Replicate is joining Cloudflare

simonw 182 days ago |

Congrats Ben and team!

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

michaelbuckbee 182 days ago | |

This feels the most like Cloudflare branching out to something closer to a traditional hosting model and further from their networking efforts (though things like workers and R2 certainly have blurred that line).

tracker1 182 days ago | | |

I'd love to work on Cloudflare's underlying email platform... from pretty early on, I've felt like the workers model could be great for handling relatively short lived tcp connections, not just web requests... in this way, all the other facilities CF offers such as R2 and D1 could be used with an email platform that could scale incredibly well. With this acquisition, it's another natural fit into such a platform for email spam filtering and classification.

Definitely exciting to see.

NicoJuicy 182 days ago | | |

Recently implemented it to only process emails where the sender is in the list of customers email and then extract info from email and do an api call.

Really easy to do it on the edge. No hassle with POP3, IMAP, ... Easy to use a subdomain of my email domain

Loc: ~40

tracker1 182 days ago | | |

Did a CF worker for an email group list which worked well enough... having each email address target have to approve themselves was about the only hiccup. Definitely nice options all around.

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.

weitendorf 182 days ago | | |

If you (or others here) are truly interested in email+tcp proxying+workers-like-computing, then I have work to offer you in the build out of my company’s network/computing platform.

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)

NicoJuicy 181 days ago | | |

It feels like more related to this: https://blog.cloudflare.com/how-cloudflare-runs-more-ai-mode...

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.

ramimac 182 days ago | |

Probably, but you can check out a more robust list here: https://blog.cloudflare.com/tag/acquisitions/

* BastionZero

* Kivera

* Baselime

* PartyKit

* Area 1

* Vectrix

* Zaraz

* Linc

* S2 Systems Corporation

* Neumob

* Eager

* CryptoSeal

* StopTheHacker

captn3m0 182 days ago | | |

Not everything seems to be tagged as acquisition. Dyte.io was acquired, and announced here: https://blog.cloudflare.com/introducing-cloudflare-realtime-...

echelon 182 days ago | |

I don't know that this was an acquisition in the sense that the Replicate investors and team made bank. I don't see a price tag, and the market for these "run model" infra companies is pretty crowded.

mritchie712 182 days ago | | |

replicate "only" raised $50m, so I'd wager the founders, investors and early employees did well here.

If they raised $500m... then yes, the wouldn't have done so well.

dagi3d 181 days ago | | |

for investors, it will highly depend on the valuation ($350m) and liquidation preferences. so, even if they "only" raised $50m, seeing almost or none return is plausible. this also includes employees who exercised their options. another story is for the founders

ilaksh 182 days ago |

From an AI integration perspective, I am hopeful that Cloudflare may be able to improve "performance on the cheap" for replicate's models a little bit.

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.

kfarr 181 days ago |

As a Replicate customer, I’m excited about this news. Cloudflare has the strategic incentive to invest in this line of business and it puts Replicate in the hands of a very good product marketing and sales team to get the volume it needs for sustainability.

pm90 182 days ago |

I was wondering if/when cloudflare would jump into doing “more AI”. Of all the paths they could have taken, this seems like a reasonable one. Hope that the dev team made some money/cloudflare stock from this transaction.

pzo 182 days ago | |

They already did and they had Worker AI for quite a while [0]. But I think it didn't catch up and they have there very old models.

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

[0] https://developers.cloudflare.com/workers-ai/models/

aitchnyu 181 days ago | | |

Is this the platform they are offering on Openrouter?

https://openrouter.ai/provider/cloudflare

pzo 181 days ago | | |

not sure but that's quite possible. I haven't used cloudflare workers ai myself because they had quite limited number of models and choices so far.

greenie_beans 181 days ago | |

you've not been paying attention if you think this is their first foray

bogwog 182 days ago | |

> Of all the paths they could have taken

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!

pm90 181 days ago | | |

Their area of expertise is in running large scale infra, which they have innovated a lot to get. They’re acquiring a company that has expertise in an area they're not familiar with, which needs large scale compute infra. It seems like their needs are aligned in this case.

coleca 182 days ago |

Great match up of technologies. Congrats to Ben and the team! I had the privilege of working with Replicate when I was at AWS, such a great team.

ProfessorZoom 181 days ago |

>cloudflare including replicate goes down the very next day

rvz 182 days ago |

A very smart move for Cloudflare and a huge missed opportunity for Render [0].

[0] https://news.ycombinator.com/item?id=42833414

echelon 182 days ago |

Acquisition or acquihire? I don't see a price tag, which seems to imply the latter?

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?

resiros 182 days ago | |

Looks more of an aquisition for me. They are integrating the 50000 models into worker AI.

datadrivenangel 182 days ago |

Replicate is my preferred go to for image models and easy proof of concepts for language models, so congratulations to them!

pancakemouse 182 days ago |

I wonder if they'll continue to support and develop Cog [0], which I've found quite useful for local deployments of models (as well as to Replicate's servers).

[0] https://github.com/replicate/cog

doctorpangloss 182 days ago | |

The one time I’ve interacted with cog, predictably, it delivered none of its benefits compared to a vanilla pyproject.toml, and all of the downsides of being, ironically, not replicable.

jumski 182 days ago |

Very smart move - I assume the Workers will get tight integration with the Replicate APIs

jonplackett 182 days ago |

I really like both these companies. Replicate just make it so easy to bundle a bunch of different AIs up into a prototype - but I think they have tonnes of untapped potential that Cloudflare could fulfil.

redwood 182 days ago |

Cloudflare is one of those companies similar to palantir where the valuation just makes you scratch your head. But bringing in AI to try and fill out the AI valuation makes some sense I suppose

JackuB 181 days ago |

I spent a moment wondering why did Cloudflare buy an on-prem tooling startup https://www.replicated.com

jay-kor 180 days ago | |

They didn't.

https://www.replicated.com -> https://replicate.com

bobnamob 181 days ago | |

The cloud is just someone else’s on-prem

antipaul 181 days ago |

Are there indicators in Cloudflare's culture or history to suggest that Replicate's strengths (docs, api, design) will remain in the long-term?

NicoJuicy 181 days ago | |

I don't think that much will change regarding to Replicate's API because of Cloudflare. It's specifically mentioned in Replicate's blog post and it's also not in their best interest.

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.

up2isomorphism 180 days ago |

Workers is a very expensive computing facility. Yeah, it is cheaper than AWS lambda but people don’t use that for everything.

tasn 182 days ago |

Congrats Ben and team. Looking forward to seeing what you build with Cloudflare's backing!

derelicta 182 days ago |

I love witnessing companies getting acquired to form monopolies.

ectospheno 181 days ago | |

Replicate was far from a monopoly and cloudflare was arguably worse off in this space. Together they still aren’t anywhere near monopoly status for running models.

flashblaze 181 days ago | | |

Who are the other ones? I'm aware of fal apart from Replicate

redwood 181 days ago | | |

Together Fireworks Hugging face Modal AWS Azure Google CoreWeave

To name a few but on hwy 101 today there were billboards for more

tvararu 181 days ago |

Great move. Congratulations to the Replicate team!

philipwhiuk 182 days ago |

> Replicate’s going to carry on as a distinct brand

Uh huh

baggachipz 182 days ago | |

Well, yeah, for a year or maybe two.

aryanchaurasia 182 days ago |

why sell if they were doing good?

jillesvangurp 182 days ago | |

Probably because their last funding round was in 2023 and they may have been looking for another round because their revenue and cost were probably not really adding up to the valuation needed for another round. Running GPU hardware at scale is just expensive

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.

aryanchaurasia 181 days ago | | |

thanks; excellent analysis and makes much more sense to me now!

gk1 182 days ago | |

You don’t know how good they were really doing and you don’t know how good the offer was.

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.

thundergolfer 181 days ago | | |

My perception (engineer at modal.com) is that Replicate lost out to Fal.ai more than they lost to Together AI and Fireworks.

ilaksh 182 days ago | |

Don't pretend that 98% of startups don't actually dream of an exit or payday.

greatgib 181 days ago |

Now you can see the great plan going on... First, positioning itself as the "gatekeeper" of AI services, then owning such a service to be able to abuse of your position...

fullofdev 181 days ago |

wowwww

thayne 182 days ago |

> all that’ll happen is that it’s going to get way better

Every acquisition announcement says something like this. It is almost never true. In fact, the opposite often happens.

tonyhart7 182 days ago | |

its better for founder because they can have their exit payout