58.1 billion, an AI unicorn founded by a Peking University alumnus has emerged

The wealth-creation myth of the AI computing power "water sellers" continues unabated.

Recently, AI cloud platform Together AI announced the completion of a $800 million Series C financing round, sending its post-money valuation soaring to $8.3 billion (approximately 58.1 billion RMB). This round was led by Aramco Ventures, a subsidiary of Saudi Aramco, with star institutions including NVIDIA, Vista Equity Partners, and General Catalyst participating as follow-on investors.

Against the backdrop of fierce competition in the large model space, why has this AI cloud platform service provider, founded just four years ago, managed to attract a stampede of industry giants to back it? The answer likely lies in two core keywords: "open-source models" and a "Chinese technical dream team".

Valuation Surges 150% in a Year, Open-Source AI Becomes a Magnetic Capital Draw

Together AI was founded in 2022, emerging almost concurrently with ChatGPT. From its inception, the company clearly defined its positioning: to become an "AI-native cloud" enterprise, with its core business focused on leasing computing power based on NVIDIA GPU clusters to enterprises, enabling developers to run open-source large models at extremely low costs.

In its early stages of development, closed-source models led by OpenAI once dominated the market. However, the industry soon recognized a harsh reality facing enterprise-level AI deployments: the exorbitant token costs of cutting-edge closed-source models were relentlessly eroding the profit margins of AI applications. This landscape is now shifting dramatically with the powerful rise of high-performance open-source models such as DeepSeek.

During the industry's transition from "closed-source" to "open-source", Together AI directly addressed this pain point. By integrating mainstream open-source models including DeepSeek, MiniMax, and Kimi, it provides enterprise clients with diverse services such as serverless inference, dedicated infrastructure, and batch inference, emerging as the biggest beneficiary of the ongoing boom in the open-source model ecosystem.

Leveraging its well-developed open-source model ecosystem layout, Together AI has created a highly cost-effective "affordable alternative" solution for open-source models. Data shows that with equivalent or even superior performance, enterprises can reduce their inference costs by 6x to 60x after accessing Together AI's computing power services.

Take its AI customer service platform Decagon as an example: after migrating all workloads to Together AI, inference costs were directly reduced by 6x. This formidable cost advantage translates directly into tangible performance orders: the latest annual booking figures show that Together AI's order volume has surged past $1.15 billion compared to the previous quarter, skyrocketing from $30 million at the beginning of 2024 to the current $1.15 billion, marking a 38-fold increase.

The 38-fold surge in annual bookings over two years has fully unlocked Together AI's commercial potential in the open-source AI space. First, with the explosion of AI Agents, enterprise AI no longer requires simple Q&A chatbots, but digital employees capable of handling complex workflows, which has generated massive volumes of token calls. Second, Together AI itself has broken through the physical bottlenecks of computing power costs through underlying technological innovations.

Through its self-developed ATLAS inference engine and cutting-edge technologies such as Speculative Decoding, the company has accelerated some inference workloads by up to 400%. This end-to-end technological optimization allows Together AI to deliver enterprise-grade inference services at costs far lower than traditional cloud vendors, building a robust technological moat.

As a result, Together AI's valuation has continued to climb. In February 2025, during its Series B financing round led by General Catalyst, Together AI was valued at just $3.3 billion. However, 16 months later, its post-money valuation reached $8.3 billion, representing a 150% increase within a single year.

Peking University Alumni and Stanford Professor at the Helm, a "Chinese Technical Dream Team" Emerges

Together AI's rapid rise is not only riding the industrial wave of open-source AI, but also backed by an emerging "core Chinese technical team".

A glance at Together AI's official website reveals that its founding team can be described as a "scholarly dream team", with Chinese members making up a significant proportion.

First, as one of the core representatives of this technical dream team, Ce Zhang, co-founder and CTO, hails from Peking University. His resume demonstrates a solid top-tier academic background: he graduated from Peking University's School of Mathematics with a bachelor's degree in 2008, subsequently earned his PhD from the University of Wisconsin-Madison, and served as an assistant professor in the Department of Computer Science at ETH Zurich.

Throughout his research, he has consistently focused on making machine learning cheaper, more reliable, and more accessible to a broader audience. He believes that AI should not be an exclusive tool reserved for giants; instead, innovations at the system layer, scheduling layer, and distributed architecture can drive down the computing power costs for training and inference, enabling individuals, startups, and traditional industries to access open-source large models at low cost.

Guided by this core philosophy, Ce Zhang led Together AI's innovations in underlying computing power scheduling and inference engines, leading the team to develop the in-house ATLAS inference engine and deeply integrate advanced distributed inference technologies such as NVIDIA Dynamo. This series of innovations has empowered Together AI to achieve context-aware intelligent routing and KV cache management, effectively eliminating redundant computation issues in large-scale computing power deployments, and ultimately delivering the impressive commercial result of annual bookings exceeding $1.15 billion.

Second, beyond Ce Zhang, the company's co-founder Percy Liang is also a top-tier Chinese scholar deeply rooted in the AI field. As a professor in the Department of Computer Science at Stanford University, Percy Liang serves as the director of Stanford's Center for Research on Foundation Models (CRFM). In the fields of natural language processing and machine learning, he focuses on the research and development of open-source foundation models and the construction of model evaluation systems.

In terms of team collaboration: if Ce Zhang built an efficient, low-cost computing power "highway" for Together AI at the underlying architecture level, then Percy Liang, with his profound academic accumulation and industry insights, acts as the "navigation system" for the company's technological implementation. Through rigorous evaluation benchmarks, he helps Together AI continuously identify optimal solutions within the complex open-source model ecosystem, ensuring the reliability and cutting-edge nature of its underlying technology.

Under the leadership of Ce Zhang and Percy Liang, alongside the company's chief scientist Tri Dao, they collectively underpin the firm's underlying competitiveness, driving Together AI to achieve exponential reductions in "computing power costs". They reinforce Together AI's underlying algorithm and system capabilities across multiple dimensions, allowing the company to maintain a leading edge in algorithm iteration, model selection, system architecture, and computing power scheduling.

NVIDIA's "Calculated Strategy": Nurturing the Ecosystem to Sell More GPUs

Looking at NVIDIA's investment move, it actually reflects the giant's underlying ambition to lay out the AI computing power ecosystem. As a follow-on investor, this is not NVIDIA's first foray into Together AI.

Back in November 2023, during the Series A financing round, NVIDIA already participated in the company's early funding. At that time, NVIDIA joined as a follow-on investor, while the lead investors of that round were Kleiner Perkins and Prosperity7, a subsidiary of Saudi Aramco.

NVIDIA's continuous investment in this company serves a clear purpose. As of today, Together AI, the enterprise favored by NVIDIA, is no longer a simple AI cloud service provider, but has become an indispensable "ecosystem node" in NVIDIA's massive computing power empire. Put simply, through its investment, NVIDIA has directly secured massive hardware orders.

Public disclosures show that Together AI's business model is built on NVIDIA's H100, H200, and even the latest Blackwell GPUs. With Together AI's annual bookings surpassing $1.15 billion, the company will inevitably have a more rigid demand for massive GPU clusters. NVIDIA's capital injection paves the way for sustained consumption of its own chips.

Furthermore, NVIDIA is attempting to break the monopoly of traditional cloud computing giants through these "Neocloud" service providers, building a more extensive computing power distribution network. For NVIDIA, this means bundling the demands of mid-to-long tail AI developers and enterprises by supporting service providers like Together AI, thereby converting them into a more stable and predictable procurement trajectory.

Of course, the deeper underlying reason is that NVIDIA is transitioning from the role of a "hardware seller" to a "revenue-sharing partner", positioning itself as a partner for AI infrastructure. On July 1st, NVIDIA officially launched its AI Compute Partnership program, clarifying its positioning as an "industry chain co-construction partner". This means that NVIDIA will not only sell hardware, but also provide credit support and revenue-sharing mechanisms, enabling enterprises like Together AI to rapidly expand their computing power scale without bearing massive upfront capital expenditures.

Under this model and new positioning, NVIDIA has already established partnerships with AI cloud service providers beyond Together AI, such as Sharon AI and Firmus. Together, they will build "AI Factories" based on the NVIDIA DSX AI Factory architecture.

The latest quarterly financial report shows that NVIDIA has delivered a remarkable performance report: single-quarter revenue reached a staggering $81.615 billion, representing an 85% year-on-year increase; net profit attributable to shareholders skyrocketed 211% year-on-year to $58.321 billion.

This figure not only shattered NVIDIA's quarterly revenue record, but also propelled the company to a $5 trillion market capitalization, fully affirming the global market's robust demand for AI computing power infrastructure. A closer look at this performance report reveals that customer diversification is one of the key forces driving NVIDIA's strong performance growth.

The financial report indicates that in its data center business, NVIDIA's surging performance demand not only comes from hyperscaler cloud vendors, but also broadly originates from AI cloud, industrial, enterprise, and sovereign AI clients. The diversification of its customer base reduces NVIDIA's reliance on any single major client, allowing it to build a more resilient growth foundation. Meanwhile, the accelerated iteration of global large models and the rapid deployment of AI applications have further driven the expansion of global AI factories and data centers, creating strong computing power procurement demand from downstream cloud vendors, AI-native enterprises, and government organizations, and reinforcing NVIDIA's positive performance outlook.

Under NVIDIA's calculated strategy, Together AI has also announced plans to expand its public cloud capacity — that is, its "computing power" scale — by 50 times over the next five years. This means that once computing power costs are no longer an industry bottleneck, the era of open-source large models may truly arrive, marking a genuine "Together AI" era for AI developers.

And this AI unicorn, led by Chinese technical talent, will continue to forge ahead, riding the global wave of AI infrastructure development.

This article originates from the WeChat public account "Dong 40 Tiao Capital" (ID: DsstCapital), authored by Chen Mei, edited by Wang Qingwu, and published with authorization from 36Kr.