Replicate is joining Cloudflare
replicate.com
replicate.com
I think this is Cloudflare's most notable acquisition yet? From Wikipedia it looks like they've previously mainly acquired smaller cybersecurity firms like https://en.wikipedia.org/wiki/Area_1_Security
If they raised $500m... then yes, the wouldn't have done so well.
Definitely exciting to see.
Really easy to do it on the edge. No hassle with POP3, IMAP, ... Easy to use a subdomain of my email domain
Loc: ~40
Would be interested in seeing an Outlook/(o365/Exchange) alternative from Cloudflare with a bit more of a feature rich offering... or at least the ability to build something like that on CF.
This is something that you can or could already do with Envoy proxy with the right amount of polish, knowhow, and elbow grease (which we have).
What Cloudflare doesn’t have are good general computing products and multitenancy supprt. But we have those, something like D1, and something like a remote tunnel that runs in your browser. So users can launch Stalwart instances painlessly and use our secret magic to render it in actual browsers without special middleware or extra state (or we could offer shared/default Stalwart to do it completely out of the box). So that would take you the rest of the way from platform-level email to application-level and user-level/managed email.
What we don’t have is the time to spare in productizing that and handling the email-specific routing in envoy, or finding capable/knowledgeable people who know all the email-specific content and skills. So hit me up if that's you, and otherwise, feel free to run wild with the knowledge that you can configure Envoy Proxy with an L4 network filter or HTTP filter that delegates to dynamically loadable/configurable wasm if you want a hackable Cloudflare workers or even a platform-level alternative (hint: store wasm code in FUSE). The L4 filter should work for email filtering.
[0] https://www.envoyproxy.io/docs/envoy/latest/intro/arch_overv...
[1] https://github.com/proxy-wasm/spec/tree/main/abi-versions/v0... (where you'd start implementing email filtering logic)
[2] https://mail.mplode.dev (demo/test Stalwart instance we're running in our dev environment. Platform magic would allow us to render an email client directly embedded in https://brilliant.mplode.dev via remote IMAP/POP3)
And doing it their way than the traditional way.
Note: Their innovation seems to lie in smaller and fine-tuned ( broader catalog) models than larger ones. So Replicate seems a perfect match.
The traditional way: Rent a GPU and run inference on an container.
* BastionZero
* Kivera
* Baselime
* PartyKit
* Area 1
* Vectrix
* Zaraz
* Linc
* S2 Systems Corporation
* Neumob
* Eager
* CryptoSeal
* StopTheHacker
Replicate has had multiple ways to deploy for auto scaling and you can just keep running periodically to keep the system in a booted and warm state, but that has always seemed like it would be too expensive for a broke bootstrapper like me so I avoided it and model popularity was a big deciding factor. Also because of that and the potential for boot up, in general I avoided it for latency-sensitive things.
I guess there is a limit to what you can do. At some point someone has to spend the money to have the resources stay ready.
But with Cloudflare, theoretically the pool of potential users goes up, and it becomes more likely for someone to have already booted your model.
At the moment I am especially interested in performant and easy ways to run models like "sensefvg/InteractiveOmni-8B" or Qwen 2.5 Omni or models that are even more all in one than that like OpenAI Realtime or Gemini Live.
Now that Ernie 5 launched with (Omni) multimodality built in, I think within six months, developers are going to start to expect speech-to-speech capability from major AI lab releases or product line ups. I feel like eventually the spatial-temporal understanding of video models will be merged in too to make the models understand the world better. But speech in and speech out is closer to being a standard expectation.
Instead of running three models for STT->LLM->TTS with a bunch of tricks like eager end of turn or speculative decoding that basically mean you run the LLM twice or on two different models, and possibly getting shut down by API rate limits, the speech to speech models are a single model that both understands and generates audio as well as text such as for function calls.
This is probably an annoying comment because I am immediately trying to increase the requirements to not only being every model for cheap, but every model for cheap in in a low latency real time streaming way. I just happen to have a contract now that has shown me that multimodal like voice to voice is much more convenient but also much more expensive and fewer options.
Replicate has been so awesome though. Within like a day of me requesting InteractiveOmni, lucataco had it up. So another annoying comment, I sure hope he got paid.
They can't innovate themselves, and rather than try to fix the reasons why (change in leadership, corporate structure, etc), they just buy a competitor, which they will most likely run into the ground. Perfectly reasonable!
Overall I think it's a good acquisition for both of them. Replicate would have to have more competitive pricing to compete with fal.ai
How many players are there in this space? Replicate, RunPod, Modal, Northflank, FAL, ... Who are the big ones? It's pretty crowded, right?
FAL was smart. They ditched the "run any model" to focus on just image and video, and now they dominate that space. They raised a pretty substantial round recently. Though I don't think there's any moat and they'll soon face competition too.
What about these vs. the routers like "Open"Router?
Cloudflare isn't solely a CDN anymore. CDN and DDOS-protection were the most logical "first" products to build based on their SDN ( Software Defined Networking).
A cloud is the next thing and there's a lot of money involved with the cloud. I see them as the only real competitor/challenger to Azure, AWS, GCE, ... because they aren't bound to regions ( less DevOps)
For example, what you might not know about Durable Objects => https://boristane.com/blog/what-are-cloudflare-durable-objec...
(https://substackcdn.com/image/fetch/$s_!-PwA!,f_auto,q_auto:...)
Revenue:
$ 85 million (2016)
$ 287 million (2019) IPO year
$ 1,670 million (2024)
$ 2,154 million (2025)
https://www.macrotrends.net/stocks/charts/NET/cloudflare/rev...
That's not a crazy valuation multiple considering their growth.
Palentir is a body shop. Cloudflare is infrastructure/cloud. Very big difference ( at least to me)
Stock ( Cloudflare - https://finance.yahoo.com/quote/NET/holders/ ): 90.87% % of Shares Held by Institutions
Stock ( Palantir - https://finance.yahoo.com/quote/PLTR/holders/): 60.09% % of Shares Held by Institutions
Defense against threats is a pretty strong centralization incentive in different kinds of networks - social, biological.
I could imagine that a lot of people are investing based on similar scenarios in their minds.
This was always their architecture, if you watched closely.
My predication 5 years ago - https://news.ycombinator.com/item?id=26821438
> And cloudflare is self hosting it on more edge locations ( i think they could even join the big 3 soon)
I feel like people like to rub that one time cloudflare messed up when I mention it but it was a gambling website and I feel like cloudflare could've better communicated it but overall its got so much less drama than the other cloud providers and its genuinely being really nice imo
But imo, cloudflare is really dirt cheap for just starting out and at scale as well especially if using cf workers
I feel like cloudflare can make a bank in enterprice section but their pricing model also feels the most saner compared to the shady tactics used by google or others with our marketing and privacy
I know that the internet is getting centralized but I feel like there are some ways of de-centralizing it, (by example archiving web pages and then seeding them, helping on internet archive or something similar as well)
As an internet user, cf feels mid but sometimes as a guy who just wants to deploy shit or basic apis, I "vibe coded" a cloudflare worker api which I actively use so much for my own purposes setting up a custom redirector and everything without paying anything at all, I think I like it.
Honestly, nothing is as good or as bad as it seems except palantir's evaluation which makes me feel like 448 pe ratio or something similar drove me nuts the other day.
Honestly now the only stable provider which doesn't have an outage is google cloud in sick twisted fate
I am wishing when google cloud has its outage so I can recommend everybody to use hetzner (yes I know its not an apples to orange comparisons but hetzner has some crazy good uptime)
What's important to know ( I think). Recently, Cloudflare released a blog post of Omni for AI inference. I think they performance tuned it better than other providers. So their costs per inference drops down a lot ( https://blog.cloudflare.com/how-cloudflare-runs-more-ai-mode... ). Since the performance is OK, they now want to expand usage and their model catalog.
Replicate is a perfect fit. Model catalog, infrastructure for larger models, more specialised tools for fine-tuning, ...
Eg. For inference, Replicate is basically just a Worker AI endpoint and easy to maintain. Fine-tuning could probably be something similar.
But then again, that's my 2 cents. It was already mentioned that Replicate will stay as a distinct brand.
To name a few but on hwy 101 today there were billboards for more
Uh huh
From what I know they were the most used inference provider by developers a few years ago, but since the Together AI and Fireworks only grew while Replicate seems to have stayed quiet. It’s a highly competitive low-margin business, so volume is critical and if you’re losing volume then you’re doomed.
Given the price was not announced, it seems investors decided to exit via acquisition instead of trying to raise a multi hundred million round consistent with a multi billion dollar valuation and a high level of ambition that investors are shooting for in this space.
If the amount was high, they'd be be bragging about it. Given that they aren't, it might be on the lower end of the 2023 valuation for the then 40M round.
Just speculating here; I don't have any more information. Basing this on my understanding of how this stuff works. This might actually be the opening round for a few more such acquisitions of the somewhat risky investments of a few years ago of companies that are probably not going to turn into trillion $ unicorns. There are lots of pretty well funded startups in this space converting investment capital into cloud GPU cost. I think some level of consolidation is overdue and might take away some building concern about over exposure in the market. Big banks and investors might be getting nervous.
Every acquisition announcement says something like this. It is almost never true. In fact, the opposite often happens.