Welcome to #107 of the AI edge.
BTC finally broke out of the 62k to 65k range it sat in all summer, running to nearly 79k before settling around 77k. More than $3 billion in shorts got wiped out on the way up.
The catalyst was a White House crypto summit. Trump got behind the Clarity Act and pushed the Senate to finally move the market structure bill, with Coinbase, Kraken, Gemini and Robinhood all in the room.
Then he mentioned the CFTC is working on bringing Hyperliquid onshore, and HYPE ripped 17% on a comment with no approval or timeline attached. A sitting president name-dropping a perps DEX was not on my 2026 bingo card.
On the AI side, Stripe confirmed it's buying OpenRouter for a reported $7.5 billion. The company raised at a $1.3 billion valuation back in May, so that's close to 6x in three months. Routing between models is turning out to be a real business, which says something about how much everyone is spending on tokens. Their CEO had spent months describing the startup as "Stripe for AI." Turned out to be a more literal pitch than anyone realized.
Let's get into this week's edition.
The Big Story: Compute You Can Actually Keep
Anyone who needs serious compute has had two options. Rent it by the hour from a provider whose pricing moves and whose capacity is never guaranteed, or buy the hardware outright and become a datacenter operator on top of whatever you were actually trying to do. The first is a bill that never stops growing. The second requires capital most teams don't have and skills most teams don't want.
B3 Labs' new product, B3IQ, is a third option.
Put 30% down on an Nvidia rig, pay the rest over five years, and the hardware is yours at the end. An H200 system starts around $54,500. The company procures and assembles it, then either ships it to you or hosts it in its Oregon facility.
The founders came out of Coinbase and started the company building for gaming in 2024, then pivoted once their own customers started screaming about GPU prices.
The Wedge
Renting is a bill. Owning is an asset: Cloud compute charges you every month and leaves you with nothing. A financed machine charges you a fixed amount and leaves you with a GPU. That fixed number is the part that matters, because it's something a grant committee or a CFO can approve once instead of re-approving forever. One of B3IQ's early customers is a PhD student at the University of Hawaiʻi researching skin cancer in people of Pacific Island descent. His grant needs predictable costs and his patient data can't legally sit on someone else's servers. Cloud failed him twice over.
Your idle GPU can pay for itself: Nobody runs a machine flat out around the clock. B3IQ lets owners rent out the downtime through its marketplace, and that money goes against what you still owe or into your pocket. Payouts come in dollars or USDC on Base, with every job recorded onchain as a receipt.
Look at who's buying: Stanford, NYU, Penn, UChicago, Dartmouth, Waterloo. Then startups training video and audio models, and healthcare and finance firms that need compute nobody else can touch. That's real demand for owning hardware.
The Fine Print
Five years is a long time to hold a GPU. Nvidia ships new chips constantly, and an H200 in 2031 won't be worth what it's worth today. Whatever the machine is still worth at the end is your problem, not B3IQ's.
Renting out idle capacity is the open question. Decentralized GPU marketplaces have tried this for years. On Akash, one such example, roughly a third of listed GPUs are actually in use at any time and lease revenue fell 44% last quarter. B3IQ's is curated rather than an open auction, and H200s pull more demand than a mixed fleet of older cards, so the comparison isn't exact. But the pattern is worth watching because plenty of people are wanting to rent out capacity, far fewer wanting to rent it.
The model makes sense on paper. Whether it holds up comes down to resale value and rental demand, and both are worth watching.
Tessara Watch: The Next Bottleneck
This week's bottleneck is a part the size of a grain of sand.
A multilayer ceramic capacitor, or MLCC, sits throughout the GPU's power-delivery network and helps absorb rapid changes in current. A single GB300 rack contains hundreds of thousands of them.
Individually, they cost almost nothing. Collectively, their value per rack is rising fast. Morgan Stanley's teardown puts MLCC content at roughly $1,530 per GB300 rack and $4,320 per VR200, an increase of about 182%.
But the interesting part is that there is no broad MLCC shortage.
Consumer-grade parts remain well supplied and headline pricing is still soft. The squeeze sits in a much narrower category: high-capacitance, low-inductance MLCCs qualified for AI accelerators.
That distinction matters because the aggregate data can look healthy while the specific grade AI systems need is getting tighter.
Murata controls roughly 45% of AI-server-grade supply and is running near 95% utilization. Taiyo Yuden's capacitor book-to-bill reached 1.7 last quarter, meaning orders are arriving materially faster than sales. SEMCO raised prices across its book by 30% in August.
And supply cannot respond quickly. New high-end lines take time to install, then still need to pass qualification before they can serve AI customers.
Simply knowing that MLCCs are tight is not enough.
It is knowing whether the constraint is getting tighter or easing, which companies have enough exposure for it to matter, and what evidence would make the thesis break.
That is what Tessara keeps live.
For MLCCs, Tessara tracks the current constraint state, the evidence behind it, the public companies that benefit or get squeezed, how material that exposure is, and the signals that would cause us to reverse the call.
This newsletter is the snapshot. Tessara is the live model underneath it. See it here:
Eastworlds reached 200 hours of humanoid teleoperation per week, becoming the largest source of Unitree G1 training data outside China while expanding commercial deployments.
IOTA (SN9) surpassed 100B training tokens with Orion-16B, now the largest LLM pretrained across globally distributed commodity GPUs.
Hippius (SN75) launched Hippius Drive, bringing end-to-end encrypted, decentralized cloud storage to the Bittensor ecosystem.
OpenRoboto (SN80) launched its Open Data Pool with Axis Robotics, giving miners access to real-world robot data to improve robotics models through open competition.
Prime Intellect released Prime Flash MoE, a Blackwell-optimized inference engine that accelerates Mixture-of-Experts models through fused CUDA kernels for faster, more efficient AI inference on NVIDIA B200 GPUs.
Teutonic (SN3) released Teutonic-I 10B, outperforming every other decentralized pretrained model tested and beating larger models like Quasar-Preview 18B and Covenant 72B across most benchmarks.
🔥 Our Weekly Top Tweets
#1 China Is Closing the AI Gap
Bloomberg's latest comparison shows Chinese frontier models rapidly closing the performance gap with U.S. labs while slashing inference costs, making it much harder for frontier labs to command premium pricing.
#2 OpenRouter Builds Its Own Frontier Model
OpenRouter launched Ox Alpha, a new frontier model built for coding and long-running AI agents, featuring a 1M-token context window and support for text, image, and video inputs.
Cheers,
Teng Yan & Arvind
Quick recap: I also publish a newsletter on the AI buildout and supply chain. 1-2 memos a week.
