#nvidia

SK Hynix's Nasdaq debut was the largest US listing ever by a foreign company. It controls roughly 60% of high-bandwidth memory, which is the actual constraint on AI compute. The harder question is whether memory has really stopped being a cyclical business, and I don't think it has.

Hyperscalers depreciate AI servers over five to six years. If the real economic life is shorter, reported profits across Big Tech are too high. Here is how depreciation actually works, what the bears get right, what they get wrong, and how to check the assumption yourself in a 10-K.

Meta puts its in-house AI chip into production in September, built with Broadcom and fabbed by TSMC. Every hyperscaler is doing a version of this. None of them are actually leaving Nvidia, and the reason is software, not silicon.

Semis ripped in the first half of 2026 and then late June cracked. The honest answer is that AI demand is real and Nvidia is no longer the trade. Those two sentences are not a contradiction, and most coverage refuses to hold both.

Dot-com companies had no profits. Nvidia earned $120 billion of net income on a 56% margin and trades at 20 times earnings. So the lazy bubble comparison fails. The real risk is somewhere else entirely, and it's more interesting.

A huge slice of Nvidia's revenue comes from a handful of customers. In the July 2025 quarter, two anonymous buyers alone were 39% of sales. The entire AI trade is a few CFOs' capex decisions away from a bad quarter.

A stock split changes the number of shares and the price per share while leaving the total value of your position untouched. So why do companies bother? The honest answer is signaling, accessibility, and a little bit of psychology.

A rigorous look at the semiconductor investment landscape: NVIDIA's moat, the TSMC bottleneck, HBM dynamics, and how to value chip companies in an AI cycle.