Semiconductor Industry News - February 2025

DeepSeek and Stargate Represent the Extreme Ends of the AI Industry - February 7th, 2025

In a landmark move poised to reshape the artificial intelligence (AI) landscape, OpenAI, in collaboration with industry giants, has unveiled the Stargate Project—a $500 billion initiative to construct advanced AI infrastructure over the next four years. This venture seeks to solidify the United States' leadership in AI by developing state-of-the-art infrastructure for future AI systems within the country.

Concurrently, the AI sector is experiencing significant disruption from DeepSeek, a Chinese startup that has rapidly ascended to prominence. DeepSeek's latest model, DeepSeek-R1, has demonstrated advanced reasoning capabilities comparable to leading U.S. AI models but achieved at a fraction of the cost and energy consumption. DeepSeek has challenged the prevailing belief that cutting-edge AI requires massive computational resources and triggered substantial market reactions, including a nearly 18% decline in Nvidia's stock value.

Stargate’s Plan Comes with a $500 Billion Investment

Days into Donald Trump's second presidency, a new, massive AI project is underway that, if it succeeds, could redefine the entire American artificial intelligence industry.

Named Stargate, this high-profile AI infrastructure project is being touted as a massive step forward in securing American leadership in AI, generating significant economic benefits globally, and creating thousands of American jobs.

SoftBank, OpenAI, Oracle, and MGX are Stargate's initial equity founders. SoftBank and OpenAI will be the lead partners for Stargate, with the former having financial responsibility and the latter operational responsibility. Arm, Microsoft, Nvidia, Oracle, and OpenAI will be the key initial technology partners, and they plan to collaborate closely on this computing system.  

According to OpenAI’s announcement, the buildout is underway in Texas, with evaluations of potential sites for more campuses before definitive agreements are finalized.

Trump has already lauded the initiative at a White House event OpenAI Chief Sam Altman and other tech executives attended as “a resounding declaration of confidence in America’s potential.”

However, the road ahead may not be as smooth as it seems. The Financial Times reports that Stargate hasn’t yet secured funding, nor will it receive government financing. Likewise, it will only serve OpenAI once it is finished.  

“The intent is not to become a data center provider for the world. It’s for OpenAI,” said one familiar with the initiative.

Another person told the Financial Times that the plan was far from developed. “They haven’t figured out the structure, they haven’t figured out the financing, and they don’t have the money committed.”

Over the last few years, Altman has been working with his team to boost OpenAI’s access to data and computing power, which is a consistent bottleneck for many large language models (LLMs). While Altman has dedicated the past year to fleshing out his infrastructure plans, the idea of announcing it at the White House came as a surprise to many familiar with the project.

Some have commented that the announcement is just due to the inauguration of a new president. While Trump has verbally supported the project, new policies could later contradict Stargate’s tentative plans.  

In the meantime, work is already beginning in Abilene, Texas. Oracle and data center start-up Crusoe will supply around $10.4 billion combined, which will go toward financing and components for the Texas site. Microsoft will provide the computing power at the Texas facility, which will power OpenAI.

The hefty number of chips required to accomplish Stargate’s dream could significantly affect the global supply chain, leading to shortage conditions and higher price trends. Companies looking to secure GPUs in the future might want to begin contacting their suppliers to see if this will affect their buy-ahead orders.  

DeepSeek-R1 Turns AI Industry Upside Down

Since the debut of OpenAI’s ChatGPT, companies and countries have been clamoring to become the next big name in the AI market. Over the last two years, the market has seen its fair share of flops as tech giants and start-ups hammered out their chatbots' flaws.  

Many considered China’s attempt at a ChatGPT equivalent, made by Baidu, a disappointment. With the U.S. passing increasingly tougher restrictions on AI, Nvidia–currently the leading player in the GPU market–products got harder to come by. It seemed impossible to make an AI model with a small amount of Nvidia chips and a tight budget when OpenAI and others poured billions into their own.  

That was until DeepSeek threw its proverbial bomb into the mix.

Industry experts have compared DeepSeek-R1, the startup’s latest AI, to OpenAI's latest AI model, o1. With only 10,000 A1 Nvidia GPUs and $5 million, DeepSeek-R1 is on par with OpenAI’s latest AI model, o1. According to Reuters, it's also 20 to 50 times cheaper to use than OpenAI’s o1 model, depending on the task.

How is that possible? Well, according to some critics, it shouldn’t be.

Alexander Wang, CEO of Scale AI, says that China, including the fund behind DeepSeek, probably has more Nvidia GPUs than they let on but, due to export controls, are staying quiet.

However, Emad Mostaque, founder of Stability AI, claims that not only is DeepSeek not faking the cost and amount of GPUs, but if they had better optimized H100 chips, they could have trained both V3 and R1 models for under $2.5 million.  

Likewise, because DeepSeek made R1 open source, that means that anyone can modify, distribute, and commercialize the AI model. Should DeepSeek be lying, others will quickly find out.  

Other critics are concerned about the impact R1’s low cost could have on the industry. Specifically, some believe the most significant threats will be to U.S. equity markets, as it calls billions of capex into question. Other experts argue that while the cost to train the model could continue to decline, the expenses that usually make up the lion’s share of the cost come from a model’s inference or the AI running in the wild.

Likewise, these technological advancements could quickly encounter a historical trend that has impacted every innovation: Jevons's Paradox. At the “heart” of Jevons' Paradox, the idea states that as efficiency rises, the cost to make a resource decreases, but the price increases due to high demand and total resource consumption.

So even if AI models became less costly, the billions poured into their upkeep and continued development will only become more expensive.  

When it comes to the future of AI, one can be sure that DeepSeek isn’t the first and won’t be the last to shake up an exploding industry. We are only beginning to scratch the surface of the possibilities associated with AI, and industry disruptors like DeepSeek are some of the primary drivers of innovation and new ideas. It will be interesting to see how the industry evolves now that it considers what can be done under constrained circumstances.