Kevin S. Xu

40:1 !!! I thought it was a typo https://t.co/giQgBTOIuT

Nebius's GPU rate hike and the Fed rate hike looks like totally disparate events, but are actually interconnected Worth unpacking why A primary reason why yields have gone up is the debt load being taken on by the AI infrastructure build out. Kevin Warsh said it himself yesterday during the presser The AI debt situation is precarious, or at least keeps market sentiment down, despite clear demand signal, because it muddles the payback period and payback quality of these massive investments. Some debt is going to investment grade companies (no problem), some are junk bonds going to less trustworthy outfits (a problem), *and* some are only borrowing on investment grade terms thanks to financing/backstop from investment grade/cash rich companies (circular financing, also bit of a problem). The Fed's rate hike is catching up to this reality. It also raises the hurdle rate (or floor) of more debt going forward that might be going to additional AI capex. (Why lend to a speculative data center when I can collect 5% while sitting on my ass?) Of course, this is only a problem, if the increasingly expensive debt is a temporary bridge, not a permanent solution, and the revenue (ideally profit) will pay for future investments. The GPU/CPU rental price is the cleanest (not the most nuanced though, imo) way to proxy whether that future is in sight. In comes Nebius's GPU/CPU price hike, the other rate hike. It shows that GPU rental prices are more than holding up, which signals a rosier profitability picture for all who are investing heavily, with some debt, to build AI factories. (CoreWeave's announcement this morning corroborated the same trend, where they are also a signing shorter contracts that commands a $40 mil/MW price) Now, these spot prices are noisy indicators, but they do show a direction of travel (to use a Fed speak) that there is light at the end of debt tunnel. Not all will make it out though, lots of silly players in the buildout just trying to catch a wave, make a quick buck, and cash out before the music spots But for the players who are doing this for the long haul, things are trending the right way, while the Fed rate hike and higher yields in general will accelerate the death of the unscrupulous ones.

Nebius is raising GPU/CPU on-demand prices...again! Second time this year $NBIS https://t.co/3kCeoPj0Si

Ciena issued new 3-year financial targets. By 2029, it looks to: - grow revenue at ~30% per year - hit ~50% gross margin - generate ~20% free cash flow margin $CIEN is an important but quiet AI beneficiary of the buildout. Its CEO, Gary Smith, took over 25 years ago, when Ciena was trading around ~$400, after hitting a high of ~$1000 a few months prior. He managed the company through the dotcom bust, the GFC (when it traded at an all time low of ~$5 bucks (!), all to the way to today. What a story!

Pacing the frontier is terrible for the AI trade, terrific for the AI investment

the overtone window of a national AI legislation is moving in real time

Fresh update to the data center moratorium dashboard Now tracking more than 300 active moratoriums of varying sizes and lengths Important to note, many if not most of these moratoriums are pre-emptive (pausing to study the topic, not directly reacting to a project in district) and not retroactive (stopping a construction already in progress) As my post on Ciena CEO's commentary shows, most data centers under construction will continue for the next two years (2026-2028): https://t.co/0ltLagZDhu The big question is what happens in 2029 and beyond

One tidbit from Ciena's earnings that should've gotten more attention: When CEO Gary Smith was asked about how all the data center moratoriums will affect his business, he shared a conversation he had with a hyperscaler customer on this issue, and his customer told him if everyone stops building data centers tomorrow, Ciena won't notice anything for 2 years $CIEN

Glad to share some of my thoughts in Bloomberg's big story on the US-China AI co-opetition Chinese AI labs' tolerance of lower profit (or lower expectations of massive profit) is an under-explored, poorly-understood angle https://t.co/mfOIy1ODfx

Qwen 3.8's license is slightly more generous than K3 Commercial agreement threshold is $50 million trailing 12 months revenue (K3's is $20m) "Branding clause" requirement is the same: 100 million monthly active or 20m monthly revenue https://t.co/9hFleVeMSn https://t.co/rPcn8qSXs4