Together AI’s $305M bet: Reasoning models like DeepSeek-R1 are increasing, not decreasing, GPU demand

When DeepSeek-r1 Initially, the prevailing fear, which prevents industry, can be achieved with less infrastructure, he said.

While leaving, this is not necessarily. At least, according to Together aiDeepsseek and open source substantiation increased the exact opposite effect: Infrastructure increases instead of reducing the need.

This increased demand helped to bring together AI’s platform and business. Today, the company led the general catalyst and announced a $ 305 million fund led by Riferity7. Together AI first appeared In 2023, in 2023, open source of large language models (LLS) to facilitate the enterprise. The company has expanded in 2024 Together the enterprise platformVirtual Private Cloud (VPC) and allows you to place AI in environments in places. In 2025, the AI ​​once again and again increases the platform again with the agency AI capabilities.

The company claims that the AI ​​placement platform has more than 450,000 recorded developers and the work has increased since more than since. The company’s customers include enterprises, as well as KREA AI, writings and AI beginnings Poke.

“Now we serve models between all the modals: language and thinking and videos and videos,” Vipul Prakash, “ViPul Prakash, the General Director General AI, Venturebeat said.

The great impact of DeepSseek-R1 has a need for AI infrastructure

The DeepSeek-R1 was the first time, for a number of reasons, for a number of reasons, a number of open source substantiates were established and the effects of a number of open sources could be placed with less infrastructure and placed in the property model.

However, Prakash explained, together with AI, DeepSeek-R1, partially grew up to increase workloads.

“This is a fairly expensive model to draw conclusions,” he said. “There are 671 billion parameters and you need to spread over multiple servers. Since quality is higher, the highest level usually needs more power.”

In addition, DeepSEek-R1 generally has longer has long-term desires that can last for up to two to three minutes. The very large user demand for DeepSseek-R1 manages more infrastructure needs.

In order to meet this demand, the AI, in the best performance, from 128 to 2,000 plugs from 2,000 plugs, was made from a service called “regulatory groups.”

How to use AI together to use instatives to organizations

There are a number of special areas used by AI’s justification models. These include:

• Coding agents: Providing models help you to separate bigger problems.
• Reduce hallucination: The cause of the cause helps to check the results of the models, so it reduces the hallucinion that is important for applications that matters.
• Improving models that are notors: Customers distillate and improve the quality of silly models.
• Opportunity to develop oneself: The use of reinforcement of strengthening with endurance models allows models to recreate the models again and again without relying on large amounts of man labeled data.

Agentic AI, AI infrastructure also manages the growing demand

Together, the AI ​​sees the growing infrastructure requirement for its users covering the EU agent.

Prakash explained the agent of the agent, which results in a user request to complete a user request, the AI’s infrastructure sets more calculation together.

Agent AI to support workloads, AI has recently acquired each other CodesdboxTechnology lives in light, fast loading virtual machines (VMS), and language models also live in language models to carry out an arbitrary, reliable code. This combines each other to reduce the delay between models that need to be called with AI agency code, to improve the performance of Agentic workflows.

Nvidia Blackwell already affects

All AI platforms face growing requirements.

NVIDIA is one of the reasons for spreading new silicons that provide more performance. The last product chip of NVIDIA is Blackwell GPU, which is now placed together.

Prakash, NVIDIA Blackwell plugs cost about 25% more than the previous generation, but provided 2x performance, he said. GB 200 platform with Blackwell platforms, especially suitable for the education and influence of the mixture of MOE) models of trained specialist (MOE) models of trained professionals (MOE). He noted that Blackwell chips are expected to be a greater performance to the result of larger models compared to smaller models.

The competitive landscape of the agent of AI

The market of AI infrastructure platforms is strong competition.

Together, the AI ​​competes competition from both the cloud providers and the beginning of AI infrastructure. All Hyperscalers, including Microsoft, AWS and Google, have AI platforms. There are a category of AI-focus players such as Grog and Samba Nova, all aims to a slice of a lucrative market.

Together, there is a complete assembly offer, including the GPU infrastructure, which has a GPU infrastructure, AI, the above program platform. This allows customers to easily establish open source models or develop their models on the AI ​​platform together. The company also focuses on research for both instability and training optimization, optimization and accelerating business times.

“For example, we serve the DeepSeek-R1 in 85 verses per second and serve in 7 verses in Azure seconds,” Prakash said. “There is a spacious expanding space in the performance and value we can submit to our customers.”