AI adds logic to reinventing the silicon wheel

MEDIATEK

17/11/2025 10:31

The author is a Reuters Breakingviews columnist.  The opinions expressed are her own.

By Robyn Mak

HONG KONG, Nov 18 (Reuters Breakingviews) - One of the biggest surprises of the artificial intelligence boom is that everyone wants to be a chipmaker. For decades, tech companies outsourced the development of semiconductors to specialists like Nvidia NVDA.O because it was cheaper and more efficient. The rise of AI has changed that calculation: giants like Alphabet GOOGL.O, Amazon.com AMZN.O, Microsoft MSFT.O and Meta Platforms META.O are all stepping up their in-house silicon efforts. There is some logic to reinventing the wheel.

Apple AAPL.O blazed a trail back in 2010 when it launched the iPhone 4 containing its own silicon. At the time, the handset maker was an outlier. But the $4.6 trillion Nvidia's dominance in developing the cutting-edge graphics processor units (GPUs) vital to training AI models has forced an industry-wide rethink.

Tech tycoons are nervous about relying on a single supplier for such a critical component. Nvidia's near-monopoly gives it enormous pricing power, too. Its data centre division is forecast to earn a 74% gross profit margin on $187 billion of sales in the year to January 2026, per Visible Alpha.

The AI arms race also requires ever-larger quantities of computing power. Despite Nvidia's rapid advances, the cost of training frontier AI models has grown 2.4 times annually since 2016, one study foundlast year. If that growth continues it could cost more than $1 billion to train the largest models by 2027, adding extra impetus for companies to find cheaper hardware.

That is why the world's largest cloud computing suppliers - Alphabet, Amazon, and Microsoft - as well as Facebook owner Meta have been developing their own custom processors, known as application-specific integrated circuits (ASICs). These semiconductors are designed for a single purpose, differentiating them from off-the-shelf offerings from Nvidia or Advanced Micro Devices AMD.O .

Alphabet-owned Google, for example, developed its Tensor Processing Units over a decade ago for a speech-to-text function. Today Google deploys TPUs for training AI models as well as responding to queries from users - a process known as inferencing. Microsoft, Amazon and Meta have followed suit. Last month, a Meta executive confirmed plans to buy chip startup Rivos to bolster the group's in-house capabilities.

One benefit of the do-it-yourself drive is that it gives companies greater control over hardware which can then be optimised and tailored for their particular cloud computing platforms and AI algorithms. Microsoft argues its "silicon to service" strategy of customising a data centre from top to bottom - including software, servers, racks and cooling systems - can help boost performance and efficiency for users of its Azure cloud service.

Another advantage is that it is relatively easy to "make" a custom chip. That's because companies leave the challenge of actually printing the semiconductors onto silicon wafers to suppliers like $1.6 trillion Broadcom AVGO.O and smaller peers like Marvell Technology MRVL.O and MediaTek 2454.TW. These companies help design and engineer the processors, work with contract-chipmakers like Taiwan Semiconductor Manufacturing 2330.TW, procure memory components and more. The AI ASIC market is set to grow as much as 55% annually to reach $60 billion in 2028, Bernstein analysts forecasted earlier this year, though that figure still pales next to the $375 billion in sales Nvidia is expected to generate in a similar period, based on Visible Alpha estimates.

 It's not only U.S. tech firms that are embracing DIY. Chinese giants like Alibaba 9988.HK and Baidu 9888.HK are developing their own semiconductors as part of the country's push to become less dependent on American technology. Carmakers are eyeing custom silicon in autonomous driving initiatives. The boom has prompted Japan's SoftBank 9984.T, which owns chip design firm Arm O9Ty.F, ARM.O, to explore a takeover of the $75 billion Marvell, Bloomberg reportedin November, citing sources.

In October OpenAI, the developer of the ChatGPT chatbot, said it will jointly develop 10 gigawatts of custom AI accelerators with Broadcom. Though the two companies did not disclose financial terms, Nvidia boss Jensen Huang recently stated that one gigawatt of data center capacity translates into $35 billion of revenue for his company. That implies OpenAI's partnership represents $350 billion of foregone sales for Nvidia.

Yet diverting sales from Nvidia does not necessarily mean the buyers are better off. A sizeable share of any potential savings OpenAI enjoys will probably go to Broadcom, which Bernstein analyst Mark Li reckons earns a gross margin of up to 60% on its ASIC business. Over time, companies may acquire enough chipmaking expertise to work directly with manufacturers like TSMC, as Apple does today. Google looks the furthest along: Bernstein's Li believes the company can handle the core "front-end" chip designs in-house and will soon outsource back-end designs and peripheral components to Taiwan's MediaTek, which charges less than Broadcom, producing annual cost savings of $5 billion.

Apple's 40% gross margin for iPhones suggests its vertical integration has paid off. The case for cloud computing giants doing the same looks less clear-cut, though. Using custom-made chips that are cheaper but less powerful might end up costing more after factoring in land, electricity costs and other infrastructure outlays.

Key to that argument is Nvidia's technological edge. A December analysis from Morgan Stanley found that the latest Blackwell GPUs can boast 1.9 times more performance per watt over Amazon's Trainium 2 chips when training AI models, based on certain assumptions. In other words, Amazon will need nearly twice as much electricity to achieve the same level of computing power. This explains why the biggest tech firms will remain Nvidia's top customers for the foreseeable future.

Even so, the economics of ASICs can work provided sufficient scale. Take AI inference, which does not require as much computing power. Installing a cluster of 24,000 Blackwell processors means upfront hardware costs of $852 million, Morgan Stanley estimated, leading to annual depreciation charges of $213 million over four years. A same-sized cluster of Google TPUs will cost $99 million in hardware, and $25 million in annual depreciation. Developing chipmaking capabilities will also enable companies to drive a harder bargain with Nvidia.

Ultimately, investors will be paying attention to the return tech companies can earn on their investments. Lower upfront and operating costs might help, but ASICs still have to be competitive in performance and generate revenue. That will determine whether it was worth reinventing the silicon wheel.

Follow Robyn Mak on X.

Broadcom's stock has outperformed Nvidia's https://www.reuters.com/graphics/BRV-BRV/gdpzjwbjepw/chart.png

Nvidia's AI-related revenue dwarfs Broadcom's https://www.reuters.com/graphics/BRV-BRV/dwpkqalyapm/chart.png

Nvidia's off-the-shelf AI chips are more energy efficient https://www.reuters.com/graphics/BRV-BRV/zdpxjrkjjpx/chart.png

(Editing by Peter Thal Larsen; Production by Aditya Srivastav)

((For previous columns by the author, Reuters customers can click on MAK/ robyn.mak@thomsonreuters.com))