50 Comments
User's avatar
Brandon's avatar

"If you go around popping a lot of balloons, you're not going to be the most popular guy in the room"

- Charlie Munger

Angsana Anderson's avatar

Better be right than popular!

Brandon's avatar

And also better to wrong than destitute.

Bazinga's avatar

Stop it, you are making too much sense....

Jokes aside, the writing on the wall is very clear....

Angsana Anderson's avatar

Indeed.

The dynamic between memory chips’ ASP and unit cost have historically been reliable indicators of cyclical inflection.

Right now, it suggests the shortage could end around Q1’27.

I discuss the details here: https://angsanaanderson.substack.com/p/where-are-we-in-the-semiconductor?r=5rl2u5&utm_medium=ios

Obsidium's avatar

Once again, it is a pleasure to read your articles. I saw that you will switch to a subscriber-only model in the near term, have you thought about having an early-subscriber discount? I would love to keep reading you but it will depend on the price you charge… thx a lot and keep on the good work!

Dr Doodad's avatar

I have heard this from credible sources as well. NVDA chips sitting in warehouses in abundance with “Dell logos”

Harrie Scarlet's avatar

Fabulous job mate

First time someone has clearly explained the dichotomy between the “insane demand “ and “GPUs collecting dust”

Can anyone show me anyone making money on the applications layer? This is bound to fail

Joey Ramos's avatar

Great read

New Tape's avatar

This needs to be read with "Paul's Dream" - Hans Zimmer playing in the bg

Lance's avatar

Open Ai is reported that they failed to meet the revenue targets. Too much risks.

Alan Reed's avatar

I really enjoy your articles and insights you ar providing to the industry. Thank you

TizyCharts's avatar

Great insights, thank you

Marcus's avatar

"Mountains of inventory collecting dust" - where is the data on this? Do you have hyperscaler GPU install rates vs shipment receipts to actually give legs to the argument? I know we have CIP balances of $51B (GOOG) / $27B (META) / $29B (AMZN), but otherwise, this is just speculation, right?

Kakashii's avatar

don’t just look at hyperscalers/neoclouds CIP but also inventories, and Nvidia’s partners inventories (e.g. SMCI and taiwanese partners)

Scenarica's avatar

SOXX up 50% in 27 days. Intel doubling. Marvell up 90%. The buy button is doing more work than the balance sheet right now.

The Anthropic delivery speed section is the most underrated part of this piece. One company's shipping cadence is forcing every other lab to burn compute they don't have to ship products they haven't finished. That's a demand spiral that has nothing to do with end-user adoption and everything to do with corporate panic. The compute shortage isn't just demand-driven from below. It's being manufactured from above by a competitive cycle that can't slow down without someone volunteering to lose.

The dot-com comparison is fair on the price action but the underlying demand is different. In 2000 the revenue was imaginary. In 2026 the revenue is real but concentrated in about four customers. Whether that distinction matters at the index level is the $10 trillion question sitting underneath every one of those ticker symbols.

Timmy Kamps's avatar

Revenue or Profit?

Byron's avatar

Does Jevons' paradox enter into this discussion with more efficient chips?

Atlas.ML's avatar

Perhaps investors should ask themselves one question.

How many Taiwan AI component suppliers are making 100% ROE right now ?

Hynix reported 70%+ operating margin. This is memory chips with each block is simply a duplicated copy of its neighbors. HBM4 is great tech advancement but can this non-logic IC semi makes 80% gross margin forever?

The quantum of seemingly endless buying in the current boom is fascinating. It breaks so many record.

And if the consensus is right about Nvidia 2028’s $580 billion revenue, does that mean that we are going to see consecutive quarters of blow out earnings from Cloud Computing service providers?

Super Bowl earnings is coming let’s see

Kenneth N. Myers's avatar

Outstanding insights. It shows that many have not considered how to integrate the laws of semiconductor depreciation into their business model in the context of technology insertion and refreshment. This naiveté will be suboptimal at best and fatal at worst, if they are able to survive to their initial deployment.

Phaetrix's avatar

Everyone sees the shortage.

The risk is how much of it is just timing.

Angsana Anderson's avatar

Many believe this time is different. They point to the SCAs Micron is signing.

But these SCAs are not new to the industry.

During the 2017-2018 boom, Siltronic signed similar long-term agreements.

But during the subsequent busts, these agreements only helped to stabilize pricing. Siltronic could not stop customers from delaying orders.

Phaetrix's avatar

Everyone points to the agreements.

The variable isn’t the contract. It’s what happens when demand slips and timing stretches.

That’s when “committed” starts getting tested.