Welcome to #108 of the AI edge.
BTC spent the week doing nothing, which is exactly what you want after a pump. It's still holding 77-82k with no give-back. Boring is bullish.
Where things get interesting is onchain. It isn't Solana or Base this time. It's Robinhood Chain, where almost every new token launch is clearing millions in market cap within hours, giving that Solana 2023-24 feel. Vlad shipped a chain for tokenized equities and got a casino instead, which is the most crypto outcome imaginable.
On the AI side, OpenAI launched GPT-6 Astra this week with a loud benchmark table. 97.6% on FrontierMath Tier 4, 95.9% on BenchCAD, and a real step up on software engineering. The team is calling it a generational leap and gesturing at the AGI word.
OpenAI has been on a good run, with Codex steadily pulling share from Claude Code. The lead keeps changing hands and nobody gets to keep it.
The other big news from the week was Nvidia buying Hugging Face for $12.93 billion. Nvidia already owned the hardware layer. Now it owns the front door developers walk through to get to it.
Jensen insists the platform stays open and that Nvidia compute won't be required to build or deploy on it, which is a generous promise to make when you already own most of the compute anyway.
With that, let's get into this week's edition.
The Big Story: The Biggest Open Training Run Ever Just Started
Training a frontier model normally means one lab, one cluster, one set of decisions about data and hyperparameters made by a small group of people who all work together. Bittensor's Subnet 3 has spent the last few months arguing that you don't need any of that.
We covered Covenant's 72B decentralized training run when it landed. A 10B model (Teutonic-I) has now beaten it, built by people competing against each other rather than coordinating.
The setup is a contest. One model holds the title at any given moment. Anyone can download it, train a better version using whatever data and hardware they can get their hands on, and submit it as a challenger. Validators run both models against the same text and the challenger only takes the title if it wins by enough to rule out luck.
Over 70 days, 2,163 challengers were tested and only 203 got through. The model that survived averaged 62.28% across 11 benchmarks against Covenant 72B's 57.55%, leading on 8 of them.
Last week the team announced that they are pointing the same machinery at a 110-billion-parameter model (Teutonic-II), ten times the size and roughly 50% larger than any decentralized training run before it. Starting from random weights, same open door.
The Wedge
Nobody is donating compute here. The network pays for every accepted improvement, currently north of $6,000 a day in TAO. Bring rented GPUs, a smarter optimizer, merged weights, whatever works. The eval decides, not a committee.
Every checkpoint ships publicly along with the final dataset. That's the part closed labs structurally can't match, and it's what makes the whole thing auditable.
Parameter count stopped being the deciding variable. A model a seventh the size won on 8 of 11 benchmarks, which suggests the search over training recipes matters more than the raw capacity you throw at it.
The Fine Print
Lots of people competed but very few won. One miner took 39% of all titles and the top three took 63% between them. Broad participation at the entry level, heavy concentration at the winning level.
Validators, not miners, decide what data the models get judged on, and they changed that mixture 15 times mid-run. That's a meaningful steering wheel sitting inside something billed as permissionless.
Most of the gain arrived in the first week, and the remaining ten weeks added six points. At 110B, showing up for that grind costs a lot more than it did at 10B. We’ll have to wait and see how it does.
Tessara Watch: The Power Grid Workaround
Everyone building AI data centers has the same problem: the electricity exists, but the connection to it doesn't.
A new grid interconnection can take years. So can delivery on a large gas turbine. But a fuel cell installation can start producing power in months, and that gap has turned an unglamorous corner of industrial hardware into one of the tightest spots in the power stack.
A fuel cell is a power plant with the fire taken out.
Instead of burning gas to make heat and then converting heat into electricity, it runs the chemistry straight into current.
Bloom Energy's version pushes air across one side of a ceramic plate and natural gas across the other, and the electrons take the long way around through an external circuit. No flame, simpler permitting on some fronts, and the output arrives as direct current, which is what the racks want anyway.
Bloom just cleared $1 billion in quarterly revenue for the first time, with margins moving up alongside it. Oracle is pushing gigawatt-scale Bloom deployments and Brookfield has widened the financing behind them.
Supply is answering, though. Bloom is taking manufacturing from roughly 1 GW to 2 GW by the end of 2026, which is why the read stops at Tight instead of Critical.
The more interesting question is what happens if fuel cells succeed. Because solving one bottleneck does not remove the physical system around it.
Fuel cells can bypass the grid interconnection queue. They still need transformers.
That creates a second-order question:
If fuel-cell deployments scale toward multi-gigawatt AI data centers, what becomes the next binding bottleneck?
Or go directly into the underlying work:
We also examine the scandium supply question hanging over Bloom's ceramic stack, the manufacturing ramp, and what evidence would cause us to change the current constraint read.
Hippius (SN75) launched Hippius Hub, an open alternative to Hugging Face that hosts AI models and containers while supporting Hugging Face APIs and standard Docker workflows.
OpenRoboto launched a tournament around its LingBot-VLA 2.0, challenging developers to improve an open robotics foundation model trained on 50,000 hours of robot data across 20 robot embodiments.
OpenMind introduced its research on PHASOR, a universal action representation that lets humanoid robots share and transfer behaviours across different hardware platforms.
404Gen (SN17) released an open-source 3JS game generator, enabling coding agents to turn a text prompt into a playable game using AI-generated 3D assets.
Prime Intellect published the technical report for Prime Agent, detailing new approaches to long-horizon agent memory, multi-agent coordination, and autonomous research workflows.
Synthdata (SN50) launched Synth Ultra, where Bittensor miners compete to predict Bitcoin prices 10 seconds ahead with sub-5ms latency.
🔥 Our Weekly Top Tweets
#1 China Leads the Physical AI Race
A new Physical AI Readiness Index ranks China first, with the U.S. placing fifth at less than half of China's overall capability. The report argues that manufacturing capacity, not better AI models, is now the biggest factor determining who leads the physical AI era.
#2 Inference Is the Next AI Battleground
a16z led a $300M investment in Gimlet Labs, arguing that AI inference is now constrained by physical infrastructure. Gimlet's multi-silicon inference cloud reportedly delivers up to 10× higher throughput per watt, turning efficiency into new compute capacity.
Cheers,
Teng Yan & Arvind
And before we close out..
Go deeper with Tessara
This newsletter gives you our weekly read.
Tessara gives you the research system underneath it.
It continuously tracks the constraints shaping the AI buildout, the companies exposed to them, the evidence changing the state, and the questions that matter next.
Instead of starting from a blank chat, you can research directly against that maintained model.
Two questions worth asking this week:
If data-center power remains constrained through 2027, which public companies gain the most operating leverage?
What evidence would tell me the HBM shortage is actually easing?
