Mar 24, 2026 7 min read

Razor's Edge: Fuck it.. Datacenter Go Brrrr

You know AI investing has gotten hard when all anyone really can talk about in terms of winners is capex stocks. I don't think I've seen a single compelling investment pitch related to AI that isn't a datacenter pick and shovel stock over the past 60 days. People have just given up. And why wouldn't they when the messaging is so confusing...

These are pretty much the gods of self-driving cars, humanoid robots, AI Agents, and the semiconductor brain powering everything telling you three different things.

Whose right?

Honestly, nobody has a clue at this point, and that's the problem.

The early cycle in AI investing was great. You cheered nvidia, all semis, netflix, reddit, spotify, hyperscalers, and even 95% of saas stocks. We loved ai for adtech and the attention economy as productivity gave u more free time. Everyone was an AI beneficiary it seemed. Consumption was back because of AI and so was IT spending and hiring in general post the covid hangover. Ai losers were relegated to seemingly economically inconsequential names like Chegg.

But things changed...

SAAS seat based names got thrown out for consumption names

Consumption names then took a hit on optimization concerns

Then they hyperscalers split into debates over who had the most consumption levg

Then all saas seat names became shorts

Then all software got dumped

Then hyperscaler consumption numbers started mattering less as capex numbers exploded

Then all internet human traffic plays became disintermediation shorts

Eventually all that was left for any analyst to talk about was their capex narrative technical work. This rack will consume more power then a city, have cables that can extend to space, and weigh more than an airplane.

Helium, copper, eml lasers, substrates, epixate, indium phopshide, mocvd, GaN, Sipho, CML, CPO, 800v, HBM, Nand and on and on...

This is fine and dandy but it really requires no work anymore. It's been a market treasure hunt for the most part and the technical understanding now has an army of folks blogging away with their AI-powered technical analysts.

Price Doesn't Matter?

The best part of all this is none of them actual wanna talk about stock prices. I got a lot of this slack with Aaoi short flip. Like congrats on reading a management slide deck back to me, but I still believe I need to value some of this crap.

Not talking about price has become a signifier of seriousness. It says: Hey, I’m focused on the technology element, I know the lingo and the fundamentals, and this is a long-term thesis. Talking about price is for traders, for flippers, for people who don’t understand the scale outs and up....

In my book, not talking about price is how you end up looking stupid.

For example, Micron now trades at roughly 3x FY 2027 EBITDA. They could probably buyback 15% of their float with next years fcf. Nvidia is the cheapest its been. So, knowing a thing or two about optics, I am going to have to call Ciena, Coherent, and Lumentum all shorts right now.

And don't give me the memory commodity argument...

Memory will be needed for every car and every robot on earth so the demand story is gonna be fine. No arguing that. As for pricing, who knows is a good answer but i do know its infinitely easier to add supply in optics. I also know the whole space is now slide decks on capacity numbers 2x to 6x in the next 18 months with pencil in fcf numbers out to 2028 as revenue 2x-4x. But this is one cohort customer market for the most part and one with dozens of suppliers capable of playing.

So, if you are basically trading the demand curve post 2027 then all optics is a screaming short now as your exposed to the same unanswered questions but with way worse supply disaster risks. And what about those hard questions...

The Market Gave Up on the Hard Questions

There’s a deeper reason capital has herded into AI infrastructure components, and it’s not just that the demand story is clean. It’s that the alternative, owning the companies where AI’s economic value actually gets created or destroyed, requires answering questions that nobody wants to touch anymore.

Think about what you’d have to underwrite to own a hyperscaler, a SaaS platform, or a services company with conviction right now. Will enterprise AI adoption follow the cloud playbook or stall at the pilot stage? Will token consumption scale with complexity, or will it collapse as frontier models get more efficient and complex tasks turn out to be decomposable into simple ones that get mapped and cached? Will OpenAI sustain its lead, or faceplant under the weight of its own cost structure? Will Google get disrupted in search or will its distribution/infra advantages prove insurmountable? How will Apple adapt when the device-level AI experience becomes the battleground? Will Microsoft eventually go its own way on models and break from OpenAI? Will hiring actually collapse across white-collar industries, and if so, what does that do to the consumption assumptions baked into every SaaS valuation?

These are the questions that determine to what extent the trillions being spent on AI infrastructure generate returns. And the market has now essentially refused to engage with them. There is no bid for “owning” the AI outcome the way there was a bid for owning the cloud outcome a decade ago. In 2014, you bought Amazon and Microsoft and Salesforce because you believed cloud adoption would compound and the platforms would capture the value. Today, the equivalent bet, buying the companies that will capture AI’s economic surplus is treated as uninvestable because the range of outcomes is too wide.

So capital flows to the one place where the analysis is simple: the components. You don’t have to answer whether AI works. You just have to believe the data centers get built. And its clear for now that a lot of them are being built. You don’t have to model adoption curves, competitive dynamics, or return on investment. You just need a capacity timeline and a transceiver-per-rack estimate.

This is intellectually comfortable and analytically lazy. The infrastructure trade has become a way to express an AI thesis without doing the actual work of understanding where the value accrues. Btw- the work is presently a work in progress so i get the laziness. It’s the equivalent of buying picks and shovels during a gold rush not because you’ve analyzed the mining economics, but because you’ve given up trying to figure out whether there’s actually gold in the ground. The picks-and-shovels framing is supposed to be a investment insight. In practice, it’s become an excuse to avoid the hardest and most important questions in the market.

And the irony is that those hard questions are exactly what determines whether the infrastructure trade works at current prices. If you can't make money in every software, adtech, media, info services biz, and hyperscaler, then why are you buying lasers and cables? The picks-and-shovels trade isn’t insulated from the demand/return question, it's a levered bet on it. And using it as a hide-out is a bad idea.

Anyway I shorted a basket of optics names today because of this. And honestly that combined with speed/performance dynamics makes it look like a appealing trade to me. Just focusing on the top line hasn't worked in many capex names, and this space is the worst when it comes to the competitive landscape dynamics. And with perf running so hot relatively with Nvda/Memory now looking like deep value stocks; i like the setup.