In the Gospel of Matthew’s Parable of the Talents, a master gives three servants talents (currency), 5 for the first servant, 3 for the second, and 1 for the third. The first two invest their talents, doubling their money; the third buries his allowance, making it inert and unprofitable.
The master praises the thrifty two and rebukes the timorous one, with the core lesson being never let opportunities, responsibilities, and resources go to waste.
But what happens when you invest and don’t earn anything to show for it (yet)? Welcome to the AI arms race.
ChatGPT, according to a JP Morgan note, is on track for $13 billion in ARR as of August, and the company is offering a secondary equity offering that values the AI frontrunner at $500 billion.
But OpenAI didn’t reach that valuation from earnings. It achieved it from spending.
OpenAI projects that it will burn $8 billion in 2025 alone on operating activities, and it expects this to balloon to $115 billion for total spend through 2029.
In another JPMorgan note, research analysts forecast that OpenAI won’t see profitability until 2029, but that’s under prior 4-year cash burn guidance, which OpenAI has since revised upward by ~$80 billion to the current $115 billion figure.
Put simply, the electro-silicon alchemy that enlivens our ChatGPT responses comes at cost – a hefty one that is outstripping profits. And this is not just an OpenAI problem – it’s an AI problem.
Mark Zuckerberg explained the dilemma hyperscalers face when scaling: Lose money now or lose the race later.
“If we end up misspending a couple of hundred billion dollars… that is going to be very unfortunate… [but] the risk is higher on the other side,” Zuckerberg said on the Access podcast this month when discussing whether or not AI is a bubble.
Zuckerberg says he believes that it could be.
Not s***.
Anyone with a brain can see that, but the question is, who laps up the liquidity and business when the bubble pops. Because like the internet, AI is not going away, will only improve breathtakingly, and will become a daily feature of the economy (physical and digital).
AI does have a profit problem, but whether or not this is unsustainable is contingent on how these companies evolve, not necessarily AI market options or how much their products earn.
Shades of AI’s future in shale
I’m focusing on OpenAI here because they are the poster child, but again, every AI company will share the CAPEX/OPEX burden. Microsoft, for example, will likely bankroll a $100 billion datacenter planned with OpenAI as part of the Stargate Project.
OpenAI CEO Sam Altman has said that the company will have over 1 million GPUs online by the end of the year. Our back of the napkin math, assuming a blend of different GPU generations, has such a deployment costing somewhere around $30 billion for the GPUs alone. Other IT hardware, electrical, and facility buildout could add another $26.5 billion (but again, OpenAI may or not be paying some or all of these costs in certain cases).
So far, companies have split their focuses into either owning infrastructure (data centers, electrical equipment, etc) or GPU clusters. Nowhere is this clearer than in the nascent litter of hybrid bitcoin mining-AI companies; Core Scientific, Terawulf, Bitfarms, Riot, Hut 8, and others are building “powershells,” datacenters which they then lease to GPU operators, while IREN and Hive run GPU themselves.
The key to sustainability, says Crucible Capital General Partner, Meltem Demirors is the same vertical integration and financialization that the oil industry experienced, except this time with compute.
““Today hyperscalers have reached the limit of the margin they can extract via integration alone,” Demirors tweeted over the weekend. “Spending $100B in capex without the ability to guarantee monetization is not sustainable.”
Demirors begins the thread discussing Meta’s decision to open up a power trading business to hedge power cost. She highlights a parallel between Meta’s trading desk and the financialization of the oil industry with derivatives trading in the 1990s.
Meta’s energy trading desk could be used to control energy prices, but it could drive income as it does with the oil industry.
The missing piece, however, are new markets for power and, eventually, compute.
“But this time, markets will look different. The need to manage the cost of energy and atoms hasn’t changed, but the tools we have are changed…there’s a MASSIVE opportunity to build new markets, new instruments, and new firms to trade supply and demand for energy and compute,” she said.
Additionally, she expects companies to vertically integrate to control costs and improve margins, just as Standard Oil did when it moved from refining up and down the oil supply chain.
Energy costs and energy availability as Demirors points out are the nagging painpoints of the AI/HPC industry. Solving these contingent issues would ease operating costs some, but companies are still going to need to evolve their business strategies to survive.
Non of this is to doom and gloom and say that AI is vaporware on the cusp of collapse. I believe this technology will revolutionize our society on a magnitude greater than the internet that delivers it. And as with prior technological advancements, from the railroad to the airplane, the radio to the internet, hydro to nuclear power, AI will have a bootstrapping phase where costs bleed profits, and while the market may get frothy and some companies and investors will lose, others will emerge as the leaders in perhaps the most transformative technology mankind has yet devised.
