and power demand, too, so good for #CEG #VST...

Full stack below.

Time to really pay attention to this custom silicon shift.

Elsewhere... #LITE #COHR #ANET win as this project will require a lot of advanced optics & interconnects and each works a lot with hyperscalers... #VRT is the primary partner for AWS in thermal management, so they'll win here, too... #MU will win b/c there's a lot of memory demand here...

Third best play: #TSM. Amazon's newest chips (Trainium3 and Trainium4) are build on TSMC's 3nm and 2nm nodes. A $100B spend agreement over 10 years guarantees massive, high-margin volume for TSMC's advanced nodes.

Second best play: #ARM. Amazon's Graviton CPUs are built on the Arm architecture. Every new Graviton core deployed (of which there will be a lot in this deal) translates into royalty and license revenue for Arm.

The best play: #MRVL. They help Amazon make those Trainium and Inferentia chips. They are Amazon's custom chip designer, if you will. #AVGO should also benefit here, though they do more custom silicon work with #GOOGL and #META.

So this 10-year, $100B infrastructure spend commitment from Anthropic doesn't mean an additional $100 billion flowing into Nvidia GPUs and that ecosystem. It means $100 billion flowing into the custom silicon stack. Specifically, the Amazon custom silicon stack.

Here's how to play it.

supercomputing.

That's because this deal is all about custom chips. It centers around Amazon's custom silicon architectures (Trainium/Inferentia/Graviton) and bypasses off-the-shelf #NVDA H100s.

The Amazon-Anthropic deal -- featuring a ~$25B investment from #AMZN and a 10-year, $100B infrastructure spend commitment from Anthropic -- represents a "crossing the Rubicon" moment in this structural shift in AI infrastructure investment from general-purpose GPU clusters to custom silicon-first AI

This is existential. We can figure out the job loss and wealth divide effects through legislation AFTER we win the race - not before.

My 2 cents.

The race to AI supremacy in the 2020s is arguably as important or even more important than the race to atomic supremacy in the 1940s. If the U.S. wants to maintain global economic and societal and political dominance, it needs to win this race. Else, China rules the world.

summer.

Why isn't Netflix leading this charge? Why is the Chinese Netflix taking the lead on this?

#IQ - often considered the #NFLX of China - said it expects AI to create the bulk of its film and shows in five years, while debuting a new AI toolkit (Nadou Pro) that can handle almost every aspect of filmmaking and which aims to release a commercially successful AI-generated film as soon as this

The U.S. needs to drop the niceties and step up its commitment to AI.

In other words... the mature-node AI chip revolution is just getting started... and this breakout in GFS stock probably has a lot more runway ahead.

They need efficient, rugged, and cheaper chips to handle local sensor data (inferencing), which then pair with the fancy Nvidia chips back at the training cluster to deliver real-time, real-world performance.

Plus, the physical AI boom will create a huge demand tailwind for these chips. Humanoid robotics, smart industrial systems, and automotive AI don't need the $40,000 liquid-cooled GPU.

The chip buyers are starting to realize that, and now, they're buying tons of mature-node chips to optimize all those Nvidia chips they bought.

Nvidia makes the brains of AI models with their H100s. But those brains cannot function at full capacity without a massive supporting cast of mature-node chips for power management, photonics, connectivity, and more.

After a long period where mature-node foundries (chips made on 12nm to 90mm processes) were sidelined by the hype surrounding #NVDA and their advanced 3nm and 5nm logic for AI chips, the AI Boom is now finally trickling down to the support infrastructure.

#GFS +10% today on reports that its peer UMC is raising prices in the second half of the year due to strong demand for its mature-node chips. The stock has now almost doubled in the last few months.

And this rally is just getting started.

My read is that there's still runway left, but things are definitely heating up in a way that suggests we aren't in the first few innings anymore.

So there are historical analogs to suggest that rebrands + stock pumps in a hype cycle do happen right at the peak.

But the best historical analog for the AI Boom is the Dot Com Boom. and in the Dot Com Boom, the rebrands + stock pumps emerged a good 12+ months before the peak.

To be sure, we had crypto rebrands and stock pumps in late 2017 -- right at the peak of that cycle. And we also had cannabis/pot rebrands and stock pumps in 2018 -- right at the peak of that cycle.

These things started to happen to early/mid 1998, about 18 months before the Dot Com Boom topped out.

April 1998 - Infomercial seller K-Tel announced that it would begin selling CDs on the internet. Stock price soared from $3 to $33 in a few days.

July 1998 - Oil and gas firm Zapata rebranded as an internet portal called Zap․com. Stock doubled in a day.

My thoughts on what the #BIRD pop says about where we are in the cycle...

Long story short, my best guess is that it's indicative that we're heating up, but not on the cusp of popping, with another 12-24 months of runway left in the AI Boom.

Let's look at the Dot Com Boom.

The custom silicon playbook looks primed for big gains in 2026/27.

#ARM #AVGO #MRVL #TSM #INTC #LITE #ANET #ALAB

models.

The custom silicon revolution is in full force. The top foundational AI model makers are increasingly building their own AI chips to reduce reliance on #NVDA. They are having success with these efforts. And they are now doubling/tripling/quadrupling down on these custom silicon projects.

A structured flowchart detailing the custom silicon investment stack, highlighting roles from clients to architects and toolmakers.

#MRVL soaring to fresh all-time highs on an Information report that #GOOGL is in discussions with Marvell to develop two new AI chips: 1) A memory processing unit designed to work alongside Google's already very-successful custom chips, the TPUs, and 2) A new TPU built specifically for running AI

Needless to say, I like this one.