April 18, 2024

Sam Altman’s chip ambitions may be crazier than feared • The Register

Opinion OpenAI CEO Sam Altman’s dream of establishing a network of chip factories to drive AI growth may be much, much wilder than feared.

As reported last month, Altman is reportedly seeking billions of dollars in funding from partners such as Abu Dhabi-based G42, Japan’s SoftBank and Microsoft, to build all those neural network accelerator factories.

Now, a Wall Street Journal report, citing even more anonymous sources, claims that the ambitious project could involve raising up to $7 billion.

This is a staggering sum that, from this vulture’s point of view, defies logic.

To put the figure in perspective, that’s almost 14 times the total revenue of the entire semiconductor market last year. According to Gartner, global semi-revenues will surpass $533 billion in 2023. And despite all the hype around generative AI, analysts expect the sales figure to grow 17 percent to $624 billion this year. anus.

But let’s say, for the sake of argument, that Altman and his partners are really that brave and can somehow come up with a quarter of America’s gross domestic product by 2023 to fund the effort. What does 7 trillion dollars buy you?

That’s enough cash to devour Nvidia, TSMC, Broadcom, ASML, Samsung, AMD, Intel, Qualcomm and all the other major chipmakers, designers, IP owners and equipment suppliers entirely, and still have them left over. billions.

While it would be fun to see Sam burn a prodigious amount of money kicking off what would be the biggest antitrust battle of the century, what he most likely has in mind to boost chip production is investing that money in chip factories and packaging. processors. In reality, we can think of many better ways to spend that kind of money, but let’s stick to chips for a moment.

That’s a lot of fabulous ones.

No matter how you slice it, $7 trillion is still a huge sum to spend on factories, even a network of them.

The cost of a state-of-the-art chip factory today ranges between $10 billion and $30 billion, depending on the size of the site and its location. Let’s say Altman’s planned facilities end up costing around $20 billion on average. At that rate, $7 trillion generates about 350 smelter sites.

The question then is: who is going to build them? These facilities are among the largest and most complex operations in the manufacturing world, requiring components and materials from countless suppliers and specially trained personnel to install, maintain and operate them.

Because of this, it is not uncommon for these facilities to take four or more years to come online and potentially much longer to bring yields to acceptable levels. There is nothing quick about building factories correctly.

In the United States, we have seen a flood of investment in domestic semiconductor manufacturing and R&D, driven in large part by a $53 billion government subsidy fund made available by the CHIPS funding bill. . However, foundry operators have already run into serious problems.

As we previously reported, a shortage of skilled workers has already delayed the development of TSMC’s factory in Arizona. TSMC has even gone so far as to send technicians from Taiwan to the United States in an attempt to get the facilities back up and running.

Last summer, the Semiconductor Industry Association (SIA) and UK-based Oxford Economics warned that the US semiconductor industry faced a shortfall of 67,000 technicians, engineers and computer scientists by 2030. fabs , puts that number between 70,000 and 90,000 in the coming years.

And that’s just for a handful of factories under development in the United States. It doesn’t take much imagination to see how 350 additional sites would be problematic on a global scale.

Flooding the market

As if that were not enough, the demand for semiconductors tends to fluctuate cyclically. Massive purchases are often followed by long digestion cycles, and increases in PC sales tend to coincide with operating system or software releases.

We assume for a moment that those hundreds of factories will not only serve OpenAI or the AI ​​world in general, but also everything adjacent to it, although it may be that Altman really just wants an endless stream of machine learning accelerators and related applications . calculate.

The memory market is only now recovering from a glut of inventory that drove average selling prices to historic lows. Meanwhile, Intel has reportedly pushed back the completion date for its Ohio factories to late 2026, blaming current weaknesses in the semiconductor market and delays in securing CHIPS Act funding.

Of course, industry gossip has yet to detail a timeline on which Altman’s rumored $7 trillion semiconductor company will unfold. It’s safe to assume it won’t happen overnight. These types of developments must be adjusted to avoid overly aggressive construction and flooding the market with too many chips.

Even if it is spread out over the next 25 years, we are still talking about a huge amount of money, enough for 14 factories a year at a cost of $280 billion annually. To reach that mark, TSMC, Samsung and Intel would need to roughly triple their capital spending and direct it all to chip plants.

Admittedly, that sounds less far-fetched, but given that theoretical timeline, why would Altman possibly need to raise $7 billion now? Typically when you see companies like Intel talk about their foundry roadmaps, they only tend to fund what’s immediately in the works.

For example, when the x86 giant announced its plan to invest $100 billion over the next decade in an Ohio megafab, it actually only committed to building two sites at an estimated cost of $10 billion each. And, as we mentioned before, even that has been delayed.

Part of a larger plan?

So maybe this $7 trillion project is really a broader plan to boost OpenAI’s ambitions. All those chips will have to go somewhere. That means you’ll not only need factories to make the chips, but also data centers to use them and (hopefully) clean energy to run it all, and that will cost you a lot of money, too.

The chips used to power AI models consume a lot of energy. A single eight-GPU Nvidia H100 node is rated at 10.2 kilowatts. Expand that up to 350,000 GPUs (that’s the number Meta says it will deploy this year) and you’re looking at a huge amount of power.

With a budget of $100 billion, just 1.4 percent of the $7 trillion budget, for GPUs, five million H100s could be purchased at a volume price of $20,000 each. For the record, that’s more than double the amount Nvidia is expected to ship in all of 2024.

Needless to say, power will be an issue. So it would make sense to save some money to address that challenge.

The good news here is that Altman has a long history of backing energy startups. Last year, Oklo, a nuclear fission startup backed by the CEO of OpenAI, announced plans to go public.

Meanwhile, on the more experimental side of things, Altman has supported Helion Energy, which is working to commercialize a modular helium-3 fusion power plant. Even though Helion had yet to prove that its reactor actually works, Altman’s involvement appears to have been enough for Microsoft to sign a power purchase agreement with the startup. The technology is not expected to be implemented until at least 2028, assuming they ever manage to get it working.

In any case, this leads your humble pirate to the conclusion that either the $7 billion used to describe the scope of Altman’s ambitions is vast hyperbole or part of a larger, more holistic plan. ®

Leave a Reply

Your email address will not be published. Required fields are marked *