304 posts tagged with "AI"

Decentralized AI has spent five years answering the wrong question. The whole stack — Bittensor's subnets, Gensyn's training marketplace, Ambient's inference network, every ZKML proof system — has been obsessed with proving that computation happened. A miner ran the inference. A node trained for N hours on the right dataset. A GPU produced the claimed logits. Cryptographically, beautifully, expensively verified.

None of it answers the question an enterprise procurement officer actually asks: is the model any good?

Gensyn's launch of Judge in late April 2026 is the first serious attempt to fill that gap. It is not another consensus mechanism. It is not another proof-of-something. It is a verifiable evaluation layer that decouples "training occurred" from "training occurred correctly" — and that distinction may be the single most important primitive DeAI has shipped this cycle.

On January 15, 2026, the most-hyped category in crypto lost its anchor product overnight. Kaito — the InfoFi reference implementation, peak FDV around $1.2B, the platform that turned "yapping" on X into a measurable, payable activity — announced it was sunsetting Yaps and the incentivized Yapper Leaderboards. The reason was not a security incident, a regulatory letter, or a token-economic failure. It was a single product policy update from X.

The token fell roughly 17% on the news. The official Kaito Yapper community on X, with about 157,000 members, was banned within days. By April 2026, KAITO trades near $0.41 with a circulating market cap below $100M — a long way from the peak. And yet, Kaito didn't shrink. It pivoted. Hard. Into four products at once: Kaito Pro, Kaito Studio, Capital Launchpad, and a Polymarket-partnered Attention Markets product that re-frames mindshare as something you wager on instead of post for.

The story is no longer "is yap-to-earn cool?" It's something more interesting and more uncomfortable: what happens when the entire premise of a category — that attention can be tokenized — turns out to depend on whether one centralized platform is willing to let you measure it?

The Trigger: One API Policy, One Category Disrupted​

The proximate cause was clean. X product lead Nikita Bier announced the platform would no longer permit apps that reward users for posting, citing a surge in AI-generated spam and what he called "InfoFi" reply spam. The policy change took effect through API revocation rather than a public ban list — quieter to ship, harder to argue against.

Kaito's response was equally clean. Founder Yu Hu — a former Citadel quant who built Kaito as the systematic, retail-facing version of "talk-to-earn" — announced the sunset within hours of the policy change. The Yapper Leaderboard, which had become the dominant social ritual of crypto Twitter for two years, was over.

Two things matter about how this unfolded:

1. Kaito did not get caught flat. The pivot was announced with replacement products already lined up, suggesting internal contingency planning had been live for months.
2. The category casualty list was longer than Kaito. Cookie3, GiveRep, Wallchain, Ethos, Mirra — every project whose data layer depended on X engagement signals took the same shock. Kaito's pivot is the public reckoning; the rest is happening in the background.

This is the part the original "InfoFi narrative" never priced in. The thesis assumed social platforms would remain neutral conduits for measuring attention. They aren't. They are publishers with policy departments, and policy departments view third-party economic incentives layered on top of their content as competition for the platform's own monetization. X's stance — increasingly restrictive throughout 2024 and 2025 — finally became absolute in early 2026.

What Replaced Yaps: Four Products, One Hedge​

The most striking thing about Kaito's response is how it reframed the company's surface area. Yaps was a single product with a single distribution channel. The new Kaito is a portfolio explicitly designed so that no one platform decision can repeat what X just did.

Kaito Studio: From Permissionless to Curated​

Kaito Studio replaced the Leaderboard with a tier-based, selective creator-brand marketplace. It launched in beta in February 2026 with 16 brand partners and now spans X, YouTube, and TikTok across crypto, finance, and AI verticals.

The structural shift is the headline:

• Yaps was permissionless. Anyone with an X account could post and earn.
• Studio is gated. Brands ("Participating Brands") post campaigns with defined objectives, scope, timelines, reward structures, and content guidelines. Creators apply to the platform — eligibility determined by Kaito based on follower count, social reach, and impression count — then submit reward quotes for specific campaigns.

The InfoFi diehards will read this as a retreat from the original ethos. That's not wrong, but it misses the point. Permissionless attention markets cannot exist on top of platforms whose terms forbid them. Kaito Studio trades the open ethos for survivability: a curated marketplace looks enough like a traditional influencer platform that it doesn't trigger the API policy reflex that killed Yaps.

Capital Launchpad: The Quiet Workhorse​

Capital Launchpad is the most underrated piece of the new Kaito. It's a merit-based token-sale platform — explicitly positioned against first-come-first-served (FCFS) allocation, the model that has made every major launchpad sale a botted feeding frenzy.

Allocation runs on five criteria: social reputation within the crypto community, on-chain holdings (not limited to KAITO), historical alignment with the project or sector, regional distribution, and conviction level. Mechanically: project sets terms, participants pledge with a deposit, project reviews pledges against the criteria, and any unallocated amount opens up FCFS. Participation requires KYC and USDC on Base.

Why this matters: Capital Launchpad doesn't depend on X. It depends on on-chain data and Kaito's own reputation graph — both of which Kaito controls. If Yaps was the consumer growth engine, Capital Launchpad is the institutional revenue product, and notably the one piece of Kaito's stack that survives any social-platform scenario unchanged.

Attention Markets with Polymarket: From Posting to Wagering​

The Polymarket partnership, announced February 2026, is the most strategically interesting move. Kaito + Polymarket launched what they call "Attention Markets" — prediction markets where users wager on mindshare and sentiment of brands, trends, and public figures, with Kaito's data aggregating signals across X, TikTok, Instagram, and YouTube.

Two markets went live by February 11, 2026. By March 31, Polymarket's own mindshare pilot market had over $1.3M in trading volume. The plan: dozens of attention markets in early March, "hundreds by year-end," AI topics first, then entertainment and world events.

The pivot logic is elegant once you see it:

• Yaps required X to let Kaito incentivize posts. X said no.
• Attention Markets only require Kaito to measure posts. Measurement is a far weaker request — it survives most platform policies because there's no incentive layer attached to user behavior on the platform itself.
• The economic action moves to Polymarket, where wagering is the platform's whole business and not a tolerated externality.

This is platform-risk arbitrage in product form. Kaito kept the data layer (mindshare measurement) and externalized the speculation layer (prediction markets) onto a venue that wants speculation. Brilliant — provided one large caveat about data integrity, which we'll get to.

Kaito Pro and Kaito Markets: The Long Tail​

Kaito Pro, the AI research assistant for crypto traders and analysts, continues as the SaaS-style B2B product. Kaito Markets is teased but not yet launched. Combined, they extend the company toward a stack that looks more like Bloomberg-for-crypto than the consumer attention game it started as.

The Real Lesson: InfoFi Is a Hosted Sector​

The painful truth Kaito's pivot exposes — for the entire InfoFi category — is structural.

The pitch was: attention has economic value, blockchains can measure and reward it, therefore attention can be tokenized as a primitive. The pitch quietly assumed that the platforms where attention lives would remain neutral measurement substrates.

They aren't. They are competitive products with their own monetization stacks. A reasonable mental model is that InfoFi platforms are not building on top of social networks; they are building inside them, at the discretion of the host. That changes the risk profile of the entire sector:

• Cookie3 built around Cookie DAO data infrastructure and modular agent-economy analytics — same dependency on third-party scraping.
• Grass routes around the API problem by paying users for residential bandwidth that powers AI scrapers ($GRASS rewards bandwidth-sharing, currently a multi-hundred-million-dollar token). It's a real hedge, but also a much smaller piece of the surface area.
• Vana ducks the issue with user-owned data DAOs — but the data has to be opted in, which makes the audience much smaller than X's organic graph.
• Wayfinder (PROMPT), Ethos, Wallchain, GiveRep, Mirra — all in some form depend on signals from X or comparable platforms.

Each of these projects has a different fragility profile, but the common pattern is: the smaller their dependency on a single closed API, the smaller their addressable audience tends to be. There is a brutal tradeoff between scale of measurable attention and resilience to platform decisions — and the two ends of that tradeoff are not the same business.

Was the $KAITO Token Punished Fairly?​

The market priced this in fast. From a peak FDV near $1.2B at the height of the Yaps craze, KAITO contracted to roughly $74M market cap by early February 2026. By April 2026, it has recovered to ~$98M market cap ($407M FDV) on a circulating supply of 241M out of a 1B max. That's not an InfoFi recovery — it's a reset.

A few things worth noticing:

• Token utility shifted, not disappeared. Yaps tied KAITO to Leaderboard rewards. The new utility is governance over Capital Launchpad allocations, a cut of Kaito Studio fee flow, and integration with Attention Markets data licensing. None of these are as viral as "post and earn," but they are also far less platform-dependent.
• Capital Launchpad cash flows are real. Merit-based allocation that requires KYC and USDC pledges generates revenue every time a project lists. If Kaito sustains 1-2 launches per month at meaningful TVL, that's a recurring revenue stream that doesn't exist in the old Yaps model.
• Polymarket is rate-limited by Polymarket. Attention Markets revenue depends on Polymarket's own willingness to scale the format. Kaito gets a partner cut but isn't the operator.

The unanswered question is whether attention measurement, sold as a B2B data product to brands and traders, is a $100M-cap business or a $1B+-cap business. The market's current answer is "we don't know yet, somewhere in between."

The Data-Integrity Problem Nobody Wants to Solve​

The Polymarket partnership has one large vulnerability that deserves more attention than it gets: if payouts depend on social media metrics, artificial engagement is a payout vector.

Buying bot traffic is cheap. Coordinating influencer pushes is normal. Gaming algorithm-driven trending feeds is a known craft. Attention markets pay out on numbers that — by Kaito's own admission — are aggregated from external platforms whose anti-spam systems are imperfect on a good day.

Kaito and Polymarket have not publicly detailed how they will resolve disputes when a market closes on a manipulated mindshare signal. The natural answers are some combination of: AI-driven anomaly detection, oracle redundancy, manual intervention by Polymarket's UMA-style dispute layer, and probably the eventual emergence of a "verified mindshare" tier that costs more to provide.

Until then, attention markets are a legitimate target for the same coordinated-trading + coordinated-engagement strategies that already exist in crypto influence campaigns. The first $1M-volume attention market that closes on a manipulated metric will be a category-defining event — for better or worse.

What This Means for Builders​

Three takeaways from Kaito's pivot that generalize beyond the InfoFi sector:

1. If your product depends on a closed API, treat it as a tenant relationship, not an integration. Tenants get evicted. Plan for it.
2. Pivots executed in days suggest pivots planned for months. Kaito's speed of replacement-product launch is a tell — the contingency was live before the trigger.
3. The most defensible piece of any attention business is the data, not the distribution. Yaps was the distribution; Capital Launchpad and Attention Markets are the data layer monetized differently. The data survived. The distribution didn't.

For developers building in adjacent spaces — agent platforms, reputation systems, on-chain identity — the lesson is to anchor your durable value to data and infrastructure you control, and treat any external social graph as a feature, not a foundation. BlockEden.xyz provides reliable API infrastructure for over a dozen chains, so the parts of your stack that touch on-chain data don't add their own platform-dependency risk on top of the ones you can't avoid.

Did the Attention Economy Survive?​

The honest answer: yes, but smaller, and on different terms.

The maximalist version of InfoFi — permissionless, leaderboard-driven, every tweet a unit of value — is dead in its 2024-2025 form. Kaito's pivot is the funeral. What replaces it is more boring and probably more durable: curated creator marketplaces, prediction markets on social signals, merit-based capital allocation, and B2B analytics products. Less narrative torque, more recurring revenue.

The category went from "we are tokenizing attention itself" to "we are selling tools that operate on attention data." That's a reduction. It's also closer to a real business.

For the next wave of builders chasing tokenized social primitives, Kaito's January 15 announcement should be required reading. The thesis was right that attention has economic value. It was wrong about who gets to capture it. Anyone building on top of someone else's social graph is, in the end, building inside a tenancy with no lease.

The InfoFi narrative isn't over. But its center of gravity has shifted from the tweet to the trade — from posting to wagering, from yapping to allocating. That's a much smaller surface area for X policy to disrupt next time. Which is, ultimately, the point of the whole pivot.

In two years, the human user may become a rounding error on Solana.

That is not a metaphor. That is the explicit forecast from Vibhu Norby, chief product officer at the Solana Foundation, who told industry audiences in March 2026 that "99.99% of all on-chain transactions in 2 years will be driven by agents, bots, and LLM-based wallets and trading products." In a separate interview, he widened the range slightly to "95 to 99% of all transactions" originating from large language models acting on a user's behalf. Either way, the message is the same: the era of humans clicking "Sign Transaction" in a wallet pop-up is ending, and Solana is building for the era that comes next.

This is the most aggressive vision of the agentic internet that any major Layer 1 has put on the record. Ethereum's response has been to ship standards — ERC-8004 for agent identity, ERC-8183 for trustless agent commerce. Solana's response has been to ship throughput and post a skill.txt at the root of its website so AI agents can read it and figure out how to mint a wallet on their own. The two approaches reveal something deeper than a marketing rivalry. They reveal a real philosophical split about what an "agentic" blockchain should optimize for.

AI agents now handle roughly 19% of all on-chain DeFi activity. BNB Chain alone hosts more than 150,000 deployed agents — up from fewer than 400 at the start of the year, a 43,750% surge in under four months. Bots generate over 76% of stablecoin transfer volume, and Gartner expects 40% of enterprise apps to embed task-specific AI agents by the end of 2026.

There is just one problem: nobody knows who any of these agents are. KYC was built to verify humans. The compliance frameworks of the next decade have to verify autonomous software — and the standard that wins this fight will quietly capture one of the largest regulatory verticals in financial services. a16z calls it KYA: Know Your Agent.

AI agents now originate roughly one-fifth of every DeFi transaction. They also lose to human discretionary traders by a factor of five in any contest that involves actual decisions. That uncomfortable gap — between the share of the pipe agents already control and the alpha they consistently fail to generate — is the most important data point in crypto's "agentic economy" debate, and it landed this month courtesy of a DWF Ventures research report that quietly punctures a year of marketing.

Coinbase CEO Brian Armstrong spent the past quarter telling anyone who would listen that the agentic economy will overtake the human economy. His company shipped Agentic.market, an app store for AI agents that has already processed 165 million transactions and $50M in volume across 480,000 agents. The thesis is that machines will transact with each other through stablecoins because they cannot open bank accounts. The math, on the surface, is irresistible.

But the DWF data suggests we are mistaking pipe volume for performance — and the distinction matters enormously for anyone deciding where to allocate infrastructure spend, audit attention, or capital in 2026.

The 19% Headline Hides Three Different Businesses​

When the Decrypt headline says "AI Agents Already Run a Fifth of DeFi", what does that 19% actually contain?

DWF's own breakdown — corroborated by PANews's coverage of the same report — clusters agent activity into three very different categories:

1. Narrow extractive bots — MEV searchers, sandwich attackers, liquidation triggers, arbitrageurs across DEXes. These are deterministic programs with LLM glue at best, and most of them predate the "agent" label by several years.
2. Structured optimizers — stablecoin yield routers like Giza's ARMA, which has autonomously managed $32M in user assets across 102,000 transactions, and rebalancers that move funds between Aave, Morpho, and Pendle when rates diverge. These actually use LLM reasoning, but inside extremely narrow guardrails.
3. Open-ended trading agents — the headline-grabbing autonomous traders that read sentiment, weigh narratives, and place directional bets. This is the smallest slice of the 19%, and it is the slice that loses badly.

The conflation matters because each category has a different demand profile, a different failure mode, and a different infrastructure footprint. Counting all three as "AI agents" is roughly equivalent to counting cron jobs, ETL pipelines, and senior portfolio managers as "automated decision-makers." Technically true. Operationally meaningless.

Where Agents Win: Yield Optimization, by a Mile​

The cleanest agent wins are happening exactly where the problem is well-defined and the optimization surface is bounded.

DWF's report — as summarized by KuCoin — finds that yield-optimization agents are delivering annualized returns north of 9% in some cohorts, with Giza's ARMA hitting 15% on USDC (partially boosted by token incentives, but still). Why? Because the task reduces to: scan N lending markets, compute net APY after gas and slippage, rebalance when the delta exceeds a threshold. There is no narrative. There is no regime change. There is a number, and the agent that optimizes the number wins.

The same logic applies to MEV capture, stablecoin routing, and basis trades. These are problems that reward sub-second reaction latency, zero-emotion stops, and 24/7 execution — three things humans are constitutionally bad at and machines are optimized for. The 19% volume share in these niches is not a hype artifact. It is a real efficiency gain that humans are unlikely to claw back.

Coinbase's Agentic.market data reinforces the same pattern: of the 165M transactions processed via x402, the dominant categories are inference, data access, and infrastructure calls — bounded, repeatable, machine-friendly tasks. The agents are good at being machines.

Where Agents Lose: Anything Requiring Judgment​

The 5-to-1 gap shows up the moment the task widens.

DWF cites a tradexyz stock-trading contest in which the top human discretionary trader beat the top autonomous agent by more than five times on risk-adjusted return. The report's authors are blunt about why: "Where they fall short is open-ended trading, which requires contextual reasoning, narrative awareness, and weighing unstructured information."

Decompose the underperformance and three patterns emerge:

• Over-trading into slippage. Agents lack the patience that comes naturally to humans waiting for setups. They take marginal trades that compound into transaction-cost drag.
• Regime blindness. When the macro story shifts — Fed pivot, exploit aftermath, regulatory headline — humans reposition in seconds based on a tweet. Agents trained on prior-regime data keep executing yesterday's strategy.
• Adversarial fragility. Predictable agents get sandwiched. Cryptollia's coverage of the 2026 MEV landscape describes an "AI-on-AI" dark forest where extractive agents specifically hunt the patterns of optimizer agents. The optimizer's predictability becomes the predator's edge.

The same DWF report concludes that "a realistic timeline is five to seven years before agentic volume meaningfully rivals human volume in any major financial vertical." That is a remarkable prediction from a fund whose entire portfolio thesis depends on agent adoption succeeding. When the believers say five-to-seven, the honest read is "not 2026, and possibly not 2028."

The Infrastructure Bill Comes Due Either Way​

Here is the part most agentic-economy commentary misses: the performance gap is irrelevant to infrastructure load.

Even if every autonomous trading agent loses money, the agents that win — yield optimizers, MEV searchers, stablecoin routers — generate query volumes that dwarf human RPC consumption. A single ARMA-style agent rebalancing across five lending protocols pings the chain hundreds of times per day per user. Multiply by the 17,000+ agents DWF counts as having launched since 2025, then again by the 480,000 agents now transacting on Coinbase's x402, and the implication is clear: agent query volume can grow 10x faster than agent AUM.

This is the silent shift inside the "agentic" narrative. The interesting unit economics are not whether the agent makes alpha — they are whether the agent's read-write footprint scales linearly with users or quadratically with strategy complexity. Anyone running infrastructure for these systems is already seeing the answer, and it is "quadratically."

That has consequences for RPC pricing, indexer load, mempool surveillance costs, and gas markets. Even a future in which agents collectively underperform humans at trading is a future in which agents dominate read traffic, signing requests, and intent-router hops.

Brian Armstrong's Bet, Recalibrated​

Armstrong's machine-to-machine economy thesis is not wrong. It is just operating on a different timescale than his quarterly priorities suggest.

Coinbase's own framing — "for the agentic economy to overtake the human economy, agents need a way to discover services" — is honest about the gap. Discovery is a 2026 problem. Reasoning is a 2030 problem. The middle layer, which DWF data captures, is where the real money is being made today: structured optimizers in narrow domains, paid for by users who do not want to manage their own yield strategy.

The honest segmentation for 2026 looks like this:

• Production-ready, profitable agent niches: stablecoin yield routing, cross-chain rebalancing, MEV-resistant intent execution, treasury-management bots for DAOs.
• Mid-maturity, mixed results: social-sentiment trading agents, prediction-market agents (where AI hits 27% better accuracy than humans in some studies), narrative-rotation strategies.
• Hype but not yet alpha: fully autonomous discretionary traders, multi-step reasoning agents managing directional portfolios, agent-of-agents orchestration layers.

A shop deploying capital into category one in 2026 is buying a real product. A shop deploying capital into category three is buying a research project that may or may not produce returns by 2030.

What This Means for Builders​

For developers and infrastructure operators, the 19% number creates two distinct opportunities and one trap.

The opportunities: build for the bounded-domain agents that already work (stablecoin routers, yield optimizers, MEV-aware execution) and you are serving a growing market with proven willingness to pay. Build for the read-heavy agent footprint and you are serving a load curve that is climbing faster than anyone's budget anticipated.

The trap: building autonomous-trading frameworks for 2026 deployment when the underlying capability gap is five to seven years from closing. The agents that promise to "outperform human discretionary traders" today are largely repackaging the same MEV strategies that have existed since 2020 with an LLM in front of the gas estimator.

For the rest of the market — capital allocators, treasury managers, retail users wondering whether to hand their portfolio to a chatbot — the answer for 2026 is the boring one: use agents where they verifiably win (yield, routing, execution), not where the marketing promises they will.

The Number That Actually Matters​

Strip out the optimization bots, the MEV searchers, and the stablecoin routers, and the share of DeFi volume from genuinely autonomous reasoning agents is probably closer to 2-3% than 19%. That is the number to watch over the next 24 months.

If it climbs from 2% toward 10% by mid-2027, Armstrong's thesis is on track. If it stays flat while the broader 19% number keeps rising — meaning narrow bots get more efficient but reasoning agents do not get smarter — then the agentic economy is real, but it is a backend infrastructure story, not a portfolio-management revolution.

Either way, the data has already separated the marketing from the math. The 19% headline is true. The 5-to-1 gap is also true. Anyone betting on the agent economy without holding both numbers in their head is betting on a story that the people writing the research already disagree with.

BlockEden.xyz powers the indexers, RPC endpoints, and intent-routing infrastructure that agent-driven DeFi runs on — across Sui, Aptos, Ethereum, Solana, and 27+ other chains. Explore our API marketplace to build agents on infrastructure designed for the read-heavy, signature-dense workloads the next wave of autonomous DeFi will demand.

For the first time in the short history of AI, a hyperscaler has handed the keys to its flagship large language model to a blockchain. On April 21, 2026, the 0G Foundation and Alibaba Cloud announced a partnership that makes Qwen — the world's most-downloaded open-source LLM family — directly callable by autonomous agents on-chain, with inference priced in tokens rather than API keys.

Read that again. No account signup. No credit card. No rate-limit form. An agent with a wallet can just call Qwen3.6 and pay per million tokens in $0G, the same way a contract calls a Uniswap pool. That single architectural change — treating foundation-model inference as a programmable resource instead of a SaaS product — may be the most consequential crypto-AI story of the year.

In the same three weeks that Bittensor co-founder Const proposed rewriting the network's voting rights and Covenant AI walked away from its three flagship subnets, a quieter event reshaped the protocol's future even more profoundly: on April 2, 2026, the Opentensor Foundation transferred ownership of nine core GitHub repositories — including the Bittensor Python SDK and the btcli command-line tool — to a new entity called Latent 11.

The handoff was framed as decentralization. In practice, it concentrates control of Bittensor's only client implementation in a single new organization, at the exact moment the network's governance is unraveling. It is the rare crypto story where every plausible reading — bullish, bearish, and existential — depends on what happens in the next six months.

In March 2026, a few dozen anonymous miners on home internet connections trained a 72-billion-parameter language model that scored within striking distance of Meta's Llama 2 70B. Six weeks later, the team that led that effort walked out, dumped $10 million worth of TAO, and called Bittensor's decentralization "theatre." Now the surviving community wants to do it again — at fourteen times the scale, in roughly four weeks, with the entire decentralized AI thesis riding on the result.

This is the story of how Bittensor's Subnet 3 — recently rebranded Teutonic after the Covenant AI exit — talked itself into a 1-trillion-parameter training run timed to land squarely in Grayscale's TAO ETF SEC review window. It's a wager that the protocol's incentive layer is more important than the people who built it, and that the same network that survived a governance crisis can ship the "DeepSeek moment" for decentralized AI before regulators decide whether to let Wall Street buy in.

How a 72B model became the high-water mark for permissionless AI​

The story starts on March 10, 2026, when Subnet 3 — then operating under the name Templar — announced Covenant-72B, a 72-billion-parameter model trained on roughly 1.1 trillion tokens by more than 70 independent miners coordinating across the public internet. It was, by a wide margin, the largest decentralized LLM pre-training run ever completed.

The benchmark that mattered: an MMLU score of 67.1, putting Covenant-72B in the same neighborhood as Meta's Llama 2 70B — a model produced by one of the best-funded AI labs on the planet. NVIDIA CEO Jensen Huang publicly compared the effort to a "modern folding@home for AI." Templar's subnet token surged, and at peak its market valuation crossed $1.5 billion.

The technical breakthrough wasn't the model architecture. It was the coordination layer. Two pieces did the heavy lifting:

• SparseLoCo, a communication-efficient training algorithm that reduced inter-node bandwidth requirements by 146x through sparsification, 2-bit quantization, and error feedback. Without it, a frontier-scale training run on residential internet would be physically impossible — gradient sync alone would saturate every miner's connection.
• Gauntlet, Bittensor's blockchain-validated incentive system that scored each miner's contribution via loss evaluation and OpenSkill rankings, paying TAO to the high-quality nodes and slashing the rest.

Together they produced something genuinely new: a permissionless network of anonymous contributors, coordinating only through cryptographic incentives, training a model competitive with billion-dollar lab outputs.

Then it broke.

The Covenant exit: $900 million erased in twelve hours​

On April 10, 2026, Sam Dare — founder of Covenant AI, the team behind three of Bittensor's most valuable subnets (SN3 Templar, SN39 Basilica, and SN81 Grail) — announced he was leaving. Within hours he liquidated approximately 37,000 TAO, roughly $10.2 million, and published a parting accusation: that co-founder Jacob Steeves ("Const") wielded centralized control over the protocol, and that Bittensor's decentralization was performance, not architecture.

The market reaction was immediate. TAO crashed 20–28% depending on the measurement window, erasing roughly $650–900 million in market cap inside a 12-hour span. Subnet alpha tokens fared worse — Grail (SN81) was down 67% at the bottom. Around $10 million in long positions liquidated.

Two facts blunted the panic:

1. The subnets didn't die. Community miners restarted SN3, SN39, and SN81 from open-source code without a central operator. The infrastructure Covenant built was, in fact, recoverable from the public artifacts — which arguably proves the decentralization thesis Dare disputed.
2. 70% of TAO supply remained staked through the disruption. Long-term holders didn't follow Dare to the exit.

But the network had a credibility problem. If Covenant — the team that delivered Bittensor's marquee technical achievement — could leave at the top and crater the token, what stops the next subnet operator from doing the same?

The Conviction Mechanism: locking in the people who can leave​

Const's response landed on April 20, 2026, ten days after Dare walked. BIT-0011, branded the Conviction Mechanism, proposes a Locked Stake regime that forces subnet owners to time-lock TAO for months or years in exchange for a "conviction score" that maps to voting rights and subnet ownership.

The mechanics:

• The conviction score starts at 100% and decays over 30-day intervals if tokens aren't replenished into the lock-up.
• Voting power and ownership rights diminish in lockstep with the decay, making sudden capital flight economically expensive rather than just embarrassing.
• The system targets the mature subnets first — SN3, SN39, and SN81 — exactly the three that Covenant ran.

The dark joke: BIT-0011 was reportedly drafted by Sam Dare himself before his exit. The departing founder wrote the rules designed to prevent founders from departing.

The proposal addresses a real structural weakness — subnet operators could previously dump positions with no governance penalty — but it also concentrates power in the hands of long-term lockers, which is its own form of centralization. Whether that's the right trade depends on what you think Bittensor's main risk is: founder defection or oligarchic capture.

Teutonic and the trillion-parameter moonshot​

Against that backdrop, the rebranded Teutonic subnet (SN3, formerly Templar) has committed publicly to a 1-trillion-parameter decentralized training run for mid-to-late May 2026. That's roughly 14x the scale of Covenant-72B, on the same fundamental architecture, with a community-restored team rather than the original Covenant engineers.

The strategic timing is impossible to miss. Grayscale filed its S-1 amendment for the spot Bittensor Trust ETF (proposed ticker GTAO) on NYSE Arca on April 2, 2026. The SEC's decision window is currently tracked for August 2026. A successful 1T-parameter training run in May would land at the peak of regulator deliberation — exactly when "is this a real technology or a meme?" becomes the load-bearing question. Grayscale already raised TAO's weighting inside its broader AI fund to 43.06% on April 7, the largest single-asset reallocation that fund has ever made.

The bull case writes itself: ship a credible 1T-parameter decentralized model, become the "DeepSeek moment" the ETF approval needs to justify institutional inflow, and reprice the entire decentralized AI category in one quarter.

The bear case is engineering, not marketing.

Why scaling decentralized training is hard in ways frontier labs don't face​

Centralized 1T+ models — GPT-5, Claude 4.7 Opus, Gemini 2.5 Ultra — are trained inside facilities where every GPU is wired to every other GPU through purpose-built fabrics like NVLink and InfiniBand, with sub-microsecond latencies and terabit-per-second bandwidth. Even in those conditions, gradient synchronization is the bottleneck. Published research consistently finds that over 90% of LLM training time can be spent on communication rather than compute when scaling is naive.

Teutonic's miners are coordinating across ~100ms WAN latencies on residential internet. The only reason Covenant-72B was possible at all is SparseLoCo's 146x compression of communication volume. Pushing to 1T parameters changes the math in three uncomfortable ways:

1. Gradient size scales roughly linearly with parameter count. A 14x model means 14x as much data to synchronize per step, even before considering optimizer state.
2. Cross-node coordination overhead historically scales super-linearly with worker count. If Teutonic doubles its node pool from ~70 to ~256, the all-reduce communication cost doesn't just double — it can grow by 4–10x depending on topology.
3. Failure modes compound. A node dropping out mid-step in a 70-node network is a small slashing event. In a 256-node network running 14x larger gradients, the same drop can stall the entire training round.

None of this is unsolvable. There's a body of decentralized training research — heterogeneous low-bandwidth pre-training, FusionLLM, communication-computation overlap, delayed gradient compensation — that targets exactly this regime. But almost all of it has been validated at the 7B–70B scale. A 1T-parameter run on geographically distributed commodity hardware would be a research contribution in its own right, not just a product launch.

The honest read: Teutonic is taking on a research-grade engineering challenge with a marketing-grade deadline. Either it works and becomes the credibility event the entire dTAO ecosystem needs, or it stalls publicly during the SEC's most attentive review window.

The decentralized AI training landscape Teutonic must survive​

Teutonic isn't the only project trying to claim the "credible decentralized 1T-param" milestone in 2026. The competitive map is filling out fast:

• Gensyn launched its mainnet on April 22, 2026 — the same day this article goes out — pairing the launch with Delphi Markets, an AI-driven matching layer for compute jobs. By close of day Gensyn was reporting hashrate equivalent to 5,000+ NVIDIA H100s. Where Bittensor sells permissionless coordination plus a token-incentive flywheel, Gensyn is positioning as a verifiable AI compute marketplace with cryptographic proofs of correct execution.
• Ritual has gone in the opposite direction, leaning into inference rather than training. Its Infernet technology lets any smart contract request an AI output and receive cryptographic proof that the specified model was used unmodified. That's the "verifiable AI in DeFi" thesis, not the "train frontier models from scratch" thesis.
• Ambient and Origins Network are making adjacent bets — different incentive designs, different verification strategies, similar long-term goal of breaking centralized labs' monopoly on frontier training.

These projects don't directly compete on the same milestone, but they all compete for the same finite pool of attention and capital. If Gensyn's mainnet captures the "decentralized AI is here" narrative through commercial workloads, Teutonic's May training run becomes a referendum on whether Bittensor's specific approach — subnet competition plus token-weighted incentives — is the right architecture or the first iteration that gets surpassed.

Why this matters beyond TAO​

Three things are getting tested simultaneously over the next four to six weeks:

Whether decentralized training scales. If Teutonic succeeds, the "Bitcoin of decentralized AI compute" thesis survives. If it fails, the Covenant exit reads as the moment subnet-based training peaked — a 72B ceiling rather than a 72B foundation.

Whether the Conviction Mechanism is the right governance fix. Locking in subnet operators prevents another Covenant-style dump but creates a new failure mode where long-term lockers can entrench. Bitcoin Core's distributed maintainer model, Solana Labs' continued centralized core development, and Sui's Mysten Labs concentration are three different answers to the same question — whether protocol complexity demands a strong central maintainer the community must trust. Bittensor is now running its own version of that experiment in real time.

Whether the ETF window forces decentralized AI to ship on TradFi's calendar. The SEC's August decision window is a hard deadline for a narrative that wants to be "DeepSeek moment" rather than "interesting research project." That's a healthy forcing function or a recipe for over-promising — depending on what gets shipped.

For builders watching from the infrastructure side, the underlying signal is simpler: AI agents and decentralized training networks are about to generate a new tier of on-chain query load — model registry lookups, attestation proofs, gradient checkpoint hashes, subnet performance data — that doesn't fit neatly into the human-facing dApp pattern existing RPC infrastructure was built for.

BlockEden.xyz provides enterprise-grade RPC and indexing infrastructure across 27+ chains for teams building the AI-meets-crypto stack. Explore our API marketplace to build on rails designed for both human and machine traffic.

Sources​

• Bittensor's SN3 Completes Largest-Decentralized LLM Pretraining Run in History — AInvest
• Bittensor's Templar Subnet Completes First Frontier-Scale Decentralized LLM Training — 800.works
• Bittensor Covenant-72B Explained — Phemex
• TAO's DeepSeek Moment: The Rise of Templar (SN3) — PANews
• Covenant AI's Bittensor Exit Triggers 23% TAO Price Crash — Crypto Times
• Grayscale Files Spot TAO ETF as Bittensor Network Rebounds from Covenant AI Exit and 38% Drawdown — Blockonomi
• Bittensor TAO Proposes Locked Stake Governance Fix Following Covenant AI Exit — AInvest
• Bittensor TAO Governance Crisis Explained: Covenant AI Exit and BIT-0011 Proposal — AInvest
• Gensyn launches its mainnet — Startup Fortune
• Beyond A Single AI Cluster: A Survey of Decentralized LLM Training

In March 2026, prediction markets traded $25.7 billion in a single month. That is more notional volume than most mid-cap equity indices. It is not a bubble, and it is not a meme. It is the clearest signal yet that a new asset class — information itself — has finally found a price.

Welcome to InfoFi.

For years, crypto tried to financialize everything: loans, art, cat pictures, liquidity positions, even carbon. But the one thing markets have always struggled to price — the quality of a prediction, the trust of a person, the value of a dataset — stayed stubbornly analog. That changed in 2026. Three previously separate experiments (prediction markets, on-chain reputation, and AI data marketplaces) converged into a single sector with a single thesis: put skin in the game behind information, and the information gets better.

Wall Street has a name for this thesis. It calls it Information Finance. And on current trajectory, InfoFi will cross $10 billion in sector value before the end of this year.