between merit and control, the trio offers a practical menu rather than a single silver bullet.
The security piece of the design puzzle is equally concrete. The authors test three approaches to blunt 51% vulnerabilities: stake caps (truncating a wallet’s effective stake above a percentile), non-linear weighting (a concave transform to flatten large stakes), and a log-transform (log(1+s), with s the stake). Across daily, weekly, and monthly snapshots, the best-performing measure—an 88% stake cap—emerges as a robust, stable guardrail. It raises the median fraction of wallets needed for majority control from roughly 1–2% to around 20%, with manageable penalties on legitimate large holders. In other words, you keep the market from being toppled by a handful of giants, without throwing the baby out with the bathwater.
That 88% cap isn’t a perfect shield, of course. It’s a design decision with trade-offs, and the authors are careful to frame it as a first-cut defense. They also caution that undermining incentives too aggressively could hollow out the very incentives that drive the AI marketplace. The real art, they argue, lies in combining modest incentive realignment with a measured suite of security protections—creating a layered defense that keeps quality outputs attractive while preserving decentralization in practice.
Why This Matters Beyond Bittensor
What makes this study gripping isn’t just the specifics of Bittensor’s tokenomics. It’s a cautionary tale about how we design decentralized systems that are supposed to produce intelligent, high-stakes outputs without turning into oligarchies. The Bitcoin comparison in the paper isn’t a victory lap; it’s a benchmark: can a deAI market match Bitcoin’s decentralization or not? The answer, at least from this dataset, is: not yet. The same line of thinking that warned about mining pools in Bitcoin also rings true here—concentration concentrates risk. If the platform’s governance and economics naturally tilt toward wealth rather than merit, the very value proposition—an open, diverse AI marketplace—can erode from within.
But the authors don’t stop there. They intentionally sketch a path forward that preserves Bittensor’s core innovation: a distributed, incentive-aligned marketplace for AI. The proposed tweaks are not abstract formulas; they’re design choices with direct implications for who gets to build, who gets to judge, and who benefits from the ecosystem’s growth. If implemented thoughtfully, these tools could nudge the system toward a healthier balance where quality matters, without surrendering the decentralization that makes deAI compelling in the first place.
There’s also a broader methodological takeaway. This is one of the few papers that treats a running blockchain project as a live laboratory, combining long-span on-chain data with econometric metrics—Gini, HHI, correlation heatmaps—and then translating results into concrete protocol changes. It’s not just about “do these numbers look good?”; it’s about “what would this look like if we changed the incentive dial, in the real world, over time?” That kind of empirical, design-forward thinking is exactly what makes the current wave of decentralized AI research feel more than theoretical fiction.
And the paper’s timing matters. The authors note that a major upgrade—the Dynamic TAO (dTAO)—has already been rolled out, replacing a centralized root-network valuation with a distributed, stake-based subnet valuation. That shift could materially alter power dynamics and decentralization. The authors wisely flag it as a ripe topic for future empirical work: how do these architecture-level changes ripple through incentives, security, and real-world use? If dTAO proves to be a lever that meaningfully diffuses power, it would be a powerful argument for the viability of truly decentralized AI markets. If not, the story remains a work in progress, full of important lessons for builders across the crypto and AI ecosystems.
One final note worth carrying forward: the study doesn’t pretend to have all the answers. It offers a disciplined, data-driven map of the terrain, a set of concrete knobs to turn, and a clear-eyed view of the tensions involved in marrying astute incentive design with robust security. In that sense, the paper is as much a blueprint for what to test next as it is a verdict on what currently works. It invites a broader conversation about how we build decentralized AI that is not only auditable and scarce but also fair, resilient, and oriented toward genuine merit—and not merely the size of one’s wallet.
In the end, the work is a reminder that the Bitcoin dream—trustless, censorship-resistant, miner-powered consensus—works beautifully for money. Translating that dream into a domain as dynamic and subjective as AI requires humility, iterative experimentation, and a willingness to accept trade-offs. The road to a truly decentralized AI market isn’t paved with a single, perfect protocol tweak; it’s paved with careful experimentation, honest measurement, and a community willing to adjust course in the name of both openness and quality. Bittensor might someday become the Bitcoin of deAI, or it may teach us lasting lessons about why that dream is harder than it looks. Either way, the journey itself is a story worth following—and a reminder that governance is as much about human choices as it is about cryptographic guarantees.
