（预览）SaaSmageddon 与未来、市场回调后的 Microsoft、Anthropic 的超级碗谎言

Hello, and welcome to a free preview of Sharp Tech. Hello, and welcome back to another episode of Sharp Tech. I'm Andrew Sharp and on the other line, Ben, how are you doing?

Irritated, Andrew.

Oh, boy. Why?

Well, you know, I was struggling a little last week with a sore throat. I got pretty sick over the weekend and had to take a sick day, which I never do.

Mm. No days off. Extracurriculars.

There's so much happening this week. I know. I feel like there's so much stuff I didn't cover, and it really irritates me. I could have really used an extra day of publishing this week.

I was feeling better yesterday, doing well, and then today I was coughing again. It's just really annoying.

Well, here's the thing: You don't sound as bad as you did at the end of last week. Last week, you really did sound like Tone Loc on all your podcasts, and now you just sound a little bit under the weather. I feel like we're making progress despite—

No, but—

—the setback here.

This is like when you get a warm stretch in February and then you get below zero in March.

Mm-hmm.

It's the return, even if you know you're almost over.

Very demoralizing.

Yes. Demoralizing, that's the word. I'm demoralized. That's the word.

Okay.

Look, it could be worse.

Okay.

I'm not as demoralized as SaaS shareholders, so there you go.

Indeed. And that's where we're going to be for most of the show today. We're going all male on this episode. There are a lot of different beats to hit, and we're starting with 2 questions on a company that lost $350 billion in market value last week.

1. Microsoft Faces The AI Reckoning

You wrote about them on Monday, and Rav says, “Andrew and Ben, Microsoft got destroyed because the company sucks at AI. Azure is an AI winner, but the rest of the business is a loser. They have had access to all of OpenAI's IP for 3 years, and yet their own products are atrocious. ChatGPT beats Copilot in enterprise settings. GitHub Copilot was usurped by Cursor, Windsurf, and now Claude Code. GitHub Copilot is stuck in 2024. Dragon Copilot already lost to Game of Thrones.

“Now Anthropic is coming for the office suite by hijacking it. This encapsulates the issue facing the entire software complex today. The leading labs control the neurons, allowing them to dictate the entire AI future product suite through mid-training and post-training RL. Moreover, their AI products will always be superior to the ones produced by software companies leveraging their APIs.

“Thus, every software company faces the threat of Anthropic or OpenAI hijacking its distribution as they convert the promise of AI into material value. Furthermore, the labs have a structural pricing advantage because LLM API calls are the most expensive part of any AI application. You cannot beat a competitor with a 70% lower marginal cost.

“Microsoft's failure to deliver a single good AI product is the red herring for all software companies. Even if you were handed the underlying LLM, you still wouldn't create a competitive product. If Microsoft can't execute, no one else will.”

Now, Ben, I read that. I have 2 notes. I think the final line—“Microsoft is the red herring for all software companies”—I think the phrase he wanted there is “canary in the coal mine,” not “red herring.” So, different animal idiom, but—

Right. We're good on the idioms, not so good on the pronunciation, but, yeah, don't test us on idioms.

That's right. Beyond that 1 note, I found myself nodding my head at basically every other point that was made there. So what do you think? What do you think of where Microsoft sits amidst all this?

Well, where to start? Do we start with Microsoft? Do we start with the broader ecosystem? I think there's 1 interesting line in here: He talked about how the AI companies are always going to be superior.

Mm-hmm.

2. Enterprise Software Solves Hidden Problems

And superior on what vector? This is maybe the single biggest question, which gets at why all these SaaS companies exist. What job do they do, in the jobs-to-be-done sense?

If the job to be done is to deliver a compelling AI experience, that's right: They're not going to be as good, definitionally, as the labs themselves. But if anyone has used enterprise software, the user experience and quality are not necessarily the selling point.

Mm-hmm.

And this has been an eternal issue in the enterprise broadly, where everyone is confused: Why do all these applications suck? How do these companies persist? What's the bit here?

The answer is that they're often solving problems that you don't see, that no one sees. That lack of visibility into the problem should give you a hint as to why they exist.

Mm-hmm.

Maybe there's a particular issue and there are 47 layers of regulation and compliance, or whatever it might be. Let's take the most extreme example: healthcare.

In healthcare, you have things like the HIPAA laws. You have all these regulations. Or let's get into prescribing drugs, which have all these interactions with other drugs and all these sorts of things.

Hanging over that are massive liability concerns, where you need to know for sure who is at fault if something goes wrong. You have to have all these backstops.

It is an extremely messy, difficult business that you can't afford to… Number 1, it's very hard to solve. Solving that problem actually takes a lot of work.

And because you're having to handle all these variables and all these edge cases, you end up with a user interface that is definitionally impossible to manage. You end up with an Epic installation and doctors having legitimate complaints about filling in 50 gazillion boxes and forms, but every 1 of those boxes and forms is downstream of someone not being liable, or there being some sort of rule or regulation, or some legitimate concern about some interaction effect.

All these rules about what could go together and what can't have to be managed.

Yeah.

And this is a problem in society broadly. People are frustrated at how difficult so many things are and why we have all these layers of stuff. Why can't the doctor just say what's wrong with me, prescribe something, and be done with it?

Mm-hmm.

Ideally, that's what would happen in 98% of cases. The problem is those 2% where stuff can go wrong if something's not caught: 1. Someone could die. 2. The liability is going to be off the charts for everyone in that stack who screwed it up.

Yeah.

What you're paying for is someone to accept that responsibility, to actually go through and get all this stuff in order.

Solve for the edge cases, sure.

Right. And the frustrating thing is this probably ends up net worse, right? It means having every doctor spend most of their day going through a horrible UI and drowning in paperwork, notes, and all these pieces instead of what they got into the business to do, which is to help patients.

Don't even get me started on the whole insurance issue. Actually, the whole insurance area. Epic is a great example. They're downstream from Obamacare. They're the ones that solved all these new issues that were brought on by government regulation, and now they're entrenched in the marketplace.

Everyone hates them because they're hard to work with, but they have this dominant position. But that is the actual thing that they are superior on.

Right.

And it's frustrating to think about and talk about because they are superior on a vector that everyone hates.

Mm-hmm.

But it's the vector that actually drives the bottom line.

Well, they're not solving the top-line problem. They're solving problems beneath the surface that people aren't even really concerned about, but they're the only ones that can solve those problems.

Many of these problems are driven by an avoidance of risk in all these sorts of issues.

Now, again, that's an extreme example, but there are all sorts of things that are like this, where there's all this muck that is solved by a company understanding the space deeply and going through and dealing with it.

They talk to a CTO or a CIO or a CEO, and they're like, “Look, you can have your employees doing this using other software or Excel or whatever it might be, or you could have all these layers of controls, and you can feel confident that—

The muck will be addressed

You're not going to sink the company, right?

Yeah.

Or whatever it might be, right? Again, this isn't a great outcome.

Mm-hmm.

The outcome is you have all these crappy experiences and interfaces, and you have the problem that we've talked a lot about on this podcast: What is all the good stuff that doesn't happen when you're dealing with all this crap?

Yeah.

But the problem is, when you're managing a company, your main concern is that the crap is visible, or the downsides are all visible. The upsides are not. And so—

Mm-hmm.

Well, and dealing with the downsides, dealing with the muck—I misspoke earlier. Epic and companies like Epic are solving the top-line problem. They're just not doing it particularly efficiently, and everybody who uses them is frustrated, but they're also addressing the muck and—

But not doing it efficiently in what way, right? It all depends on what you—

Well, fair, yeah.

So everyone in tech—I took an extreme example with healthcare. But there are so many processes. I was talking about this with Benedict Evans, and I thought he had a very good framing of this: Why do companies have, like, 400 SaaS apps or whatever it might be? It's pretty nuts.

Mm-hmm.

And the reality is, if you have some sort of functionality where it's being done regularly, here's a good example. I was just having a discussion with Domin about—

Okay.

Interview scheduling. I interview people in different time zones. And so—

How do you put that in the calendar? Yeah.

How do you put that in the calendar? Exactly. He was taking an approach that I didn't agree with. The reason I didn't agree with it is, number 1, it's his job to think about that. That's why I'm outsourcing it. But number 2, I might change the calendar event sometimes.

Mm-hmm.

And by definition, I have a lot of stuff on my plate, a lot of stuff I'm thinking about, and I want to make sure we have a system that minimizes the possibility that I mis-schedule something because it's in the wrong time zone.

Yeah.

So in this case—sorry, Domin, I'm going to throw you under the bus here—he was setting the time zone based on the guest's location.

Mm.

And I'm like, "This is stupid." They're not changing—

It's my podcast.

No, no, the point is that there's a certain sense to it, right?

Yeah.

Because whenever I communicate with people, I'm always communicating in their time zone.

Mm-hmm.

When I'm texting with them, figuring out a time. I've learned a long time ago, especially when I was in Taiwan, I'm the weirdo here, okay? I'm used to managing time zones. I can deal with it. It was always a very hairy—

I always appreciated it while you were in Taipei, you know? Very gracious of you.

It was always a very hairy few situations whenever the time zone changed, because Taiwan did not have daylight saving time.

Mm-hmm.

Taiwan does not have daylight saving time, and they do here, and the likelihood of a mistake happening in those few weeks—just very fraught. I'd get anxious about it. So—

But the point is, it's your calendar, so put the time zone that you're using. Is that right?

That was what I said. The reason is, I know this is an issue, so I definitely should, and I usually do, pay close attention to the time zone that the event is in.

Mm-hmm.

But the highest likelihood of a mistake being made is me. I'm at my son's basketball game, someone says, "Can we change a thing?" I look down, I change the calendar, I don't notice the time zone, now we're mis-scheduled and it's a big disaster, right?

Yeah.

That would happen. So we need to have a process that assumes I'm going to screw up and minimizes—

Mm-hmm.

That issue, okay? There are 8 gazillion processes like this in every enterprise, where it's actually pretty clear it's not that hard to do, but what you're concerned about is the one fat finger or the one time someone isn't paying attention, and it's done.

Yeah.

And this was the point Benedict was making: A lot of these buttons or forms in your crappy enterprise application are capturing an institutionalized process. This is something that's done repeatedly, so we're going to encode it in code to make sure that the chances of it getting screwed up are very low.

Mm-hmm.

And that complexity—the dealing with that crappy whatever it is—is offloaded to the employees who complain about it or whatever it might be. Or you can look at it from the outside and say, "Why are you paying $50 a seat for this application? You could just do the—It takes 3 minutes to do it right."

Mm-hmm.

And it's like, yes, but—

Well, that's not their core competency either.

One screw-up—

Yeah.

One screw-up costs you a bunch of money.

Yeah.

And so, broadly speaking, this is why I wanted to zoom in on this point in this email, which I think was a good email.

Mm-hmm.

But this "superior on every vector" is superior on every vector 98% of the time because the LLM just does it better. And look, you have a natural-language interface. I could say, "Oh, change my interview to X, Y, Z," and then it's changed. The problem is, if you gain a small amount or save a small amount of money over time, that can be completely undone by one screw-up.

By the 2%.

And that's right.

Which these companies are already paying to address with the SaaS that they're already paying for. That all makes sense to me, sure.

So this is sort of the—I wasn't actually sure where we'd go with this conversation, but I guess we're starting with a defense of software to a certain extent, particularly in the face of LLMs, which are amazing. By their very nature, there's a thing I wanted to put in this rundown. I don't know if you added it later.

I did add it.

We'll sort of circle back to this.

I'm a little skeptical of it. We'll get there in an hour or so.

We'll get there in an hour, yes. This is professional podcasting. We're giving you a tease for later on in the episode. They make mistakes.

Yeah.

They hallucinate. They are probabilistic entities. The whole point is they're right the vast majority of the time, and because they are loose with it in this probabilistic way, they create this, quote-unquote, "superior experience."

Mm-hmm.

But embedded in that superior experience is the possibility of error.

Yeah.

And a huge amount of software is about eliminating the possibility of error.

Mm-hmm. Fair enough. Well, before we go further down that road with some of the other SaaS companies who are implicated by everything we've seen this week, I do want to focus on Microsoft specifically. One thing I appreciated about your article earlier in the week was the callback to where we were 2.5 years ago, with Microsoft looking like a clear winner from the AI era. You had the integrity to cite your own optimism rather than cite one of, like, 1,000 other people who were saying the same thing 2.5 years ago.

Look, I don't get to quote myself endlessly when I'm right if I don't get to quote myself when it's not looking so good, so—

That's right.

You have to take the L sometimes as well. And to me, it was just a testament to how quickly all this has moved, because all those conversations about Microsoft and the boundless optimism feel like they happened 5 years ago at this point. But in general, this is the biggest software company on the planet. What should Microsoft be trying to do? Where should they want to be in 10 years? Are they going to be a platform, something else? What's the roadmap for that company?

Well, to get to Microsoft specifically, I think the other thing I disagree with in this email is I would have framed it as: Why do we think Microsoft was going to be good at this when they've sucked at products forever?

For a long time.

Right? It's like he's expressing surprise that Microsoft—

Yep.

That's exactly right. Which is, paradoxically, the reason to still be optimistic about Microsoft and sort of assume they're just the next big tech company to go through this cycle.

Mm-hmm.

We went through it with Google, went through it with Apple, went through it with Meta. Amazon's been a low-level bit of concern, although I think there are some aspects of this discussion that are actually good for Amazon, which we can get to in a little bit.

Okay.

And now it’s sort of Microsoft’s turn to be facing skepticism. The thing is, what is the implication of software becoming dramatically cheaper to produce? I think for all the defense that I just laid out, which a lot of people have laid out, you don’t get to stop there.

Mm-hmm.

You don’t get to say, “Nothing’s going to change because software is really important.” When a fundamental input completely changes, things are going to change.

Okay.

3. The Internet Changed Newspapers

So this is where I went back to content and sort of the internet, the ’90s internet. You start out and the internet seems like it’s great for everyone. It’s great for The New York Times. It’s great for The Washington Post.

Mm-hmm.

Another entity in the news this week. The Washington Post doesn’t just publish for people in the—what do you guys call it?

The DMV, that’s right.

Hilarious. Hilarious that the Washington, D.C., area calls itself the acronym of what people associate with a horrific experience.

The worst example of bureaucratic excess and largesse. Absolutely.

It cracks me up endlessly. What is it? D.C., Maryland, Virginia.

Virginia.

It’s not even a good acronym.

Yeah.

That’s the thing.

Well, listen, it’s a sensitive time for residents of the DMV because our team just traded for Trae Young and Anthony Davis and is apparently aiming squarely for the middle over the next 3 or 4 years after allegedly tanking to build a championship team. So I can’t even mount a spirited defense of the DMV label here. But the Georgetown DMV, the actual DMV, is better than it has been in my entire life. So I’m at least grateful for that as a D.C. resident.

No, I heard it’s great. I heard things like snow removal are excellent. Great city services.

Okay. So we’re down bad right now. It is what it is.

Yeah, it sounds like your snow removal is being run by the DMV. That’s sort of what I understand is going on.

Probably, yep.

Anyhow, the DMV—I even forgot where I was going with this. Oh.

Oh, The Washington Post.

Content. Yeah, The Washington Post. Your initial take on the internet is like, “Wow, The Washington Post doesn’t just get to serve the DMV; they can serve the entire country.”

Mm-hmm.

It’s amazing. Our total addressable market just went from a few million, or however many people are in the DMV. I’m just saying “the DMV” as many times as I can in this segment. But their addressable market is astronomical. It’s the whole world. This is great.

Everyone in the world will subscribe.

The problem is that expansion of the market applies to every single publication. Most pertinently, it applies to The New York Times. If you want the story of what happened to The Washington Post, it’s that everyone subscribes to The New York Times.

Yeah.

That’s right.

Well, in this new expanse—

You’ve got to stake a claim, guys.

Power laws predominate. People aren’t going to subscribe to multiple papers that are not particularly differentiated from one another and pay $20 a month for each one. You’re going to choose 1 subscription, and most of the people who are in that space have chosen The New York Times, which frankly has a more distinct point of view. That probably helps them retain customers.

Well, yeah, they’ve been the big winner of this space generally. We’re going to come back to The New York Times because I think it’s actually a very interesting analogy here.

Okay.

The point being, what was the input that changed? The input that changed was the cost of distribution. Newspapers actually were like manufacturing companies with a trucking business.

They printed newspapers, delivered them, and put them on your doorstep, in a newspaper box, in a stand, and all these sorts of things. That was actually their business. That whole thing they felt constrained them—“I can’t serve the whole country”—was actually what protected them. They had local monopolies in their geography.

Once that input went to 0—the cost of distribution, because you’re online—that was actually, in the long run, completely value-destructive because now you’re competing with everyone. Over time, you’re not just competing with all other publications; the cost of entry went way down.

Mm-hmm.

So suddenly you’re competing with me.

Yeah.

The Washington Post is literally competing with me because people can only read 1 thing at a time.

Mm-hmm.

And if they’re subscribing—

Exactly, yeah. They’re competing with you on subscription prices as well.

That’s right. From the eyeball perspective, you’re also competing with Facebook, TikTok, and all these sorts of things.

I actually think this is the 1 point I put in there that sort of refuted my thesis. At the beginning, I was like, “Look, AI’s not replacing software,” for all the reasons we talked about.

Yeah.

4. User Content Dominates Media

But you go to the content 1, and it’s like, actually, the content people mostly consume today is all user-generated content. It’s not professionally produced content.

Mm-hmm.

The thing is, it took 30 years. Maybe my whole defense of software will be moot in X number of years because the AI will actually get good enough to do all that sort of stuff, right? That’s very plausible. I think it’s going to take longer than people think, but then again, you can say that about lots of stuff that moves very quickly.

Mm-hmm.

Anyhow, that aside, the point—

Well, wait. Before we move on, can I actually read an email that’s further down in the rundown here but is related to this particular point on content?

5. Software Enters A New Cycle

Marshall says: “One of the best parts of Ben’s analysis is its roots in the history of technology. It provides valuable context and helps to ground the discussion in long-term, durable dynamics as opposed to what ends up being ephemeral. For example, when some bulls were claiming we were all gonna buy 3 Peloton bikes during the peak of COVID.”

I was probably a little too optimistic about Peloton. I’m just pushing on my sore spots here, but continue.

There you go, owning the Ls along the way.

“Given this, I’d love for Ben to provide some context on what I perceive as a sea change in the way he seems to be talking or writing about 2 areas of technology: semiconductors and software. Has AI fundamentally changed the characteristics of these industries in a permanent and sustainable way relative to the past few decades?

“My working hypothesis is that—

Oh, great job. I saw that word coming. I’m like, man, I’m not sure I’m ever gonna say this.

Deep breath.

Yeah.

Cyclicality. There you go.

A lot of Cs in there.

It will eventually return to the semiconductor business, cyclicality, and the software business—

Look at you, just showing off at this point.

Yeah, there you go.

Amazing.

Though diminished from its halcyon days—

Oh, look at that, another 1. Amazing.

—it will regain its luster, at least somewhat.

“In your recent piece on Microsoft, you compare the impact of generative AI on software to the impact of the internet on newspapers. Though the analogy is apt to a degree, I think it may underappreciate the stacked and multifaceted nature of many software moats.

“While code writing becomes much easier due to Gen AI, and I imagine some switching costs get reduced, there are a variety of other moats enjoyed by leading software businesses that the newspapers did not have once their distribution monopolies went away. Is that fair, do you think?”

Yeah, no, totally. I was thinking about the content analogy all weekend when I was thinking about this piece, and I was a little hesitant to go there for this exact reason, because the defensibility of newspapers actually ended up being quite shallow. It really was just geography.

Yeah.

To his point, there’s a lot more that goes into software, and so I think that’s a valid point. The issue I wanted to push on, though—number 1, I just threw in the user-generated content bit. I don’t think anyone in 1993 fully thought that actually 99% of most people’s media consumption—maybe setting aside TV to a certain extent—was going to be user-generated content.

Yeah.

Most people don’t read newspapers at all. They don’t read books. They don’t read magazines.

They are watching Instagram. If they're literary, they're reading Twitter and reading some stats, right?

Well, they're watching YouTube. The YouTube views that were reported this week were like 200 billion or something like that.

Yeah.

It's absolutely insane.

No, that's the thing. I keep talking about YouTube being the biggest threat to Netflix, and I think I asked Greg Peters about this. Netflix is being modest. They're only talking about YouTube on TVs. If you talk about YouTube on phones and computers—

100%.

It's a gazillion times higher. I was actually questioning it. I'm like, "Your situation is much worse. Maybe it's time to be talking about this, the reality of this." So, anyhow, the implication of this is you might be underestimating AI and its ability to—why couldn't AI actually discern all those rules and all those issues on the fly?