home...

Week 15 - Dual Use Technologies - Chips, China, and AI

September 2024 (2942 Words, 17 Minutes)

This piece is part two of a three-part series on American industrial policy at the cutting edge of technology. This week, we’ll examine the latest moves the US has made to retain its dominance over China in the AI space, particularly in the realm of chips.

AI

The American tech sector has seemingly secured dominance in the AI space (as of time of writing). Each of the “Big Seven” American tech companies has their own extremely well-resourced AI-related venture, with some making it a keystone of their business strategy moving forward. Google has Gemini, Amazon has made significant investments into GPUs for AWS EC2 instances and unveiled Amazon Bedrock, Apple has the upcoming Apple Intelligence, Meta has the Llama open-source model suite, Microsoft has made cozy bedfellows with OpenAI and rolled out Copilot, Tesla (well, really Musk) has xAI, and of course Nvidia’s stock graph went nearly vertical by selling them all GPUs.

GPUs, or Graphical Processing Units, are specialized hardware originally developed for computing the vast matrix multiplications used in graphics processing. A typical CPU in a computer is able to do practically any operation, whereas a GPU can do only a few types of operations, typically elementary matrix operations, thousands of times faster than a CPU by running calculations simultaneously.

The semiconductor industry was shaken by Nvidia’s meteoric rise to global importance as the demand for GPUs increased. Given Nvidia’s stock price, I’m surprised the company is still running. If I were a senior Nvidia engineer with five years seniority, I would have already sold out and bought a cabin in Wyoming with two big dogs, a horse, an illegally imported 2020 Toyota Hilux, and absolutely no internet connection, so I can never again accidentally see old ISIS beheadings on Twitter in public or asked to be a CTO of a San Francisco-based AI startup by a 19-year old Stanford waitlister from Hungary who doesn’t know how to swing a hammer.

Back on topic: China hasn’t simply submitted to an American-led AI landscape. Baidu, China’s Google, has rolled out their own Chatbot named ERNIE 4.0. The Amazon of Asia, Alibaba, has also developed an open-source 110B parameter model called Qwen 1.5. However, Chinese companies lack the GPUs that their American counterparts have in surplus and are unable to manufacture them domestically. While consumer-grade GPUs that slot into a desktop are much smaller and simpler than the server GPUs used in enterprise applications, they can indicate a company’s sophistication and ability to deliver high-end enterprise chips. Nvidia has the best consumer-grade GPUs on the market and is the industry leader in the GPU space. The best of China’s domestically produced gaming GPUs match up to where Nvidia was a full decade ago.

We’ll return to hardware in a second, but imagine you’re an American policymaker. Perhaps you’re part of the Congressional-Executive Commission on China and have read this report from Australian National University that details how the development of data collection and processing facilitated by AI could render the stealth of nuclear-powered ballistic submarines useless, effectively making the ocean transparent. You now have a very provocative and convincing reason to ensure continued American dominance in the AI space that will make any General turn their head.

You now have two goals to pursue for furthering American dominance in the AI space. You could ensure that the United States is attracting the best AI experts in the world, and you could restrict your adversaries’ access to GPU hardware. Attracting talent is the easy part, especially with plenty of influential people in the tech sector making their dissatisfaction with the visa process for highly skilled workers known. American soft power has resulted in 50 million immigrants calling my country home and bolstering our economy, in turn making the US a more attractive place to emigrate to. The same cannot be said for China, which, despite its population of over 1.4 billion, only has just over 1.4 million foreigners, about 600 thousand of that number being from the Special Administrative Regions of Macau and Hong Kong. As far as I’m concerned, my friend in Singapore bought this Chappell Roan shirt after seeing me wear it in a Discord call for the same reason that the US will never be short of the best AI talent: the US is the center of the world culturally and economically.

In terms of hardware, American companies still lead the pack. Juggernauts like Nvidia and AMD are years ahead of Chinese products. This leads to the question, “Why don’t the Chinese import GPUs?”. The answer lies in American export controls restricting GPU sales to adversaries. In my previous article, I discuss the International Traffic in Arms Regulations (ITAR), a set of regulations used by the Government to restrict the export of dual-use technology to further American national security objectives set by the Department of State. Dual-use technologies are items that have both military and civilian use cases. A big chunk of the restricted items in ITAR are listed in the US Munitions List, like aircraft, submarines, training equipment, and ammunition. The Export Administration Regulations (EAR) accomplishes a similar goal but has a broader scope which includes economic goals, not only national security goals. As of yet, ITAR does not contain guidance on anything artificial intelligence related, including the dispersal of models or GPU hardware. The EAR has instead become the instrument used to control AI proliferation. The Department of Commerce recently updated their regulations to ban the export of GPUs past a certain performance threshold to adversaries. This ban affected the Nvidia’s A100 GPU with its measly 19.5 TFLOPS (Tera-Floating point operations per second, basically 19.5 trillion calculations a second) and the H100 with its beefy 67 TFLOPS. Note that these two units were released 34 months apart. In 34 months, Nvidia was able to more than triple the performance of their flagship enterprise product, which is no small task.

GPUs are just about the most complicated products in semiconductors. Some products like DRAM (Dynamic RAM) are completely commoditized. Despite their simplicity relative to other semiconductor products, DRAM still requires incredibly complex supply chains and niche expertise. Over the last few decades, multiple companies have developed the capability to manufacture RAM profitably and have created a healthy ecosystem that benefits the consumer with plenty of commodity components to choose from. A RAM chip from SK Hynix is essentially the same product as RAM from Micron. If you’re older than 30, you probably remember days when four gigs of RAM sticks costed multiple hundreds of dollars. Now, you can buy 128 gigs worth for the same price thanks to the competition to provide better products. Other sectors, like CPUs, have more complex market dynamics, but also have multiple competitors like Intel, AMD, and various ARM-architecture manufacturers driving forward innovation.

The GPU space is much different. Nvidia is so far ahead of the curve that it’s almost not worth discussing its competitors. This has been reflected by Nvidia’s stock hitting orbital velocity since Q1 2023 while Intel and AMD, Nvidia’s competitors in the AI space, have seen a decrease in price. Nvidia has the monopoly in GPU chips simply by offering the best performance in the space, capturing 80% of the market in Q4 2023. No one is capable of designing chips like Nvidia.

This isn’t the whole story. Taiwan Semiconductor Manufacturing Company (TSMC) actually builds the majority of Nvidia’s AI chips. TSMC is the only company on the planet that possesses the ability to manufacture the most advanced leading-edge chips used in everything from iPhones to F-35s. Nvidia uses a “fabless” business model, where they don’t build the chips in-house, only providing their designs to a fab partner like TSMC or Samsung. This business model came about as chip manufacturing became increasingly capital-intensive and required more specialized knowledge. This has allowed Nvidia to focus on designing the best GPU chips without putting forward the capital required to spin up their own fab. Taiwan’s TSMC is a very attractive partner because unlike Intel or Samsung, TSMC doesn’t design anything themselves, removing intellectual property theft from the table. This compartmentalization has won TSMC 62% of the market share in the foundry space, with Korea’s Samsung in a distant second with 13%. The best chemists, physicists, machinists, engineers, and factory managers of Asia are on TSMC’s payroll, making TSMC one of the only places that have the technical experience in running cutting-edge semiconductor tools at scale.

While TSMC is not state-run, the company has an incredibly close relationship with the Taiwanese Government. The presence of most of the company’s fabrication capacity on the island has thrust Taiwan into global importance. Chinese, Asian, European, and American chips are fabbed on the island, and disruption to this supply chain would be catastrophic. The Department of Commerce projected that a full Chinese takeover of TSMC in Taiwan would increase the cost of chips in the US by 59 percent, which would lead to incredibly high prices attached to final products, not to mention the disruption this hypothetical would cause to national security-critical supply chains. Taiwan has used this leverage to gain support from the world’s democracies as the keystone of the information-age economy. While the US shockingly does not formally recognize Taiwan’s independence to preserve relations with its big red brother and the States’ largest trade partner, the Americans have a long history of selling weapons and training to the Taiwanese. Most of this kit as of late is the kind of weapons and platforms that would be used to fend off an amphibious assault against a large naval force like Switchblade drones, Reaper drones, Patriot missile systems, surface-to-air missiles, and F-16s. While it is impossible to know the future, it seems likely that the United States and her allies in Korea, Japan, and the Philippines as well as the rest of NATO would come to the defense of democratic Taiwan in the event of a full-scale invasion from China. The Taiwanese have cleverly made the island a critical link in the supply chains that run the democracies of the world, and in return, they have received protection from a China that seems to be increasingly possessive of the democratic and independent nation of Taiwan.

We’ve gone over sticks, let’s talk about some carrots. In 2022, Congress passed the CHIPS and Science Act, a hefty bill that authorizes funding for various federal science agencies and mechanisms to support the domestic manufacture of semiconductors. This bill instructs the Department of Commerce to incentivize investment in bringing every step of the semiconductor value chain stateside, including fabrication, assembly, testing, and packaging. It also provides funding for everything from advanced materials science, infectious disease research, and quantum cryptography, and provides funding for 2 SLS missions to be launched a year until a human mission to Mars is possible. While taking notes on the bill, I wrote, “Pretty sure half of these are fancy flashlights for the physics nerds to play with.”

The CHIPS and Science Act has made $39 billion available to companies that bring chip manufacturing to the states. During a recent episode of ChinaTalk, Jordan Schneider discussed the CHIPS act with former director of national security at the CHIPS office Benjamin Schwarz. Ben discussed a small deal worth about $35 million with defense contractor BAE to produce a Monolithic Microwave Integrated Circuit (MMIC) chip that is vital to the electronic warfare systems on the F-35 platforms. He described it as a “smart use of U.S. government funding”, as a small injection of cash was able to solve the shortage of these critical components to our most dominant systems.

At a much larger scale, the CHIPS office announced $6.6 billion in grants to TSMC to bring manufacturing to Phoenix, Arizona. Currently, TSMC plans to build three large fabs producing both four nanometer and two nanometer nodes. Once completed and running, this will be the first time in decades that the United States domestically produces the leading-edge chips. I’ve seen these buildings in person, and their scale is absolutely massive. While the water-conscious reader might recoil at the thought of setting up one of the most water-intensive modern manufacturing processes in dry Phoenix, TSMC has shot for a 90% water recycling rate. Phoenix is also uniquely positioned to hold this specialized and high-skilled industry, as it has seen a population boom instead of decline after the pandemic, already has a highly skilled population in both engineering and software, has the largest university in the country from ASU that would be able to fulfill the demand for semiconductor technicians, it has relatively cheap skilled and unskilled labor, and its geographic proximity to the ports of California makes importing and exporting overseas that much easier.

While this investment can be read as a move by the American Government to wean itself off of the Taiwanese chip industry so as not to get pulled into a war in the event of a Chinese invasion of Taiwan, my opinion is that this helps Taiwanese independence in the critical near future when China is most likely to invade. TSMC, and thus the Taiwanese Government, has shown its willingness to work with the United States in critical industries, and is likely to use this to pursue the same level of allyship that the United States give to Japan and South Korea. Some estimates say that the US is two to three decades away from complete chip independence, whereas the some of the most well-informed estimates for a cross-strait war timeline posit that an full-scale invasion of Taiwan is possible as soon as 2027 to 2030, in the first few years after the 21st National Congress of the Chinese Communist Party. Taiwan is walking a balance between ensuring that they can count on American support in the event of an invasion by allowing the Americans to develop a copy of their biggest bargaining chip.

By restricting China’s access to the best American-Taiwanese hardware and bringing semiconductor manufacturing stateside, the US government has ensured American dominance AI and supported democracy in the Pacific, at least for the foreseeable future. There are chips slipping into China through direct smuggling or by passing through shell companies, but this has always been an issue with export regulation. Given China’s recent history of intellectual property theft, it’s nearly guaranteed that some chips are already in the hands of a darling defense contractor in China being reverse engineered. Too bad the best Chinese engineers are already designing the next Nvidia chip in Palo Alto to be sold directly to American Big Tech.

But what about using the models that have already been built? If I were an operative running an influence campaign to try to swing the American election this year, I would simply use ChatGPT. It turns out, the US has taken actions against this vector as well. OpenAI kickstarted the AI revolution by releasing ChatGPT on November 30th, 2022, bringing life to a stagnant Silicon Valley focused on SaaS vaporware and Web3. It’s become known to the tech community that Microsoft has had a roughly 49% stake in OpenAI since 2019, a bet that brought Microsoft back to the cutting edge after more than a decade of UX updates to Windows and a cloud business that has the same capability set as the three other hyperscalers. Generative AI has complemented Microsoft’s existing suite extraordinarily well, and the market seems to know this. The stock price has increased by north of 30% since Copilot was announced as Microsoft has leaned into AI. Anecdotally, everyone I know at Microsoft is interested in either AI or security. Microsoft also has incredibly deep relations with the US Government, capturing 85% of the public sector office software market, and providing Azure cloud services to the Government. While OpenAI remains independent, that 49% Microsoft stake has probably come with some gentle suggestions for business strategies, like opening the company to the biggest customer in the country.

On June 13th, OpenAI appointed former head of the NSA General Paul M. Nakasone to its board of directors. General Nakasone led the world’s most advanced and well-funded intelligence agency from 2018 until his retirement from uniform on February 2nd, 2024. This resume means he is just about the best person to choose to effectively represent Uncle Sam’s national security interests in this revolutionary technology. General Nakasone has been busy; on July 8th, OpenAI announced they would block Chinese IPs from accessing their APIs, effectively closing off the country to their product. We’ll briefly revisit the implications of that in a second. On August 29th, NIST’s Artificial Intelligence Safety Institute announced that the Institute would “receive access to major new models from [Anthropic and OpenAI] prior to and following their public release.” This means that the US Government has free-reign over OpenAI’s unreleased models like Sora, as well as anything else OpenAI has cooking up. OpenAI has attracted the best AI engineers in the Valley and has proven to be more innovative than legacy big tech, so inserting it into the Pentagon’s expansive list of private collaborators alongside Microsoft and Amazon is a no-brainer. In the last few months, OpenAI has seemingly turned into a modern-day Skunkworks for AI, and rumors permeate San Francisco that many of the engineers at OpenAI have had to obtain clearance.

Blocking IPs from China likely will only prevent some malicious use of Generative AI from the nation. Every person I know from China grew up using one of a number of specialized apps to dodge the country’s censorship system to access the open net and would easily be able to dodge this form of blocking by using any reputable VPN. Restraining access to free information because of location goes against the gut impulses of many Americans, even if their homeland is an adversary to the US. ClosedAI’s OpenAI’s deep connection to Microsoft and recent cozying to the US Government obviously informed this decision, and those shareholders do not share the same reluctance to exclude nearly a billion and a half people from this technology.

While the United States inches toward chip independence and continues the long American tradition of throwing billions of dollars at Silicon Valley in exchange for exclusive access to world-shaking technology, perhaps we should consider whether future steps must be taken to ensure that the best AI stays in the hands of the world’s democracies and whether the current steps are a bridge too far, excluding the Chinese people from the cutting-edge.