Nvidia gets almost half its revenue from just 4 customers. Here's who they might be

In July, Google-parent Alphabet reported sales of nearly $85 billion for the second quarter — smashing Wall Street’s expectations by $840 million.

“Our strong performance this quarter highlights ongoing strength in Search and momentum in Cloud,” chief executive Sundar Pichai said in a statement. Alphabet’s Cloud revenue surpassed $10 billion for the first time, while its operating profit was over $1 billion.

“Our AI infrastructure and generative AI solutions for cloud customers have already generated billions in revenues and are being used by more than 2 million developers,” Pichai said.

And while AI Overviews, the AI search feature rolled out to Google Search in May, initially went awry, generating wrong and even dangerous information, Pichai said the company had made “great progress,” and increased engagement for users between 18 and 24.

Google announced in July it was making its fastest, most cost-efficient model, 1.5 Flash, available in the unpaid version of its Gemini chatbot in over 40 languages and more than 230 countries and territories.

With 1.5 Flash, Gemini will have “quicker and more helpful responses,” Google said, adding that users would notice improvements to Gemini’s reasoning and image processing abilities.

In August, Google brought back Gemini’s ability to generate images of people, after having to pause the feature in February, when users pointed out that Gemini was generating historically inaccurate images of people, including racially diverse Nazi-era German soldiers.

Months after Nvidia unveiled its next-generation, highly sought-after Blackwell chips, Amazon Web Services, the largest cloud computing provider in the world, “fully transitioned” its order of Nvidia’s Hopper chips for Blackwell because the “window between Grace Hopper and Grace Blackwell was small,” the company told The Financial Times.

In April, Amazon chief executive Andy Jassy said in an annual letter to shareholders the company is “optimistic that much of this world-changing AI will be built on top of AWS.” While Amazon spent about $1.5 billion on Nvidia’s H100 chips in 2023, the company has also built its own chips: Trainium for training its AI models, and Inferentia, a chip for inferencing.

Meanwhile, the company is reportedly working on an AI chatbot, internally called “Metis,” to rival OpenAI’s ChatGPT. The chatbot will be accessible through a web browser, and will be powered by one of the company’s internal AI models, Olympus, Business Insider reported, citing unnamed people familiar with the matter as well as an internal document. Olympus is reportedly more powerful than Amazon’s publicly available AI model, Titan.

While Amazon changed the virtual assistant game when it launched Alexa in 2014, the company’s new AI-powered Alexa is not even close to being ready, Fortune reported, citing unnamed former employees. Amazon reportedly doesn’t have enough data nor access to the chips needed to run the large language model (LLM) powering the new version of its virtual assistant. It has also reportedly deprioritized the AI-powered Alexa to focus on building generative AI for its cloud computing unit, AWS.

However, an Amazon spokesperson told Quartz its former employees are incorrect and uninformed on its current Alexa AI efforts, and that the Amazon Artificial General Intelligence team has access to both in-house Trainium chips and Nvidia graphics processing units, or GPUs.

In March, Amazon completed its $4 billion investment in AI startup Anthropic — its largest investment in an outside company ever.

In July, Meta released its open-source Llama 3.1 models that it says are its “most capable” to date. The largest model, Llama 3.1 405B, has 405 billion parameters, or the variables a model learns from training data that guide its behavior. Llama 3.1 405B rivals its closed-source competitors from OpenAI and Google in “state-of-the-art capabilities,” including general knowledge, math, and translating languages, Meta said. The release also included upgraded versions of its 8B and 70B models, which were introduced in April.

Meta said it trained the model with over 16,000 of Nvidia’s H100 GPUs, or graphics processing units. Meanwhile, Nvidia announced an Nvidia AI Foundry service for enterprises and nation states to build “supermodels” with Llama 3.1 405B.

Meta’s Llama models are used to power its AI chatbot, Meta AI, which is available on Facebook, Instagram, and other platforms. Mark Zuckerberg, Meta’s chief executive, told Bloomberg the company is working on Llama 4.

In September, Microsoft unveiled “the next wave” of its Copilot artificial intelligence tools in its suite of work apps.

The AI-powered features include Business Chat, which databases web data, work data, and business data into a new tool called Copilot Pages, where human and AI-generated data can be edited, added to, and shared between work teams. Microsoft chief executive Satya Nadella said the new tool was part of his “daily habit.”

Microsoft also launched Copilot in Excel with Python, making it possible to use the programming language to work with data in Excel with natural language, not coding. The new Copilot also includes AI-enabled features for Powerpoint, Outlook, Word, OneDrive, and Teams, which Microsoft co-founder Bill Gates has said he uses “a lot.” And the company introduced Copilot agents, or AI assistants, that can be used to automate certain business tasks.

In March, Business Insider reported that some Microsoft customers were complaining that Copilot was falling short of OpenAI’s ChatGPT. However, employees said Microsoft 365 Copilot had mostly positive feedback, but some customers using older versions of Microsoft’s suite of business tools were expecting the more advanced Copilot to work with it, leading to the unfavorable comparisons with ChatGPT.

In May, Microsoft announced Copilot Plus PCs, or its laptops equipped with AI hardware and support for AI applications, that would be available from its major laptop partners, including Dell (DELL) and HP (HPQ). The company said in May it had released 30 “responsible AI” tools with over 100 features to support AI applications developed by its customers on its Azure cloud computing platform.

Microsoft’s shares reached a record high in March after it announced its Copilot for Security tool, which it called the AI industry’s “first generative AI solution” for security and IT professionals; it’s trained on “large-scale data and threat intelligence,” including over 78 trillion security signals that the company processes daily, Microsoft said.

Tesla, which its founder and chief executive Elon Musk has touted as an AI company, has worked on its Full Self-Driving (FSD) tech for years. The electric vehicle company has been accused of exaggerating its driver-assistance system, but in July, analysts said its latest update was encouraging.

“There have been 90+ updates to the FSD system in the last four years, and fully autonomous Teslas still don’t exist,” Alexander Potter, an analyst at Piper Sandler, wrote. But “judging by user reviews on the X platform, version 12.5 is revolutionary.”

In September, Tesla unveiled its Actually Smart Summon feature for its S3XY electric vehicles. The feature allows Tesla owners to order their car to drive to their location or another destination via the Tesla app. Another feature, called Dumb Summon, allows owners to move their vehicle forward or backward through the app.

Tesla also has a humanoid robot called Optimus in development, which it put on display at China’s World Artificial Intelligence Conference in July. The second-generation robot, which is not yet at full-scale production, is “capable of performing tasks that are unsafe, repetitive or boring,” the company says. In June, Musk claimed that robots will eventually power the company to a $30 trillion market cap.

Musk, who also founded an AI startup called xAI, announced in September that its training cluster, called Colossus, is powered by 100,000 Nvidia H100 graphics processing units, or GPUs, and is expected to double in size to 200,000 chips, including 50,000 of Nvidia’s more powerful H200 chips, “in a few months.” In July, he posted a poll on his social media platform X, asking if Tesla should invest $5 billion in xAI, “assuming the valuation is set by several credible outside investors.”

When it released ChatGPT in November 2022, AI startup OpenAI kicked off the current generative AI boom that has seen major tech companies, including its partner, Microsoft, spend billions on Nvidia’s chips. Despite its startup status, OpenAI is undoubtedly one of the leaders in the race, with its models consistently outperforming those of its competitors.

In September, OpenAI released its series of “reasoning models” that it said can “think” through problems before responding — “much like a person would.” The model series, which is available now as o1-preview through ChatGPT and the company’s API, can handle more complex tasks and problems in science, coding, and math than earlier OpenAI models. The company also announced OpenAI o1-mini, a smaller and cheaper version of the new model, which can help developers with coding tasks.

In May, the startup released a flagship model called GPT-4o, which it called a step towards much more natural human-computer interaction.” The model can accept combinations of text, audio, images, and videos, and generate combinations of text, audio, and images, OpenAI said. The company added that the new model was 50% cheaper and twice as fast as its predecessor, GPT-4 turbo.

Sam Altman, OpenAI’s chief executive, said in March that he thought its GPT-4 model “kind of sucks.” However, he said ChatGPT-4o “feels like magic to me.”