With just hours to go before Anthropic PBC is due to raise some $65 billion in what will be the largest initial public offering in history, I can’t shake the feeling that it all feels like one giant roll of the dice on the AI bubble and praying a seven turns up.
Co-founder and Chief Executive Officer Dario Amodei made what will be one of his final pitches to investors on Friday following weeks when Anthropic’s target valuation has fluctuated between $1 trillion and $2 trillion. “We’re very paranoid,” he told JPMorgan Chase & Co. CEO Jamie Dimon during a livestream, when discussing efforts on safety and model alignment.
Holding back on new models and products due to safety concerns, Amodei warned, could be painful to investors in the short run. And what of things that might be painful in the long run? One risk stands out. “Compute is is the deepest business dilemma that the company faces,” he said, referring to the the hardware and power needed to train and run AI systems. After all, just like with any product, even best AI is useless if you don’t have the capacity to get it into the hands of paying customers.
See more: Anthropic IPO on Tap, NVIDIA Updates, & Other Investor Days Shaping Markets
Amodei’s rationale is this: Anthropic relies on more of a pay-as-you-go model than its competitors. This is good for flexibility, as it can adjust compute spend to meet demand. But it leaves the company exposed — more so than its competitors committing vast sums of money and resources to build their own infrastructure to be used as they see fit. Meta Platforms Inc. CEO Mark Zuckerberg, for instance, has often told his investors he would prefer to spend too much than too little, and worry about what to do with any excess capacity later.
If Anthropic ever faces a shortage of available compute, things could quickly spiral downward. “If you buy too much, of course, you can financially overextend yourself,” Amodei said. “If you buy too little, then you can't support inference for customers, and even more importantly you can't do the R&D” to build new frontier models, he added.
This issue isn’t theoretical. There was “increased frustration” — Jefferies analyst Brent Thill noted — at the beginning of the year among Anthropic’s clients over regular service outages. Anthropic’s uptime was worse than OpenAI’s. To put it right, Anthropic entered into a deal with rival Elon Musk’s SpaceX to use its Colussus data centers, utilizing what capacity Musk doesn’t want to use for his own AI projects. This arrangement will cost Anthropic as much as $3.1 billion a month starting in December. Either side can walk away from the deal with just 90 days notice come March — not the kind of terms one might want when dealing with the mercurial Musk.
Of the $901 billion that Anthropic has committed to spend on computing include agreements with its largest rivals in the AI space; deals that will need renewing or renegotiating down the line when the competitive landscape may have intensified. Alphabet Inc.’s Google might have found greater success with its Gemini AI and seek to use more of its own chips and servers for itself. (Of course, the opposite might also be true. If Google’s efforts fall short, they might be all too glad to cut Anthropic a better deal.)
In its IPO filing with the Securities and Exchange Commission, Anthropic acknowledges these dependencies. Changing from one data center operation to another is not straightforward. “Migrating workloads to alternative providers would be costly, time-consuming, and operationally complex,” the filing warns, “and may not be feasible within timeframes that would avoid material disruption to our model development roadmap or our ability to serve customers.”
If faced with a computing crunch, Anthropic may need to make tougher choices on how to divvy up available capacity for training or for inference. Trade offs in either direction would come with consequences for Anthropic maintaining its lead in enterprise AI. It’s certainly far easier for customers to switch out an unreliable AI model in favor of a more dependable or more advanced one — and that’s where Anthropic’s key vulnerability may lie.
Yet, despite the existential risk of being cautious to a fault, Amodei told investors the company would continue to act conservatively. "We buy relatively little compute for relatively low cost ahead of time, and then if things do not go as well relative to expectations, we’re well covered,” he said. “If things go really well relative to expectations, then we have an opportunity to buy additional compute on the spot market, which so far has always been available and which we believe will continue to be available.”
That’s all well and good if he’s right — with this likely record-breaking IPO, he’ll certainly have the cash balance to work with. But if he’s wrong, it’s a big problem.
A message from Advisor Perspectives and VettaFi: Discover something new! Click here to register for our upcoming webcasts.
Bloomberg News provided this article. For more articles like this please visit
bloomberg.com.
More Innovative ETFs Topics >