Welcome to #97 of the AI edge.
BTC retested last week's $79.5k high and got rejected again. $80k remains the wall. Still holding above $75-77k though, so the structure hasn't broken. But until that level flips, we're stuck watching the same ceiling.
On the AI front, OpenAI is having a moment and not the good kind.
The Elon Musk lawsuit went to trial this week, which means weeks of dirty laundry aired in public. Reports also surfaced that they missed 2025 revenue targets, with CFO Sarah Friar reportedly warning leadership that spending commitments are at risk if growth doesn't accelerate.
And to top it off, Anthropic has officially passed them in annualized revenue as of April. Model lead gone. Revenue lead gone. The company that defined the AI race is suddenly chasing it. Sam Altman is either cooking something massive behind closed doors or the walls are closing in.
Meanwhile, the money behind AI keeps accelerating. Three major hyperscalers reported earnings this week and every single one raised or reaffirmed higher capex. Meta bumped guidance to $125-145B. Alphabet guided $180-190B. $500B+ in committed AI infrastructure capex, and the question every AI investor should be asking is where it actually flows.
I tracked every print this week on Tessara Research, my AI infrastructure newsletter. Capex broken down by vendor, regime scoring across HBM, packaging, and power, names exposed at each chokepoint. If you trade the AI trade and you've been guessing at the supply chain side, this is the gap-fill. Free, 2-3x a week. Subscribe →
With that, let's get into this week's edition.
The Big Story: Open Source Agents Are Beating the Labs
Oro went live on Bittensor recently, a decentralized arena where developers submit autonomous AI agents that compete head-to-head on real-world tasks. The first battleground is shopping. Miners build agents, validators evaluate them independently in sandboxed environments, and the best performers earn emissions. All this without an application process or gatekeepers. Ship your agent and let the benchmark decide.
The early results are hard to ignore. Since it went live, open-source agents competing on Oro have already topped GPT-5.4 on ShoppingBench, a grueling evaluation suite built on over 2.5 million real products. Forty-five agents crossed that line without billion-dollar research budgets or closed-door training clusters. Just global builders iterating daily with skin in the game.
The Wedge
Every major lab is pouring resources into proprietary agent development. Oro is testing whether open competition with direct financial incentives can outpace them. Three weeks in, the scoreboard says yes. Over $85k in rewards already distributed to builders with everything fully open source.
The thesis is simple. If a task can be benchmarked, Bittensor can commoditize it by incentivizing a global pool of miners to chase the highest score at the lowest cost.
The founder, Shardul Bansal, was one of the earliest engineers on the Bittensor protocol. So this isn't a new team parachuting into an unknown ecosystem.
The Fine Print
Shopping agents are a clean starting point because the tasks are measurable and well-defined. The harder question is whether this model translates to messier, less benchmarkable agent use cases.
Beating GPT-5.4 on a specific benchmark is a strong headline. But benchmarks are just snapshots. Frontier labs iterate fast and the gap can close quickly. Oro needs sustained outperformance, not just a single leaderboard moment.
Open source has always been chronically underfunded despite being the more successful development model. Oro is trying to fix that by turning competition into compensation. If it works, the playbook changes for every open-source AI project that comes after it.
FractionAI launched Index, a platform that turns plain-English trading ideas into live perpetual futures agents, letting users backtest, iterate, and deploy AI trading strategies directly into the market.
TIG Foundation launched Prometheus, a system for collaborative algorithm mining and monetization at scale without traditional gatekeepers.
Hermes Agent introduced Hermes Curator, an automated system that tracks, consolidates, and prunes user-created skills weekly, keeping agent workflows clean, optimized, and manageable without touching pinned or built-in skills.
Lium pivoted to an agent-first compute model, enabling AI agents to autonomously create accounts, fund GPUs, deploy models, and report UX friction.
Ethy V2 introduced revenue sharing, allowing users to stake $ETHY and earn a portion of real product revenue.
USDai launched $CHIP staking, offering boosted Allo points (up to 10x) and incentives for early participants.
HeyElsa integrated x402 payments, enabling agents to combine reasoning workflows with native onchain financial execution.
Bittensor activated Treasury Wallets at the protocol level, unlocking new capital coordination mechanisms across the ecosystem.
📡 From Tessara Research this week
The biggest AI infrastructure earnings week of 2026 just closed. $500B+ in capex tested across AMZN, GOOGL, META, MSFT, Samsung, AAPL, ETN, and PWR.
Before the prints, I laid out the megawatt thesis. What I’m watching out for every name, where the binding constraint should move next. Read it here
The post-mortem on what landed and what broke drops on Tessara this week. Subscribe to Tessara Research →
🔥 Our Weekly Top Tweets
#1 Autonomous Drug Discovery Agents Go Live
This team launched a fleet of autonomous agents running a 9-gate peptide discovery pipeline 24/7, publishing every decision onchain and automatically paying for wet-lab validation from its own wallet, removing human handoffs from target to synthesis.
#2 You’re Probably Using Only 8% of Hermes
A deep dive into 15 overlooked Hermes feature. From persistent memory and session branching to multi-model routing, cron automation, and custom slash skills. All that turn it from a chat interface into a full-fledged agent operating system.
Cheers,
Teng Yan & 0xAce
P.S. If you want the supply-chain side of every AI capex print, my other free newsletter Tessara Research has you covered. Memos, earnings previews, regime scoring across HBM, packaging, and power. [Subscribe →]
P.P.S. If you’re a startup doing serious work in AI or robotics and want to work closely with us, reply to this email and let’s chat.
