Together bet of AIS $ 305 million: Models such as deepseek-r1 increase and do not lose weight, GPU demand

When Deepseek-R1 First, the prevailing fear emerged that shook the industry that advanced reasoning could be achieved with less infrastructure.

As it turns out, this is not necessarily the case. At least according to information Together aiThe rise of deepseek and open source argument has had the exact opposite effect: instead of reducing the need for an infrastructure, it increases it.

This increased demand has contributed to promoting the growth of the platform and the AI’s business of AI. Today, the company announced a funding round of $ 305 million, which was directed by General Catalyst and is shared by Prosperity7. Ai first appeared 2023 with the aim of simplifying the use of open source large language models (LLMS). The company expanded in 2024 with the Together Enterprise platformWhich enables AI support in virtual private cloud (VPC) and local environments. In 2025, the AI ​​together with argumentation cluster and Agentic AI skills grows their platform again.

The company claims that its AI deployment platform has more than 450,000 registered developers and that the company has risen 6 times over the course of the year. The company’s customers include companies and AI startups such as KREA AI, Capties and Poke good.

“We now serve models in all modalities: language and argument and pictures as well as audio and video,” said Vipul Prakash, CEO of Together AI, to Venturebeat.

The enormous influence Deepseek-R1 has the demand for the AI ​​infrastructure

Deepseek-R1 was considerably disturbing for several reasons for the first debate. One of them was the implication that an open source argumentation model for open source argumentation with less infrastructure could be built up and used as a proprietary model.

However, Prakash said that the AI ​​together has partially expanded its infrastructure to support the increasing demand for workloads with deepseek-r1.

“It is a pretty expensive model that you ran on,” he said. “It has 671 billion parameters and you have to distribute it over several servers. And because the quality is higher, there is generally more demand at the top, which means that you need more capacity. “

In addition, he found that Deepseek-R1 generally has longer inquiries that can take two to three minutes. The enormous demand for Deepseek-R1 continues to lead to more infrastructure.

In order to satisfy this demand, Ai has introduced a service, which he refers to as a “argumentation cluster”, which provide dedicated capacities in the range of 128 to 2,000 chips to carry out models with the best possible performance.

How Together AI helps companies to use AI Argumenting Ai

There are a number of specific areas in which the AI ​​together sees the use of argumentation models. This includes:

• Coding agents: Models of argument help divide major problems into steps.
• Reduction of hallucinations: The argumentation process helps to check the expenditure of models and thus reduce hallucinations, which is important for applications in which the accuracy is of crucial importance.
• Improvement of non-boundary models: Customers distill and improve the quality of non -existent models.
• Enable self -improvement: The use of reinforcement learning with argumentation models enables models to improve themselves recursively without relying on large amounts of human -marked data.

The Agent -KI is also increasing to increased demand for AI infrastructure

Together, AI also sees an increased infrastructure demand because its users accept the Agent -Ki.

Prakash said that agents workflows, in which a single user requirement leads to thousands of API calls to do a task, set more calculation according to the infrastructure of AI from AI.

To support Agentic Ai Workoads, AI recently acquired together Codes sandboxwhose technology offers slight, quickly bending virtual machines (VMS) to carry out any, safe code in the Ai cloud together, in which the voice models are also located. This enables Ai to reduce the latency between the agent code and the models called, which improves the performance of agents workflows.

Nvidia Blackwell already has an impact

All AI platforms are faced with increased requirements.

This is one of the reasons why Nvidia keeps performing a new silicon that delivers more performance. The latest NVIDIA product chip is the Blackwell GPU, which is now being used at Together AI.

According to Prakash, Nvidia Blackwell chips cost around 25% more than the previous generation, but 2x the performance. The GB 200 platform with Blackwell chips is particularly suitable for training and the conclusion of the mixture of expert models (MEE), which are trained on several infiniband-connected servers. He noted that Blackwell chips are expected to offer a larger performance climber for the inference of larger models compared to smaller models.

The competitive landscape of the agents -KI

The market for AI infrastructure platforms is very competitive.

Together, AI competes both by established cloud providers and through AI infrastructure -startups. All hyperskallers, including Microsoft, AWS and Google, have AI platforms. There is also an emerging category of AI-focused players such as Groq and Samba Nova, all of which strive for a slice of the lucrative market.

Together, AI has a full stack offer, including the GPU infrastructure with software platform layers above. In this way, customers can easily build up open source models or develop their own models on the composite AI platform. The company also focuses on research work to develop optimizations and to accelerate the terms for both inference and training.

“For example, we serve the Deepseek-R1 model with 85 tokens per second and Azure serves it with 7 tokens per second,” said Prakash. “The performance and costs that we can provide to our customers is a fairly expanded gap.”