Here is the full transcript of Rob Toews’ talk titled “AI’s Single Point of Failure” at TED conference.
Listen to the audio version here:
TRANSCRIPT:
The following statement is utterly ludicrous. It is also true. The world’s most important advanced technology is nearly all produced in a single facility. What’s more, that facility is located in one of the most geopolitically fraught areas on Earth, an area in which many analysts believe that war is inevitable within the decade.
The Central Role of TSMC
The future of artificial intelligence hangs in the balance. The Taiwan Semiconductor Manufacturing Company, or TSMC, makes all of the world’s most-advanced AI chips. This includes Nvidia’s GPUs, Google’s TPUs, AMD’s GPUs, the AI chips for Microsoft, Amazon, Tesla, Cerebras, SambaNova, and every other credible competitor. Modern artificial intelligence simply would not be possible without these highly specialized chips.
Little wonder, then, that Time magazine recently described TSMC as, “The world’s most important company that you’ve probably never heard of.” Nvidia CEO Jensen Huang put it more colorfully, saying, “Basically, there is air … and TSMC.” TSMC’s chip fabrication facilities, or fabs, the buildings where chips are physically built, is located on the western coast of Taiwan, a mere 110 miles from mainland China.
Geopolitical Tensions and AI
In this map, Taiwan is shown in orange and China is shown in green. Today, China and Taiwan are nearer to the brink of war than they have been in decades. Many policymakers in Washington predict that China will invade Taiwan within the next five years.
A China-Taiwan conflict would be devastating for many reasons. Aside from the heavy human toll, one underappreciated consequence is that it would paralyze the global AI ecosystem. Put simply, the entire field of artificial intelligence faces an astonishingly precarious single point of failure in Taiwan.
Amid all of the fervor around AI today, this fact is not widely enough appreciated. If you are working on or are interested in AI, you need to be paying attention.
The Semiconductor Industry
How did we get here and what can we do about it? Let’s start with a brief, whirlwind overview of the chip industry. Semiconductors, or chips, are the most complex object in the world that humanity knows how to mass-produce. Making semiconductors requires the world’s purest metals, the world’s most expensive machinery, legions of highly specialized engineers, and atom-level manufacturing precision.
It is important to distinguish between two different types of chip companies. First, fabless chip makers, which design but do not manufacture their own chips. And second, foundries, which manufacture chips designed by other companies. Almost every well-known chip company today is fabless, from Nvidia to AMD to Qualcomm. These companies do not produce their own chips.
Instead, they design chips, and then they rely on foundries like TSMC to actually manufacture those chips for them. There are only three companies in the world today that are capable of manufacturing chips anywhere near the leading edge of semiconductor technology: TSMC, Samsung, and Intel. Of those three, only one can reliably produce the world’s most advanced AI chips, including chips like Nvidia’s H100 GPUs. That one company is TSMC.
As of this morning, TSMC’s market capitalization was 470 billion dollars, making it the 13th-largest company in the world, larger than ExxonMobil, JPMorgan Chase, or Walmart. How has TSMC become such a dominant force? The short answer is that powerful economies of scale exist in the world of chip fabrication, leading inexorably to winner-take-all dynamics. Making advanced semiconductors requires tremendous upfront and ongoing capital expenditure.
The Global Semiconductor Landscape
In 2021, TSMC announced that it would invest 100 billion dollars over the next three years to continue expanding its fabrication capabilities. No other company in the world can justify that level of investment. TSMC can, because of the sheer volume of chips that it produces, far more than any other company in the world. A related dynamic that helps explain TSMC’s unassailable position is what has come to be known as the TSMC Grand Alliance.
TSMC has invested heavily over decades to develop deep partnerships with dozens of companies across the semiconductor supply chain, from software providers like Cadence to equipment manufacturers like ASML to chip designers like Nvidia. In turn, these companies have developed their own products in accordance with TSMC’s roadmap, leading to powerful lock-in.
In summary, a combination of economies of scale, network effects, and unrivaled specialization have made TSMC irreplaceable and have made the entire world deeply, precariously dependent upon it.
Geopolitical Strategies and Implications
This brings us to the present, delicate geopolitical moment. Last October, the Biden administration took the dramatic step of banning the export of all high-end AI chips to any entity in China. The rationale behind these measures was clear. To leverage US control of the global semiconductor supply chain as a choke point to handicap China’s AI capabilities. The US government is currently formulating expansions to this policy.
At the same time, the US is taking steps to reduce its reliance on chip fabrication facilities located in East Asia. In late 2022, TSMC announced that it would invest 40 billion dollars to build two new state-of-the-art fabs in the United States, in Arizona. The first of these two fabs is slated to begin production in 2025. Bringing advanced chip production to US soil will help mitigate the AI industry’s absolute dependence on Taiwan-based fabs.
The Silicon Shield Theory
But the Arizona fabs will not solve everything. Their production capacity will be modest, representing less than five percent of TSMC’s total global output. And the most advanced semiconductor production capabilities and technologies will remain in Taiwan. So where might things go from here? Let’s briefly consider a few possibilities on this three-dimensional chessboard.
Potential Outcomes
Let’s start with the optimistic scenario. Taiwan’s central role in the global semiconductor industry is often referred to as its “silicon shield.” The basic theory is this: because China depends so heavily on Taiwan for the chips that it needs to keep its own economy running, China will stop short of invading Taiwan and putting TSMC’s production at risk. And because the rest of the world is likewise so dependent on TSMC, the United States and other powers will go to great lengths to protect the island and defend its sovereignty.
Under this theory, while China may continue to build out its military and engage in cross-strait saber-rattling, it will stop short of kinetic action against Taiwan. But the silicon shield is just a theory, not a guarantee. What would happen if China were to move decisively to retake Taiwan? TSMC’s fabs would almost certainly be rendered inoperative.
The Consequences of Conflict
It is conceivable that the Taiwanese, or even the US military, would preemptively destroy the fabs in order to prevent the CCP from taking control of this valuable strategic resource. Even if the physical buildings were to remain undamaged after a Chinese invasion, it is unrealistic that the CCP would be able to continue operating the fabs to produce cutting-edge chips.
Keeping leading edge fabs running requires ongoing and deep partnership with organizations across the global semiconductor ecosystem, as well as a steady inflow of materials, equipment and services. These would be denied to an invading power.
Let me say this one more time: if or when China invades Taiwan, TSMC’s fabs will, in all likelihood, go offline. This will mean that no more Nvidia H100s or any other cutting-edge AI chips will be able to be produced anywhere in the world.
What would this mean for the world of AI? After TSMC, the company best positioned to step up and produce cutting-edge AI chips is Samsung. Samsung is currently the only company in the world, other than TSMC, that is capable of producing three-nanometer chips, today’s cutting-edge technology. But Samsung’s production capabilities are far inferior to TSMC’s today.
In a best-case scenario, it would take Samsung years to scale up to TSMC’s current AI chip yields and volumes. This brings us to America’s former chip champion, Intel. It was hardly a decade ago that Intel’s chip-manufacturing capabilities were the envy of the world. But in recent years, Intel has fallen behind.
Conclusion
The company struggled mightily in its transition to both ten-nanometer and seven-nanometer node technologies, even resorting to outsourcing some of its leading edge production to TSMC. Under CEO Pat Gelsinger, Intel aspires to regain its chipmaking supremacy with an ambitious plan to leapfrog TSMC and begin producing two-nanometer chips in 2024. Whether this ambitious plan will actually prove achievable, however, remains to be seen.
Before we despair too much, let us note a couple of encouraging points. First, keep in mind that a considerable stock of AI chips already exists in the world. And even in a worst-case scenario, these chips would remain in use. Second, while the most advanced AI chips, like Google’s TPUs or Nvidia’s H100s, can only be manufactured in Taiwan, there are many fabs around the world, from the US to Europe to Israel, that are capable of producing lagging-edge logic chips at scale.
Though they are far less powerful than today’s leading AI chips, these previous-generation chips could be used in a pinch to support some AI computing workloads. Ultimately, though, it would be devastating for humanity to lose its ability to produce the chips that power today’s cutting-edge artificial intelligence. Progress in AI would be profoundly disrupted. Let us hope that diplomacy prevails.