Aramco Just Put $800M Into the 'Neutral' AI Cloud You Were Going to Self-Host On. The Independence Layer Now Has a Landlord Too.
Aramco Just Put $800M Into the 'Neutral' AI Cloud You Were Going to Self-Host On. The Independence Layer Now Has a Landlord Too.
On July 1, Together AI raised $800M at an $8.3 billion post-money valuation, led by Aramco Ventures, with Nvidia, General Catalyst, and Vista Equity Partners in the round. That's up from a $3.3B valuation about sixteen months earlier. The company says annual bookings crossed roughly $1.15B last quarter, it serves customers including Cursor and Cognition, and it has locked in more than 500 megawatts of compute to fund something like 50x capacity growth over five years. The pitch is exactly what a lot of us have been repeating to each other: run open models (DeepSeek, MiniMax, Kimi) for less than you'd pay OpenAI or Anthropic for closed ones.
Here's the thing I keep coming back to. Together AI is the company plenty of solo operators point to when they say "I'm not locked into one lab; I've got a neutral open-model cloud as my fallback." That fallback just took $800M of sovereign money and announced plans to become a hyperscaler. The independence layer got a landlord.
The short-term read is genuinely good for you
Let me not be a cynic about the parts that help. More capacity, real money behind open-model hosting, and a well-funded competitor leaning on OpenAI and Anthropic pricing are all tailwinds for a one-person business renting inference. When a neocloud commits 500 megawatts and raises at $8.3B specifically to host open-weight models cheaply, the direct consequence is downward pressure on what you pay per token and more headroom before you hit a capacity wall.
The customer list tells you the product works at the level that matters. Cursor and Cognition are not hobby projects; if Together's hosting is good enough to sit under tools shipping to serious developers, it's good enough to sit under yours. And the open-model angle is the real prize: for two years the honest answer to "can I self-host something competitive?" was "not quite," and Together's whole business is that the answer flipped. Using them to run a capable open model at a fraction of closed-API cost is a legitimately good move today.
So if the story ended at "cheaper open-model inference with a big balance sheet behind it," I'd just tell you to go use it.
The part that should make you keep your abstraction
It doesn't end there, because "vendor-neutral" open-model hosting is still a vendor. That sentence sounds obvious written down, but it's the exact thing people forget when they pick a fallback and mentally file it under "safe."
Look at what this raise actually concentrates. A single neocloud, now carrying sovereign-fund money from Aramco Ventures, built on a deep Nvidia hardware dependency, with a 50x growth plan that only pencils out if the capital and the GPU supply keep flowing. That is a specific, named set of risks (geopolitical, financial, supply-chain) attached to one company. It happens to be the company you chose precisely to hedge against attaching your business to one company. The neutrality was the marketing; the concentration is the reality.
I'm not saying Together is fragile. I'm saying it's a bet, the same way OpenAI and Anthropic are bets, and treating it as the neutral ground beneath the bets is a category error. If your architecture assumes your fallback provider is somehow outside the risk model that made you want a fallback, you don't actually have a hedge. You have a second favorite.
What I'd actually do
Use the cheap open-model hosting. Just don't let it become the new thing you're locked into.
Concretely, the value isn't Together specifically: it's that model access flows through a seam you control. Keep your inference calls behind a thin provider abstraction so that "who serves this model" is a config value, not a hard-coded assumption threaded through your app. Wire up at least two paths: Together for the cheap open-model calls, and a second option that's a different neocloud or a self-hosted deployment of the same open weights. You don't have to run the fallback in production; you have to have tested that you can flip to it in an afternoon if Together's pricing, availability, or ownership changes in a way you don't like.
The reason this matters more now, not less, is that the whole appeal of open weights was that you could take them elsewhere. Together hosting DeepSeek or MiniMax doesn't remove that option (the weights are still open), but only if your architecture is built to exercise it. The company just raised $800M to make leaving frictionless to never think about. Keep the seam that lets you leave anyway.
The honest counter-take
I'll argue the other side, because I might be overcooking the risk. For the vast majority of solo operators, Together AI is going to be a boringly reliable, cheaper way to run open models for years, and all this hedging is insurance you never claim on. A better-funded provider is generally more stable, not less: the $800M makes an outage or a shutdown less likely in the near term, not more. If you spend a week building elaborate multi-provider failover for a one-person side project doing modest volume, you've optimized for a risk that will probably never fire while your competitor shipped features.
That's fair. So calibrate to your stakes: if inference is incidental to your product, just use Together and move on. But if your business genuinely depends on model access (if an availability change or a pricing reset would actually hurt you), then the cost of keeping one tested alternate path is small, and the comfort of "vendor-neutral" is not a substitute for it. The lesson isn't distrust Together. It's that abstraction beats loyalty, even loyalty to the provider whose entire brand is that you don't need to be loyal to anyone.
Author
Lukas
@lukcombinator