“When will Baseten offer sandboxes” has been a common customer question. And we’ve taken a huge step in answering it: Baseten has acquired Blaxel. The nature of the median inference call has changed in the past year. Models stopped just answering questions; they act now. A coding agent writes a patch, runs the test suite, reads the failure, and tries again, dozens of times to close a single ticket. Every one of those loops needs somewhere to execute code. And the agent itself needs its own runtime. And so do RL rollouts. The list goes on. Running some code in a container isn’t hard. The hard problem is storage (state), networking (milliseconds-long cold starts), and security (VMs). That's what Blaxel has built: stateful, pause/resumable, VM-isolated sandboxes that scale to thousands of instances in seconds. Zooming out, tomorrow’s cloud isn’t one of inference, training, or sandboxes. It’s one where these 3 primitives come together cohesively to meet users where they are and to unlock value that’s greater than the sum of the parts. That’s been the north star that Baseten and Blaxel share (they were in fact going to build inference and training). I couldn’t be more excited to do it together with Paul Sinai and team. Welcome 💚 baseten.co/blog/blaxel-is-jo…

I'm excited that I can finally talk about this: the talented Parsed team is joining Baseten. They’re a team of 10 RL experts with a mix of research (Oxford and Cambridge PhDs) and industry backgrounds, with a track record of shipping state of the art models. And they share our mission: freeing companies from Big Token. There’s a world in which AI is done in 4-5 labs and the rest of us are token consumers, getting bigger and bluer plaques for it. We don’t want that to be the case and we’re doing something about it. The recipe is simple: Hands-on RL experts + Baseten’s training infrastructure + Baseten’s inference = owning the intelligence layer This recipe unlocks a few things for our customers: 1. Models that improve over time: more inference results in better models. And better models mean owning the intelligence layer and creating moats. 2. Massively improve economics: at scale, it’ll always be cheaper (5-10x) to pay for the underlying inference compute than to pay-per-token for closed models. 3. Control over uptime: 99.9% (3 nines) uptime isn’t achievable on closed models today, as their status pages admit. Owning the models means getting 99.99% (4 nines). 4. Control over latency: closed models are inherently optimized for high throughput and not low latency. And the market is moving towards low-latency use cases at a fast pace. The Parsed founders go into much more detail in their post. Leaving it in the comments. And please DM me if you have post-training needs.

We closed our series D at $2.1b. It happened 8 months after our series C, which seems too fast until you consider the facts: 2-years worth of growth in 8 months, virtually 0 customer churn, healthy margins, and QoQ NDR numbers that are considered top-tier YoY. The market demand for low-latency, elastic, and reliable inference in insatiable. And our early focus on building a devex to support custom models has truly paid off: people want power and visibility over how their models are optimized, deployed, and served; not a blackbox inference API. And on the latency front, companies that just a year ago said “latency isn’t a top priority” are working with us to shave off 10s of milliseconds because of the impact on both their UX and bottomline. The reason our latencies are unmatched is because we rejected the two extreme approaches: a) go all in on an oss lib like vllm, or b) build our own inference stack from the ground up. The promised land is in the middle: leverage the best of the different oss frameworks and contribute back, while building alpha on top of it (e.g. speculation engine, custom kernels, distributed kv storage, rewrite hot paths in rust, etc). Baseten is 6 years old today. Cheers to another 6. I'm sure it only gets harder from here.