Baseten Doubled to $11B in 90 Days Selling "AWS for Inference." Here's When a Solo Builder Should Actually Rent It.
Baseten is in talks to raise $1 billion at an $11 billion valuation (more than double the $5 billion it carried about three months ago, when NVIDIA and others backed a $300 million round). The company's annualized revenue went from roughly $200 million at the start of Q1 to around $600 million by the end of it. The pitch underneath the numbers: inference will be two-thirds of all AI compute demand by year-end 2026, and the layer that serves it is the place to be.
Set the funding aside for a second, because the valuation isn't your problem. The useful question is the one Baseten's existence forces every builder to answer: when you ship something with an AI feature, where does the model actually run: somebody's API, a managed inference platform, or your own hardware?
What "AWS for inference" actually means
Baseten calls itself that for a reason. It rents NVIDIA GPU capacity and wraps it in tooling so you can deploy, customize, and operate models (mostly open-weight ones) without running the boxes yourself. Its customers include Notion, Cursor, Writer, and HeyGen: companies serving real production traffic that decided they wanted control over the model and the serving stack, but didn't want to be in the datacenter business.
That's the middle of a three-tier spectrum, and naming the tiers clearly is most of the value here.
The three tiers, and the crossover points
Tier 1: Call a frontier API. OpenAI, Anthropic, Google. You send a request, you get tokens, you pay per token. Zero infra. This is where every solo builder should start and where most should stay. The model is somebody else's problem; you buy capability by the call.
Tier 2: Managed inference. Baseten, and platforms like it. You bring a specific model (usually open-weight, sometimes fine-tuned) and they serve it on rented GPUs with autoscaling, monitoring, and a deploy story. You don't touch hardware, but you own the model choice and the serving config.
Tier 3: Self-host. Your own GPUs, cloud or physical. Maximum control, maximum cost, maximum operational burden. You're now running an inference service as part of your business whether you wanted to or not.
The crossover from Tier 1 to Tier 2 happens for one of three concrete reasons, not because managed inference sounds more "serious":
You need a specific open model the big APIs don't host: a fine-tune, a niche model, an Apache-licensed one you want to control. You're hitting volumes where frontier-API markup genuinely hurts, and a cheaper open model on rented GPUs pencils out even after the platform fee. Or you have latency, data-residency, or customization requirements the public APIs can't meet.
If none of those is true, you don't graduate to Tier 2. You stay on Tier 1 and you're right to.
The crossover from Tier 2 to Tier 3 is rarer still and almost always about either extreme scale economics or a hard compliance requirement that won't let your weights touch someone else's infrastructure. For a solo operator, Tier 3 is usually a mistake dressed up as control.
Tie it back to the self-host question
A few days ago the conversation was about NVIDIA's Nemotron 3 and when self-hosting an open model finally makes sense. Managed inference is the answer to "I want to run that open model but I don't want to run GPUs." It's the bridge between "just call an API" and "rack your own hardware." That's genuinely useful: it means the choice isn't binary, and you can get the control of an open model without the operational tax of Tier 3.
But a bridge is only worth crossing if you need to get to the other side.
The honest take
Most solo builders never need Tier 2, and the Baseten round will make some of them think they do. A $11B valuation and "inference is two-thirds of compute" creates a gravitational pull toward feeling like you should own your serving stack. Resist it until a real constraint pushes you there.
Here's the test I'd apply. If your AI feature works fine on a frontier API, your margins survive the per-token cost, and you're not blocked by model choice or data rules, you are done: calling the API is the correct architecture, not a beginner's compromise. Adding a managed inference layer before you have a reason is the same mistake as adopting Kubernetes for a site with forty users: infrastructure cosplay that costs you the time you should be spending on the product.
The builders who actually belong on Baseten know exactly why. They're running a specific open model at a volume where the math flipped, or they have a customization or latency need the public APIs can't serve. If you can't name your reason in one sentence, you're not there yet, and the round's size doesn't change that.
Where I'll grant the bull case: if inference really does become two-thirds of compute and open-weight models keep closing the gap with frontier ones, the Tier 1 → Tier 2 crossover will come earlier for more builders than it does today. The economics of serving your own open model get better every quarter. So this isn't "never": it's "not until you can say why," and for most one-person shops that day hasn't arrived.
Author
Lukas
@lukcombinator