Open transcript
It feels like there's a new database startup every week. From PlanetScale to Terso to Neon to Convex, there are just so many different options. At least there were. But recently, it feels like we're ending up with less options, not more. From PlanetScale killing the free tier and focusing entirely on enterprise customers, from Fauna shutting down, and now Neon being acquired. Yes, Neon, everyone's favorite Postgres database provider, has now been acquired and is part of Databricks. This is a fascinating acquisition and there's a lot to talk about with it. I'm very excited to go in-depth on what this means, both for Neon users and the Neon company, more importantly, for the whole database ecosystem and the innovation that we've been seeing in database platforms over the last decade or so. There's a lot of context that's important to understand for what's going on here. And if you're trying to decide which database to use, I think these things will be useful as you figure out the ecosystem and where things are at. That all said, none of these companies are paying for this video and someone has to pay my editor. So quick word from today's sponsor, then we'll dive right in. Today's sponsor's a web browser. Yes, really. But not one that you or I would use. One for AI. I know this sounds insane, but hear me out because BrowserBase is a phenomenal product. I actually wanted this before the whole AI wave was a thing because setting up Puppeteer and all of the chaos around Chrome in the cloud when you're trying to hit it via serverless, it's not fun. BrowserBase was originally built to solve that. And it turns out that tech is useful for a hell of a lot of other things, including, of course, using it with AI. They now have full support for the computer use model from OpenAI. They have a fun little demo of the computer use stuff that I want to show off quick because it's just so cool. They have a bunch of examples here. You can give it your own questions too. Let's just ask it to review a pull request on GitHub. This is the message it sent by default, just the one that it had pre-populated. Find the most recently opened non-draft PR on GitHub for BrowserBase's stagehand project and make sure the combination evals and the PR validation passed. So here it spun up a virtual Chrome instance on their servers, so you don't have to worry about it. And now it's making API requests to that endpoint to do things in the browser. And we're just being beamed back the results from it. So it can do everything from search Google to scan random websites, all the stuff you would want a browser to do. Now your AI can do it too. Fun thing we're seeing here, it's completing the CAPTCHAs for you. Yes, really. Kind of insane that it can just do this. And I haven't pressed anything, by the way. It's just inferring all of these things based on the task I gave it. Now that it's on the stagehand repo, it's actually finding the right tabs to click. You can see it's telling you which position it's clicking on the screen. It's going in, taking a look at this most recent PR. Now it's going to hop over to the code tab and give it a review. It's operating a computer. So if you have something that you want to do that doesn't have an API, you can now build an API for anything accessible on the web. It's pretty nuts. You want to give BrowserBase a shot? It's free to get started. Check them out today at soydev.link slash BrowserBase. Neon and Databricks. Before we dive too deep into the post, I should give a little bit more context on who these players are. Neon is an interesting company. Neon wanted to rethink Postgres. If you're not familiar with Neon, the TLDR is serverless Postgres. The longer form version is that the founder of Singlestore, who you might know because they previously sponsored one of my tutorials. Singlestore is an interesting product. It's very expensive and hard to set up. But if you need the features it has, it seems like it makes a lot of sense. That said, for use in things like tutorials and small projects, 90 cents an hour or like 50 plus dollars a month is not an easy sell for a lot of people. So interesting role. You can also self-host, which is cool. But yeah, the reason I bring up Singlestore has very little to do with them, though, is that their CEO left and started Neon. He started Neon because Singlestore was a little more MySQL-y and had focused entirely on enterprise. And he wanted to focus more on this new era of small to medium stage companies, side projects, all the things that Singlestore realistically wasn't great for because he saw a growing market in these smaller teams building things. And there's an important detail here that I think we need to understand. I hope I don't have to go too in-depth here. We'll see how y'all feel when I say it. Postgres kind of sucks for serverless. I know, bold statement, but the issue with Postgres is connections. You're effectively forced to manage individual connections to the Postgres instance, and those take resources. When you're using serverless environments like Lambda, like Vercel, like Netlify, even like Cloudflare, the connections you form are ephemeral. You make a connection, you complete the request the user has, and then you close the connection. And that means that at any given time, the number of connections you have formed is roughly the number of things you have people doing at that moment. So if you have 10 users making a request at the same time, you have 10 connections. If one user is making 10 requests at the same time, they're causing 10 connections. If your database is limited on the number of connections it can have, good luck, have fun. I learned about this the hard way, way back when I was trying to use the free tier on Heroku alongside Vercel. And just in my own testing, I was able to hit connection pool limits and watch my stuff plummet. And the option that was recommended to me was to spend like 400 bucks a month on their private connection pooling solution that also would have been relatively limited. This has caused problems for a lot of people betting on serverless, including companies like Supabase, because they've had to build their own connection pooling layer that still has relatively low ceilings for how many connections can be formed on a given Postgres instance that results in high throughput serverless workloads not working great in a Postgres environment. And to be clear, MySQL doesn't magically solve this. However, a lot more work has gone into pooling and separation in MySQL because as an architecture, the pieces are a lot more separated than in Postgres where everything's kind of tightly tied together. There's also companies like PlanetScale that provide an HTTP layer. So instead of querying my DB via a native connection to it, I can send a post request to an HTTP endpoint that handles all the pooling and connection management for you. Works really well for serverless. I found PlanetScale to be a phenomenal solution for that reason. So Neon wanted to solve this as well. Not just because serverless is hard, but there was a lot of DX problems that using traditional Postgres could cause you. One of the biggest ones that I think Neon actually got really, really right was preview environments. Preview environments are one of those things that like once you have it, you can't really go back. If I go to any of the current pull requests on T3 chat, I can hop over to this import chunking PR where Mark fixed the imports. And I can click this visit preview button. And this will bring me to a link that is a preview environment. I have to enter a little password for it. And now that we're in, we have a fully working preview environment that has everything we need to test the changes that were just made. You'll see it's a URL specific to our project. The point here being, it's really useful as part of our review process to actually test the thing the person made without having to download the whole project, have to switch to that branch, install the dependencies, run the dev command, and then go play with it. Now you can just click visit preview. And once you have a workflow like that, it's really, really hard to leave it behind. And that's a huge part of what Neon wanted to figure out is those types of modern DX expectations that traditional Postgres made very difficult to do. Because if you had a branch that added something to your database schema, you can't connect it to your traditional staging environment, you probably have to go spin up a new database directly. And each individual DB is quite expensive due to, again, everything being so tightly tied in Postgres. So one of the things Neon focused a lot on was branching and making any individual DB deployment as cheap as possible. You might be curious how they did that. The answer is interesting. This is their open source Neon database, serverless Postgres, as it says right there. Don't know how much you guys know about Postgres, but it was not written in Rust. Neon is a full rewrite of Postgres focused on all of these things with the goal of, to the best of its ability, making something that is Postgres compatible while also being much, much easier to use in these new environments and ways. An open source alternative to Aurora Postgres, which is also AWS, believe it or not, has seen a lot of the same problems and tried to build their own solutions on the AWS side. Neon's an open source alternative that also has their own cloud. And that cloud has historically been quite cheap. The free plan gives you 10 projects, half a gig of storage, and then a ton of compute hours. For 20 bucks a month, you get 100 projects, 10 gigs of storage, and auto scaling up when your databases need more throughput. It made a lot of sense as a product. Their uptime was horrifying. I think it's gotten better since I don't care to check because they don't report their uptime in ways that I consider respectable. But they built something legitimate with the goal of making Postgres work better in these modern environments. They also scaled up really aggressively. This was the biggest concern I had with Neon from when I first started talking with the team and the founder. I'm going to drop some numbers and have you guys guess some things. They're deploying 3,000 dBs daily. They have 750,000 of databases under their management right now. How many employees do you think they have? For reference, T3Chat does a few hundred thousand messages a day and it's two employees marking me. And if I recall, PlanetScale has under 40 employees. Might even be smaller than that. So here are chat's guesses. 5, 3, 10, less than 50, 20, 25, 40, 3,000. Mostly numbers under 50. They had 130 fucking employees. 130 people. Do you understand how insane that is? I remember when they were still really early and they had like 90 employees while having like less than a fifth the traffic of PlanetScale, which had like 40. It's massive, massive overhiring, in my opinion. So napkin math, even if we assume a lower salary, because they're obviously not all ends, we'll assume 90,000 base salary times 130 people. Just salaries alone, that's 11.7 million a year just paying your employees. That's kind of insane. If you come from the single store enterprise world, it makes sense to have a ton of employees. But in modern startups, it doesn't at all. And to contrast with PlanetScale, they got roasted for doing layoffs a while back. Kind of crazy. This was over a year ago. Yeah, I thought there might be direct numbers in here. There aren't. But when PlanetScale killed the free tier, the goal was to make the business profitable so it will be there forever. So you don't have to worry about your database running out of funding and going out of business. Like Fauna just did a few weeks ago. The goal was to make the database something you knew would still be there next year. So big businesses that were iffy on making this type of bet would be able to do so more reliably and with more trust. So again, getting your database company profitable is hard. And doing that with that large number of employees is nearly impossible. So let's look at the current database startups that we have. I'm going to start with PlanetScale because they're the one I'm the most familiar with and like know the most about. Indos and characteristics about PlanetScale are individual DBs are expensive, scales incredibly well, best performing option even considering AWSs. Especially now that PlanetScale has Metal, they have one of the best products on the market. It is going to be really hard to compete with them if you're looking for something to set and forget and never need to touch again. Next, we could look at something like Terso. Terso has very different characteristics. If you're not familiar, Terso is a SQLite database platform and they are hilariously cheap. Their free tier lets you have 500 databases. The reason for that is individual DBs are hilariously cheap because they store the SQLite database as like cold storage in S3 and recover it when a request is made. A DB that's doing no traffic costs them zero. It is effectively free. That allows them to do crazy things in terms of generosity of the number of databases that you can have set up. But the scale is weak. I always be able to make it buckle relatively easily just under some basic tests. And reliability is not there. Of all the database companies, Terso has had some of the most embarrassing flubs. They recently started defaulting users to their new AWS deployment instead of their fly deployment. And they lost all the data on the AWS deployment a few weeks in. So everyone who had clicked the next button three times and just went with the defaults lost all their data. Prior to that, they had a bug with the restoration of the S3 file where they could restore the wrong one on a request and certain users were getting access to other people's data because when they tried connecting to their DB, someone else's database was restored on the node. So if you're really trying to make something reliable, they're not a realistic option, sadly. So we have, if we were to like make a range here, we have Terso on like the beginner side project side, planet scale at the enterprise, like ready to go for scalability side. Neon was trying really hard to be an in-between here. Neon wanted DBs to be relatively cheap, enough for branching. Branching being if you make schema changes, you can have a branch with the changes. And then once you've decided it's good after your pull request is done and all that, you can merge the branches changes in to the main deployment. Scales, fine. Reliability is improving. As far as I know, they haven't had massive data loss instances or data exposure stuff like Terso has, but they had a lot of downtime over their first two years. And it's gotten better, but it's still from what I've heard, not great. So Neon's this weird in-between of these. If we were to think about this a little differently, instead of just thinking of this as like, how do they differ technically? Let's instead think about how much money it costs to run them and how much money they could potentially make. So if a given PlanetScale database costs the minimum 20 bucks a month to run, they need to make serious money on these users. So if it's 20 bucks a month, they can charge 30. With the new metal stuff, they can charge much, much higher amounts, like literally like 500 plus a month with no issue because they are so far ahead of the competition and can do things that like you normally spend thousands for. They can get away with that. No user on PlanetScale is free anymore. So every user is paying them. So anytime somebody spins up a database that costs money, PlanetScale charges them more than it costs them. So every new user is guaranteeing they make some money. Things are a little different for Neon. I have no idea what their cost basis looks like. I could try running their DB stuff and like get a gut feel myself, but I haven't had a chance to do that just yet. But if we assume it's like even 10X or 100X cheaper, it's 20 cents a month per DB that someone's running on Neon. The average Neon customer pays them $0 a month. So every time a new customer spins something up on Neon, it's purely at cost for them. If you combine that with the amount of money they are spending on employees and marketing and all the other things they're doing, their costs are probably absurd. Terso is similarly hilariously cheap, probably even less. I'll say it's like 5 cents a month for them to run things. But I am the only person I know that's paying for Terso right now. I just don't know a lot of people who are doing it. And they've been experimenting all with their tiers to try and make money on a cheaper tier. If I've learned anything about this with products like Upload Thing, you will never make a successful business charging devs $10 a month. You need to charge a lot more than that in order to make your thing profitable. So even though PlanetScale has the highest cost per user, they are by far the most profitable, have the best margins, and the best long-term upside. But there's something that buffers this a lot. Fundraising. Fun fact about PlanetScale's history. The CEO of PlanetScale, Sam Lambert, was actually a Sequoia partner. Sequoia is a huge investment firm, clearly was part of how he got involved. He was an investor in PlanetScale before he joined and then became CEO. Sam deeply understands the fundraising world. And he saw back in 2023 that the likelihood they could raise another huge round at PlanetScale was low. So these are all entirely fake numbers. I just want to give you guys an idea. If PlanetScale raised $10 million on a, I don't know, $80 million valuation, they now have $10 million in the bank that they could spend and use to eat costs, justify free tier, do marketing, sales, all those things. And they gave up an eighth of the company in the process. What happens if they run out? Once that $10 million hits $1 million and you see that your runway is down to just a few months, you're going to be out of money in six months or so. You could raise more money. But if your profit at this point was, I don't know, $800K a year, and now your profit's $1.2 million a year, and you got that gap over two to three years, that's not a whole lot of growth. And your costs are still absurd. It wouldn't make sense to invest. And that would mean they have to do a thing called a down round. Down rounds are terrifying. When you get more money on a lower valuation, your company is now worth less. And this $10 million, especially if it was on a safe or some type of MFN equity, this is no longer an eighth of your company. It's now a sixth of your company. And you just sold another sixth. So you went from giving up an eighth of your company to giving up a third of it very quickly. Sam saw the writing on the wall. He knew that since he raised on a really high valuation, the likelihood of doing another higher one was low. And the reason VCs invest on these high valuations is they think eventually the company can sell or IPO for a much bigger one. So if you put in this 10 mil at the 80 mil val, and they end up being worth 800 mil or a bill, you just turn 10 mil into 100 mil. That's why these companies are investing. But if it doesn't look like there's a chance of this being worth a billion dollars in the future, you can't really raise money at that point. So if your costs are going up and your revenue isn't going up fast enough to compete, and you don't see a path to raise more money, you have to look for other options. And the option PlanetScale looked for and found was to make themselves profitable. If they can't raise their revenue fast enough, and they don't want to raise more money, what if we reduce costs instead? If you have the three metrics, which are money in the bank, money that you're making, revenue, and costs, you can't increase money in the banks, you can't raise more. Making more money costs more money. So how about you reduce costs? That was the choice that they made that resulted in them being able to not only survive, but thrive. They're doing better than they've ever done. They're able to build things they couldn't have built before. And the results, awesome product, that's super reliable, being used by everybody from T3 chat to Square and Cash app. It's kind of crazy how much stuff, both data and now money, flows through PlanetScale. And Sam's at the point where he is like anti-fundraising. So why am I talking about all of this here? Because Neon didn't just fundraise. Well, if we think about these three things, which are money in bank, costs, and revenue, and we think about where Neon's at for all of these, we'll change the order a little bit to make this flow better. Costs, as we've discussed, aren't great. The engineering side in particular and the employee side is absurd. I'd be surprised if they were spending less than $12 million a year on payroll alone. Revenue, I can't imagine it's particularly high. As we saw in their pricing page, the only options that are under the fold here are $0 a month and $19 a month. And then they have the scale, which is $70 a month, and the $700 a month for business. But they're not printing. They're not printing. I know that much. There's almost no way Neon is making a ton of money right now. Due to the recent surge in AI app builders, things like V0, Lovable, Bolt, Replit, etc. Many of them use Neon as the default database because the free tier is generous. They could spin up a bunch of databases using Postgres. It made a lot of sense for them, which has caused a massive surge in the number of these databases that are being opened. How many of those people that are generating these apps do you think end up becoming paying customers for Neon? My guess is it's a very small number. They have 4x more databases created by CodeGen than by humans. As of February, that ratio is getting worse, not better. So costs going up massively. Revenue, probably not moving a whole lot. Maybe it's bumped a little because they're getting companies like Replit to pay them for some amount of the allocation. Money in Bank, almost certainly going down fast. Neon has so far raised $129.6 million. That is insane. That means they've either given up over half the company or their valuation is something crazy like a billion dollars, which I don't see how they could ever be a billion dollar company personally. So yeah, it's kind of absurdly, I don't want to say overvalued because I don't know what the actual valuation was. The 19 bill number here, that's Databricks, not them. But $129.6 million raised is absurd. And that's the only reason they could justify the 130 employees that they had. Yeah, this is my problem with Neon is it always felt like they were operating two tiers past where they were. They'd be competing in the startup space, but hiring as though they were an enterprise company. And it always, it felt like a, like they were building a castle out of like cards that was just waiting to collapse. And I was starting to think it would. And my guess is that Nikita went to go raise more because he went to all these VCs like, hey, we're the database that all these AI companies are using. We need more money. They're like, oh, cool. New business, probably what, 100 mil valuation? And they're like, oh no, we're raising on a 2 billion valuation to which the VCs looked, looked at their numbers, looked at their costs, looked at Nikita and said, what the fuck are you talking about? So if you can't raise without giving up a ton of your business or doing a down round, which when you've already raised 130 million, good luck, their revenue is not growing fast enough. Their costs are going up like mad. What are your remaining options? That's when acquisition happens. The thing Neon had that makes them valuable is the ability to, at early to mid stage, deploy tons of Postgres databases for relatively cheap. Databricks, they're a unified open analytics platform for building, deploying, sharing, and maintaining enterprise grade data analytics and AI solutions at scale. Their data intelligence platform integrates a cloud storage and security in your cloud account and manages and deploys cloud infra for you. So it lets you point things at random databases so you could do data analytics, warehousing, all the things that your data teams need. They've been around for a while. They make a ton of money. I have a few friends that work at Databricks. They've never paid me or even reached out, but they're a legit like enterprise-y data analytics company. Databricks doesn't host databases, at least in the traditional sense. Like they're not the thing you write to when a user signs into your site. They're a company that helps you build the analytics platform around your existing database. But they're also very clearly trying to get into AI. As you can tell by the nonsensical homepage. Your data, your AI, your future. Yeah. Yeah. You get the point. Databricks is a very enterprise-y business. It's trying to figure out the AI thing. They don't have a database platform in the traditional sense. They have a whole lot of money, a whole lot of enterprise customers, not a whole lot of way to grow in the small to medium-sized stuff. Neon has a whole lot of growth in the small to medium-sized, no money, and no enterprise customers. It's actually kind of a match made in heaven in that sense. Databricks now has a way to introduce their offerings to smaller companies as they start to scale. And Neon now has a way to not go out of business. The core point being Neon needed the stability of enterprise and money. Databricks wanted more early-stage stuff and more AI-focused everything. So let's look at the actual post that Databricks put out. We're excited to announce that we've agreed to acquire Neon, a developer-first serverless Postgres company. Actually, okay, one more tangent just because I have thoughts on Nikita as an individual and I think this will help showcase what I'm talking about. We're excited to announce that we have agreed to acquire Neon. It's clearly what's happening. They have a ton of money. They are spending the money and spending some equity in order to buy the Neon business. Neon and Databricks. This is the post on the Neon site. Today, we are thrilled to announce that Neon will be joining forces with Databricks. The comms I've seen from Nikita have mostly avoided the word acquisition. At the very least, it's multiple sentences in. The posts I saw on Twitter and whatnot were very much, Neon is partnering with Databricks. The phrasing I have used that's a little rude, but it's the best I have for this is what I call cosplaying startup or cosplaying CEO. It's when someone really wants to run a big company. They see what all these other big companies do and they copy those details instead of the actual thing. The better terminology would probably be cargo cult. I always forget the word. The concept is people see the outward action someone does. Like for my YouTube channel, you see the goofy thumbnails and titles and you assume that's why I'm successful. So you copy those parts, but you don't actually see where those came from and why I do them. You just see the thing so you copy it. Successful businesses hire a lot of people. So if you want to be successful, you should hire a lot of people, right? No, not at all. The reason successful businesses hire a lot of people is because successful businesses have a lot of problems and they begrudgingly have to hire to solve them. These are common things I see in founders and startups and when I invest in an early stage company, I often make them promise that they will not more than double in size by the end of that year. I actually just made this promise with two different startups I was investing in. I made them guarantee to me that they would have less than 10 employees by the end of the year and they wouldn't go spend this money over hiring and destroying themselves in the process. Neon did not do that. Neon was very much built as a almost like a sequel to single store, the big enterprise database company and as a result, I think Nikita might feel a bit of shame that he went the acquisition route. It is kind of playing it out a little differently as a result. I'm sure this is not what he wanted to do but it was the best path to preserve as much of the business and the money and the investors' interests as possible. Let's go back to the official post from Databricks. Neon's team engineered a new database architecture that offers speed, elastic scaling, branching, and forking. Capabilities make Neon great for devs and also great for agents. Today we'll deliver an open serverless database foundation for developers and AI agents. Oh look, that same 4XM we were talking about earlier is in here too. AI agents are creating four times more databases than humans on Neon. This paragraph kind of says it all. You think of AI agents as your own massive team of high-speed junior devs potentially being mentored by seniors that it's not that surprising the same capabilities that the Neon team was focused on that made it great for those junior devs also make it great for these AI agents. That is Databricks' whole vision here. Previously, they were fully focused on enterprise because that's where the money is but the thing that is changing right now is that what was previously stereotyped as like junior, beginner, small stage stuff is happening more and more at these big businesses and those small things are becoming more and more profitable thereby are more and more willing to spend on products like Databricks. So if you want a company like Microsoft to use Databricks and you are struggling to get the big Microsoft products to use it, Microsoft now has a bunch of small teams vibe coding their way through random stuff that they're dogfooding and testing out. If one of those happened to use Neon and then it does well and then they need analytics on it, Databricks would be a way now to get this one small team at this big company to start using their new product and now Databricks can get in at these places that traditionally this would have been how you get into a small startup. Big companies are becoming a bunch of small startups within them. This is more and more the trend. Facebook's been on top of this forever now but we're seeing more businesses do the same thing where there are small vertical slices that build their whole product top to bottom and those teams are getting smaller and faster than ever using these AI tools. If Databricks wants to survive this new wave of small to medium enterprises this is a great path there and they knew their existing product made no sense for it. All of a sudden this change is happening and this is kind of a hedge to bet where if these smaller teams and smaller projects and junior engineers that are making these decisions start to make real money at these big businesses Databricks now has a way to keep them in family and pull more people in as a result. I've seen this a few times now where effectively different businesses built a thing focused on a specific segment of the market that accidentally made a really good thing for AI. Some of my favorite examples of this are Convex who admittedly is a sponsor not of this video but they sponsor random videos on the channel Cloudflare and now of course Neon. Convex does this well because Convex puts everything inside of a single directory in your code base. They are truly infrastructure as code. The schema for the T3 chat rewrite this is the whole thing is just here as a file. If you were to get this code base and you wanted to connect Convex you don't need access to our deployment because there are no switches we hit in their dashboard or anything. You literally just run npm run dev it asks you to authenticate with Convex if you haven't yet and then it asks which project is part of or if you want to make a new one and then it provisions the whole thing. The magic there is that AI is really bad at navigating a dashboard like AWS and remembering which switches are hit where and what state things are in. AI is really good at taking a bunch of text and making changes to it. So if you can actually properly do the thing we wish Terraform did where your entire deployment and all of its characteristics just live in a directory of text files ideally source code that's type safe like it is with this. This helps a lot with me as a dev not making mistakes as I iterate but it helps even more for AI because now it has a type system enforcing all of the behaviors it's doing and deployment is just a command it runs rather than a bunch of steps it has to get right. Even with Terraform things are not competent in that side you still spend a lot of time doing weird bullshit and hooking things together. Convex is kind of the perfect backend for AI because everything just is a directory like this and stuff like triggering updates on client is comically less work too due to their sync engine. Convex by building a really good simple thing for front end devs didn't want to do the inner workings of backend they built the perfect backend for AI app builders. Cloudflare built the perfect infrastructure for doing inference for generating LLM responses. Cloudflare's workers are a glorified switch statement that's what I've called them forever. That's both why they're so magically cheap and spin up so quick and it's also why they suck at things like SSR because the amount of CPU on a given worker is really low. Cloudflare workers are isolated in the V8 runtime so it's similar to how tabs are isolated in your browser to an extent which means that when one tab is waiting for some data to come in other tabs now have more CPU as a result Cloudflare doesn't charge you based on how long a request took it charges you based on how much CPU did you use during that request so if I am generating an AI response from a slower model like DeepSeq and it takes 70 seconds to generate if I do that on Vercel I'm charged for 70 seconds if I do it on Cloudflare I'm charged for like 40 nanoseconds because it's spending most of the time waiting for the next chunk to come in from the provider they built the perfect infra for infinitely scalable AI inference Vercel's catching up with the stuff they're doing with fluid compute but you're still being built on a wall clock in the end if I have a request that takes 70 seconds the lambda's up for 70 seconds I can stuff another 100 requests in that 70 second window but it's not going to get close to the 40 nanoseconds that I'm charged for on Cloudflare due to the difference in their compute model that compute model made workers garbage for so many things they're not running nodes you lose access to everything from like file system to connection like management they didn't even have async local storage until recently I don't even know how that implementation works I'm almost scared to figure it out they're too slow to do things like SSR well so you ideally want to generate the assets ahead of time and just use Cloudflare as like a redirect again switch statement for getting data to and from things but the price is hilariously cheap on there dropped in chat that Cloudflare SSR sucked really hard for them it was three second time to first byte that means when the user makes a request it took three seconds for them to start seeing HTML because they were generating the HTML on the worker some amount of that is because the worker isn't necessarily close enough to the DB so if it's fetching data it's going to be slower but a lot of it's just that the CPUs in the amount of like throughput you're getting on the CPU on a worker is relatively low this is focused on running simple JavaScript scripts so yeah it's not the best thing for generating web pages but it's phenomenal for taking a request sending it somewhere else with an API key and streaming the result back to the user so checks and balances with all of these things Neon is similar in these ways where they built the thing focused on like junior devs and simple workloads that also is very beneficial for AI if an AI has to sit and wait for five to 15 minutes for your Postgres instance to spin up you have a terrible experience being able to spin up a database in seconds is cool as a junior getting started trying out the platform it's way cooler as an AI agent that needs to spin it up in order to work on a problem or do something like a preview environment so I can see the changes the AI made all three of these companies spent years building a thing that was useful to an extent but relatively niche in how valuable it could be and they were more parts of an ecosystem rather than a bigger solution that are now 10x more valuable and way bigger than they ever could have been because AI happens to work really well with the mental models that these tools recommended that's just a thing I've been noticing is there's a lot of these tools that made some sense before that make way more sense due to AI and Neon is one of them but their costs are just beyond unrealistic so there isn't really a path for them to raise a whole bunch of money and keep growing they're just going to keep losing money but if Databricks wants to play in this space because Databricks built a product that isn't ready for the AI space what they built doesn't have the weird benefits that these things do so they're acquiring one of those weird benefits they're acquiring one of the platforms that benefits greatly from the AI wave in order to have some of that benefit for themselves I think I covered everything I wanted to in this one the state of databases is fascinating right now I don't have direct recommendations I hope this helps you better understand where things are at what the benefits and negatives of each of these platforms are and hopefully you can avoid a disaster similar to Fauna and if we look at this as Neon protecting their users and their data knowing that raising more money was unrealistic and profitability was too awesome but if we look at this as the end of database startups things get a little bit scarier I loved the 2020s surge of these database providers trying to rethink how we operated with our data and how we thought about SQL I think that era might be over and I think the acquisition of Neon shows just how over it is let me know what you think am I overreacting here or is there really a change going on in this industry curious how y'all feel until next time keep querying I'm going to see you
