“One AGI to Rule Them All”

Something fishy is going on…

Yesterday, news of U.S. Department of Justice (DOJ) actions against NVIDIA rattled the share price of the world’s most valuable semiconductor company.

Subpoenas were issued to NVIDIA – and other related companies – in search of evidence that NVIDIA is somehow guilty of violating antitrust laws.

While not a formal complaint, the premise is that NVIDIA makes it too difficult for its customers to switch to another graphics processor unit (GPU) supplier – namely Advanced Micro Devices (AMD).

Not surprisingly, AMD’s share price was up after hours last night and up this morning on the news.

The other concern is that NVIDIA provides priority to certain customers when allocating supply of its GPUs, which are in high demand.

Is there any substance to these theories?

Or is this DOJ action nothing but a red herring?

The details are important.

They speak not only to NVIDIA but the entire semiconductor industry. And, for that matter, much of the high-tech industry.

A Great Development Environment Is Not an Antitrust Matter

The claims that NVIDIA makes it too difficult to switch to another GPU provider largely center around NVIDIA’s proprietary software, known as CUDA – short for Compute Unified Device Architecture.

Many think that semiconductor companies just sell chips – i.e. the hardware – while other companies develop software that runs on those chips.

But complex semiconductors require software, which can include drivers, kernels, compilers, and application programming interfaces (APIs) needed to run software applications on those semiconductors.

This is an area I’m deeply familiar with due to my time working at Qualcomm and NXP Semiconductors.

Whether those chips go into smartphones, televisions, payment terminals, or other wireless devices, there is a software layer that acts as a bridge between the hardware and the software applications.

That’s why NVIDIA developed CUDA in 2006.

As GPUs became more complex and the graphics industry started to leverage more of the parallel processing capabilities of GPUs, software was necessary to improve performance and accelerate software applications.

Naturally, software developers learned how to work and program with NVIDIA’s CUDA, given the widespread use of its semiconductors.

And the same is true with AMD’s GPUs.

In 2016, AMD launched ROCm – which originally stood for Radeon Open Compute platform – for the same reasons that NVIDIA launched CUDA. The big difference is that AMD made ROCm open-source software and NVIDIA has kept CUDA proprietary.

But neither ROCm nor CUDA locks customers into using one GPU platform versus the other.

Both software platforms are designed to be developer-friendly and available to all, and they have fantastic support resources available to anyone working with either of the duopoly GPUs.

It’s worth noting that this kind of industry dynamic is not limited to just semiconductors.

For example, in information technology, Cisco has its Internetworking Operating System (IOS), which runs on Cisco’s IT hardware systems. And Juniper Networks has its own Junos OS to do the same. Software engineers are familiar with and certified to work with both operating systems.

This is another example that I’m intimately familiar with having been a senior executive at Juniper Networks.

While there were a larger number of Cisco-certified software engineers consistent with Cisco’s larger market share, most of my customers had both Cisco and Juniper products in their network architectures.

We competed on product performance, pricing, and service/support, but not on the software operating systems.

As for the issue of prioritizing certain customers over others with regards to allocating production…

Well, that’s certainly true, but it’s not an antitrust violation.

Demand-Supply Dynamics

The semiconductor industry is known for its cycles.

Going from periods of oversupply – and high inventories resulting in lower prices – to periods of incredible demand – and having to figure out how to best allocate production amongst a large number of customers – is normal. And it’s an impossible job to keep everyone happy.

Having worked for years at both Qualcomm and NXP Semiconductors – two of the largest semiconductor companies in the world – there are always periods where demand exceeds production and decisions about product allocation have to be made.

And strategic customers that materially impact your business almost always get preference over others.

To whatever extent possible, semiconductor companies try to ensure their products are being sold to customers who will put them to use as opposed to selling them to a company that might be stockpiling in hopes of reselling at higher prices.

Again, allocating production when demand exceeds supply is not limited to the semiconductor industry. All we have to do is think back to all the supply chain issues during the pandemic to understand how companies were allocating their production everywhere due to undersupply.

This is why the claims that the DOJ is investigating make no sense at all.

Any business faced with the same demand/supply dynamics would be forced to make the same kinds of decisions.

It begs the question… what is the real motivation here?

What underlying concern or concession is the DOJ really after?

It would be naïve for us to think there isn’t one.

Hard-Earned

In the last decade, NVIDIA spent an incredible amount on research and development – $37.7 billion since fiscal year 2014 to be specific. And it will spend over $10 billion more this fiscal year (ending January 31, 2025) on R&D.

The reality is that NVIDIA’s stock was largely a sleeper for most of the last two decades. It wasn’t until early 2016 that NVIDIA started to get very interesting. That’s when I first noticed that its semiconductors were being utilized in increasing amounts for machine learning and artificial intelligence.

20-Year Chart of NVIDIA (NVDA)

NVIDIA placed a huge bet on its vision of the future. And a decade ago, even NVIDIA couldn’t have said with 100% certainty that its vision of the future would be correct.

But it was.

And it was perfectly positioned for the breakthroughs in artificial intelligence, resulting in its current rise.

This is entirely my speculation, but I suspect that the DOJ (i.e. U.S. government) is less concerned about what’s happened to date…

And far more concerned about what’s coming in the next two years.

There Must Be an Endgame

As we explored in The Bleeding Edge – AI’s Need for Speed a few days ago… this summer, OpenAI demonstrated its most advanced artificial intelligence (AI) – codenamed Strawberry (formerly Q*) – to U.S. national security officials.

And Q* had been previously rumored to have demonstrated some signs of artificial general intelligence (AGI). It’s OpenAI’s next generation of AI after GPT-4.

As with most governments, politicians, and bureaucrats… power, control, and money tend to be the driving motivations behind most actions.

And the power that comes with having control over the development of – or use of – an artificial general intelligence is just about on par with Ash Nazg…

Roughly translated, that’s the One Ring in J.R.R. Tolkien’s Lord of the Rings.

There are already concrete plans to build out not just one but multiple $100 billion-plus AI data centers explicitly to develop AGI. Not years from now, but as soon as possible. We’ve been documenting these developments right here in The Bleeding Edge.

Elon Musk and his team at xAI just announced that they brought online Colossus, now the world’s most powerful AI training cluster. It’s comprised of 100,000 NVIDIA H100 GPUs. And the most incredible part – it was done in just 122 days.

Perhaps even harder to comprehend is that Colossus will double in size in the next few months. Musk and his teams, as usual, find ways to do things that others believe are impossible.

Now, it doesn’t take much creativity to imagine that the current U.S. government might not like the idea of Elon Musk beating everyone to AGI and possessing “the ring,” without the government having access to such incredible “power.”

“Why shouldn’t the U.S. government control such a powerful – and potentially dangerous – technology? It’s a matter of national safety and security!” Again, it’s not crazy at all to imagine such a stance.

So whether it’s preferential access to GPU supply (perhaps to be allocated to the Department of Energy, which oversees national supercomputers)…

Or special access to NVIDIA’s proprietary software (to gain more control over the performance of the GPUs)…

Or perhaps some level of control over GPU allocation decisions (so that the government can put the GPUs in “the right hands”)… there is some kind of endgame.

Whoever is behind this latest DOJ action wants something…

The antitrust case is just a red herring – a distraction, a diversion – from what’s actually happening.

The only thing NVIDIA is guilty of is being ridiculously successful because of the massive risks and investments it made over the last decade. That kind of long-term thinking and strategy is rare. The company earned its success.

And just like Gollum both hated and loved the One Ring, there are those who are jealous of NVIDIA’s success… and also want the power that its technology will bring.

One ring to rule them all.