Next month, Amazon is expected to announce widespread availability of its proprietary AI chips, Trainium2, which are used to train large language models. Amazon’s foray into chip design demonstrates a wider trend for Big Tech hyperscalers including Microsoft, Meta and Alphabet to enter the business of semiconductors.
Big Tech has been panic buying AI chips, since GenAI’s emergence in 2022, to reduce reliance on the world’s leading AI chip designer, NVIDIA. In November 2024, Nvidia’s 90% monopoly on the supply of AI chips saw it overtake Apple to become the world’s most valuable company, as its stock price almost tripled throughout the year.
Go deeper with GlobalData
Driven by the AI boom, Amazon’s cloud business AWS posted $10.4bn in operating income for Q3 2024, up 50% year-on-year, and $27.4bn in revenue, up 19%. Amazon CEO, Andy Jassy, commented on the company’s planned expenditure for 2025: “the majority of it is for AWS, and specifically, the increased bumps here are really driven by generative AI.”
Jassy noted that Amazon’s AI business is growing three times faster than AWS did at the same stage of its evolution. Without AI chips, there is no generative AI development. And given that AWS has long been Amazon’s biggest profit centre, it is no surprise that Amazon and its Big Tech competitors are in a race to ensure AI chip supply.
Amazon is developing two types of AI chips, Trainium2 and Inferentia2. The Trainium AI accelerator is a machine learning chip built for deep learning training of models with more than 100b parameters. Inferentia was developed as a high-performance low-cost AI chip for deep learning and GenAI applications, according to GlobalData’s Artificial Intelligence Executive Briefing (fourth edition).
Big Tech chip designs
Meta’s in-house, custom AI accelerator family, the MTIA (Meta Training and Inference Accelerator) is designed for the company’s internal workloads and is used in conjunction with GPUs to improve performance and efficiency and decrease latency. In April 2024, Meta announced the latest chip in its MTIA family, Artemis. The chip will be manufactured by TSMC and Meta hopes it will replace Nvidia’s AI chips, according to GlobalData’s AI Executive Briefing.
How well do you really know your competitors?
Access the most comprehensive Company Profiles on the market, powered by GlobalData. Save hours of research. Gain competitive edge.
Thank you!
Your download email will arrive shortly
Not ready to buy yet? Download a free sample
We are confident about the unique quality of our Company Profiles. However, we want you to make the most beneficial decision for your business, so we offer a free sample that you can download by submitting the below form
By GlobalDataGoogle has two AI chips, Tensor Processing Units (TPUs) and Axion Processors. The Cloud TPUs are custom-designed AI accelerators, used to train large AI models. They are designed specifically for neural networks and to scale cost-efficiently. In April 2024, Google announced its Axion AI processor which is claims will be 30% faster than the leading ARM CPU. In May 2024, Google unveiled its sixth generation TPU, Trillium. Google claims the chip will be 67% more energy efficient than TPU v5e with 4.7 times more peak compute performance per chip.
In November 2023, Microsoft unveiled its first custom AI chip, Azure Maia 100, built for its Azure cloud service and is said to optimise GenAI tasks. Microsoft also unveiled its first microprocessor built in-house for cloud computing, the Azure Cobalt 100. The company is the last of the big three cloud service providers to develop their in-house chips, according to GlobalData’s Artificial Intelligence Executive Briefing.
Big Tech’s AI chip manufacturing risk
A semiconductor supply chain squeeze following the Covid-19 pandemic remains a grave warning against overreliance on a single company for semiconductors, in that case TSMC which manufactures 95% of the global supply of advanced semiconductors.
While Big Tech’s effort to avoid soaring AI chip prices and, sometimes, up to 12-months delay in orders being completed, the risk still stands even if Big Tech companies design their own chips. Chip design can be brought in-house, but Big Tech still needs a reliable manufacturing supply and packaging chain to access the volume of chips required for innovation, growth and globally competitiveness. TSMC’s expertise in AI chip fabrication means that it remains the sole manufacturer for 95% of the world’s advanced chips.
Chinese encroachment across the Taiwanese straits and a number of domestic factors in Taiwan including energy blackouts, seismic activity and workforce fluctuations, mean that Big Tech cannot eliminate the risk of another global semiconductor supply squeeze.
The Biden administration’s $53bn CHIPS ACT, signed in August 2022, offered investment incentives for semiconductor manufacturers. TSMC has since invested $65bn in three Arizona fabs, Samsung $17bn in a Texas fab, with Intel expected to invest over $90bn by the end of the decade in four new fabs across Oregon, Arizona, Ohio, and New Mexico.
According to US Semiconductor Industry Association estimates, in May 2024, the country is on course to triple its semiconductor manufacturing capacity by 2032. Questions still remain around whether this level of production will be sufficient for rapidly advancing AI, as well as concerns about short-term shortages.
Is Big Tech DIY on AI chips a long-term trend?
While Big Tech ramps up its AI chip design operations, the trend for designing chips in-house actually started around 2010 and is here to stay, according to GlobalData senior consultant analyst, Mike Orme. Apple ended its relationship with US chip maker Intel as supplier of its iPhone processors in favour of starting its own captive chip design operation at that time, notes Orme.
Apple’s walled garden reputation extends to chips. The company’s chips are not for sale and are optimised for its own operating systems. “Intel hadn’t been able to meet its demands on the energy efficiency front. The result is that Apple has built a world-beating captive chip operation enhanced now by its high performance, energy efficient M Series chips for its Mac laptops,” said Orme.
Around the same time, Google developed its own TPU family of accelerator chips, which now challenge Nvidia in terms of performance, notes Orme. “Indeed, when training its Apple Intelligence LLM, Apple chose TPUs over Nvidia’s GPUs,” he adds.
Increasingly at play, is Moore’s Law – the idea that compute power doubles exponentially every two years along with the increasing number of transistors held on a microchip. According to Orme, a hardware market for general purpose processors is being progressively replaced by a software-first, application specific, proprietary chipset model by those that can design their own, namely, Big Tech.
