Welcome to #105 of the AI edge.
This week Bitcoin held the 60-65k range all week while stocks bled red and the S&P slid session after session. Crypto steady as equities sink. We've seen this pattern before, and it rarely ends with everyone shaking hands. Our restlessness continues.
On the AI front, Moonshot released Kimi K3, billed as the most open frontier model going and the first open model at 2.8 trillion parameters.
It topped a coding benchmark this week, edging past Claude Fable 5 on front-end tasks, though it still trails overall and the weights don't ship until July 27. China's labs keep chipping the gap down month by month.
And SK Hynix, the Korean memory chipmaker behind most of the world's high-bandwidth memory feeding AI servers, went live on the Nasdaq. It raised around $26.5 billion, the largest US listing ever by a foreign company and the second-biggest behind SpaceX.
Shares popped 14% on day one, then its Seoul stock had its worst day in nearly twenty years a few sessions later. A record debut and a near-record crash in the same week. Peak timing.
With that let’s get into this week’s edition.
The Big Story: The World's Largest 3D Dataset Came From A Subnet
3D generative AI has a supply problem. Teaching a model to produce usable 3D objects, the kind games, AR headsets, and virtual worlds are built from, takes an enormous library of examples to learn from, and those examples are slow and costly to make because a human has to model each one by hand. No team of modelers can churn them out fast enough to feed a hungry model.
This week 404, the team behind Bittensor Subnet 17, said it has cleared that wall. Its new repository holds more than 21.5 million 3D assets across roughly 40TB, and the team claims that beats every other open-source 3D dataset put together. The twist is who built it. No studio, no in-house pipeline, just a swarm of independent miners.
The Wedge
The assets are synthetic, meaning generated by machines rather than sculpted by people. It's also exactly the repetitive, parallel grind a decentralized network is suited for. Miners across the subnet generate assets, grade each other's output, and filter the weak ones, all pushed along by token rewards.
Volume alone isn't the pitch. The dataset ships open-source and attribution-ready, with metadata, usage rights, and ownership attached to each asset, which is the unglamorous paperwork that decides whether a studio can legally train on it at all. It's aimed at gaming, AR/VR, and techniques like Gaussian Splatting and NeRFs, two methods for turning flat photos into full 3D scenes.
It answers a real shortage. The bigger the "world models" teams want to build, systems that simulate 3D space rather than just text, the more 3D data they need, and the open supply has been thin. This drops a large pile of it into the commons at once.
The Fine Print
Every headline figure here is the project's own. No outside group has confirmed the assets are as high-fidelity as claimed, and 21.5 million mediocre models would still fill 40TB. Size is easy to measure. Usefulness is not.
Synthetic data cuts both ways. Train on enough machine-made assets and a model can start absorbing the flaws in that data instead of the richness of the real world.
Decentralized networks have always been good at brute-force parallel work, and generating millions of 3D assets is about as brute-force as it gets. Whether this turns into infrastructure the 3D world actually leans on or just the biggest pile nobody picked through comes down to the one thing a subnet can't mint: adoption.
Tessara Watch: The Old Fab Inside The Chip Shortage
Everyone is watching GPU supply.
This week, the tighter bottleneck may be the chips that make GPUs usable at all.
Analog power-management chips sit between the grid and the processor, converting, regulating, and protecting electricity before it reaches the silicon. As AI racks move beyond 100 kilowatts and processors draw more than 1,000 amps at under one volt, every additional watt requires more power-management content.
Demand is rising faster than supply.
Analog Devices told customers in early July that lead times for some components had stretched to roughly six months. Pricing across the category has also moved sharply higher. Once one of these components is designed into a system, replacing it is slow, expensive, and often impractical.
The supply problem is harder to fix than it looks.
Many of these chips are produced on older 8-inch wafer lines. The industry has spent years directing capital toward newer and more profitable leading-edge fabs instead. New 8-inch equipment is scarce, additional capacity can take years to qualify, and Samsung is closing mature-node capacity as demand accelerates.
The capacity that is coming is concentrated in China, creating a second constraint around tariffs, sourcing rules, and supply-chain dependence.
This week’s Tessara read
Constraint: Mature Node Wafers
Status: Tightening
What is driving it: AI rack power density, long lead times, limited 8-inch capacity
Companies exposed: Texas Instruments, Analog Devices, Monolithic Power Systems, and several others
Next checkpoint: Texas Instruments earnings
The important question is not whether analog demand is strong.
It is who benefits, who absorbs the higher costs, and how long the shortage can persist.
Tessara maps the full chain, scores each company by exposure and materiality, and updates the read as supplier filings, earnings calls, and capacity announcements arrive.
Tessara is the research terminal for the physical AI buildout.
It tracks constraints across compute, memory, packaging, networking, data centers, cooling, and power, then maps the public companies exposed to each one.
Pro members get:
the live bottleneck board
company exposure maps
pre-earnings research and forecasts
Last week, Tessara forecast that TSMC would raise full-year capital expenditure guidance to $60–64 billion. TSMC raised it to $60–64 billion three days later.
Every forecast is timestamped before the event and graded after it.
Bittensor activated Conviction, a protocol-level staking mechanism that rewards long-term subnet commitment and enables community-driven ownership transitions.
Vidaio (SN85) launched Sentinel, a surveillance platform built on its AI video compression engine, while integrating its models with ManakoAI to reduce bandwidth costs for edge vision deployments.
Chutes (SN64) achieved fully non-blocking decentralized training with Parallax, matching centralized training within just a 0.6% quality gap.
TAO.com rebuilt the original Bittensor Chrome wallet, adding subnet token trading, portfolio tracking, staking, and real-time pricing in a redesigned interface.
Eastworlds partnered with Unitree Robotics to accelerate embodied AI deployments, combining robotics hardware with Eastworlds' data and deployment infrastructure across Southeast Asia.
Stillcore Capital published its State of Subnets (July 2026) report, offering an in-depth look at capital flows, performance, and trends across the Bittensor ecosystem.
🔥 Our Weekly Top Tweets
#1 Robinhood Chain's Agent Economy Accelerates
Robinhood Chain has already processed $77M+ in agent volume, with 2,100+ agents launched and $1.3M+ earned by builders, highlighting the rapid growth of agent-native consumer finance.
#2 The Agent Shift Inside OpenAI
Agents are rapidly becoming part of everyday work inside OpenAI. Since August 2025, Codex now generates 99% of engineering work, with adoption also reaching 91% in finance, 89% in recruiting, and 88% in legal.
Cheers,
Teng Yan & Arvind
Quick recap: I also publish Tessara Research, focused entirely on the physical AI buildout and the companies exposed to its constraints. 1 - 2 research memos each week.
