Welcome to #99 of the AI edge.
S&P hit a new all-time high this week. BTC chopped between 80-82k with a quick wick down to 79k. Range still holding. We'll take it.
The bigger story was Trump's first China trip in a decade. He didn't go alone. Musk, Jensen, Larry Fink, and a roster of US CEOs flew in with him.
Jensen called it "one of the most important summits in human history." Translation: if this lands, Nvidia gets back into the second-largest compute market on earth. The one geopolitics has been quietly bleeding them out of for years.
Speaking of Jensen's competition, Cerebras, the inference chip company, finally IPO'd this week. The IPO was priced at $185 but it opened at a whopping $350. Being in the middle of a semi bull market certainly didn't hurt.
Now we get to see if the price action holds up over the next few months or if its going to be like the rest of the IPOs in recent years: pump on opening, then down only.
With that, let's get into this week's edition.
The Big Story: Your Agent Just Got Its Own Visa Card
Virtuals Protocol shipped EconomyOS this week. In under two minutes, developers can now spin up an autonomous agent that comes pre-loaded with its own Visa card, crypto wallet, email address, and verifiable identity.
One suite. No stitching together a dozen different services that were never designed for non-human users. Until now, agents that needed to transact were borrowing human credentials or mostly keeping it onchain. Your card. Your email. Your login.
EconomyOS gives them their own. An agent can now shop, book travel, handle email, and settle payments without ever touching its creator's personal accounts. Tokenization is also now optional through the console, lowering the barrier for teams that want to build productive agents without launching a token first.
The Wedge
The full stack bundles neatly on top of ACP, the same Agent Commerce Protocol that powered the robot-to-robot delivery loop we covered a few weeks back. Virtuals keeps layering infrastructure on infrastructure. Each release makes the last one more useful.
The barrier to deploying an economically autonomous agent just dropped to near zero. Virtuals is positioning itself to capture builders the moment they show up.
A Visa card is a bridge the crypto world rarely builds. Agents paying onchain is one thing. Agents swiping a card at any merchant that accepts Visa is an entirely different level of real-world reach.
The Fine Print
Giving autonomous software its own payment credentials is powerful and risky in equal measure. The guardrails around spending limits, identity verification, and fraud prevention will matter more than the feature set itself.
Virtuals keeps shipping at pace but adoption metrics are what count next. Infrastructure without builders using it is just a tech demo. The question is whether developer activity follows the tooling.
An agent with its own card, wallet, email, and identity isn't borrowing a seat at the economic table anymore but instead, its bringing its own.
Constraints Watch: The Bottleneck Beneath the AI Bottleneck
Crypto-AI is a downstream bet on the AI buildout. Every agent building on Virtuals, every inference token traces back to the same physical supply chain. When that chain hits a wall, the whole stack notices.
This week it hit one. Photonics stocks (AAOI, LITE, COHR) tanked after earnings last week, then ripped to fresh all-time highs by Tuesday. The sector is one of the big choke points of the AI buildout right now. Underneath it sits an even tighter constraint…
Indium phosphide (InP) lasers.
InP laser is what actually emits the light inside the high-speed optics networks wiring up NVIDIA's GPU clusters.
All three companies flagged it on their calls. Lumentum (LITE) is sold out through 2027. Coherent (COHR) has bookings stretching into 2028. Capacity expansions are coming, but new laser fabs take roughly two years to spin up at minimum.
And the bottleneck is already migrating upstream to InP substrate (the raw wafers lasers are built on) suppliers like AXT, where hyperscalers are reaching past the laser makers to lock in supply directly. Same playbook that played out in HBM, one tier deeper.
NVIDIA's $4B stake in LITE and COHR earlier this year is starting to make a lot more sense.
If you want to learn all about what’s driving the photonics bull run this week - there’s a full breakdown in this week's Chokepoint (our AI infra newsletter)
USDC is now live in NEAR’s AI Agent Market with Confidential Intents, enabling private, dollar-denominated agent payments across 34+ networks.
RESI is launching a Bittensor subnet focused on off-market U.S. property valuations, targeting real estate without active price discovery.
General Tensor acquired Backprop Finance, consolidating a major share of Bittensor’s subnet trading volume.
GenLayer launched arguedotfun on Base, an AI-powered argumentation market using AI-native consensus to evaluate debates and settle pools.
Peaq introduced “Initial Machine Offerings” with CoinList, enabling tokenized, yield-bearing robot ownership for global investors.
Akash Network kicked off its Homenode pilot, letting RTX GPU owners lease idle hardware for rewards.
OpenTensor announced Subtensor upgrades “Conviction” and “TAO Flow V2,” alongside highlighting new subnet teams expanding network utility.
Nous Research released Token Superposition Training (TST), achieving 2–3× pretraining speedups at matched FLOPs without changing model architecture.
Fetch.ai launched AEVS, providing tamper-evident, publicly verifiable receipts for every agent execution.
Score Subnet 44 unveiled “Detect-fire,” a real-time vision agent for fire detection in environments like fuel stations.
🔥 Our Weekly Top Tweets
#1 Claude Recovers 5 BTC After 11 Years
After ~3.5 trillion failed password attempts, a forgotten seed phrase from a college notebook matched with an old wallet file Claude uncovered and unlocked 5 BTC that had been inaccessible for over a decade. It did not crack the encryption but it connected the dots. Total compute cost was about $15.
#2 Autonomous Agents Break the nanoGPT Speedrun Record
Prime Intellect let Codex and Claude autonomously iterate on the nanoGPT optimizer track for two weeks, running ~10,000 experiments across 14k H200 hours. The agents beat the human baseline and pushed the step frontier down to 3345 steps, showing that LLM agents can meaningfully optimize training loops, though still relying on human priors to discover new directions.
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.
