A popular view of generative AI is that it is unjustifiably expensive, chronically wasteful, rarely useful, and forced on the general public for ideological reasons even as it makes the services they depend on worse. Governments will certainly have to deal with it.
This is what Barclays says:
The first wave of AI is in full swing, powered in large part by billions of hyperscale dollars. The concern with Nvidia, and ultimately with the second wave of AI ecosystems, is where the next bag of dollars will come from once hyperscale capex cannot shift further to AI or grow meaningfully year over year.
In recent months we have seen a growing initiative from countries around the world to quickly educate themselves and stay at the forefront of the powerful potential of AI. In practical terms, this amounts to public announcements of spending plans of hundreds of millions and even billions of dollars from several countries (Saudi Arabia, Singapore, Germany, Britain, India), which will go towards supporting the AI hardware ecosystem .
The AI cash burn in the private sector is already state-sized. The combined investments of Amazon, Meta, Google and Microsoft will amount to approximately $200 billion this year, according to Bernstein Research.
These sunk costs must provide some sort of return by the time the depreciation expense hits the income statement. Continued acceleration in capital growth depends on companies finding something the public wants to buy. Their need for income may soon become urgent, and the concepts so far are not encouraging.
But because political leaders care more about one-upmanship than ROIC, taxpayer-subsidized AI can continue to flourish even if the corporate bubble bursts.
Barclays estimates in a recent note that if countries outside China matched the US’s $4 billion in AI spending, scaled up to GDP, it would add another $3.5 billion:
For Nvidia, this amounts to barely a monthly turnover. The most important thing is the replacement cycle.
Barclays estimates that hardware purchases will become obsolete within two years. As AI hardware becomes larger and more expensive, total annual government spending could easily exceed $25 billion very quickly:
Overall, we view AI as the most powerful catalyst for technological progress, but also as a major security risk as hostile countries increase their capabilities, ultimately justifying our estimated expenditures and giving us confidence that the numbers should be significantly higher.
The US took an early lead because its government is relatively enthusiastic about AI. In September, a Federal Use Case Inventory was released identifying more than 700 potential applications, and earlier this month a Senate roadmap for artificial intelligence policy proposed a $32 billion R&D budget.
While such figures may prove fanciful, the larger costs are much less scrutinized. The Senate roadmap does not include defense, which appears to be responsible for almost all current U.S. federal AI spending.
A government procurement study published in March by the Brookings Institution found that the US Department of Defense was aggressively ramping up AI investments in 2022. Based on maximum potential contract value, just over $4 billion of the $4.56 billion in AI procurement costs last year was for the defense agency, Brookings calculates.
Senate Majority Leader Chuck Schumer has said the US AI defense budget should be increased by about eightfold. The exact purpose of all these investments will remain classified information.
Countries that follow America’s lead will want to build something in-house that is at least equivalent to OpenAI’s GPT-4, Barclays says. Last year’s best technology represents “the minimum starting point for countries trying to stay at the forefront of AI, for both economic and security purposes.”
The cost of processor blades for such an installation will cost $600 million at today’s prices, plus the same for interconnection, storage, energy costs, and so on. What such a scheme will not do is frighten the enemy. That requires us to stay on the cutting edge of AI, which is a… lot more expensive.
GPT-4 training would have required 25,000 accelerator cards, while GPT-3 – released less than three years earlier – required only 1,000. The grid below gives an approximate idea of current all-in construction costs in increments of ten thousand accelerators or XPUs.
If hardware cost inflation continues at its current rate, the cost of a single best-in-class AI computing cluster could easily exceed $5 billion, Barclays says:
The Info Wars arms race will escalate so quickly that only about 15 countries can afford to participate, Barclays says. And for those who can pay, there is no option to opt out, the report says, because “AI capabilities have become one of the most important, if not the most important, national initiatives worldwide”:
In our view, the global development of AI applications will undoubtedly become a national security issue, just as the government sees domestic leading chip manufacturing, and under the lens of the approximately $39 billion CHIPS Act passed several years ago, we see plenty room in the government budget for higher spending on new clusters and more advanced hardware as soon as they are immediately available.
Additionally, we believe that new AI/compute investment plans from Saudi Arabia ($40 billion AI investment fund according to the New York Times), Singapore, Germany and even India could push policymakers to take action sooner rather than later to drive more robust AI to write. convert investment plans into policy
So buy Nvidia, Barclays tells customers. The shares may look expensive, with a lot of risks related to sanctions and antitrust measures, but officials don’t know better than to buy servers off the shelf:
We see NVDA as the biggest beneficiary of Sovereign AI, given its already dominant share of the merchant AI accelerator’s installed base, performance leadership, and developer community preference. We also see the Sovereign AI market as a strong potential user of the company’s full rack solution [ . . . ] given the lack of technical knowledge and resources among government agencies needed to put together customized solutions around vendor hardware. Overall, we view Sovereign AI’s expected spend as complementary to the entire AI ecosystem, and therefore believe it will trickle down to the broader AI ecosystem as well.
And surely. Why not. Once a trailing PE gets above 300x, anything is possible.
Being an ESG darling was a big part of last year’s Nvidia buying case. A year earlier it bordered on the shitcoin bubble, and before then it was mostly about Cyberpunk 2077 frame refresh rates. Now it’s a buy because it’s the de facto arms supplier for World War II GPT.
A common thread connecting Nvidia’s customers and shareholders is that they don’t know what they’re buying or why they need it, but they’re confident they should have it. An international arms race for billions of dollars and tiger-repellent rocks would fit this description perfectly.
Read further:
— ‘Sell Nvidia’ (FTAV)