SED News: CoreWeave IPO, Anthropic’s MCP, and Microsoft Turns 50

0:00Welcome to the pilot episode of SED News, a new podcast series from Software Engineering Daily. Join host Gregor Vann and Sean Falconer as they break down the week's most important stories in software engineering, machine learning, and developer culture. In this episode, Gregor and Sean discuss the CoreWeave IPO and the company's recent acquisition of weights and biases, dig into Anthropics model context protocol, surface highlights from Hacker News, and reflect on Microsoft turning 50. We'd love to hear what you think of the format. reach out on BlueSky at atsoftwaredaily or on X at atsoftware underscore daily at Gregor Van or at Sean Falconer.

0:53Hi and welcome to SED Weekly News. This is a new format that we're trying out at SEDaily where we're going to be sort of digesting some of the week's events from software. It's myself and I've got Sean Falconer, who I'm sure you all know. Say hello, Sean. Hey there. Hey, everyone. Hey, Gregor. So hopefully as listeners, you've probably heard both myself and Sean normally interviewing guests. And today we're going to be really just talking about some of the things that we've been seeing in the last couple of weeks in sort of the mainstream news around software and also things like hacker news, kind of trying to bring to the surface things that we see day to day, but don't quite make it into the sort of regular episodes with our guests on software engineering daily.

1:35So yeah, I mean, just to roll into this, how's your week been, Sean? It's been good. I was traveling at the head of this week. And also, one thing I would like to point out, Gregor, is this is the first time that I'm hearing your voice at normal cadence. Usually I'm listening to this. That is a great point. Yes, yes. So it's a little off-putting. I don't speak as quickly as the recordings make out. So yeah, and likewise for you, Sean. Yeah. And how about you? What's going on in your world? Yeah, just I'm normally in Singapore, as I often mention in the episodes. However, this week I am in the highlands of Scotland, which I'm from Scotland originally.

2:09So quite nice to be recording this from up here, kind of where I come to do some thinking and clearly record new formats of software engineering daily as well. So but it's been good. It's been nice to do some dev up here and just kind of keeping on top of what's going on in tech. So I guess with that, maybe we should roll into, these are more like the mainstream headlines that we're looking to cover. So one of the big ones that came up this week was CoreWeave doing an IPO. Did you catch on that, Sean? Yeah. I mean, I feel like CoreWeave's been in the news a number of times recently. They had public acquisition of Weights and Biases.

2:44I've had Weights and Biases, the founders on SED in the past. So they're a company I've known about for a long time. I thought that was an interesting acquisition. And now this IPO at a time when I feel like a lot less companies actually are going public. So I think that's an interesting move from their spot. This was an interesting one. The financial news kind of picked this one up, but I think it's definitely interesting from just being on the software side. At the end of the day, CoreWeave, my understanding is they had a lot of GPUs left over from their crypto days. So there's one of these sort of crypto to AI pivot companies, I believe.

3:21Yeah, there's a couple of those out there. Exactly. Yeah. So they have a lot of GPUs, NVIDIA being the supplier of those. But curiously, NVIDIA is also a customer of CoreWeave. And I think maybe what the financial news kind of couldn't, you know, hurt a few things and they couldn't quite understand why NVIDIA would be a customer of CoreWeave, CoreWeave renting out these chips. And I think maybe as software developers we can probably put the pieces together a little bit faster on that one where nvidia we did an episode with nvidia and they are delivering in inference for their like products but i imagine they're not actually running that themselves all the time and i i guess core weave is now stepping in to be that kind of infra that runs the actual chips rather than nvidia having data centers themselves so i thought that was kind of interesting financial news picked it up on it kind of in a big way because there's a lot of debt and so i think to your point sean you were saying like not a lot of ipos are happening i expect this is maybe just a necessity for core wave at this point they have a lot of debt it sounds like and maybe no other way to kind of handle that the other thing that was brought up was quite interesting was the fact that microsoft is also quite a big customer of core waves and i think a really good piece of commentary was made about that which was just that like Microsoft CoreWeave is kind of like an off balance sheet asset for them, which is, I mean, this is very financial, but I think it's kind of, again, interesting in software terms, where you've got a tech company that Microsoft could clearly run this themselves, but have decided like, well, here's this other company doing the thing that we could do, but actually it's very costly.

4:58Why don't we just let them do it? And then if it fails, then we can come in there later, but it was never on our balance sheet. I think that's kind of interesting. Yeah, there's some weird, I guess, like business strategy mixed up in all the technology that's going on here. And I think that's why really the like the big fang tech companies are such a I don't know, like the beast to take on in the world of business is just they had their hands in so many things in so many different ways. And when they get to a place where they can kind of turn on that engine of growth that they need to, they just have kind of like unlimited rocket fuel to throw at those types of things, which is, I think there's pros and cons to all that.

5:34In terms of like CoreWeave's IPO, and then also these like relationships with NVIDIA and Microsoft, what impact did that have in terms of their public offering? I think that was the story looked good from that perspective. I think a lot of people that were, I guess, looking to invest in this kind of just saw it as a plus. They're like, oh, wow, NVIDIA is a customer. Microsoft's a customer. Fantastic. And I think it's actually only the sort of financial news that's maybe taken another look at that and gone, well, why are they customers and why are they not doing this themselves? That's where I think from the software side, we know why renting GPUs, that's a whole business in itself.

6:10There's a reason why Microsoft and NVIDIA wouldn't particularly want to do that themselves. Microsoft probably more so because they're already a cloud provider, but NVIDIA starting to get caught up in that side of things as well, the service side, that makes sense to me that they would be a customer of CoreWeave. But I think it was that debt thing is what kind of sunk this over anything else. Mm-hmm. And is CoreWeave the biggest company or the most well-known company that's associated with AI that's gone public in the last couple of years? Are they the first one to make a move like this? Obviously, there's a ton of really hot AI companies, but most of them, from what I can recall anyway, are all still operating privately.

6:52yeah no i think that's a great point yeah i mean off the top of my head i can't think of any major ai company that we would actually sort of say is core ai of course we've seen a few kind of say yeah we're blah and we're blah plus ai or the ai for such and such but i think this is one of the first ones where their whole play has been we are an ai company we're only ai and obviously as we know the history is slightly different it was we're crypto and now we're ai right yeah but yeah i mean this is the point like any other company that we could probably just kind of rattle off they're all so private i'm like you know open ai announcing 40 billion investment like not that long ago yeah i wonder if you know obviously there's a huge variance in terms of what it means to be an ai company like are you running inference and gpus or are you doing something that's you know more software based you build the models all these types of things like i wonder if there will be because the ipo didn't go that well if there'll be any sort of downstream negative effects to the companies that are still running privately in terms of just the macroeconomics around this?

7:53Does it hurt valuations? Does it hurt their timelines of potentially going public? Yeah, for sure. I mean, I think I am interested in always the financial side of these things, but more so in the technology. And I think my take on this is just that the average investor is still institutional, are still pretty clueless, really, on a lot of tech and AI. And so, yeah, I think unfortunately, CoreWeave not working out So at least just if you look at just the headline, oh, it went down and such and such. Yeah, unfortunately, this will have a knock on effect for companies that are doing great things.

8:25But just there's still this sort of outside of the tech world. Is AI a fad? Is AI just the next crypto boom thing? I think on the tech side, we can safely say it is different this time. At least that's what it feels like. But yeah, unfortunately, to the slightly outside world, it might just have a knock on effect at the moment until we maybe see another company IPO. Yeah, I guess it might kind of give some of the naysayers fuel to their fire that complain about what is like, you know, tokens are too expensive, inferences too expensive. Is there really like a business model here and stuff like that?

8:59But to me, that's a little bit like complaining about the cost of transistors in 1972. Like the cost of these things will go down over time and then the economics change drastically. So sort of in relation to this, one of the things I wanted to chat about that also came out recently on the AI front was meta releases Llama 4. Yeah. As far as I understand, they skipped over Llama 3, but we just went straight to 4. And I think there was a time when a model announcement was a huge deal. Now there's like a model announcement every day of the week. It could be Sunday afternoon on a holiday. It's like Christmas Day and someone's releasing a model.

9:33But I think what was interesting about the Llama 4 was one, this was kind of long anticipated. I think Llama 2 was showing its age in comparison to some of these other models that have been released. And essentially, there's three model versions, there's Scout, Maverick, and Behemoth. And some of the, I think, the key things that they have that differentiate from their prior model offerings is multimodal support, which is a big deal. I think obviously, like all the sort of premier models are now multimodal in some fashion. Huge context window. So scouts like 10 million tokens, something crazy like that.

10:06Mixture of experts architecture, which has kind of become the industry standard for a lot of the foundation models, but Meta had not used that in any of their open models. And then it's just inexpensive from its performance perspective. So some interesting things there. You know, I'm hearing, I think, mixed reception. Obviously, there was a lot of like hype around it when it came out. People excited about the sort of low cost, the size of these context windows and stuff like that. But other people that I trust that are deeply involved in the world of AI say that from their perspective, there's not necessarily a lot of innovation there aside of sort of the context of Windows size and the fact that it's multimodal.

10:42yeah i mean i guess what in the last also to your point you know models every week yeah we've seen gpt 4.1 come out and that sort of strangely has superseded 4.5 and they sort of sort of rolled that one back or something to that effect and 4.1 is now the shiny one and multimodal i believe is part of that yeah i mean what's your take on we're definitely seeing general models winning at the moment there's a company harvey which is a legal ai company and i think it's kind of starting to come to light that sort of their highly tuned model for law is not quite hitting the mark. Because I would say I've had some behind the scenes comments from lawyers saying they go home and use GPT.

11:24So what's your take on Islamophore? Is that in this direction where just a bigger, better general model is winning the day? Yeah, I mean, I think there's two parts of this. So there's the open weight models. And there's sort of certain like, criticism that comes along with those where, I mean, the advantage of the open weight models is you can pick it up, run it wherever you want. A lot of people use those to fine tune, but I think a lot of people feel like it requires fine tuning. At least that has been the criticism of Llama 2. I haven't played around with Llama 4, so I can't necessarily say from my own experience, but definitely there is a substantial performance difference that I've seen from using Llama 2 sort of base model versus using any of the other sort of premier models of like GPT or Claude and stuff like that.

12:09It's just clearly a difference in performance. And there's reasons why you might want to use that model, but a lot of people end up having to like fine tune it to get it to the performance that they need for the specific application. So there's sort of that contrast of the open weights versus the other models. And there's also these two worldviews or sort of debate that's going on in the industry of, is it going to be a world where we bring the data to the model or do we bring the model to the data. And it's not a hundred percent clear what the right way is right now. But I think if you're looking for the best performance via a simple API call, you're probably going to get that from your open AI model, quad sonnet 3.7, like those types of models, kind of just out of the box, give you the best performance.

12:53And I think a lot of people start there and they build their application there because they know they can get decent performance. And then over time, they might start using one of the open weight models where they try to fine tune it to get it to the performance that they want or adjust prompts and stuff like that. Got it. Where are you running Lama? Well, I mostly use the smaller models so I can run those locally. And then also the company that I work for Confluent, we have an early access version of one of the smaller Lama models that you can use via like native inference. So if you're running something on Confluent Cloud, you can basically run native inference directly on Confluent Cloud.

13:28You can also call out to the bigger models but one of the models that we support sort of locally is the seven billion parameter model for mama too nice yeah and just to go back to that point you made bringing data to the model or model to the data i think that's a great point it's something i've been thinking about more one example i thought about this week was probably one other people have thought about which is like stack overflow and at what point does like stack overflow for example stop having the data to train the models so where did the models suddenly get their like code and debugging data from because no one's producing that anymore.

14:00I think it's an interesting one. I mean, that's the big challenge that a lot of people have pointed to is like, is there more public information that these models can suck up? And there was some news that came out last year or late last year about how the promise of some of the next generation models haven't lived up to the promise. Like essentially the performance gains are starting to slow down, even though they're accessing more information and they're trying to figure out like, Where do we go and get more data? There's a lot of research and testing around how much of this can we synthetically generate, but there's also pros and cons there.

14:31If you're using a model to generate synthetic data and then training on it, does that degrade some of the performance over time? At some point, the public information on the internet is going to be surpassed by essentially the majority of it's going to be AI generated versus human generated. And what does that do? And I think like the sort of next generation and maybe the arms race of around AI is like, who has the best data sources? All these companies that have been around for a long time are sitting on mountains of probably really high value human generated data. So is there a startup out there somewhere that becomes like a marketplace for data sharing that help train these models or some other way to tap into those types of sources?

15:13And the other thing is, you know, if I'm a big company that's been around for a long time and I have that data in, I want to start to leverage AI. That's sort of my value to the equation. That's my proprietary information that I want to hold on to. So maybe I want to use that as part of some fine tuning process or at least use it as part of whatever sort of rag process I'm doing. Yeah, completely. So we're going to move on. So we've done the main kind of headline topics. And now we're going to move on to a sort of all-encompassing topic that's sort of been in Hacker News a lot in the last couple of weeks, been in the main news, I guess, even.

15:48But it's certainly a topic that probably none of us have been able to escape from, which is MCP, MCP servers. So as I mentioned, I'm sure as a developer, you're listening to this. You must have seen something on MCP in the last while, LinkedIn, Hacker News, etc. Unless you've been under a rock somewhere. Right, exactly. I was just thinking, yeah, exactly. And unless you've been under a rock, MTP just seems to, I mean, as soon as it hits LinkedIn, that's like classic thing, you know, if your hairdresser starts talking to you about X, then it's probably, so I wouldn't be surprised if my hairdresser started asking me about MTP servers next week.

16:20So I think the key thing is that a lot of people have different ideas and definitions around what MTP even is. We'll get into, I think the one that's been bandied around a lot is this is the USBC of interacting with AI services. And yet then there's lots of other people that say that's a rubbish way to describe it. So I think let's just start there. Sean, do you want to just walk us through this? What is MCP? What is an MCP server? Yeah. So MCP, which stands for Model Context Protocol, was announced from Anthropic last November. And Anthropic pushed us forward, but it's not an Anthropic technology.

16:57It's essentially a proposal for an open standard. And the problem that we're trying to solve was that when you're building like an AI agent, part of an agent typically is tool use. And a tool could be a function that's going and executing something, some sort of a deterministic process, or it could be a function that's going and gathering data from some sort of place. And before we had something like MCP, every agent framework that you're building on, or if you're doing this from scratch, you're kind of doing that in a new way. You know, you're going and writing code to talk to an API, or you're going and writing code to talk to a database server or whatever it is, which is fine if you're doing three tool integrations and that's the end of it.

17:36But as people build more and more agents, you want to get away from having to go and essentially write a lot of bespoke code that you then becomes a potential technical debt in your stack because you have to manage that and stuff like that. It's just like the value of using an API gateway or something like that, right? To interconnect to different systems. So what Anthropic did is they're trying to solve that problem. So they They propose this standard, which is based on a client server model, where if I have an MCP client and I'm a data provider like Slack or something like that, or Google Drive, I can create essentially an MCP server that adheres to this protocol, stand that server up somewhere.

18:13And now that client can talk to that server and gather data from Slack and stuff like that. And the client, which could be essentially an agent, doesn't need to know anything about Slack protocol or what that is. is essentially just expressing in natural language what it is that it needs to gather. And the servers can define, they can have multiple tools defined, and there's essentially this standardization around how you expose that. So you have tools, you have something called resources, which are essentially data that you're providing. And then you can also provide prompts that are predefined as well that these clients can use to interact with.

18:44So that was announced in November. I was really excited about it when it first announced. I read it, I understood what it meant potentially for the industry. And even in my day job at Confluent, working on AI, that was something that I have worked on Champion. We actually released an MCP server open source project a month or two ago that allows somebody to talk to data in Kafka, also manage Confluent Cloud directly through natural language as a tool interface. And you can plug that into whatever MCP client you want. But what's really interesting is, so MCP, its success, like any standard, depends on adoption.

19:21Because if there's one MCP server out there, it's not much of the standard. So it had very steady, significant growth of a lot of open source projects and stuff like that. But really, I feel like over the last month, it really exploded because OpenAI came out with their new agent framework that supports MCP. So even though they're arguably anthropic as maybe their biggest competitor, rather than OpenAI coming up with their own version of MCP standard, they said, we're going to support this as well. Then AWS and Bedrock also came out with their agent platform, and they announced support for MCP.

19:54And then last week at Google Next, Google announced their new agent framework, and they're also supporting MCP. So you have these really, really big players in the industry that are also supporting the standard, which I think has really elevated it into the stratosphere of conversation around Aon. and there's like platforms like smithery for example where you can go and basically just find mcp servers that do whatever you want it to do but i think that's maybe been some of the criticism as well which is sort of like who's running these servers and like who's behind them because i think that was maybe the slight mental switch i had to make when say looking at something like smithery which was i was like okay where's the repo and it's like there's no repo and then oh okay so then we've got to think about okay well hey here's one that's like this can help you interact with Google Workspace.

20:42It's like, okay, but then my API keys are going where exactly and this kind of thing. So is that something you've kind of, I guess, hit up against like as you've been using them? I mean, not so much of my use, but I understand that. It's kind of a little bit of the wild west, like any sort of new technology. It's clearly not great if people are just like standing up MCP server. I mean, maybe that's fine if you're doing your hobby project and stuff like that, but obviously you have to be aware of what you're consuming. If it's just some random person standing up a server that's going to go and talk to your Google Workspace.

21:13Clearly, where's the API key staying? What's their logging? What's the agreement that you're making about data sharing stuff? People get so scared about the idea of sharing something with an open AI model. I'd be a lot more scared about sharing my information with a random MCP server that's running in some data silo in Northern Russia or something. I think these are things that you have to be thinking about. Just like when you go and you grab a random GitHub repo or NPM resource or whatever it is, you have to be thinking about, especially in production scenarios, like what is the potential supply chain issues that could happen and stuff like that.

21:50Yeah, I think that's a good delineation running MCP locally versus production. And I kind of get the impression that at the moment, MCP for the average person, average developer, it's probably more of a local endeavor. The easiest way I understand to get up and running with MCP at all is, for example, download the Cloud actual native app and then have it authorize into any given tool like Slack or Google Workspace or whichever. But at least that's all running locally as well as you probably trust Anthropic at this point with any kind of data like that. And then outside of that, sure, you can run any kind of local MCP setup and not worry too much about maybe where things are going.

22:33But then in production, I think that's where there's maybe just a ton of question marks at the moment. Yeah. And I think that over time, obviously, like the number of MTP servers and clients that are available, like on GitHub as open source projects is like exponential, right? And obviously, that's, I think, like wire technologies, that's like the tip of the spear. but over time, more and more people who have, I don't think Snowflake has this yet, but like, let's say Snowflake, as a Snowflake user, you can expose your warehouse as an MCP server because you're building some agent, right? So there'd be, I think, these more trusted sources, like I'm already a Snowflake user.

23:08I'm going to use their built hosted MCP server to talk to my Snowflake rather than some third-party server that's running somewhere where I need to give them my credentials to talk to my Snowflake and stuff like that. So I think it'll be these like more trusted versions of these servers more and more available from the regular companies that you would expect. Yeah. And I mean, just then in terms of actually, what does this enable? I think you touched on when we were just chatting earlier about agent to agent and how that's been mentioned pretty recently. So maybe just talk us through that. Yeah.

23:41So agent to agent was another announcement from Google last week. And it's complimentary to MCP. I see there's like a million articles the talk now about explaining the differences between these things. So I think like MCP was really focused on the tool. So to use your analogy or the analogy that you alluded to earlier, like, you know, being this USB, it's really about like a standardization or like how I go and gather data or how I execute some sort of function. And what agent to agent is trying to do is also propose an open standard with solving a different problem. It's actually a problem I wrote about a month and that Google's trying to address is essentially, it's probably a future problem where as enterprises adopt more and more agents and more and more agentic software, I have my Salesforce agents, I have my Glean agents, and I have my Cortex agents.

24:31And then maybe I'm building my own agents. I'm creating these islands of essentially independent agents that don't talk to each other. So just like we've created tons and tons of data silos, we're essentially creating a future of these AI intelligence silos where they have no way of communicating to each other. So in the article that I wrote, I talk about this problem and I've proposed a solution through data streaming technology. And what Google has done is they propose a standard to solve this problem. And essentially, that's this agent agent protocol. So it's more focused on how do we take these crews or meshes or swarms of independent agents and make it so that they can all talk to each other.

25:08So I can go and I could build an agent using Microsoft Autogen and I could build one in WayneGraph. And presumably if my Glean agents also support this, I could have them all have conversations with each other. It sounds like a great, I don't say dream in the sense it's not possible. I just mean, it sounds like a great sort of vision and a great future. What kind of applications can you, you know, applications to, to kind of the real world, if you like, can you see this helping with? I think they're trying to address a problem that maybe doesn't fully exist right now because most businesses, at least the ones that I talk to day to day in my job, are just kind of like, I just want an agent.

25:42I just want to solve this. I want to do loan underwriting or claims processing. We have a bunch of forms that we need to fill out and we want to automate it. So they're starting there. This is kind of solving, I think, a future problem. But that future problem, if you believe the vision that at some point enterprise is going to be having thousands of agents running around, doing various times tasks, or software is some sort of agentic workflow, then this is going to be a real problem probably in the next few years. And I think fundamentally what it is, and I talk a little bit about this in my article, is it's kind of like what HTTP was to the web, where you have the standard of protocol that suddenly websites can talk to each other.