My disappointment related to this is that I still think Snowpack's philosophy was the right one for JS present and especially for JS future: don't bundle for development at all because ESM support in today's browsers is great and you can't get better HMR than "nothing bundled"; don't bundle for Production unless you have to (and have performance data to back it up).

I know Vite inherited the first part, but I still mostly disagree with Vite on switching the default of that last part: it always bundles Production builds and switching that off is tough. Bundling is already starting to feel like YAGNI for small-to-medium sized websites and web applications between modern browser ESM support, modern browser "ESM preload scanners", and ancient browser caching behaviors, even without HTTP/2 and HTTP/3 further reducing connection overhead to almost none.

But a lot of projects haven't noticed yet because the webpack default, the Create-React-App default, the Vite default, etc is still "always bundle production builds" and it isn't yet as obvious that you may not need it and we could maybe move on away from bundlers again on the web.

Vite is driven by pragmatism. They are bundling for prod because it is pragmatic.

If they do not bundle for prod, then initial load will cause a flood of network request and cause slowness due to a lot of network requests

I think a lot of people assume this is the case for every website/web application without actually watching their network requests and having hard data on it. Like I said, small to medium sites and applications in some cases start to be more efficient with ESM modules than ever before (because it is the ultimate code-splitting) even before you get into new tools like HTTP2 and HTTP3 which significantly reduce connection overhead for network requests entirely and make a lot of old bundling advice outdated at best.

Plus, ESM modules always load asynchronously without DOM blocking so in any and all "progressive enhanced" sites and apps, it may feel slow to hydrate parts of it, but you generally aren't going to find a better performing experience while it hydrates. The page itself won't feel "slow" to the user because they can start reading/scrolling faster.

Some of it comes down to your framework, of course, and if you can afford SSR or are using a "progressive enhanced" style of site/application. For one specific counter-instance I'm aware of, Angular's NgModules dependency injection configuration mess routes around ESM's module loaders and creates giant spaghetti balls the browser thinks it needs to load all at once. I can't see recommending unbundling Angular apps any time soon as long as they continue to break the useful parts of ESM style loading with such spaghetti balls. Other frameworks, especially ones with good SSR support/progressive enhancement approaches are increasingly fine to start using unbundled ESM on the modern web.

I thought http2 made the network requests a non-issue (at least for 100 files or so). The only advantage of the larger bundle is though compression.

Even compression is maybe somewhat a non-issue with Brotli compression options (now supported by every major browser and often "required" for strong HTTP2+ support) and its hand-tuned dictionary for things like common JS delimiters/sequences. With Gzip you do sometimes need large bundles before its automatic dictionary management algorithms pick up things like that. In theory, Brotli always starts from a place of more "domain knowledge" and needs less "bulk" for good compression.

Have you ever tried to load a development build of a project with 3000 modules in Vite? It’s like a webpack build all over again every time you load the page.

I didn't have much of any trouble in Snowpack with even the largest projects I worked on but at least with Snowpack it was easy enough to target "freeze" a sub-tree of modules into Snowpack's "web_modules" with esbuild. I don't know about Vite in that case as I haven't been convinced to actually try Vite despite Snowpack's shutdown endorsement. Recent stuff for me has either been projects still using webpack for platform reasons or ESM with hand-targeted esbuild bundles for specific module trees (libraries) without using Vite or Snowpack above that. About the only thing I feel I'm missing in those esbuild workflows is a good HMR dev server but the simplicity of using fewer, more targeted tools feels handy to me.

The folks working on Snowpack didn't just give up. They transitioned over to working on Astro. It was a very positive move. I see great things coming out of that project as we move into a more "server rendered" future.