TheChek News Flow

For years, Apple has had a bumpy relationship with Nvidia. But the iPhone maker’s AI ambitions are testing how completely it can shun investments in Nvidia’s products.

Like most big tech companies with artificial intelligence ambitions, Apple has little choice but to use chips from Nvidia, whose graphics processing units are practically a de facto standard in the development and running of AI software. But Apple is working on ways to spend less on Nvidia chips without undermining its success in AI.

Instead of buying boatloads of Nvidia chips as its tech peers do, the iPhone maker mostly rents access to them from cloud providers like Amazon and Microsoft. In some cases—for example, to train its biggest models—Apple has rented AI chips designed in-house by Google rather than Nvidia. And in its boldest move yet to avoid relying on Nvidia, Apple is working with Broadcom to design an AI server chip expected to be ready for mass production by 2026, news of which The Information first reported earlier this month.

Apple’s Nvidia allergy appears to stem partly from its frugality and a desire to own and control the key technological ingredients for its products to avoid giving others leverage over its operations. But leaders within Apple have also quietly nursed grudges against Nvidia for nearly two decades, stemming from business disputes that originated during the era when Steve Jobs was Apple’s CEO, according to 10 former Apple and Nvidia employees with direct knowledge of the relationship.

In 2001, for example, the two companies forged a major partnership when Apple included Nvidia’s chips in its Macs to give the computers better graphics capabilities. At the time, Jobs told a packed audience at an event in Tokyo that they had a “great relationship.”

Behind the scenes, though, their dealings were tense. In a meeting around that time between Jobs and a senior Nvidia executive, the Apple CEO remarked that Nvidia products contained technology copied from Pixar, the computer animation studio Jobs then led and had a controlling stake in.

The Nvidia executive pushed back on the idea, saying Nvidia held more graphics patents than Pixar and should sue the animator instead. For the remainder of the meeting, Jobs pretended the Nvidia executive was no longer in the room, according to a person who witnessed the incident.

Analysts say some of Nvidia’s biggest customers by revenue include Google, Microsoft, Amazon, Meta Platforms and Tesla. Apple, despite its size and reach, isn’t even among Nvidia’s top 10 customers, they say. Some of that reflects the nature of the business Apple is in.

Unlike Google, Microsoft and Amazon, Apple doesn’t sell cloud computing services to other companies, so it doesn’t have to satisfy demand from clients who want access to Nvidia chips. (All three of those cloud providers have designed their own AI chips as well, though the response from customers has been mixed in some cases.) Meanwhile, for companies like Meta, OpenAI and Tesla, AI is becoming or is already a more fundamental part of their businesses.

It’s unclear whether Apple will be able to hold off indefinitely on making big purchases of Nvidia chips. The company has scrambled this year to respond to the rise of AI applications, such as OpenAI’s ChatGPT, by weaving new AI features into products like the iPhone.

But Apple has been compelled to provide customers with the option to use ChatGPT for tasks its own models still can’t handle. This has put pressure on the company to train bigger and better models, which would require even more access to high-end GPUs.

This year, Wall Street analysts have focused on the possibility that Apple might need to spend a lot more on AI than it has been. They’ve been badgering Apple CEO Tim Cook about whether the company plans to increase investments in data centers and chips to be more competitive.

Cook has downplayed claims that Apple’s AI features will significantly increase the company’s costs, telling analysts some of the expenditures will be borne by the third-party cloud providers it rents servers from. Still, Apple’s work with cloud providers doesn’t come cheap: The company ranks No. 1 in The Information’s database of cloud spenders, with billions of dollars in estimated annual cloud bills.

Spokespeople for Apple and Nvidia declined to comment.

A Demanding Customer

The chilly relations between Apple and Nvidia continued after the tense 2001 meeting between Jobs and the Nvidia executive.

In subsequent years, Nvidia executives viewed Apple as a demanding customer that had unreasonable expectations given the small size of its Mac sales relative to Windows personal computers, said former Apple and Nvidia employees with direct knowledge of the relationship. Nvidia CEO Jensen Huang, in particular, was unwilling to dedicate a lot of resources to the partnership with Apple, given how little revenue Nvidia received from it, according to people familiar with his thinking.

At the same time, Apple was then in the midst of a comeback with the success of the iPod and was becoming accustomed to getting its way with suppliers, the former Apple employees said. It saw Nvidia as especially difficult to work with.

Nvidia chips were power hungry and generated lots of heat, both of which made it difficult for Apple to include the chips in its laptop designs. Apple wanted Nvidia to build custom graphics chips for its MacBooks but grew frustrated by its unwillingness to do so, multiple former Apple employees said.

Another one of Apple’s frustrations: The company’s engineers spent months trying to persuade Nvidia to send engineers to China to debug its graphics chips as Macs came off a production line. They also had to twist Nvidia’s arm to write software that would allow Apple to test Nvidia chips on Macs rather than Windows PCs, according to a former Apple employee.

‘Bumpgate’

One of the biggest flashpoints in their relationship occurred in 2008, when Nvidia produced a faulty graphics chip that ended up inside computers from Apple, Dell and Hewlett-Packard. The problem came to be known as bumpgate because it involved balls of lead-based solder called bumps that were overheating and causing parts of the chip to crack, leading to graphics failures.

Initially, Nvidia refused to take responsibility for the problem and wouldn’t fully reimburse Apple and other PC makers for repairs, former Apple and Nvidia employees with direct knowledge of the issue said. Apple was forced to extend warranty coverage for certain MacBooks with the faulty chips and was unable to get full compensation from Nvidia, which upset Jobs and other executives, according to former Apple employees.

From Nvidia’s perspective, the company had no contractual obligation that required it to make Apple whole, according to a former Nvidia executive with direct knowledge of the issue.

The dispute played a role in Apple’s decision to switch to another supplier: AMD, whose graphics chips didn’t perform as well as Nvidia’s. Apple had more leverage over pricing with AMD and also was able to persuade it to design custom chips for its MacBooks. Unlike Nvidia, AMD accommodated Apple because it was eager to win its business, according to former Apple and AMD employees.

Nvidia’s unwillingness to bend over backward for Apple might have been justified, as later events showed. By 2018, Apple was spending only around $200 million annually with AMD to supply GPUs for its computers—a fraction of the nearly $6.5 billion AMD earned in revenue that year, according to a person with direct knowledge of the figure. And in 2020, Apple began phasing out its use of AMD GPUs, releasing its first MacBook with a GPU it had designed in-house.

Licensing War

Apple executives also clashed with Nvidia in its attempts to demand licensing fees for Apple’s use of graphics chips on mobile devices like the iPhone.

In the early 2010s, Samsung, Qualcomm and Apple were either using their own designs for the mobile GPUs in their smartphones or licensing technology from mobile GPU designers such as Imagination Technologies.

Huang believed many of the techniques used to render graphics were based on patents Nvidia owned for GPUs inside computers, so the company began approaching Apple, Samsung and others to demand they start paying licensing fees for its mobile GPUs, according to court documents and former Apple and Nvidia employees familiar with the matter.

Nvidia’s initial request for a licensing fee incensed Apple’s top leaders—especially Dan Riccio, the company’s former senior vice president of hardware engineering. That further soured Apple on using Nvidia’s products, according to multiple former Apple employees. Apple didn’t agree to the demand, two former Apple employees said.

Eventually, Nvidia sued Samsung and Qualcomm in 2014 for violating its patents and asked the U.S. International Trade Commission to block shipments of Samsung’s mobile phones and tablets into the U.S. The company likely hoped to use victories in the cases as leverage against Apple, according to current and former Nvidia executives.

Instead, Nvidia lost the ITC case against Samsung and ultimately settled its court cases against that company and Qualcomm.

Self-Driving Chip

Around 2015, when Apple began developing a self-driving car, one component was conspicuously absent from it: an Nvidia GPU.

Instead, the company’s top leaders decided to design their own chip for the vehicles to handle inference, the object recognition and decision-making trained AI models perform—in this case, to autonomously pilot a car. At the time, Nvidia’s chips were the industry standard in autonomous vehicles.

Apple used its chip, code-named Tinos, in dozens of test vehicles, some of which could contain as many as 16 chips, each running different AI models to help with self-driving, said four former Apple employees who worked on its self-driving–car program.

Apple couldn’t entirely avoid Nvidia chips, though: Its engineers still needed them to train the AI models for its autonomous systems, renting them through Amazon’s cloud computing service, one of the people said. At one point, Apple was Amazon’s largest customer for renting Nvidia GPUs in North America, the person said.

Apple ultimately scrapped the car project earlier this year, moving many of its AI engineers to work on Apple Intelligence, the family of AI features in iPhones and other devices.

Google Versus Nvidia

In 2018, when Apple hired John Giannandrea to lead a new organization for AI and machine learning within the company, one of the first things his new team asked for was more Nvidia GPUs to assist with training AI models.

Prior to his arrival, Apple had purchased a small number of its own Nvidia GPUs for AI training that were scattered throughout different parts of the company, including the teams that developed Apple’s Face ID facial recognition system and the chip design team, which used Nvidia chips to simulate iPhone chip performance. The scarcity of the Nvidia chips meant employees faced long wait times to get access to them.

Giannandrea, who had previously worked at Google, instead encouraged the use of an AI chip Google had designed in-house called a tensor processing unit, said multiple current and former Apple employees involved in AI. Google had developed the processor in the mid-2010s in part to reduce its dependence on Nvidia. In the years to come, Apple trained some of its largest and most important models using Google TPUs rather than Nvidia chips, it said in a research paper it published earlier this year.

The decision partly reflected the pedigree of much of Giannandrea’s team, many members of whom had previously worked at Google and were familiar with using the TPUs for machine learning, the former Apple employees said. But it also was partly due to a shortage of Nvidia GPUs over the past two years, as they are in high demand among customers who rent or purchase them to train AI models, some of the employees said.

This month, Apple said it was evaluating Amazon’s in-house–designed AI chips for training as well.

Still, relying on Google and Amazon for AI chips could over time just make Apple dependent on a different set of companies. Apple’s solution is to build its own AI chip to power its internal servers, an unusual move for a company so focused on designing chips for consumer devices.

Recently, there have been some signs that Apple’s relationship with Nvidia isn’t entirely acrimonious. In March, Huang briefly featured Apple’s Vision Pro mixed reality headset during a presentation on Nvidia’s 3D software tools. Last week, Nvidia and Apple announced that they had collaborated on a research project that sped up the operation of large language models on Nvidia chips.

And while Apple executives have historically expressed frustration with Nvidia, current and former Nvidia executives and employees describe it as a mostly one-sided beef. Nvidia executives remain open to collaboration and believe Apple could become a significant customer in the future, current Nvidia executives say.

However, other signs suggest Apple is exploring ways to keep its distance from Nvidia. This summer, Apple freed up some chip engineers in Israel after canceling a high-end Mac chip, redirecting them to work on the new AI server chip, code-named Baltra after a small island in the Galápagos, The Information previously reported.

The involvement of Broadcom in that effort suggests Apple might be interested in eventually using the chip for training models in addition to inference. That’s because one of Broadcom’s strengths is in chip networking, which involves linking chips so they can work fast and in unison, a major requirement for training AI models.

Such an effort would further reduce Apple’s need to buy or rent training chips from Nvidia, along with Google and Amazon.