（预览）Microsoft的平台生存方案、Meta与市场的放行、局势感知缺失

Hello, and welcome to a free preview of Sharp Tech. Hello, and welcome back to

We are in person. This is Sharp Tech.

Welcome back to Sharp Tech.

And when I get the-

Here we are

... when I get the Zoom open or Riverside open, cracks me up every time. How are you Anthony?

Live and in person. I'm good. It's great to be in Madison. It's balmy outside relative to what it was in January, and we're here for some company meetings. There's an AI show-and-tell that we have planned that I'm already pretty stressed out about.

Because you have no AI in your life. Are you going to demonstrate looking up NBA stats?

Yeah, demonstrate copy editing, Sharp text posts, and whatnot. You guys are on a whole other level. We're also going to a Brewers game later today, which I'm very excited about. It will be my first baseball game in probably 5 or 6 years. So here we are.

Well, you just asked me, "Are we going to tailgate?" and I was just flabbergasted.

A total noob East Coast elite question. Of course we'll be tailgating.

It's funny, we were just talking about tailgating on Dithering, and I was expressing how East Coasters say they tailgate, and I find it very underwhelming. In D.C., they don't even bother, so it's good to know.

Don't even bother; just show up at game time. Absolutely. Well, for now, before the tailgate, we're going to review big tech earnings, which came down last week, and we're going to start with 2 companies that are encountering very different reactions from the market.

First up is Microsoft. The market is very happy with Microsoft. The company was up 15% and added $450 billion in value in a single day last week—the biggest single-day jump in market history, according to Bloomberg. Microsoft is not pushing toward the frontier whatsoever at this point. They haven't been since 2024. What do you think of what they're doing instead?

If you zoom out, I knew earnings were going to be last week. I would come back, have all these things to digest, and I think all 3 updates this week were actually interconnected.

Mm-hmm.

I argue we should have done 1 big post trying to tie them all together.

An omnibus. Okay, yeah.

Yeah. I was trying to look at how every different hyperscaler is slightly different in where they are. Do they have a cloud business? Where are they relative to the frontier? What's their focus there, and what is their overall threat from AI?

So the 1 I didn't get into as much as I wanted to was Amazon. We're going to get to them in a little bit, but Amazon's on 1 extreme. They have a total cloud business, and I think their core business is not very threatened by AI.

Yeah.

AI is not delivering my packages.

Durable physical infrastructure for Amazon.

It's a good place to be. Yes.

Yeah.

I think Amazon's core business, their e-commerce business, is fine—arguably better than ever. Their cloud business is obviously doing very well. Amazon is also a company that is more accustomed to, and okay with, tighter margins, which I think is 1 of the concerns about hyperscalers in the long run. Once compute is finally everywhere, who's going to have the best margins? Are margins going to compress? They've always thought about having lower-cost structures.

They live in that world.

Right. We talked about that last episode. So Amazon is 1 extreme. Then you have Microsoft.

Mm-hmm.

Microsoft is taking a strategy that's similar to Amazon in some respects, particularly in the fact that they're not on the frontier. They're building their own models, but they're not building models that are going to be as good as the models from OpenAI—

Right.

—or from Anthropic. But they're serving those models.

And they got off that train a couple of years ago with OpenAI.

They weren't really on that train. They were on that train to the extent that they were supporting OpenAI.

Yeah.

But what happened—the way to think about what happened to Microsoft and OpenAI is, Microsoft said to OpenAI, "We'll keep your inference business."

Mm-hmm.

"So the business where it's actually customers using your models—your training business—you've got to go find someone else."

Yeah.

And so that was the whole shift to talking to Oracle, talking to SoftBank, and building these mega data centers. That's the train that they got off of.

Their software business is, I think, very threatened by AI.

Mm-hmm.

We're going to talk about vibe coding a little bit. This will be part of the AI show-and-tell, but we are not representative of a large enterprise, to say the least. There's an aspect of us experimenting with AI Astrotechery that's important just to understand the product.

Yeah.

But in a very tangible example, we were mostly still using Teams for 1 specific use case, which was podcast workflows. We're totally off it.

Mm-hmm.

We made our own software, right? Now, again, that's a very small example, but you could imagine that writ large—

Sure.

—at least in the fullness of time. So they have a cloud business that in some respects is similar to Amazon, but they have a core business that is very much threatened.

Mm-hmm.

You move over, and now we're at Google. Google obviously has a big cloud business. I'm sure we're going to talk about that in a little bit. They are also seeking to be on the frontier, and they do have a business that's threatened by AI, which is search generally. We've spent a lot of time talking about that.

I think there's more to say about it than I wrote about this week in that regard, or than I mentioned in passing. So we'll talk about Google in a little bit, but they're kind of in the middle.

Then you get to Meta. Meta is maybe between Amazon and Google in that they're seeking to be on the frontier. They don't have a cloud business; that's the difference. They are threatened by AI, but I think less acutely—

Than Microsoft.

—than Microsoft or even than Google.

Yeah.

Right? You can see very clearly how AI is a threat to search.

Mm-hmm.

It's harder to see how AI is a threat to social networking—

Social networks, yeah.

—or entertainment, or whatever it is. But my thesis is, number 1, it takes up a lot of time—

Mm-hmm.

—which is a detraction. So anything that takes up time is a detraction from Meta in the long run.

My overall thesis is that anything digital is fundamentally threatened by AI.

Yeah.

The opposite of the Amazon thing: The more physical you are, the safer you are.

Mm-hmm.

The more digital you are, the more threatened you are. That's the gradient of these entities.

I thought it was interesting to talk about all 4 of them this week, considering those aspects, even if I wasn't super explicit about it, and what the differences are between them, how the market has responded to them, and all these bits and pieces. So that's the overall context of Microsoft—

Where we're at, yeah.

Yeah, and Microsoft in this case.

With Microsoft specifically, they want to be middleware. Is that right? And to what extent is that a function of necessity, because that's the only option that makes sense for them at this point?

Yeah. The way to think about Microsoft is exactly that. It's funny because I wrote an article a few years ago comparing Microsoft to IBM.

Mm-hmm.

It's actually 1 of my favorite articles. Some articles stand out because they were so painful to write. I remember I wrote this 1 at an Airbnb in Madison before we had a place here. It took me hours to get it out, and I posted it at about 3:00 in the afternoon instead of 6:00 in the morning, when I wanted to.

Oh, boy.

I went through Lou Gerstner, who says "elephants can't jump," or something like that. It was his autobiography. I was talking about the issue IBM faced in the '90s, where they're this behemoth that seems way behind the times. Investors are demanding they break up.

He went the opposite direction. He's like, "We should not break up. You don't understand. We're not actually good at anything. So all these pieces you want independent are not going to do well in the market."

They'll all fail independently. Sure.

"What we're good at is being big."

Mm-hmm.

Actually, being big is useful because you're kind of good at everything, but you're not really great at anything.

Okay.

Or you're acceptable at many things.

Yeah. That certainly describes my Microsoft experience over the last 20 years.

This was IBM in the '90s. The internet's coming along. You have all these big companies that are like, “How do we handle—what do we do with the internet?” And IBM was like, “We'll take care of it.”

What IBM did was build up a huge consulting arm, which, by the way, Microsoft talked a lot about building up. They actually had a name for it. They've always had a big sales force, but the whole forward-deployed engineer thing. I should look up what the name was. It came up on the earnings call.

That does make a lot of sense for the next 10 or 15 years here.

Well, everyone's doing it for good reason, right? Palantir really has led the way here. And then they also built a bunch of middleware.

Mm-hmm.

Basically, they built all these layers for companies to have websites and e-commerce and all these bits and pieces that we talked a bit about last week. There are still companies that are on IBM because they've been on there since the '60s. There are a lot of companies that have been with IBM since the '90s because they just built in all these layers to get you online, to get you all the things that you needed to do that, and it was a very smart strategy that basically gave IBM an extra 20 or 25 years of life.

Yep.

And the framing of that article was: once you've been a monopoly, you've kind of lost the ability to be great.

Okay.

Because you get fat and flabby, and life's too easy. And what Lou Gerstner did was realize, “Look, we don't have the chops to succeed by being the best at any individual thing. What we're good at is doing everything because that's what we've been doing for a long time.” The analogy there was: look, this is the Microsoft path. You're not going to succeed by being the best at stuff.

The leader in any area.

You can do everything, and it's good—actually, being big is your biggest asset.

Yeah.

And I put that in the context of having been, at that point, a strong endorser of Satya Nadella's approach to deprioritizing Windows and turning it into a services layer. Yes, they were late relative to Amazon, late to the cloud, but most everyone else was late to the cloud. So they and all their customers were coming to the cloud together.

I wrote that in the context of a keynote where he spent a long time talking about pen or whatever, their digital ink, their ability. I was like, “What are you talking about?” You are not going to win on product differentiation. You're going to win by being good at everything, and I think Microsoft has continued down that path by and large. You fast-forward to AI, and now that comparison, I think, is becoming very, very tangible.

Okay.

So Satya Nadella's been posting on Twitter all these articles. It's like he's coming for my job these days.

I know. You said a couple of weeks ago he's coming for your job. He really is. He's very, very active on Twitter with these essays that he's putting out, and they're always these long tweets, and I wish he would put them on a website somewhere because I just can't read that much on Twitter. But he's clearly spooked by the moment and trying to counter the Anthropic messaging.

Yeah. I get the sense that spooked happened maybe a year or so ago. One thing that was interesting: I was at the Build keynote in June. Usually, at a Microsoft keynote, Satya Nadella would do the first 10 or 15 minutes, frame what's going on, and then hand it off to a parade of executives who would do the rest of the keynote. This keynote was all Satya Nadella.

Maybe I over-indexed, but part of my take this week is that I'm right to index on the meta of these things: what executives say actually matters.

Mm-hmm.

What I took away from that keynote is that he felt the need to say, “I need to take a much firmer hand and much more control of this company, because we're actually in bigger trouble than people realize.”

Yeah.

And I suspect a lot of it was that they have this whole software business, right? You can make the software that Microsoft does.

Mm-hmm.

Now, are you going to make it—

We just did.

Are you, are you going to make it actually useful and dependable, and are companies going to want to waste time maintaining it, et cetera, et cetera? All these are open questions, and we've litigated that, and we'll continue to litigate it probably for years to come. But Anthropic and OpenAI are coming for Microsoft's business.

Mm-hmm.

Day-to-day productivity: they want your agent to do all that. They want your agent to do your email, your agent to do your documents. Why do we have documents? Why can't we just ask the AI what it is, right?

Yeah. They're both very, very expensive for enterprises as well. And as those costs continue to rise, you could envision a future where enterprises are choosing between spending on Microsoft or spending on OpenAI as a platform or Anthropic as a platform.

Right. This is the thing about tech in general. Even right now, when stuff seems to be moving so quickly, one of the fun things about analyzing tech is how many decisions actually take years to play out. We've talked about this in the context of chips, right? Intel was doomed. I wrote in 2013, “Intel's in big trouble.”

Took a solid 10 years.

Yeah. It took 10 years for that to actually manifest, but it was totally right. Intel actually needed to make changes. I thought I was too late in 2013, and I was, I think, too late. But it would've been better to make changes in 2013 than to wait until 2021 or whatever.

Yeah.

Pat Gelsinger ripped the Band-Aid off. Again, there are pluses and minuses to his tenure, but he certainly ripped the Band-Aid off and said, “We have to shift our model over time.” But this applies to software too. By the time your business is threatened—or it shows up in the results—it's way too late.

Mm-hmm.

And so I got that sense from Nadella. What he's been talking about in his blog posts is a bit of both: an articulation of the threat to Microsoft and trying to convince everyone else, “This is a threat to you as well.”

Totally.

These companies are trying to understand your business. They're trying to figure out what you actually do with your computer.

And they will replace you.

So they can replace you.

Yeah.

This gets to the question of—

It's explicit counter-messaging relative to—

Oh, for sure.

Anthropic. Yeah.

But one of the big things with the AI models, and the question that remains open in my mind, is: will these models generalize? Can you learn one thing and then know how to do other things? I'm on the “I don't see evidence of generalization yet” side.

Mm-hmm.

There's all this amazement about these math discoveries, right? Math is like coding. It's a knowable thing. You're discovering things that exist, but it's a bounded space to an extent, in a way that maybe some other things aren't. What you need to do is get data, see what works, and verify whether it worked or did not work. So at the end of the day, it might be a moot question.

Mm-hmm.

Maybe you don't generalize; you just gather data on everything in the world, right? We'll get neural links in everyone's head and gather the thought processes, and then we'll actually be able to do everything. But this is the threat: we know for sure that if you can get enough data and verify, then you can do it.

Mm-hmm.

And that's the threat he's putting his finger on. The more you use OpenAI or Anthropic and use their harness—you use Claude Code, you use Cowork, you use Codex—the more they're gathering everything. And even if they're saying, “Oh, we're not using your data to train the model,” that's actually not the important part. The important part is all the metadata around it. What tools are you using? What things worked? What things didn't?

Mm.

And if you're using their user-facing software, you're giving away the crown jewels of your company, which are how people actually do work, what works, what doesn't, all the metadata around—

You're training the disruption threat that—

That's right.

Yeah.

That's right. And so what Microsoft is proposing to do is reduce models to being processors.

Mm-hmm.

They want to recreate the world where Microsoft Windows is the layer everyone uses, and underneath that is an Intel chip, an AMD chip, or whatever it might be. What does that chip do? It gets sent stuff by the operating system, does some processing, and sends something back. It's the operating system that's the orchestrator, figuring everything out and passing things to developers on top of that.

Microsoft is saying, “Don't use their harnesses. We're going to build a harness, or we're going to give you the tools to build your own harness, such that these models are basically stateless.”

Mm-hmm.

They're just things you send something to and get something back from, and they have no other context. All they get sent is exactly what they need to know to do a computation, and then it comes back. They're very explicit about it. That's what they're seeking: to reduce the model companies to being processors.

The model companies are trying to be—

Everything.

—everything.

Right.

And so that's the direct confrontation that's going on.

Yes. When you look ahead, is this the right strategy for Microsoft? It may not work.

Well, this is where it's interesting to put Microsoft in contrast to Meta.

Okay.

How do you survive in a world where you're not on the frontier?

Mm-hmm.

What they're saying is, “We will be the layer between the frontier, where you can choose—”

You don't have to worry about the models. It's going to be model-agnostic. You have your own harnesses with us, and you don't have to fork over proprietary data that puts your business at risk.

That's right.

It makes sense.

I think it makes sense. It's a very reasonable strategy. The question, and I think the challenge for Microsoft, is most fraught right now because in any new technology, it doesn't work well enough, right? The AI still isn't good enough for a lot of these workflows.

The more you integrate and the more you do all the pieces around it, the better the stuff works. This is classic Clayton Christensen theory about integration versus modularization. The idea is that early on in a technology's life cycle, it's not good enough, so the more you do, the better. The integrated solution wins at the beginning.

In the long run, modularization takes over, where you want different suppliers and you care about competing on parts.

You care about cost, yeah.

There was a huge wave of, “Oh, these models are so expensive.” That felt a little astroturfed over the last few months, to be totally honest.

Hmm.

That is a narrative that Microsoft absolutely wants to feed, because they want you thinking, “We can't double down.”

“We can't burn through the budget in 2 months on these models.”

That's right.

And so you think about it: “A harness is good because we can use different models. You can do sort of an X, Y, Z.” It's kind of funny because I'm not sure how much I buy the multimodel idea. If you ask a really smart model a very easy question, it's going to give you a very quick and fast answer, right?

Mm-hmm.

A lot of the model capability is more of a ceiling than it is a floor.

Good point, yeah.

There's an aspect here where it feels like Microsoft loves the “expense, use lots of models, expensive models, cheap models” narrative, because that fuels the question, “How do I do that?” Well, you need a harness, right? You need a harness that has a router that chooses the right model for the job.

And they're the harness company at this point.

There are cost differences, to be clear. I'm not trying to overstate or underplay that. But it also is a narrative that feeds what Microsoft wants people to think.

Mm-hmm.

You can't build directly on these. On the flip side, a lot of companies are going to find this appealing. They don't want to try to figure all this out. It's like the cloud: Microsoft's their good friend. What's going on here?

Exactly.

Figure it out.

They've been working with Microsoft for 35 years at this point.

Right. There's also a question, though: How much better is it actually if you're working directly with the frontier model companies?

Right.

Can you actually build stuff that you can't build otherwise? Can that stuff actually accelerate your business such that your competitor is so worried about having multiple suppliers and not getting locked into the model companies that their products suck—

Mm-hmm.

—compared to yours because you gave in? You're like, “That's right, I'm all in on OpenAI, or I'm all in on Anthropic, and I'm so much faster and accelerating so much more quickly.” And, oh, by the way, they have this feedback loop that's actually extending their lead.

The idea is that Microsoft is going to be making a harness that is lowest-common-denominator functionality—

And is comparable to what Anthropic and OpenAI are offering.

That's right.

Yeah.

Now, we'll see. I don't know, right? At the end of the day, does the harness actually matter that much?

Yeah.

There's a thing where you can feel this with the models. When I started doing the vibe-coding stuff, it was with 5.5 on extra high.

Mm-hmm.

Extra high? Okay.

No, for 5.5, not 5.6.

Oh, okay, okay.

For 5.5, there was this Superpowers skill package someone put together that forced it to do this planning and this review, and someone recommended it to me. I was already doing planning and reviewing, but I still thought it was kind of useful.

You get to 5.6, which overdoes this stuff on its own.

Mm-hmm.

You layer on this stuff, and it was out of control. Part of the reason it was out of control was that it was doing all this crazy stuff. Basically, the point is that as the models got better, they internalized more and more of the functionality that came from the harness.

Right.

There's a harness called Pie. They had a really interesting blog post this week talking about their whole approach: to do the minimum amount necessary in the harness, and make it possible for you to add on the specific tools that you need.

Everyone's overengineering this such that 3 months down the road, a new model comes out and you're actually making the whole situation worse because your harness is competing with your model.

Mm-hmm.

You did all this work upfront to overcome the model's limitations, but then the model got better. Why? It's all wasted effort, right?

Right.

You should just let the model do everything.

It's funny, because that's a world that's actually—possibly—good for Microsoft.

Mm-hmm.

If it's all about the model and it doesn't matter what harness you use, then, yeah, use Microsoft's harness. Use your harness.

There's also a world where Microsoft's harness is so overengineered and so careful to try to do things and not let the models get the data they need that they're not even coming close to tapping into what the models can do.

Right.

The models would be better if you gave them way more stuff, right?

You're gating the performance.

Yeah.

Yeah.

All this stuff is kind of unknown. This was so interesting. I think what they're doing is logical. There are a lot of benefits that come from being the incumbent.

Mm-hmm.

There's a reason why, as a technologist, you could ask, “Why would any company use IBM's solution to build a webpage in the 1990s?” And yet it worked phenomenally well.

Lots of companies did.

Right?

Yeah.

And so—

Well, it will be interesting. Obviously, it's all sort of TBD, but you look at the number of companies that signed the open letter, for instance. You look at the weekly HFC series—

Oh, I didn't get a chance to write about the open letter on Stratechery.

—from Satya.

Hilarious. That thing cracked me up.

Well, it's clearly a reaction to massive uptake for Anthropic and OpenAI, and the concern that that inspires everywhere else to go down that—

Oh, and the funniest thing is OpenAI signing the letter.

Oh, I didn't see that OpenAI signed it.

Oh, they did.

That's wonderful.

The most two online company in history.

Don’t worry, guys. We’re with you.

Yeah, we got it. My finger’s up in the wind. I guess we better sign the letter. Of course, they don’t want these open-model competitors.

That’s fantastic.

But do you know what I mean?

Like, that’s sort of an indication of the real concern.

For sure.

Yeah.

The fact that the whole industry is lining up to sign this letter is not an indicator of strength for the companies signing the letter. You don’t sign open letters because you’re winning. You sign open letters because you’re freaking terrified.

Because you see these entities swallowing everything. They can do stuff you never expected, right? You see, you talk about Apple and Apple not seeing the memory crunch coming, as we’ve talked about. Or you see even this response. Has Tim Cook vibe-coded an app? I’m guessing probably not.

And I’m not saying you have to vibe-code an app to understand AI, but there is an extent to which what these models can do is so far beyond what I think most people—

Can appreciate.

—really understand.

Yeah. Sure.

But someone like Satya Nadella gets it, right? And when you see this with NVIDIA, too—Jensen Huang, the sort of guy pushing this letter, is important enough for him to get on Twitter, right?

Exactly.

One of the worst decisions a human can make.

Well, to the extent that we lack perfect visibility into how enterprises are using these tools today, I think the reaction from the incumbents across tech—the non-Anthropic and OpenAI incumbents—says a lot about what companies are using these things for.

Right. So NVIDIA feels like they’re in the catbird seat, right? Everyone uses NVIDIA chips. Why are they writing an open letter? Because in a world where there are only 2 companies that buy chips—

Yeah.

—NVIDIA’s the one that gets ordered around.

Margins get compressed.

Yes.

Right.

They will have the scale, the capability, and the cash flow to make their own chips. And it’s the same thing for all the hyperscalers.

Yeah.

In a world where there are only 2, and all you’re doing is competing to serve OpenAI and serve Anthropic, those margins are going to get compressed a lot.

Mm-hmm.

You’re just going to be a financing vehicle to help them build these sorts of things out. And so the whole question is, why are VCs in on this thing? Is there even a startup ecosystem if you just use the models to do everything?

Every startup gets replaced within a year or 2 if it succeeds.

That’s right.

Well, shifting gears, you wrote on Monday with respect to Meta that you came away from the Meta call a bit alarmed. So, on a scale of 1 to 10, how alarmed are you by the enterprise pitch from Meta? Or is annoyance a better way to put it with Meta?

So the Meta and Microsoft comparison, I think, is super interesting.