本文由半导体产业纵横（ID：ICVIEWS）编译自technologyreview
What will the chip of the future look like? Who will make them and what new technologies will they unlock?
Due to the vigorous development of artificial intelligence, the chip world is on the cusp of a huge tide shift. There is an increasing demand for chips that can train AI models faster and operate on them from devices such as smartphones and satellites, allowing us to use these models without revealing private data. Governments, tech giants, and start-ups are all vying for a piece of the growing semiconductor market.
Here are four trends to watch in the year ahead that will define what the chips of the future will look like, who will make them, and what new technologies they will unlock.
CHIPS operates around the world
Two of the world's largest chipmakers, TSMC and Intel, are racing to build campuses in the desert that they hope will become the seat of America's chipmaking prowess. One common denominator of these efforts is funding: In March, President Joe Biden announced $8.5 billion in direct federal funding and $11 billion in loans for Intel's nationwide expansion. A few weeks later, TSMC announced a $6.6 billion investment.
The awards are just part of the subsidies that poured into the chip industry in the U.S. through the $280 billion CHIPS and Science Act signed in 2022. This funding means that any company that ventures into the semiconductor ecosystem is analyzing how to restructure its supply chain to benefit. While much of the funding is aimed at boosting U.S. chip manufacturing, there is still room for other players to apply.
The U.S. isn't the only country trying to localize part of its chip manufacturing supply chain. Japan itself has spent $13 billion, equivalent to the CHIPS Act, Europe will spend more than $47 billion, and earlier this year, India announced $15 billion to build its own chip factory. According to Chris Miller, a professor at Tufts University and author of "The Chip Wars: The Battle for the World's Most Critical Technologies," the roots of this trend can be traced back to 2014. Since then, China has begun offering huge subsidies to its chipmakers.
"This has created a dynamic where other governments have concluded that they have no choice but to offer incentives or see businesses move manufacturing to China," he said. This threat, coupled with the proliferation of artificial intelligence, has led Western governments to fund alternatives.
Miller said the funding is unlikely to spawn entirely new chip competitors or become the largest chip maker through restructuring. Instead, it will primarily incentivize leading players like TSMC to put down roots in multiple countries. But funding alone wasn't enough to do that quickly — TSMC's efforts to build a factory in Arizona bogged down in missed deadlines and labor disputes, and Intel similarly failed to meet its promised deadlines. It is unclear whether the equipment and workforce of these factories will be able to match the level of advanced chip manufacturing that these companies have abroad, whenever they are put into operation.
"Supply chains are only going to slowly shift over years or even decades," Miller said. But the situation is changing. ”
More AI at the edge
Currently, most of our interactions with AI models like ChatGPT are done through the cloud. Relying on the cloud has some drawbacks: on the one hand, it requires internet access, and it also means that some of your data is shared with the mockuper.
That's why there's a lot of interest and investment in AI edge computing, where the process of AI models happens directly on your device, such as a laptop or smartphone. As the industry becomes more committed to AI models, helping us understand a lot about the future, all of this requires faster "edge" chips that allow models to run without sharing private data. These chips face different limitations than data center chips: they often have to be smaller, cheaper, and more energy-efficient.
The U.S. Department of Defense is funding a lot of research into fast, private edge computing. In March, its research arm, the Defense Advanced Research Projects Agency (DARPA), announced a partnership with chipmaker EnCharge AI to create an ultra-powerful edge computing chip for AI inference. EnCharge AI is working on a chip that enhances privacy but runs with very low power consumption. This will make it suitable for military applications such as satellite and off-grid surveillance equipment. The company expects to ship the chips in 2025.
AI models will always rely on the cloud for certain applications, but new investments and interest in improving edge computing can bring faster chips to our everyday devices, leading to more AI. If edge chips become small enough and cheap enough, we may see more AI-powered "smart devices" in homes and workplaces. Today, AI models are largely confined to data centers.
Naveen Verma, co-founder of EnCharge AI, said, "Many of the challenges we face in the data center will be overcome. I expect people to focus on the edge areas. I think this is critical to enabling AI at scale. ”
Big tech companies join the chip manufacturing space
Companies are paying high computational costs to create and train AI models for their businesses. For example, models that employees can use to scan and aggregate documents, as well as external-facing technologies such as virtual agents. This means that the need for cloud computing to train these models is increasing dramatically.
The companies that provide most of the computing power are Amazon, Microsoft, and Google. For years, these tech giants have dreamed of increasing their profit margins by making chips for their own data centers, rather than buying chips from companies like Nvidia. Amazon began its efforts in 2015 with the acquisition of startup Annapurna Labs. Google launched its own chip TPU in 2018. Microsoft launched its first AI chip in November, and Meta launched a new version of its own AI training chip in April.
This trend could cost Nvidia an edge. But in the eyes of big tech companies, Nvidia doesn't just play the role of a competitor: Regardless of the cloud giant's own internal efforts, their data centers still need their chips. This is partly because their own chip manufacturing efforts can't meet all of their needs, but also because their customers want to be able to use top-of-the-line NVIDIA chips.
"It's really about giving customers choice," says Rani Borkar, head of Azure hardware at Microsoft. She said she can't imagine a future where Microsoft will provide all the chips for its cloud services: "We will continue to maintain strong partnerships and deploy chips from all of our chip partners that we work with." ”
While cloud computing giants are trying to wrest some market share from chipmakers, Nvidia is also trying the opposite. Last year, the company launched its own cloud service so that customers can bypass Amazon, Google, or Microsoft. As the battle for market share unfolds, the year ahead will be about whether customers see Big Tech's chips as similar to Nvidia's most advanced chips.
Nvidia vs. startups
Despite Nvidia's dominance, there has been a wave of investment going to startups that aim to surpass Nvidia in some areas of the chip market in the future. These startups all promise to deliver faster AI training, but they have different ideas about which computing technology will enable this, from quantum to photonics to reversible computing.
Many companies, such as SambaNova, Cerebras and Graphcore, are trying to change the underlying architecture of the chip. Imagine an AI accelerator chip that constantly needs to move data back and forth between different regions: a piece of information is stored in the memory area, but it must be moved to the processing area, where the calculations are performed, and then stored back in the memory area for safekeeping. It all takes time and effort.
Improving the efficiency of this process will provide customers with faster and cheaper AI training, but only if the chipmaker has good enough software to allow AI training companies to seamlessly transition to new chips. If the software transformation is too unwieldy, model makers like OpenAI, Anthropic, and Mistral may stick with large chip makers. This means that companies that take this approach, such as SambaNova, are spending a lot of time not only on chip design, but also on software design.
Onen, founder of chip startup Eva, proposes a deeper change. Instead of using traditional transistors, which for decades have provided greater efficiency by getting smaller, he uses a new component called proton-gated transistors, which he says Eva is specifically designed for the needs of AI training. It allows devices to store and process data in the same place, saving time and computing energy. The idea of using such components for AI inference dates back to the 60s of the 20th century, but researchers have never been able to figure out how to use it for AI training, in part due to material barriers – it requires a material that, among other qualities, is able to precisely control conductivity at room temperature.
One day, in the lab, "by optimizing these numbers, and being very lucky, we got the material we wanted." Onen said. After months of trying to confirm that the data was correct, he founded Eva, which was published in the journal Science.
But in this space, many founders have promised to topple the dominance of leading chipmakers, but have failed, and Onen admits that it will be years before he knows if the design works as intended and whether the manufacturer will agree to produce it.
*Disclaimer: This article was created by the original author. The content of the article is his personal point of view, and our reprint is only for sharing and discussion, and does not mean that we agree or agree, if you have any objections, please contact the background.