① Increasingly evident signs indicate that the global race for dominance in the burgeoning artificial intelligence market is driving technology giants to adopt or emulate business models akin to IBM's approach in the 1960s... ② These "AI powerhouses," including Google, Meta, Microsoft, and Amazon, are at various stages of independently developing custom AI chips.
Cailian Press News on February 9 (Editor: Xiaoxiang) Increasingly evident signs indicate that the global race to dominate the burgeoning artificial intelligence market is driving tech giants to adopt or emulate business models akin to those from the 1960s... $IBM Corp (IBM.US)$ ...
Including $Alphabet-C (GOOG.US)$ 、 $Meta Platforms (META.US)$ 、 $Microsoft (MSFT.US)$ and $Amazon (AMZN.US)$ These 'AI giants' (hyperscalers), including companies such as Google, are at various stages of independently developing custom AI chips. These chips will be deployed in their data centers to provide computational power for cloud services and software products. The most advanced among them, Google, is reportedly even in talks to sell its TPU chips to Meta Platforms, a move that would place it in direct competition with leading chipmakers like NVIDIA...$NVIDIA (NVDA.US)$...
These developments have prompted industry analysts to predict that the custom AI chip market could grow to $122 billion by 2033.
Moreover, the supply of proprietary components by these AI hyperscalers is no longer limited to chips—Jonathan Atkin, an analyst at Royal Bank of Canada Capital Markets, noted in a recent report to clients that Microsoft and Amazon are actively investing in “dark fiber,” referring to optical fiber cables that have been laid but remain unused.
Atkin wrote that while Google and Meta also own their own fiber optic cables, they still purchase from third parties. These cables are essential for connecting the companies’ data centers and serving enterprises that rely on these centers.
Many industry insiders point out that the trend of cloud service providers producing their own hardware components to support core software products signifies a return to vertical integration in Silicon Valley—a model pioneered by oil and steel magnates in the late 19th century and later adopted by IBM during the digital revolution.
Replicating IBM from 60 Years Ago?
In the 1960s, IBM, one of the most successful vertically integrated companies, produced its own hardware components for large computer systems. IBM’s strategy was based on the idea that manufacturing proprietary components would enhance the performance and profitability of its end product (IBM mainframes)—at a time when there were concerns about potential shortages of early computer parts.
This strategy proved effective: Economist Carliss Y. Baldwin noted in her book Design Rules that by 1985, IBM accounted for more than half of the total market capitalization of the computer industry.
Of course, all of this eventually collapsed. In the 1990s, with the decline in semiconductor manufacturing costs and the rise of software giant Microsoft and chip leader Intel, IBM's once-powerful 'moat' was eroded. By 2000, the company no longer claimed to be vertically integrated.
Just as the rise of computers propelled IBM toward vertical integration, the proliferation of artificial intelligence since the advent of ChatGPT in late 2022 is pushing today’s cloud computing giants onto a similar developmental trajectory. Particularly, the high cost and supply shortages of NVIDIA chips have driven tech giants to accelerate their plans for developing proprietary AI chips. These custom chips are not only less expensive but also better optimized for integration with their own software systems.
"Hyperscale cloud providers have realized that relying on a single supplier for AI computing poses significant strategic risks," noted Jay Goldberg, an analyst at Seaport. "As a result, they now have a strong strategic incentive to develop their own chips."
Custom chips emerge in abundance
It was reported that Meta began testing its self-developed AI training chips last year and recently acquired the chip startup Rivos to expedite its custom semiconductor development process.
Google’s TPU chip technology has advanced to such an extent that Anthropic, OpenAI, and even rival Meta have signed major cloud service agreements with it to gain access to these chips.
After prolonged delays, ... $Microsoft (MSFT.US)$ Amazon also released its new generation Maia 200 chip in January this year.
During a recent visit by industry insiders to Amazon’s chip lab in Austin, Texas, and an adjacent testing center, the company showcased its latest UltraServer cluster. This server cluster is equipped with Amazon’s newest generation of proprietary AI chips, Trainium, CPU processors Graviton, and customized network cables and switches connecting these components.
Despite $Amazon (AMZN.US)$ Remote data centers still primarily offer AI computing services based on NVIDIA GPUs rather than proprietary accelerators, but this tech giant is increasingly emphasizing the advantages of its self-developed hardware. $NVIDIA (NVDA.US)$
Paul Roberts, Chief Technology Officer of Amazon Web Services, revealed that compared to GPU-based inference workloads, the Trainium3 chip can deliver up to a 60% cost-performance advantage for cloud customers. "Market validation shows that custom chip solutions—compared to general-purpose GPUs—can achieve remarkable improvements in energy efficiency through dedicated processors and accelerators."
As the AI data center boom begins to be affected by power constraints, such energy-saving advantages will become increasingly prominent.
Beyond chips, from large-scale adoption of optical circuit switches (OCS) like those introduced by Google, to tech giants making massive acquisitions of energy and power companies, this comprehensive integration implies that future competition will no longer be solely about model superiority but rather a holistic contest over 'who possesses more physical resources (electricity, networks, land, chips).'
This fervor for 'full-stack' vertical integration is most extreme in Musk's xAI company. In an effort to completely摆脱 external supply chain constraints, xAI has not only built the world’s largest supercomputing cluster, Colossus, in Memphis at an unprecedented speed but has also extended its integration efforts to the energy infrastructure. By creating a closed loop of 'self-developed models + proprietary supercomputing + self-supplied power,' it aims to bypass public grid capacity limitations. There are even rumors that it plans to deploy computing power into orbit using SpaceX's Starship, achieving a truly physical closed loop.
However, Seaport analyst Goldberg believes that the trend toward vertical integration is nearing its 'limit,' and not all tech giants will ultimately succeed.
"If you want to design a leading chip, that is an enormous expense," he pointed out. "Only a handful of companies can afford it."
Editor/Rocky
