Can Cerebras Take on NVIDIA?

May 31, 2024
Can Cerebras Take on NVIDIA?

A lot of work is being done on your behalf.

Our readers might not be seeing it, but as Jeff alluded to in last week’s AMA, we’ve been working very hard behind-the-scenes on a series of big announcements, which will be coming down the pike in the next couple of weeks.

We appreciate your patience.

And in the interim, we hope you’ll continue enjoying Jeff’s daily episodes from the outermost limits of high-tech.

As Jeff often says, the radical pace of change in tech continues to accelerate by the day — even more so now with the help of artificial intelligence.

It's hard to believe it’s all happening right now — here in our lifetimes.

We're going to get to see how it all unfolds, and to a certain extent, be a part of it.

What a race. What a sensational spectacle.

We have so much to look forward to.


Cerebras Chips

Jeff: I’d be interested in your perspective in terms of how Cerebras chips would be suitable in AI platforms (or not)? I think their current applications are largely somewhat customized and are mostly orientated towards medical solutions (i.e., what I’m really curious about is how their chips, which intentionally have very large footprints (think frisbee) — but contains an absurd amount of transistors — compare to NVDA’s. And would this company be a good takeover target? Is their business plan (aka not being a slave to miniaturization) not as viable or as universally appealing as I think? Thanks! — G.L.

Hello G.L.,

Cerebras is definitely one of the most exciting artificial intelligence companies to watch at the moment. It is a private, fabless semiconductor company that designs semiconductors specifically for machine learning and artificial intelligence applications.

To your point, the Cerebras “chip” isn’t really a chip at all, it’s an entire wafer upon which many chips have been etched into.

Source: Cerebras
Source: Cerebras

Above is Cerebras’ latest semiconductor, the CS-3 wafer-scale engine (WSE). It’s called that because the CS-3 is an entire wafer, as compared to cutting up a wafer into thousands of individual semiconductors after having been etched into the silicon wafer.

For further comparison, we can see the image below of the CS-3 side by side with the largest GPU from NVIDIA:

Cerebras v Nvidia
Source: Cerebras

For an easy comparison of the two products side by side, we can see the dramatic difference in performance between Cerebras and NVIDIA’s H100.

Cerebras v. Nvidia Chart
Source: Cerebras

All of the above numbers are true. Cerebras really has a large competitive advantage on a product-to-product basis.

But the nuance is that it’s not really an apples-to-apples comparison. We can loosely think of the Cerebras CS-3 as equivalent to about 50 NVIDIA H100’s connected together.

Cerebras definitely has some cost and performance advantages with this wafer-scale approach. It designed a powerful architecture in a single product, capable of training a 24 trillion parameter large language model. That’s pretty incredible.

It’s worth keeping in mind that Cerebras is targeting data centers with such a large product. Unlike an NVIDIA GPU, which can be used in a desktop computer or even a laptop, with Cerebras that would be impossible. Cerebras’ products are like workhorses that run computationally intense loads in datacenters for training artificial intelligence models and for inference (when running AI).

Cerebras isn’t a company just focused on the medical/health industry. Its technology can be used, and is used, for any kind of machine learning (ML) or artificial intelligence (AI) software. Its products are used by a variety of industries, basically for anything that can benefit from the optimization and discovery that can come from AI/ML.

The one advantage that NVIDIA has over any of the other AI-focused semiconductor players is its extremely well-developed software language for its GPUs — CUDA — and its developer community. In order to work with Cerebras, developers have to learn how to use the product and associated software. 

As for your question, yes, absolutely, Cerebras is a takeover target.

With that said, Cerebras is looking to go public this year. Its last-known valuation was $4.25 billion, but that was from a priced round back in late 2021. I’m confident that the valuation will be much higher as it approaches the public markets this year.

The real question will be whether or not Cerebras is acquired before or after going public…

Insurance Risk

It’s my understanding that if an autonomous vehicle gets into an accident, the owner/driver is responsible. I don't think that I would be silly enough to put my financial responsibility in the hands of a software company until they take all of the insurance risks. What's your opinion? — Gary W.

Hello Gary,

This is an interesting topic right now, given the incredible increases in car insurance that we’re all experiencing.

Just this year in the U.S., car insurance jumped 26% compared to last year. The prices are unbelievable, but the reasons are well-understood.

The current government’s fiscal, economic, and monetary policies are responsible for the persistent inflation that we are all experiencing. That inflation has done two things. It has increased the cost of auto parts and labor, and it has driven up the levels of uninsured and underinsured motorists.

The sad reality is that uninsured drivers have reached all-time highs at 14%. Many can’t afford to pay insurance premiums, but still need their cars. And the percent of underinsured drivers is even higher. That has driven up insurance premiums for everyone else, along with auto parts and labor increases.

It’s not sustainable at all. And it’s going to get worse in the near term.

Longer-term, autonomous driving will dramatically bring down the costs of insurance. Tesla has already demonstrated how much safer its software is compared to human drivers. The data will be even more compelling by the end of the year. Waymo has strong data, as well.

But to your point, who will actually be responsible for paying the insurance?

  • Short-term: If a consumer purchases a self-driving car, they’ll still have to have car insurance.
  • When a consumer rides in an autonomous ride-hailing service, either the ride-hailing service (like Uber or Lyft) would need insurance, or the insurance would be borne by either the company that made the autonomous car, or the company that made the autonomous software that runs on the car.
  • Long-term: As autonomous cars gain regulatory approval and are sold/leased without steering wheels (i.e. there can’t be any drivers), either the company that made the autonomous car, or the company that made the autonomous software that runs on the car will be responsible for the insurance.

It will take more than a decade for the benefits of autonomous driving technology to have a large impact on insurance costs, but it will happen. 94% of all automobile accidents are the result of human error. Once the industry solves that problem, there will be dramatically less accidents, fewer claims, and a much lower cost of insurance. 


Dear Jeff, I disagree with your comments that Ivermectin is useful for COVID-19 infections. 2 placebo controlled trials on ivermectin did not show a benefit: New England Journal of Medicine, 3/30/22, vol 386, No 18. JAMA 2/20/23. I feel it is important that the public understand that physicians not supporting Ivermectin use are basing their opinions on legitimate medical studies. — Brian G.

Hi Brian,

I appreciate you writing in about this given the controversy and importance of this issue.

There is extensive peer reviewed medical research that has demonstrably proven the efficacy of Ivermectin both as a therapy and for early treatment of the COVID-19 virus. The issue is that when we conduct a generic search for this information, the research is almost impossible to find unless we know the specific name of the study — which very few people ever would.

And scientists in the U.S. were discouraged and even not permitted to publish scientific research about Ivermectin that demonstrated its efficacy. Even medical journals would reject peer-reviewed scientific research, and they continue to do so.

So much money was being thrown around by pharmaceutical companies and the U.S. government that forced “compliance” to support the desired narrative. And we should remember, if the CDC/FDA were to admit that Ivermectin was an effective therapy for COVID-19, they would not have been able to have approved the COVID-19 mRNA experimental drugs for Emergency Use Authorization.

It is so deeply corrupt, at the cost of so many lives, and it continues to be so.

As for Ivermectin, here are some randomized controlled trials that demonstrated efficacy:

We could go on an on. In total, there are more than 60 high-quality studies that showed statistically significant improvement in outcomes from Ivermectin.

Ivermectin isn’t perfect for COVID-19, but varying levels of efficacy were proven consistently. And it is safer than Tylenol, unlike the mRNA “vaccines.” And yet, U.S. physicians weren’t permitted to prescribe Ivermectin before or after the Emergency Use Authorization was given, and pharmacies weren’t permitted to fill subscriptions for those doctors that did prescribe it. Horrible.

And Brian, you actually raised a key point of how the whole system was gamed for profit at the cost of life. The NIH/CDC aggressively pushed and incentivized medical boards across the country to not prescribe Ivermectin, wait until the Emergency Use Authorization, and then push the mRNA “vaccines.”

Physicians, for fear of losing their jobs or damaging their career, took their marching orders from their medical boards based on their “opinion,” which was provided to them by a government health agency which provided sloppy, heavily biased, and corrupted trails as “legitimate evidence” that Ivermectin didn’t work and couldn’t be prescribed.

The entire medical system failed us, at immense loss of life.

It is so difficult and time consuming to proactively research what’s happening right now, very few have the time or willingness to do so. But at the moment, that’s the only way I know to be well-informed and have a good enough understanding to make decisions for ourselves and our families. 

And I’m always happy to share my research with my subscribers and hope that it is useful.

Thanks again,


Previous Post Next Post