Reporter Wang Ziwei
The Magnificent 7 of American technology is now catching up with each other on basic models. As one of the seven giants, Amazon's cloud services department is steadily moving forward in cloud services, providing customers with further services and products.
From March 14th to March 16th, Uwem Ukpong, Vice President of Global Customer Technical Support and Service at Amazon Cloud Technology, visited the Shanghai office. During an interview, Uwem Ukpong told the Daily Economic News that in 2024, the focus of Amazon Cloud Technology's layout will be on Amazon Q and Amazon Bedlock.
Unless unexpected, both are related to generative AI.
Amazon Bedlock is Amazon's cloud hosting service, which has been integrated by many large model service providers, including the recently popular Claude 3. And Amazon Q is a universal generative AI assistant, released by Amazon Cloud Technology at the re: Invent conference at the end of November last year.
From the observation of reporters, Amazon Cloud Technology has seen a watershed in 2023. Before the first half of 2023, Amazon Cloud Technology mainly provided basic products such as cloud services, and had more cooperative relationships with service providers for basic models. From the financial report, it can be seen that Amazon Cloud Technology (AWS) has made a significant contribution to Amazon's overall profits;
Since the second half of last year, Amazon has been vigorously and aggressively developing large models. For example, investing $4 billion in Anthropic, a large model company that can compete with OpenAI, and investing $100 million to establish a generative AI innovation center. The independently developed large model Amazon Titan is also more customer-oriented and has engaged in deeper cooperation beyond cloud services with some large model service providers.
Generative AI enables enterprises to embrace the cloud
Time and momentum are also important. Since the outbreak of generative AI in 2023, the biggest feeling for cloud service providers is that there is no need to drive enterprises to go cloud.
"On a global scale, many companies (in the past) have been very slow in adopting cloud technology. When you talk to these companies about the cloud, they will remain silent and do not have a firm belief in making this decision. However, now they are facing pressure from different stakeholders and shareholders to make a difference in generative AI. When client companies need to embark on a journey of generative AI, it means they need to go to the cloud." Uwem Ukpong, who is responsible for Amazon Cloud Technology's global customer technical support and services, has been extremely busy in the past year. Today, Amazon Cloud Technology's global services cover 51 countries and regions worldwide.
What left a deep impression on him was a railway company in the United States. Over the past two years, the company has been in frequent communication with Amazon Cloud Technology, but has yet to make a decision. Until the outbreak of generative AI, this company was rapidly advancing cloud migration.
In fact, such reactions have become the norm for businesses, which has also made AWS the "cash cow" of Amazon.
According to Amazon's (Nasdaq: AMZN) Q4 2023 financial report, its revenue increased by 14% year-on-year to $169.961 billion, operating profit was $13.209 billion, and net profit was $10.624 billion, a significant increase of over 37 times compared to the same period last year's $278 million net profit; Amazon Cloud Technology achieved a revenue of $24.204 billion in the quarter, while its operating profit reached $7.2 billion in the quarter.
For the full year of 2023, Amazon's operating profit was $36.852 billion, while Amazon Cloud Technology's operating profit for the full year of 2023 reached $24.631 billion. Ultimately, Amazon's net profit for the year reached $30.425 billion, compared to a loss of $2.722 billion in the same period last year.
After experiencing a slowdown in revenue growth last year, Amazon Cloud Technology has experienced three consecutive quarter on quarter growth since the second quarter of last year. Prior to this, the revenue of cloud computing business had increased month on quarter for 34 consecutive quarters.
Amazon CEO Andy Jassy reiterated after this financial report that generative AI services remain a relatively small business for Amazon, but the company believes it will bring in "billions of dollars" in revenue in the coming years.
Three layer service architecture
"For over 20 years, artificial intelligence and machine learning have been the focus of Amazon's attention. However, we didn't call it generative AI before. In fact, Amazon's retail business has heavily adopted AI technology," Uwem Ukpong said.
He told the Daily Economic News that currently, Amazon Cloud Technology is based on generative AI and provides a three-layer service architecture, including applications built using basic models, tools built using basic models, and infrastructure for basic model training and inference.
Firstly, cloud infrastructure services. Over the past five years, Amazon Cloud Technology has continuously increased its investment in self-developed chips, constantly breaking through performance and price limits to support workloads such as machine learning training and inference that have extremely high requirements for this. The Amazon Training and Amazon Inferentia chips can provide extremely low cost training of models and running inference on the cloud. In addition, in terms of GPUs, Nvidia has taken the lead in implementing multiple advanced computing capabilities on Amazon Cloud Technology.
This layer is the flagship product of Amazon Cloud Technology and a key focus of Amazon Cloud Technology reinforcement. At present, Amazon's cloud technology infrastructure covers 105 availability zones across 33 geographical regions, and plans to build 6 new regions and 18 availability zones in Malaysia, Mexico, New Zealand, Saudi Arabia, and Thailand.
Second, the model layer. Based on the Amazon Bedlock cloud hosting service, enterprises are allowed to access a series of foundational models from text to images on top of Amazon Bedlock through APIs, such as AI21 Labs, Antitopic, Cohere, Meta, Stability AI, Mistral AI, and Amazon Titan foundational models. And support customers to customize based on private data differentiation.
Thirdly, the application layer, which is a generative AI driven application. Currently, the main focus is on Amazon Q. Amazon Q was released at the re: Invent conference in November last year, which is a new type of generative AI supported assistant that can be customized according to customer business needs, specifically designed to meet office scenarios. Various information repositories such as code repositories, websites, databases, documents, emails, etc. can be connected to assist businesses or individuals in providing historical data retrieval, generation, content writing, and other services through natural language dialogue.
Uwem Ukpong stated in this interview that companies have been actively going cloud over the past year, which has reduced the operating costs of Amazon Cloud Technology's data centers. But Amazon Cloud Technology also faces the pressure of assisting customers in initiating Proof of Concept (POC) for generative artificial intelligence.
Uwem Ukpong has repeatedly mentioned POC, which is more like the commonly mentioned "finding landing scenarios" in China. The department he manages has generative AI innovation centers worldwide. He said that the Generative AI Innovation Center and the professional engineering services provided by Amazon are the services that help customers solve the last three kilometers of POC scenario landing.
"On the one hand, we help customers establish POCs, assist them in concept certification, and enable them to use their AI services. On the other hand, in the process of doing POCs, we also have a deeper understanding of customers' needs. We collect experience from POCs around the world, and then send it as feedback to Amazon's cloud technology software development team, empowering future product development."
Uwem Ukpong said that he has observed that the financial industry is currently one of the most solid and large-scale industries using generative AI. "Policies need to be customized according to the actual situation of individual users. Previously, manual processes required two weeks to create a policy, but now with generative AI, this process only takes one day."
In addition, the application of generative AI in the home appliance and e-commerce industries is also relatively leading. For example, traditional industrial concept design, e-commerce marketing material image generation, etc.
2024: Competing for Cloud Services
Musk explained that artificial intelligence may become smarter than anyone else this year and may surpass the intelligence of all humans by 2029. If it is indeed as Musk said, then these five years are crucial for any cloud service provider.
The industry consensus is that using generative AI must first go to the cloud. Domestic cloud service providers have started to compete for major customers by lowering prices.
On February 29 this year, Alibaba Cloud announced that it would reduce some public cloud products deployed in Chinese Mainland. The price reduction involved more than 100 products, with an average reduction of 20% and a maximum reduction of 55%. After the price reduction, the price of Alibaba Cloud's core products broke through the lowest price on the entire network. JD Cloud immediately followed up and announced that the core products of the entire series will continue to participate in network wide price comparisons, including computing, storage, network and other products. The "specific cloud vendors" for price comparisons include Alibaba Cloud, Huawei Cloud, and Tencent Cloud.
In May last year, cloud vendors began a round of price reductions to seize this great opportunity: Tencent Cloud announced that several core cloud products would be reduced in price starting from June 1st last year, with some product lines experiencing a maximum reduction of 40%; Mobile Cloud also announced a price reduction in May, with a maximum reduction of 60%, becoming the first operator cloud to announce a price reduction.
When it comes to the advantages of Amazon's cloud technology, Uwem Ukpong emphasizes technology and ecology. "The biggest differentiation of Amazon Cloud Technology is its wide range of model choices. We are putting more LLMs on Amazon Bedlock to provide customers with a rich and free choice space."
Uwem Ukpong also revealed two points: firstly, Amazon has chosen to invest $4 billion in the large model company Antiopic, which has released Cloud 3 that can compete with GPT-4, mainly using Amazon Training and Amazon Inferentia chips for training and deploying future basic models. After industry evaluation, Claude 3 is currently the most leading large-scale model in the industry in terms of performance, accuracy, speed, and cost.
"Amazon Cloud Technology has the technology to develop its own chips. After partnering with Aerospace, we can directly obtain a lot of first-hand feedback from Aerospace, which can be used for the next stage of chip technology improvement," Uwem Ukpong said.
In addition, Amazon's self-developed large model and Amazon Titan's basic model have a wide range of applications, mainly providing support for high-performance image, multimodal, and text models for a wide range of generative artificial intelligence applications, such as content creation, image generation, and search and recommendation experiences.
Uwem Ukpong told the Daily Economic News that in 2024, Amazon Cloud Technology's service and product focus will be on Amazon Q and Amazon Bedlock.
Amazon Q is a new generative AI support assistant developed by Amazon Cloud Technology, which can be customized according to customer business needs and is specifically designed to meet office scenarios. Amazon Q provides information and advice to employees, helping them simplify tasks, accelerate decision-making, and solve problems, as well as stimulating creativity and corporate innovation in their work.
He also stated that in 2024, the application of generative AI will become increasingly complex. "Last year's POCs were mostly just dialogue assistants, but this year's applications will become more complex and profound. We will see more content generation, data insights, and even future prospects prediction."