Welcome to #104 of the AI edge.

BTC spent another week testing nerves, dipping back below 58k before closing the week north of 61k. We've run that round trip enough times to know the shape of it, but the mood underneath is shifting.

Memecoin trenches are printing $100m tokens again, crypto x AI fundraises are slowly waking back up, and risk appetite is creeping back into the room. Still cautious, still one eye on the door, but creeping.

Speaking of comebacks, Claude's Fable 5 is open to everyone again after being blocked for non-US nationals on national security grounds. Word in some corners is that a few capabilities got quietly trimmed along the way, though Anthropic hasn't confirmed or denied anything. Guess we'll never know.

OpenAI wasn't about to let that be the week's story. They pushed their own new and powerful family of models into limited preview this week, named Sol, Terra and Luna.

Serious firepower, wrapped in a naming scheme nobody ran past a crypto trader first. Anyone still carrying scars from 2022 is going to break into a light sweat every time they load up Terra or Luna.

With that let’s get into this week’s edition.

Before a prompt reaches an AI model, something has to decide whether it should get through. That job belongs to an input guard, a small classifier that reads the incoming request and flags the unsafe ones before an agent acts on them.

This week Trishool, a decentralized red-teaming subnet on Bittensor (SN23), released HaloGuard 1.0 and says it's now one of the strongest open-weight guards in the field.

What's turning heads is the size. Guards get graded on an accuracy score out of 100 that rewards catching real threats while not flagging harmless prompts, and higher is better.

HaloGuard's smaller model scores 90.9 and its larger one 92.1, the best any open-weight guard has posted, which puts it at SOTA. And it got there while running 10 to 30 times smaller than the models it beats. ShieldGemma, at 27B parameters, sits at 70. So a model a tiny fraction of the size is landing 20 points ahead of it, and near the ceiling those last points are the hardest to win.

The weights are already on Hugging Face, out of a subnet with a market cap of only a few million dollars. Whether or not the paper survives outside scrutiny, a decentralized network shipping a result like this on that budget is exactly what the sector keeps promising and rarely delivers.

This week the bottleneck in the AI buildout is power, and the fastest fix for it has quietly turned into the scarce thing.

US electricity demand is rising for the first time in two decades, and the grid is the slow part. Getting a new data center wired into the grid now takes years, so the hyperscalers stopped waiting and started building their own power on-site.

The machine they reach for is a heavy-duty gas turbine, which burns natural gas to spin a generator the size of a locomotive, around half a million pounds, enough to power half a million homes. Right now, it's the only firm, always-on power you can stand up on a two-to-three-year clock. Nuclear takes five years or more. Renewables can't carry the base load alone.

Only three companies build turbines this big, GE Vernova, Siemens and Mitsubishi, and their order books have gone vertical. GE alone closed last year with roughly 80 gigawatts booked, its lines full into 2029, most of it chasing data-center power.

The scarcity shows up in price too with the turbine equipment now running about 195% above where it sat in 2019.

Here's why it stays tight. A turbine can only ship as fast as its hardest single part, and that part is the first row of blades, the ones sitting right behind the flame.

Each blade is a single unbroken crystal of nickel, grown by cooling molten metal so precisely that the whole thing hardens as one continuous grain, leaving no boundaries for heat to crack along. It's closer to growing the silicon in a chip than to ordinary metalwork, and a single batch can take over a year.

Which is where it gets interesting. The entire layer comes down to two American casting houses. One is a materially cleaner way to own the blade than the other, and at these prices only one of them is still worth chasing here.

Which name, and why, is the whole of this week’s Tessara Research newsletter.

Or…skip the writeup and open the full bottleneck board instead.

Our read comes directly from Tessara’s research terminal: every public company tied to their bottlenecks, scored by direction, materiality, and exposure quality as the tape moves.

Cheers,

Teng Yan & Arvind

Quick recap: I also publish a newsletter on the AI buildout and supply chain. 1-2 memos a week.

Subscribe to get it every week in your inbox →