Join our daily and weekly newsletters for the latest updates and the exclusive content on AI coverage. Learn more
When Deepseek-R1 For the first time, the dominant fear that rocked the industry was that advanced reasoning could be made with less infrastructure.
It turns out that this is not necessarily the case. At least, according to AI setThe rise in deep and open-source reasoning had the exact opposite effect: instead of reducing the need for infrastructure, it increases it.
This increased demand has contributed to fueling the growth of the AI platform and affairs together. Today, the company has announced a series of financing for the series B of $ 305 million, led by General Catalyst and co-directed by Prosperity7. Together, AI is emerged first In 2023 in order to simplify the use of the company of the models of large open source language (LLMS). The company extended in 2024 with the Set of the corporate platformwhich allows the deployment of AI in virtual private environments (VPC) and on site. In 2025, together, AI again increased its platform with reasoning clusters and agental AI capacities.
The company says that its AI deployment platform has more than 450,000 recorded developers and that the company has increased in total in annual shift. Company customers include Pika Labs.
“We now serve models in all the modalities: language and reasoning and images and audio and video,” said Vipul Prakash, CEO of Together Ai, in Venturebeat.
The huge Deepseek-R1 impact has the IA infrastructure demand
Deepseek -R1 was extremely disruptive when she made her debut for the first time, for a certain number of reasons – one of which was the involvement that an open source model of cutting -edge source could be built and deployed with less infrastructure than an owner model.
However, Prakash explained, together, the AI increased its infrastructure in part to help support the increase in the demand for workloads linked to Deepseek-R1.
“It is a model quite expensive to execute inference,” he said. “It has 671 billion parameters and you have to distribute it on several servers. And because the quality is higher, there is generally more demand on the upper end, which means that you need more capacity. »»
In addition, he noted that Deepseek-R1 usually has longer term requests that can last two to three minutes. Huge demand from users for Deepseek-R1 more stimulates the need for more infrastructure.
To meet this request, together, AI has deployed a service which it calls “reasoning clusters” which offer a dedicated capacity, ranging from 128 to 2,000 chips, to perform models to the best possible performance.
How together AI helps organizations to use AI Reasoning
There are a number of specific areas where together AI notes the use of reasoning models. These include:
- Coding agents: Reasoning models help break more important steps.
- Hallucinations reduction: The reasoning process helps to check the outputs of the models, thus reducing hallucinations, which is important for applications where precision is crucial.
- Improvement of non -referring models: Customers distill and improve the quality of unrecoverated models.
- Activate self-improvement: The use of learning to strengthen with reasoning models allows models to impregnate recursively without relying on large amounts of data marked by humans.
The agency AI also stimulates increased demand for IA infrastructure
Together, the AI also notes an increase in infrastructure demand while its users adopt an agency AI.
Prakash explained that agent workflows, when a single user request translates into thousands of API calls to perform a task, grants more calculation on the AI infrastructure.
To help take care of AI agency workloads, together, AI has recently acquired Box of codeswhose technology provides light and rapid virtual machines (VM) to execute arbitrary secure code in the AI cloud together, where language models also reside. This allows AI together to reduce latency between the agent code and the models that must be called, improving the performance of agent workflows.
Nvidia Blackwell already has an impact
All AI platforms are faced with increased requests.
This is one of the reasons why Nvidia continues to deploy a new silicon that offers more performance. Nvidia’s latest product chip is the Blackwell GPU, which is now deployed in Ensemble AI.
Prakash said NVIDIA Blackwell fleas were about 25% more than the previous generation, but provide performance 2x. The GB 200 GB platform with Blackwell Chips is particularly well suited to the training and inference of experts’ mixing models (MOE), which are formed on several servers connected to infiniband. He noted that Blackwell fleas should also provide a greater increase in performance for the inference of larger models, compared to smaller models.
The competitive landscape of agentic AI
The IA infrastructure platform market is fiercely competitive.
Together, AI faces competition from established cloud suppliers and AI infrastructure startups. All hyperscalers, including Microsoft, AWS and Google, have AI platforms. There is also an emerging category of players focused on AI such as Groq and Samba Nova which all aim for a lucrative market.
Together, AI has a complete offer, including the GPU infrastructure with layers of software platform on the top. This allows customers to easily build with open source models or develop their own models on the AI platform together. The company also focuses on research in optimization development and accelerating execution times for inference and training.
“For example, we serve the Deepseek-R1 model to 85 tokens per second and Azure serves 7 tokens per second,” said Prakash. “There is a fairly enlarged gap in the performance and costs that we can provide to our customers.”
