Hi Traders,

Two 30-billion-parameter models landed in 48 hours.

Meta shipped Muse Glimmer on Monday. NVIDIA shipped Nemotron 3.5 Lightning this morning.

Both run on a single graphics card in a laptop or a workstation. Both published their weights under a license that lets you modify them and point them at your own data.

One thing actually changed this week: AI stopped being something you rent by the token and became something you own and run in your own building.

Nobody Buys Less of a Thing That Got Cheaper

The lazy read says this is deflationary. Cheap local models, less cloud, less spend.

NVIDIA answered that in the same release. They shipped a routing library called NeMo Switchyard alongside the model, and it hands routine agent work to the cheap local model while saving the frontier model for the hard 20 percent.

The numbers they published aren't small. Ramp matched frontier performance on its own coding benchmark while cutting cost 58 percent and runtime 33 percent. LangChain cut costs 74 percent across 145 multi-step tasks by sending just seven percent of calls to the expensive model, giving back six percent on accuracy to do it.

Cost per task collapses. Task count explodes. That's more compute at a lower price, which is how every technology cycle since the steam engine has worked.

Every Agent Needs a Room to Sleep In

Now follow it downhill.

A 30-billion-parameter model has to physically sit in memory. Meta gave us the number: Muse Glimmer needs more than 55 gigabytes at full precision, and they squeezed it under 20 by compressing the weights. That lands you on a machine with 24 or 32 gigabytes.

NVIDIA's model doesn't dodge this. It only activates three billion parameters per token, but all 30 billion have to be sitting there waiting their turn.

So count the machines that don't have it yet. The June Steam survey puts 16 gigabytes as the most common configuration in the world at 41.57 percent, and 32-gigabyte systems went backwards last month because memory got too expensive to add.

Every endpoint. Every workstation. Every always-on agent. I haven't seen that demand curve in a single supply forecast.

Where I'd Rather Sell Than Buy

The name to sit with is Micron Technology (MU).

Memory was already the bottleneck in the datacenter. Micron just printed $41.5 billion in quarterly revenue at an 84.9 percent gross margin, and it's locking customers into multi-year supply agreements that pay up front.

Supply can't answer this quickly. Micron's new Idaho fab won't turn out wafers until the second half of 2027, and a memory plant takes years to go from groundbreaking to real volume. This week stacked a second demand curve at the edge on top of the one everybody's already fighting over.

Here's what makes it a Turbo Income name. MU trades near $877 after touching $1,250 in June, and 30-day implied volatility (the market's price on how much a stock will swing) sits around 72. Fat premium on a stock that already took its haircut.

You don't have to buy the highs to collect on it. A credit spread (you sell one option, buy a cheaper one behind it for protection, and pocket the difference) pays you while you wait for the thesis to show up.

MU is one name. The basket behind it runs wider: storage, the hardware makers selling the on-prem boxes, the power to run them, and the security layer that has to watch agents nobody can see. NVIDIA named CrowdStrike among the first companies customizing this model, which tells you that last leg isn't theoretical.

Turbo Income members get the full list, plus the spread structures to collect on them while the market catches up. Tap this link to join.

Time to get strategically greedy.

Here for a good time AND a long time,

Hans