Taiwan has the goose that lays the silicon eggs, but now America is learning to breed our own. The fastest AI chip on earth was designed in California, funded by American investors and celebrated in Washington as a…

Taiwan has the goose that lays the silicon eggs, but now America is learning to breed our own.

The fastest AI chip on earth was designed in California, funded by American investors and celebrated in Washington as a national triumph. It is manufactured in Taiwan.

Cerebras IPO’d on May 14th, raising $5.5 billion and nearly doubling on its first day of trading to close at a $66 billion valuation — their processor, roughly the size of a dinner plate, runs AI ten times faster than anything Nvidia sells.

Taiwan is on top of the supply chain, with trade secrets in the brains of senior Taiwan Semiconductor Manufacturing Company employees, never written. They alone produce the chip wafers with the tiniest size, with Korea behind them, then China.

A wafer is the base of all processors, memory, hard drives, the nanometers represent how finely they chisel little groves that encode 0’s and 1’s. There are American-based semiconductor companies like SkyWater Technology and GlobalFoundries that can produce lots of 22nm wafers.

Cerebras still uses expensive Taiwan-based chips but more configurations and American specialty-chip companies can be funded and tried for less capital than you might think. This can create abundant fast AI for local models, investment opportunities and a manufacturing boom.

I’ve used Spark, the model OpenAI made for Cerebras chips, it’s fast but gets ahead of itself, milliseconds matter now, as they innovated on computer use timing. In AI words are broken down to “tokens”, warmly is two tokens, ‘warm’ and ‘ly’, tokens per second (tps) is the speed, weaker graphics cards are slower, smaller models are faster. To get 50% faster people pay double for /fast mode, but Cerebras is 10x faster.

Advertisement

A 10x speed advantage is more dangerous for rogue AI and means cybersecurity professionals and their AI agents need the same edge. The only thing that stops a bad AI with speed is a good AI with more speed. The AI race became more literal. This very danger is the business risk that employees will have to train for in managing their own locally run defender agents.

Current specialty chips target the biggest models but I recently ran Bonsai, by PrismML, an 8 billion parameter model that fits on my bottom-of-the-market Nvidia card from 2023. It trades big numbers for 0’s and 1’s to make an agent just capable of following precise instructions fit. The only problem is speed; we need chips designed for lean models.

If the Securities and Exchange Commission loosens Crowdfunding rules, start-ups could raise enough from the public to make out a new chip design, based on American silicon. We do not need Taiwan’s precision, ten times larger can work.

The problem is not processors or memory, it is bandwidth, and American talent can solve that. We could see $500-$1000 mini-PCs designed for fast AI, running American-made efficient AI-optimized chips.

We can sell it to the world and taste that 1950s industrial advantage again.

Will this create enough jobs? Fabless chip companies, like most tech startups, do not need massive headcount.

The job engine is the supply chain: wafer substrates, foundry capacity, packaging and testing, board assembly, thermal systems, power electronics, firmware and driver teams, validation labs, repair, integration and developer support, all of which can be onshored.

Advertisement

But that is just the beginning.

AI Alignment Researcher and Cyber Security as formal job titles are specialized and amount to a few thousand and under two hundred thousand jobs respectively.

In the 1980s being good at VisiCalc, the Excel precursor, could be a distinct job. Now it is just a skill in every grocery chain’s logistics office. In the future, tens of millions of people will apply alignment and cyber skills to managing the assets of everyday businesses.

It all hinges on the speed advantage of American AI chips.

Patrick Dugan is an independent AI researcher and founder of MoralityLab.

The views and opinions expressed in this commentary are those of the author and do not reflect the official position of the Daily Caller News Foundation.

All content created by the Daily Caller News Foundation, an independent and nonpartisan newswire service, is available without charge to any legitimate news publisher that can provide a large audience. All republished articles must include our logo, our reporter’s byline and their 📰DCNF affiliation. For any questions about our guidelines or partnering with us, please contact [email protected].