Welcome to #100 of the AI edge.
BTC slid back down to 76.5k this week after hovering in the 79-82k range the last couple of weeks. Looks like we're grinding toward the lower end of the range. Hoping for a bounce, but 80k feels further away than it did seven days ago.
The AI front had a plot twist nobody saw coming. Andrej Karpathy, one of the most respected AI researchers alive and a co-founder of OpenAI, joined Anthropic this week.
Karpathy has done a remarkable job staying neutral in the frontier lab wars, even with his OpenAI history. But picking a side now, and picking Anthropic, had to have stung Sam & Co. a bit.
Speaking of OpenAI, the WSJ dropped an exclusive that they're preparing to file for an IPO, possibly as early as May 20. That obviously didn't happen, but the race to see which of the two big labs goes public first is very much alive.
Makes you wonder if Karpathy has any OpenAI allocation. If so, he might be one of the happiest guys in AI when both IPOs eventually land.
With that, let's get into this week's edition.
The Big Story: Machines Just Started Paying Their Own Bills
Until now, every robot that needed compute, storage, or data had a human somewhere footing the bill. Subscribing to APIs. Provisioning cloud resources. Managing payments. The machine did the work. The human handled the money.
peaqOS just changed that with the launch of Scale. Robots and machines running peaqOS can now independently find services, purchase them per request from their own onchain wallets, and settle payments across multiple chains. A drone that needs extra processing power buys compute on its own.
A robotaxi stores its sensor data on decentralized storage without asking anyone. A humanoid secures its own IP onchain. Every transaction is autonomous, within operator-defined spending limits and permissions.
The marketplace launches with twelve services out the gate, spanning frontier AI inference from Anthropic, OpenAI, and Google, semantic search, document parsing, translation, captcha solving, and decentralized storage.
The Wedge
peaqOS has been building this stack in layers. First came Activate, giving machines their own identity and wallets. Then Qualify, which assigned each machine a dynamic credit rating. Scale is the third piece, connecting those identities and wallets to an actual marketplace where machines spend, earn, and compound value. Each function makes the previous ones more useful.
The financial layer goes beyond just spending. Machine Money Markets let robots earn yield on idle revenue, borrow against future earnings, and post collateral.
For Web3 infrastructure protocols that have been building compute, storage, and data networks for years, this opens up an entirely new demand side. The customers aren't humans anymore. They're machines that need these services at scale to operate.
The Fine Print
Twelve services at launch is a starting point. The value proposition for operators depends heavily on how fast that catalog grows and whether the services available actually match what their machines need day-to-day.
Cross-chain settlement and autonomous spending sound clean on paper. In practice, edge cases around failed transactions, disputed payments, and agent misbehaviour at machine scale will test how robust the guardrails really are.
Every week the gap between "machines that work" and "machines that operate as independent economic participants" gets smaller. peaqOS just collapsed another chunk of it. The robot economy keeps graduating from pitch deck to production.
Constraints Watch: $9,500 → $115,000 per AI rack?
Everyone's been watching the power grid. Substations, transformers, interconnection queues. That layer still matters. But a new chokepoint is forming inside the data center itself, and it's all about power semiconductors, the tiny components that convert, regulate, and deliver electricity from the building to the GPU.
AI racks are moving from 120 kW to 600+ kW. At those power levels, traditional silicon components that handle voltage conversion start losing efficiency. Too much heat, too much energy wasted. The industry is shifting to two advanced materials: silicon carbide (SiC) and gallium nitride (GaN). They switch faster, run hotter, and lose less power in the process.
The dollar shift is staggering. ON Semiconductor, one of the major global players in power semiconductors, put it plainly on their earnings call: a current-gen 120 kW data center rack has roughly $9,500 worth of power semi content. A next-gen 800V rack has $115,000. That's 12x on the same unit of compute.
ON’s Earnings Takeaway on Tessara
Demand is already showing up. ON Semi's AI data center revenue grew over 30% quarter-on-quarter and they're now rationing parts to customers. Wolfspeed, a specialist in SiC chips, saw AI applications grow 30% sequentially on top of 50% the quarter before.
The constraint? Growing silicon carbide crystals is still slow and expensive. New fabs take years to build. Qualification cycles for high-voltage parts can stretch past 12 months. Nothing is fully binding yet, but every quarter of 30%+ AI growth tightens the screws.
If you want to learn more about how power semiconductors will be the next big bottleneck in AI - there’s a full breakdown in this week's Chokepoint (our AI infra newsletter)
Fetch.ai released “Collabs,” enabling structured AI-to-AI handoffs with transparent activity tracking.
Venice began rolling out agentic chat to new users, delivering private-by-default, multimodal AI interactions.
Crypto.com added $TAO and subnet staking via Yuma validators, bringing Bittensor yield access to its 150M+ users.
Chromia launched ATBASH, an on-chain policy engine that routes and verifies sensitive agent actions, adding programmable safety and accountability to decentralized AI systems.
Aethir introduced CARA, an AI due diligence agent generating instant reports on funding, teams, tokenomics, and competition.
Fraction AI launched Index, turning plain-English trading ideas into backtestable and deployable AI trading agents.
Hippius Subnet (SN1) burned over $5M worth of tokens, reducing supply by 11% and signaling strong on-chain economic activity.
🔥 Our Weekly Top Tweets
#1 Base's Agent Economy in Numbers
217k+ unique buyers and 32k+ unique sellers on Base's agentic marketplace in the last 30 days. Agents are transacting at scale and the demand side is outpacing supply nearly 7 to 1.
#2 Google Hits 3.2 Quadrillion Tokens Per Month
Google's monthly token volume jumped 7x year-on-year, from ~480T to over 3.2 quadrillion across all its surfaces. The scale of inference demand is entering territory that's hard to even comprehend.
Cheers,
Teng Yan & Arvind
Quick recap: I launched The Chokepoint, a new weekly newsletter on the AI supply chain. Free, every Tuesday on Tessara Research.
