Welcome to #98 of the AI edge.
BTC finally broke the 80k wall this week, hitting a high of $82.5k. The range that held us hostage for weeks is behind us. And just as things were looking good, Saylor floated the idea of selling some BTC to pay a dividend. We really can't have anything nice.
On the AI front, the semiconductor bull market is running hot. Most semi stocks are climbing and every new earnings call pushes them higher. Intel has practically been printing new all-time highs every week for the past month.
And the reason is becoming obvious. The demand for compute is real. It’s real enough that Anthropic closed a deal with Elon Musk's SpaceX this week for massive compute capacity. Rate limits across all Claude products are set to increase.
For context, this is the same Elon who once tweeted "Anthropic is misanthropic." When you're agreeing to infrastructure deals with the guy who publicly trashed you, the demand for compute has clearly outrun pride. Dire times call for dire partnerships.
With that, here's the question every AI investor should be asking: where is all that compute getting bottlenecked? I’ve spent 500+ hours diving into AI infrastructure this year, and I just shipped The Chokepoint, a new weekly newsletter on the AI supply chain where all the action is happening in the markets. Edition #1 dropped Tuesday.
The headline finding: power just overtook memory as the binding bottleneck, the largest weekly move on our entire constraint board. The rent is migrating from the fab to the substation.
Free. Every Tuesday on Tessara Research. Get it here →
With that, let's get into things.
The Big Story: Autonomous Spending Just Hit Mainnet
Back in August 2025, we published a deep dive on KiteAI when it was still running on testnet. The thesis was straightforward: AI agents can reason, but the moment they need to spend money, book a service, or settle a transaction, everything stalls. The infrastructure for autonomous economic actors simply didn't exist. We called it the missing layer.
This week, that layer went into production. KiteAI launched its mainnet, an EVM-compatible L1 built on Avalanche, purpose-built for agent-driven payments. Alongside it, they opened Kite Agent Passport to the public.
Passport lets a user grant an AI agent permission to spend within defined limits, a budget ceiling, a time window, approved assets, and per-transaction caps. The agent operates freely inside that box. No credit card exposed. No private key handed over. No human approving every half-cent API call.
The demo that ships at launch tells you where this is headed. You tell an agent "buy me snacks on Amazon under $25." It searches, filters, applies your rules, completes the checkout, and logs every step.
The Wedge
KiteAI is deliberately not chasing TVL or DeFi liquidity cycles. They're measuring success by transaction count, active passports, and settlement volume. In a space full of ghost chains inflating vanity metrics, that positioning stands out.
The passport system solves a problem every agent builder hits early. Giving an agent a wallet without restrictions is reckless. Approving every micro-transaction manually defeats the purpose. Scoped delegation with cryptographic guardrails is the working middle ground .
The stack already integrates with major agent harnesses like Claude Code, Cursor, and Cline out of the box. One install line and your agent gets spending capabilities inside whatever environment it's already running in. That's the kind of low-friction onboarding that actually gets developers to show up.
The Fine Print
The hardest part is convincing real-world services to plug in. If major platforms like Shopify, DoorDash, or travel APIs don't integrate, agents are transacting in a closed loop. KiteAI's success is as much a business development story as a technical one.
The timing might be perfect. Building agents has never been easier, the narrative is rapidly going mainstream, and every week another team ships a live use case. The race to own the payment rails for the agent economy is heating up.
KiteAI just planted its flag.
Yotta Labs launched a multi-cloud, multi-silicon orchestration layer, offering a unified API to reduce AI infra fragmentation.
Chutes.ai entered a collaboration framework with Nasdaq-listed AIxCrypto (AIXC) to explore decentralized compute support for real-time, multi-agent environments.
Ridges AI (SN62) unveiled its Q2 roadmap, adding BYOK for miners, screener auto-scaling, variance reduction, and an Infinite SWEBench for synthetic tasks.
Heurist integrated with Visa Intelligent Commerce via Inflowpay, enabling agents to discover skills and settle payments autonomously mid-task.
Allora Network integrated zero-knowledge proofs (via Brevis) into its reputer layer, enabling end-to-end verifiable AI inference and provable model weight updates.
🔥 Our Weekly Top Tweets
#1 Virtuals Dominates Base Agents
43% of all agentic transactions on Base now flow through Virtuals Protocol, making it the clear infrastructure layer powering the majority of onchain agent activity.
#2 RL Beats Opus on Sheets Tasks
Ramp partnered with Prime Intellect to build “Fast Ask,” an RL-trained subagent for spreadsheet retrieval that outperforms Opus by +4% on exact match accuracy at Haiku-level latency, optimizing directly on deterministic finance tasks like invoice IDs, dates, and cent amounts.
Cheers,
Teng Yan & 0xAce
Quick recap: I launched The Chokepoint this week, a new weekly newsletter on the AI supply chain. Edition #1 went live Tuesday. Headline: power just overtook memory as the binding bottleneck. Free, every Tuesday on Tessara Research. Read it →
