Depth丨Large models that grow on the cloud will eventually change the landscape of cloud computing

Compared with the prosperity of the previous two years, the cloud computing industry has faced difficulties and growth has been weak in the past two years.

Driven by the boom in large models, cloud computing vendors have become one of the main beneficiaries and are seen as opportunities to achieve overtaking in corners.

It is expected that large models will drive a sharp increase in AI computing power on the cloud, gradually narrowing the gap between AI computing power workloads and general computing power, and bringing new opportunities to cloud vendors.

According to estimates from A16Z, a well-known venture capital firm in Silicon Valley, application layer vendors spend approximately 20% to 40% of their revenue on inference and model fine-tuning. This revenue is usually paid directly to cloud vendors or third-party model providers.

And third-party model providers also spend about half of their revenue on cloud infrastructure. As a result, 10%-20% of total generative AI revenue flows to cloud providers.

On the other hand, OpenAI’s research shows that the amount of computation increases exponentially in the training of the largest artificial intelligence models.

The current growth rate of computing power required for large model training remains at a doubling rate of 3-4 months, far exceeding the growth rate of Moore's Law.

To sum up, a large amount of funds for generative AI eventually flowed to infrastructure, and AI large model training brought considerable direct revenue to cloud vendors.

At the same time, generative AI has promoted the upgrade of the entire cloud service industry and brought new market increments.

The traditional cloud ecosystem is centered on software and services. During the rise of cloud computing, a group of sales, software, service, consulting, and integration partners naturally gathered around cloud vendors, forming a [cloud native] ecosystem.

The tacit understanding between cloud vendors and their partners is that they only provide basic public cloud services (such as computing, storage, network, database), and the partners are responsible for the rest.

The advantage of this ecology is that it has clear boundaries and clear division of labor. However, over the past many years, there have been many shortcomings in the ecological construction of China's cloud market.

The level of application modernization in the Chinese market is not yet high. Integrated projects and human outsourcing are the current mainstream forms. The problems this brings are high service costs and low collaboration efficiency.

In the era of big data models, the new cloud ecosystem will have multi-tenant platforms and AI native applications as its core.

Application software transformed by AI technology will play a more critical role in the entire ecosystem.

Atomized, single-point, and fragmented AI applications will be richer and more valuable.

Its value is mainly reflected in solving the main problems of the previous generation cloud ecosystem, namely reducing service costs and improving service efficiency.

The large-model stage cloud ecosystem with AI as the core is gradually becoming mainstream and gradually replacing the old cloud ecosystem.

This historical process is determined by the continuous increase in the proportion of AI computing power and the continuous decline in the cost of AI computing power.

According to OpenAI's forecast, the cost of training large AI models will continue to decline rapidly, and is expected to drop to only $600,000 by 2030.

As the number of large models increases, many companies have realized that it is not wise to train large models independently. Instead, it is more practical and efficient to conduct secondary development based on existing general large models.

This view is consistent with the response strategy to the phenomenon of "reinventing the wheel" that is prevalent in the current field of large models.

In the context of the emergence of a large number of AI native applications, in addition to early model training, cloud service providers' profit margins will come more from providing developers with powerful basic large models and charging AI fees for various business scenarios and user needs. Applied Reasoning Fees.

In order to achieve this goal, cloud service providers need to ensure the stability of computing services and the quality of inference experience to stand out in the fierce market competition.

Eventually, MaaS will become a new basic service and rely on new IT infrastructure to further subvert the current cloud computing market structure.

In the past few years, steady-state services were usually deployed on hosts and required absolute stability, while sensitive services were more often built on the cloud to meet changing needs.

On this basis, there are currently two main cloud migration models: one is flat migration to the cloud, which has lower costs and ensures data controllability, but at the same time has low availability;

The other is to completely transform the underlying architecture into a cloud-native solution. Although the cost is high and the cycle is long, it can bring long-term benefits to the enterprise. However, few companies can afford to suspend servers for architectural changes.

Currently, major cloud service providers are actively launching large model platforms to reduce the cost of training and using large models.

Baidu launched the Wenxin Qianfan large model platform, specifically for B-end enterprise users. The platform provides one-stop large model customization services, including data management, automated model customization and fine-tuning, and inference service cloud deployment. It also provides Wenxinyiyan enterprise-level inference cloud services.

Tencent Cloud has launched a new generation of HCC high-performance computing cluster, using the latest generation of Tencent Cloud Xinghai self-developed servers and NVIDIA H800 Tensor Core GPU to provide ultra-high interconnect bandwidth of up to 3.2T.

Alibaba Cloud launched [Magic Community], which is a user-oriented model service platform. The platform is committed to lowering the threshold for using models and providing adapted APIs for multiple scenarios.

Huawei Cloud's one-stop AI development platform ModelArts provides comprehensive optimization support for the training and inference of the Pangu large model, including computing optimization, communication optimization, storage optimization, and algorithm optimization, and has become an important basic platform resource for the Pangu large model.

Volcano Engine, a cloud service platform owned by ByteDance, has launched an intelligent recommendation-high-speed training engine, which adopts an optimized design integrating software and hardware to support efficient training of 100GB-10TB+ ultra-large models.

Under the influence of large models, the transformation of cloud computing from CPU cloud-based in the Internet era to GPU cloud-based in the AI ​​era has become an industry consensus and inevitable trend.

On the lower-level chip side, in addition to NVIDIA, manufacturers such as Qualcomm, Intel, and Arm have also begun to incorporate chip design and production for large model training and inference into their work schedules to prepare for the possibilities of the next era.

However, in addition to changes in chip types and quantities, the impact of large models on cloud computing manufacturers is now deeper.

Major cloud vendors must pay sufficient attention to research and investment in large model-related capabilities.

The support of general artificial intelligence is the general trend and is also a task that major cloud vendors must complete.

In this regard, any negligence or deficiency may lead to the inability to gain a foothold in the fierce market competition.

Therefore, major cloud vendors should regard it as a necessary core competency and continue to increase investment in order to respond to changing market demands and competition.

As new demands emerge, it is necessary to build a new ecosystem that matches them and formulate corresponding new strategies.

Reference for some information: Finance Eleven: "Large models are about to change the cloud ecosystem", Lei Feng.com: "It's the end of the year, let's talk about cloud computing and large models", China News Weekly: "On the cloud, Seeing the future of large models", Bitnet: "What will the ever-increasing popularity of large models bring to the cloud computing market? ", Technology Cloud Report: "Will generative AI usher in a new decade of cloud computing? ", 21st Century Business Herald: "In the era of large models, the dual identities of cloud vendors"

The articles and pictures published on this official account are from the Internet and are only used for communication. If there is any infringement, please contact us for a reply. We will process it within 24 hours after receiving the information.

Recommended reading:

For business cooperation, please add WeChat:

18948782064

Please be sure to indicate:

"Name + Company + Cooperation Requirements"