With the tightening of cloud spending on Internet enterprise software, slowing growth is gradually becoming a dark cloud over the heads of cloud vendors. the emergence of ChatGPT breaks this bottleneck, and AI will reshape software. Software companies, the customers of cloud vendors, are aggressively embedding the AI capabilities of big models into existing workflows to accomplish higher levels of automation.
With new cloud customers drying up, software companies are no longer going to the cloud for the sake of going to the cloud, but are seeking to improve productivity with AI. This is the biggest incremental addition to the cloud market over the next decade. Arithmetic infrastructure is the absolute dividend beneficiary of the big model. A cloud computing industry player who has been practicing for more than a decade elaborated on Geek Park. With such a prospect, several overseas cloud service giants-Microsoft, Amazon, Google, and Oracle-have been quick to make changes. Over the past few months, the cloud giants have dropped real money to develop big models, make strategic investments, and develop their own AI chips……. The era of big models is booming, and they have already targeted a new generation of AI software customers.
The former rivers and mountains are far from unbreakable, the cloud market is rapidly reshuffling, and the giants have opened a new curtain of competition. After all, the fall of Big Brother in the mobile Internet era is just around the corner, and Nokia went from 70% of the cell phone market in its heyday to no one’s business in a few years, just a matter of making the wrong decision. As for the big model, the cloud industry is rapidly forming a consensus: this time AI is by no means a small variable, and from the industry’s rapid development speed, the current leading players may also be left behind. 2023 is already halfway through, this article will be centered on several major overseas cloud giants to sort out what is the key to competition among cloud vendors today.
1. R&D of AI-specific chips, can’t give all your life to NVIDIA
After the advent of the big model era, for cloud service providers, the scarcest resource today is arithmetic power, or AI chips. Investing in the lowest level of infrastructure – AI accelerated chips – has also become the first focus of today’s competition among cloud vendors. Scarcity and cost are considered to be the primary reasons for cloud vendors to accelerate self-developed chips. Even Musk, such a powerful big man in the technology circle, has commented that this thing (NVIDIA GPUs) is more difficult to get than pharmaceuticals, and has secretly bought 10,000 cards from NVIDIA for his own AI company, X.AI, and has collected a lot of Oracle’s idle equity.
Such a level of scarcity is reflected in the cloud giant’s business, which directly corresponds to the loss of business from the necklace. Even Microsoft, which started first, has been exposed to rumors that due to the shortage of GPUs, the internal AI research and development team has implemented a GPU rationing system, various new plans are delayed, and new customers have to queue up for months on Azure. Even venture capital organizations to grab the project, have to rely on the hands of NVIDIA chip inventory. In order to N card, all forces to the point of doing everything. Another name for scarcity is expensive. Considering that the large model of the computing power demand has increased tenfold, the card will only be more expensive. Recently, an investor said to Geek Park, that at the beginning of the year 80,000 for a single card A100, now has been speculated to 160,000, can not. Accordingly, the cloud giants’ tens of thousands of cards to pay the NVIDIA tax will only be an astronomical figure. The most popular Microsoft has the most right to say whether it feels good to have your life hanging in someone else’s hands. A month ago, The information exclusively reported that Microsoft set up a 300-member team to speed up the pace of self-research AI chips, The codenamed Cascade server chip may be launched as early as next year.
Not only because of the necklace, cloud vendors self-research chip, there is another layer of reference – GPU is not necessarily the most suitable for running AI chip, self-research version may optimize specific AI tasks. It’s true that most advanced AI models are currently powered by GPUs, which are better at running machine learning workloads than general-purpose processors. However, GPUs are still viewed as general-purpose chips, not truly processing platforms native to AI computing. As Fargawa Institute’s A Crack in the NVIDIA Empire points out, GPUs are not made for training neural networks, and the faster AI evolves, the more these issues are exposed. Relying on CUDA and various technologies to magically change one scene at a time is an option, but not the optimal solution.
Amazon, Google, and Microsoft have been developing chips called ASICs – Specialized Integrated Circuits – that are better suited for AI. The Information interviewed a number of chip industry practitioners and analysts and concluded that NVIDIA GPUs helped train the model behind ChatGPT, but ASICs are generally faster at performing these tasks. NVIDIA GPUs helped train the models behind ChatGPT, but ASICs typically perform these tasks faster and use less power. As you can see in the chart above: Amazon, Microsoft, and Google have all elevated chips to in-house importance, developing two types of chips for their data center divisions: standard compute chips and chips specifically designed to train and run the machine learning models that power chatbots like ChatGPT.
Currently, Amazon and Google have developed customized versions of ASICs for key on-premise products and have made these chips available to customers via the cloud. Microsoft has also been working on developing custom ASIC chips to power large language models since 2019. Some of the chips developed by these cloud providers, such as Amazon’s Graviton server chip and AI-specific chips released by Amazon and Google, are already matching the performance of chips from traditional chipmakers, according to performance data released by cloud customers and Microsoft. Google’s TPU v4 is 1.2-1.7 times faster than NVIDIA’s A100 computer while consuming 1.3-1.9 times less power.
2. Strategic Investment Race: Giants Spend Money on Customers
In addition to R&D chips, the second key point of competition among several overseas cloud giants is foreign strategic investment to grab AI customers and AI projects. Compared with venture capital, the giants’ strategic investments have an absolute advantage, and the collaboration between OpenAI and Microsoft serves as an excellent model, opening up a precedent for big models and strategic investments. This is because the resource barriers required by big models and related applications are extremely high, and only money, limited money, is simply not enough to grab AI projects. After all, Google, Microsoft, AWS, Oracle, or NVIDIA can not only write huge checks but also provide scarce resources such as cloud credits and GPUs.
From this perspective, the grabbing of projects and customers is happening among the cloud giants, with no other rivals. They are playing a new game – seeking promises from AI companies that their cloud services will be used instead of their competitors. Microsoft sits as OpenAI’s exclusive cloud provider, paying OpenAI’s huge cloud bill in exchange for a host of enviable benefits such as equity in OpenAI and preferred access to its products.
Microsoft’s competitors are also scrambling to win over other AI customers. These cloud providers are offering steep discounts and credits (credits) to AI companies to win their business. Critical voices have pointed out that this is akin to buying customers, although the practice of taking equity stakes in future or current customers is not uncommon in the enterprise software space. Oracle has also offered hundreds of thousands of dollars worth of compute credits as an incentive for AI startups to rent Oracle cloud servers, The Information reported earlier.
Google is probably the most aggressive of the big cloud vendors, offering AI startups a combination of cash and Google Cloud credits in exchange for equity. Earlier this year, Google invested $400 million in Anthropic, one of the main startup challengers to OpenAI. Google Cloud said in February that it had become Anthropic’s preferred cloud provider.
Recently, Google invested $100 million in Runway, an AI company in the field of Vincentian video. But before that, Amazon AWS touted Runway as a key AI startup customer. In March, AWS and Runway announced a long-term strategic partnership to become their cloud provider of choice. Now, Runway appears to be one of Google’s pawns in its showdown with Amazon, as Runway is also expected to rent cloud servers from Google.
Earlier, Google Cloud also announced partnerships with two other popular AI companies: Midjourney, in the field of text-to-life mapping, and chatbot app Character.ai, which was previously a key cloud customer of Oracle. It’s too early to tell whether these deals will help Google catch up with its bigger cloud competitors, AWS and Microsoft, but Google Cloud is on the offensive.
Of the 75 (AI) software companies in The information database, Google provides some cloud services to at least 17, more than any other cloud provider. Amazon is close behind, with at least 15 companies using AWS for cloud computing. Microsoft and Oracle, on the other hand, provide cloud services to six and four companies, respectively. Of course, using multiple clouds is also customary in the industry, with at least 12 of these 75 companies using a mix of multiple cloud providers.
3. The Big Model, the Real Key to Winning
Calculation power and war chests are the early high ground in this cloud war. But in the long run, the big model is the real key to winning or losing the market competition. Microsoft’s collaboration with OpenAI is the key to its success, coupled with the excellent engineering capability of the Microsoft team, GPT-4 was embedded in the Microsoft family bucket within a few months. In the past six months, Microsoft first utilized the preferential use of OpenAI products and price reductions for enterprise software products to capture more of the cloud market. Then it relied on the price increase of the product line upgraded to Microsoft 365 Copilot to gain more revenue.
According to the research of Yunqi Capital, Microsoft’s bottom layer model basically relies on OpenAI, and after accessing the big model, Microsoft began to sell Teams, Power BI, Azure, and other application layer products packaged at a lower price. Microsoft CFO Amy Hood told investors in April that OpenAI will bring revenue to Azure as more people start using OpenAI’s services.
The latest reports indicate that Microsoft is charging some Office 365 customers an additional 40 percent to test the AI features – which automate tasks such as writing text in Word documents and creating PowerPoint slides – and that at least 100 customers have already paid up to $100,000 in flat fees. Less than a month after launch, Microsoft has generated more than $60 million in revenue from Microsoft 365 Copilot’s AI features, the data suggests. In stark contrast to Microsoft, Amazon Cloud, once the leader, is facing a tougher challenge today after falling one step behind and one step behind on the big model.
AWS had been an early developer of AI cloud services, with a presence from around 2016. But customers didn’t find those services very useful, including facial recognition, converting text into realistic speech, and chatbots in raw form for tasks like customer service.AWS also launched SagaMaker, an AI digital tool for use by the engineering community in 2017 that helps them develop and use machine learning models, which at one point was AWS’s dominant AI product.
However AWS’s AI offerings have failed to keep up with the wave of big language models in the years since, and since November 2021, Microsoft has been selling AI products developed based on the GPT family of models for use by enterprise customers. Meanwhile, Google has grabbed major AI startups as cloud customers and is selling proprietary AI software to its cloud customers. Even Oracle, the laggard of the cloud, has its own advantages when it comes to providing computing resources to AI startups.
AWS, a laggard, is trying to catch up, and in April it announced a cloud service that lets customers integrate big models from Stability, Anthropic, and AI 21 Labs as a base for their own products. In return, AWS will take a portion of its revenue and share it with these partners.
Google, on the other hand, got an early start but was late to the party. As the biggest player in the field of large models, Google’s response to the release of ChatGPT was not fast, quickly releasing Bard, a conversational intelligent robot, and PaLM 2, a new-generation large language model, in response. As a result, the launch event flopped, and the pace of subsequent product releases was not ideal, in stark contrast to Microsoft’s strong engineering capabilities.
Finally, it is worth mentioning that Oracle, which fell out of the forefront of the cloud market a long time ago, is unexpectedly trending backward in this wave of enthusiasm. Oracle has been a longtime laggard in the cloud space, but has been surprisingly successful in leasing cloud servers to high-profile AI startups that compete with OpenAI. Part of the reason is that the Oracle Cloud can run complex machine learning models more economically than Amazon Web Services or Google Cloud, according to The Information. Oracle Cloud’s approach to the AI circuit appears to be similar to that of AWS, which has developed its own AI software to sell to customers but will also sell access to open-source AI software as well as products from other AI developers. Additionally, some of the people familiar with the matter said that Oracle has already begun testing OpenAI products to enrich its product line for B-side customers, including human resources and supply chain management software, though Oracle is more likely to develop its own software for this purpose, with future AI capabilities that could help Oracle customers quickly generate job descriptions and schedule meetings between recruiters and candidates, though the company is still in the process of deciding which products to improve first.