Taiwan is the powerhouse behind silicon production, yet the United States is starting to explore its own capabilities.
The world’s fastest AI chip, developed in California with American financial backing, has been celebrated in Washington as a significant achievement. However, its manufacturing still takes place in Taiwan.
Cerebras went public on May 14, raising $5.5 billion, and saw its valuation nearly double on its first trading day, closing at $66 billion. Their processor, about the size of a dinner plate, operates AI tasks ten times faster than what Nvidia offers.
Taiwan leads the semiconductor supply chain, with trade expertise embedded in the minds of experienced engineers at Taiwan Semiconductor Manufacturing Company. They produce chip wafers at the smallest sizes, leaving Korea and China trailing behind.
A wafer serves as the foundation for processors, memory, and hard drives. The nanoscale measurements reflect how finely the tiny grooves are cut to encode data. There are American semiconductor companies like SkyWater Technology and GlobalFoundries that can manufacture a considerable number of 22nm wafers.
Despite relying on expensive chips from Taiwan, Cerebras is considering expanding configurations and supporting American specialty-chip companies, which might require less funding than one might expect. This could foster the development of rapid AI capabilities for local models and spur a manufacturing resurgence.
I’ve had the chance to use Spark, the model OpenAI designed for Cerebras chips. It’s impressive but can be somewhat overzealous; in AI, timing is everything. Terms like “tokens” are central, where ‘warmly’ breaks down into two tokens. The speed, measured in tokens per second, varies depending on the graphics card—smaller models tend to run faster. Some folks will pay extra for faster modes, but Cerebras delivers a tenfold advantage.
This tenfold speed is not just a technological milestone; it heightens risks associated with rogue AI. Cybersecurity needs to keep pace—after all, to counteract a fast AI, you need an even faster one. This shift makes the AI race feel all the more urgent, and employees will need to prepare for managing their own protective agents.
While specialty chips focus on the largest models, I tried Bonsai by PrismML, an 8 billion parameter model that operates on my basic Nvidia card from 2023. It excels at number crunching but speed remains an issue; we require chips tailored for more streamlined models.
If the SEC eases crowdfunding regulations, startups could gather enough public support to develop new chip designs using American silicon. Precision like Taiwan’s isn’t necessary if we can manage something ten times larger.
The barriers lie not in processors or memory but in bandwidth, and American ingenuity could address that. Picture $500-$1000 mini-PCs powered by efficient AI-optimized chips made in the U.S.
This could allow us to compete globally, bringing back a taste of that 1950s industrial dominance.
Will this lead to significant job creation? Companies focused on chip design often don’t require massive workforces.
The real employment boost could stem from the supply chain: wafer substrates, foundry capabilities, packaging, assembly, thermal systems, power electronics, and support teams—these can all be brought back to the U.S.
And that’s just scratching the surface.
Formal roles like AI Alignment Researcher and Cyber Security are specialized, each comprising a few thousand jobs, with under two hundred thousand in cybersecurity.
In the 1980s, expertise in VisiCalc could secure a job, but today it’s merely a skill in many logistics roles. In the future, millions may apply alignment and cyber skills to handle everyday business needs.
The future of American AI chips will hinge on achieving that speed advantage.
