Daily TEA – AI Bills Eat Salaries, CCP-China Kills Meta’s $2B Manus Deal

AI compute vs payroll, harness shelf life, Manus blocked, Binance agentic wallet, deterministic vs predictable code

Hello, dear TEA-mates! Here is what you need to know today.

1. 💸 AI Compute Is Now More Expensive Than Employees At Some Companies

IT budgets are getting torched as some firms spend more on AI than on salaries. Bryan Catanzaro, VP of applied deep learning at Nvidia, told Axios that for his team, compute costs far exceed employee costs. Uber’s CTO already blew through his full 2026 AI budget on token costs, and Swan AI CEO Amos Bar-Joseph publicized his Anthropic bill on LinkedIn while declaring he is scaling with intelligence, not headcount. Gartner projects worldwide IT spending will hit $6.31 trillion in 2026, up 13.5% year over year. The piece flips the usual narrative: instead of AI replacing expensive humans, AI compute itself is becoming the line item that breaks the budget. (Read More)

🫖 TEA For Thought: “Well, it takes some time and money to set up. But you have to look at the long term, right?”

2. 🪤 Today’s Harness Is Tomorrow’s Prompt

Tanay Jaipuria argues that every “harness” we build around LLMs (RAG pipelines, JSON parsers, OCR layers, multi-agent orchestrators, browser scripting, voice cascades) is a workaround for a current model limitation, dressed up as architecture. He walks through harnesses already eaten: chat-with-PDF was killed by 1M-token context windows, structured output became a single API parameter, OCR pipelines died once vision models started reading receipts directly. He bets against today’s harnesses too: AutoGen and CrewAI swarms, Playwright-based browser agents, and Whisper plus LLM plus ElevenLabs voice stacks. His conclusion: keep building harnesses, but build them cheap enough to throw away once the next model release absorbs them. (Read More)

🫖 TEA For Thought: “If whatever you’re going to be building will eventually get eaten, what exactly is it that we can build now? Something every builder should be thinking about.”

3. 🚫 CCP-China’s NDRC Vetoes Meta’s $2B Manus Acquisition After Months-Long Probe

CCP-China’s National Development and Reform Commission ordered Meta and Manus to unwind their roughly $2 billion deal, one of the country’s most aggressive cross-border interventions in the AI sector. Manus, an agentic AI startup founded in Beijing in 2022 by Xiao Hong, Yichao Ji, and Tao Zhang, relocated to Singapore mid-2025 before Meta announced the acquisition in December 2025. Around 100 Manus employees had already moved into Meta’s Singapore offices by March, with CEO Hong reporting directly to Meta COO Javier Olivan. Hong and Chief Scientist Ji are reportedly under exit bans preventing them from leaving mainland China. Meta said the transaction complied fully with applicable law and expects an appropriate resolution. (Read More)

🫖 TEA For Thought: “Very, very unfortunate. There are a lot of great AI scientists in China, but I guess there’s no freedom in CCP-controlled China. When there is no democracy, there is no freedom for innovation. If there is innovation, it has to be controlled by the party.”

4. 🤖 Binance Launches Agentic Wallet For AI-Driven On-Chain Transactions

Binance unveiled Agentic Wallet, a keyless, isolated sub-account inside Binance Wallet that lets AI agents trade, transfer, and manage assets within preset parameters without touching the user’s primary funds. Features include spending limits, token restrictions, whitelisted-address-only transfers, and a real-time agent activity dashboard. Launch supports BNB Smart Chain, Solana, Base, and Ethereum, with one wallet per user, and is compatible with OpenClaw, Claude Code, and Cursor agent frameworks. Winson Liu, Global Head of Binance Wallet, framed the product as extending the Binance AI experience from the exchange into Web3. A 15-day promo offers up to 20 gas-free transactions per user, capped globally at 200,000 transactions, alongside waived service fees during the period. (Read More)

🫖 TEA For Thought: “Agentic wallet is what every fintech crypto company is trying to build. Infrastructure is not ready yet.”

5. 🎲 LLM-Assisted Coding Is Not Deterministic. Does It Matter?

Vrypan separates determinism (same inputs always produce same outputs) from predictability (we can foresee the result with our tools and time). Weather is deterministic but only narrowly predictable. Casino odds are non-deterministic at the event level but statistically predictable. He argues human-written software was never deterministic at the production stage either: developers, like LLMs, are problem-solving processes operating under uncertainty, and end users only care whether the resulting program behaves correctly most of the time. The industry already built tests, staging, observability, rollbacks, and reproducible builds to handle non-determinism. He points to DO-178C, the aviation safety-critical software standard, which is objective-oriented rather than method-prescriptive, as a framework that fits both human and LLM coders. (Read More)

🫖 TEA For Thought: “This speaks to the paradigm shift from the previous computer and internet era, where all the code is deterministic, when everything is binary, 0 or 1, to tokens, which are unpredictable and predictable at the same time, but definitely not deterministic. What matters now, besides verification, is also the infrastructure and system to make sure that whatever output is generated by agents or coding agents can be verified, tested against, and validated before release.”

🛠️ Tools of the Day

mattpocock/skills — Skills for real engineers, lifted straight from Matt Pocock’s .claude directory. 28.5K stars.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:
Twitter/X