Chip race: Microsoft, Meta, Google, and Nvidia battle it out for AI chip supremacy – The Verge
Filed under:
By Emilia David, a reporter who covers AI. Prior to joining The Verge, she covered the intersection between technology, finance, and the economy.
The rise of generative AI has been powered by Nvidia and its advanced GPUs. As demand far outstrips supply, the H100 has become highly sought after and extremely expensive, making Nvidia a trillion-dollar company for the first time.
It’s also prompting customers, like Microsoft, Meta, OpenAI, Amazon, and Google to start working on their own AI processors. Meanwhile, Nvidia and other chip makers like AMD and Intel are now locked in an arms race to release newer, more efficient, and more powerful AI chips.
As demand for generative AI services continues to grow, it’s evident that chips will be the next big battleground for AI supremacy.
Jun 4
Emilia David
As the AI craze continues, even the microcomputer company Raspberry Pi plans to sell an AI chip. It’s integrated with Raspberry Pi’s camera software and can run AI-based applications like chatbots natively on the tiny computer.
Raspberry Pi partnered with chipmaker Hailo for its AI Kit, which is an add-on for its Raspberry Pi 5 microcomputer that will run Hailo’s Hailo-8L M.2 accelerator. The kits will be available “soon from the worldwide network of Raspberry Pi-approved resellers” for $70.
May 30
Emma Roth
The Ultra Accelerator Link (UALink) Promoter Group, will work to create an open standard to help AI accelerators “communicate more effectively” within data centers and boost performance. Other members include AMD, HP, Broadcom, and Cisco — but not Nvidia, which has AI chip-linking tech of its own.
[TechCrunch]
May 22
Sean Hollister
Nvidia just made $14 billion worth of profit in a single quarter thanks to AI chips, and it’s hitting the gas from here on out: Nvidia will now design new chips every year instead of once every two years, according to Nvidia CEO Jensen Huang.
“I can announce that after Blackwell, there’s another chip. We’re on a one-year rhythm,” Huang just said on the company’s Q1 2025 earnings call.
May 22
Sean Hollister
Sales jumped 262 percent in Q1 2025 to hit a record $26B in revenue, of which nearly three-quarters ($19.4B) was data center compute — especially its Hopper GPUs for training LLMs and generative AI apps, says Nvidia. Gaming only accounted for $2.6 billion revenue this quarter.
Nvidia’s expecting record revenue again next quarter — $28B. Shovels in a gold rush, people.
May 14
Wes Davis
CEO Sundar Pichai just announced new Trillium chips, coming later this year, that are 4.7 times faster than their predecessors, as Google competes with everyone else building new AI chips. Pichai also highlighted Axion, Google’s first ARM-based CPU, which the company announced last month.
Google will also be “one of the first” cloud companies to offer Nvidia’s Blackwell GPU starting in 2025.
Correction: Axion was announced last month, not last year. Also, corrected the spelling of Axion.
May 9
Emilia David
Apple plans to start its foray into generative AI by offloading complex queries to M2 Ultra chips running in data centers before moving to its more advanced M4 chips.
Bloomberg reports that Apple plans to put its M2 Ultra on cloud servers to run more complex AI queries, while simple tasks are processed on devices. The Wall Street Journal previously reported that Apple wanted to make custom chips to bring to data centers to ensure security and privacy in a project the publication says is called Project ACDC, or Apple Chips in Data Center. But the company now believes its existing processors already have sufficient security and privacy components.
May 7
Richard Lawler
Apple — like Google, Meta, Microsoft, OpenAI and everyone else this side of Nvidia — is reportedly working on custom server hardware to power AI models as it prepares to introduce a slew of new features.
Over the past decade, Apple has emerged as a leading player designing chips for iPhones, iPads, Apple Watch and Mac computers. The server project, which is internally code-named Project ACDC—for Apple Chips in Data Center—will bring this talent to bear for the company’s servers, according to people familiar with the matter.
Apple watcher Mark Gurman followed up saying a similar-sounding project was canceled and it doesn’t make sense anyway: it would be too expensive, lack differentiation, and Apple prefers on-device AI.
Update: Added Gurman’s rebuttal.
[WSJ]
May 6
Emilia David
The Biden administration is taking applications for $285 million in federal funding — allotted from the $280 billion CHIPS and Science Act — seeking companies to “establish and operate a CHIPS Manufacturing USA institute focused on digital twins for the semiconductor industry.” The plan for the CHIPS Manufacturing USA institute to establish a “regionally diverse” network to share resources with companies developing and manufacturing both physical semiconductors and digital twins.
Digital twins are virtual representations of physical chips that mimic the real version and make it easier to test new processors before they’re put into production to find out how they might react to a boost in power or a different data configuration. According to the press release, digital twin-based research can also leverage tech like AI to speed up chip development and manufacturing in the US.
Apr 30
Sean Hollister
AMD also says an AI PC refresh cycle will help PCs return to growth in 2024, and that 150 software vendors will be developing for AMD AI PCs by year’s end. The company’s top priority is ramping AI data center GPUs, though, which are “tight on supply.” New AI chips are coming “later this year into 2025,” too.
Apr 15
Emilia David
OpenAI’s Batch API now lets users upload a file of bulk queries to the AI model, like categorizing data or tagging images, with the understanding that they won’t need immediate attention. Promising results within 24 hours lets them run when there is unused compute power, and keeps those pricey GPUs humming around the clock.
Apr 10
Emilia David
Meta promises the next generation of its custom AI chips will be more powerful and able to train its ranking models much faster.
The Meta Training and Inference Accelerator (MTIA) is designed to work best with Meta’s ranking and recommendation models. The chips can help make training more efficient and inference — aka the actual reasoning task — easier.
Apr 9
Emma Roth
Intel first introduced its Gaudi 3 AI accelerator last year, but now the company has revealed more details on performance. When compared to the H100 GPU, Intel says its Gaudi 3 accelerator can deliver “50% faster time-to-train on average across the Llama2 models” with better efficiency.
The company also says the Gaudi 3 AI Accelerator will be a “fraction of the cost” of Nvidia’s pricey H100. It will become available to companies like Dell, HPE, and Lenovo in the second quarter of this year.
Mar 29
Lauren Feiner
Reuters reports the list could strengthen the Commerce Department’s existing restrictions on US tech shipments to Chinese chip factories. The US government has voiced national security concerns about letting China access US technology to grow its own capabilities.
US companies have complained it’s difficult to know which Chinese factories produce advance chips and are subject to the restrictions, Reuters says.
[Reuters]
Mar 25
Emilia David
CNN gained entry into the training facility where TSMC teaches engineers to design and operate the machines that build semiconductors. The hope is to train engineers to “seed” factories it’s building in the US, Japan, and Germany.
TSMC told CNN that it needs “to hire thousands more” employees to staff facilities around the world as demand for advanced chips grows.
Mar 19
Richard Lawler
That’s what this NYT report says a16z and Saudi Arabia’s Public Investment Fund are considering, which may explain the rumors that Elon Musk (who denied them), Sam Altman, and others looking to the Middle East to fund their AI dreams.
It also recalls last fall’s news that the US government added an “additional licensing requirement” for Nvidia and AMD AI shipments to unspecified Middle Eastern countries.
Mar 18
Sean Hollister
Nvidia’s must-have H100 AI chip made it a multitrillion-dollar company, one that may be worth more than Alphabet and Amazon, and competitors have been fighting to catch up. But perhaps Nvidia is about to extend its lead — with the new Blackwell B200 GPU and GB200 “superchip.”
Nvidia says the new B200 GPU offers up to 20 petaflops of FP4 horsepower from its 208 billion transistors. Also, it says, a GB200 that combines two of those GPUs with a single Grace CPU can offer 30 times the performance for LLM inference workloads while also potentially being substantially more efficient. It “reduces cost and energy consumption by up to 25x” over an H100, says Nvidia, though there’s a questionmark around cost — Nvidia’s CEO has suggested each GPU might cost between $30,000 and $40,000.
Mar 6
Emilia David
A federal grand jury has indicted a Google engineer, Linwei Ding, aka Leon Ding, for allegedly stealing trade secrets around Google’s AI chip software and hardware on March 5th, before he was arrested Wednesday morning in Newark, California. Deputy Attorney General Lisa Monaco said in a statement that Ding “stole from Google over 500 confidential files containing AI trade secrets while covertly working for China-based companies seeking an edge in the AI technology race.”
Much of the stolen data allegedly revolves around Google’s tensor processing unit (TPU) chips. Google’s TPU chips power many of its AI workloads and, in conjunction with Nvidia GPUs, can train and run AI models like Gemini. The company has also offered access to the chips through partner platforms like Hugging Face.
Mar 5
Joanna Nelius
JEDEC Solid State Technology Association has released details about its new standard, which it says is better positioned to handle the demands of gaming, networking, and AI.
JESD239 GDDR7 is the first JEDEC standard DRAM to use the Pulse Amplitude Modulation (PAM) interface for high frequency operations. Its PAM3 interface improves the signal to noise ratio (SNR) for high frequency operation while enhancing energy efficiency. It also says GDDR7 has double the bandwidth over GDDR6, up to 192 GB/s per device, and double the number of channels.
We don’t expect to see GDDR7 out in the world until Nvidia and AMD release next-gen GPUs, which could come before the end of 2024 but is a long way from being confirmed.
[www.jedec.org]
Feb 27
Emilia David
Intel vice president David Feng said during Mobile World Congress that as part of the push to put AI into everything it builds, it will produce 40 million CPUs for AI PCs this year and 60 million in 2025, reports Nikkei Asia.
The “AI PC” concept includes Microsoft’s new CoPilot button plus Intel Core Ultra processors with built-in GPUs and neural processing units for AI models, which are now available as part of its vPro platform for business laptops.
[Nikkei Asia]
Feb 27
Emilia David
With over 600 statements of interest received, Commerce Secretary Gina Raimondo acknowledged today that the amount requested is more than twice the $28 billion the government has budgeted to invest.
We have decided to prioritize projects that will be operational by 2030. There are worthy proposals with plans to come online after 2030 that we say no to in order to maximize our impact in this decade…We anticipate that our investments in leading-edge logic chip manufacturing will put us on track to produce roughly 20% of the world’s leading-edge logic chips by 2030, up from the zero percent we produce today.
The CHIPS Act originally had $52 billion in subsidies to boost US semiconductor manufacturing, but it’s not nearly enough to catch up by itself — industry leader Taiwan Semiconductor Manufacturing Company (TSMC) earmarked $44 billion in 2022 just to expand its existing capacity.
[U.S. Department of Commerce]
Feb 23
Emma Roth
Nvidia has officially hit a $2 trillion market capitalization, making it the first chipmaker to cross the threshold. It’s also the third US company to reach over a $2 trillion valuation, right behind Apple ($2.83 trillion) and Microsoft ($3.06 trillion).
The California-based company has seen rapid growth over the past year due to its leadership in the AI chip market. Nvidia’s market cap reached $1 trillion only less than one year ago, and it left both Amazon and Alphabet in the rearview mirror as it became a $1.83 trillion company earlier this month.
Feb 22
Wes Davis
Intel will be producing custom chips, designed by Microsoft for Microsoft, as part of a deal that Intel says is worth more than $15 billion. Intel announced the partnership during its Intel Foundry event today. Although neither company specified what the chips would be used for, Bloomberg noted today that Microsoft has been planning in-house designs for both processors and AI accelerators.
“We are in the midst of a very exciting platform shift that will fundamentally transform productivity for every individual organization and the entire industry,” said Microsoft CEO Satya Nadella in the official press release.
Feb 21
Emilia David
As Nvidia reports its Q4 2023 earnings, CEO Jensen Huang says:
Accelerated computing and generative AI have hit the tipping point. Demand is surging worldwide across companies, industries and nations
That demand, and Nvidia’s dominance in AI chips, powered the company’s record earnings of $60.9 billion for the full year 2023 — a 126 percent increase from last year. Its Q4 revenue of $22.1 billion is an increase of an astounding 265 percent over the same period last year.
[NVIDIA Newsroom]
Feb 21
Emilia David
The open-source Gemma models are optimized for “the installed base of over 100 million Nvidia RTX GPUs,” installed in PCs around the world, in addition to Nvidia’s ubiquitous AI chips like the H100.
The models will also be part of Nvidia’s Chat with RTX demo, which lets AI models run locally and access users’ files to generate answers to prompts.
[NVIDIA Blog]
Feb 21
Sean Hollister
Intel has said 2025 is the year it leads the world in chips again (TSMC begs to differ).
Beyond that lies Intel 14A — the company’s smallest node yet, thanks to High-NA EUV. But Intel is hedging its bets with a tick-tock 18A successor, where the P in “18A-P” stands for performance jump. What does 14A stand for? Intel didn’t say. “We think this is the next leadership technology and we don’t want to give anyone something to shoot at,” Intel VP Craig Orr tells me.
/ Sign up for Verge Deals to get deals on products we’ve tested sent to your inbox weekly.
The Verge is a vox media network
© 2024 Vox Media, LLC. All Rights Reserved
