If Amazon wants to charge you for the personalized AI advantage it wants to develop over rivals Google and Microsoft, then it needs to have the best possible models running on its local accelerators. Just as Google is already doing with its own Gemini LLMs on its own TPUs and as Microsoft will eventually do with OpenAI’s GPT on its Maia accelerators.
And that’s why another $4 billion investment in Anthropic last week by Amazon, the parent company of the Amazon Web Services cloud and the biggest beneficiary of the massive spending you’re all making on this cloud while you pay a premium for a high-end computer product, was absolutely predictable. This fact and others are why we can expect to see tons of money flowing between Amazon and Anthropic and Microsoft and OpenAI for years to come – at least until Anthropic and OpenAI become rich enough to create their own AI accelerators and build their own infrastructure.
There are broad, holistic strategies at work here between model builders and cloud builders. We explained some of this back in August in Sugar Daddy’s Boomerang Effect: How AI Investments Are Blowing the Clouds when we reviewed the financials of Microsoft Azure and Amazon Web Services and talked about their parent companies’ $13 billion and $4 billion investments in OpenAI and Anthropic, respectively. And we wondered aloud how much of the increase in cloud computing spending in 2024 was due to investments being made in AI startups by Amazon and Microsoft. We think this is an important number and we also think that, given this, this figure should be disclosed in their financial statements. Google parent Alphabet has invested $2.55 billion in Anthropic so far and is hedging its bets on large language models.
Wall Street is just happy that someone is still investing in AI and that this boom town is still growing. Eventually, all of these investments in AI will need to provide a return on investment, and so far people are generally hoping that this will be the case, while being somewhat concerned about the impact that AI will have on the knowledge economy.
We delved deeper into Anthropic in September 2023, and we’re not going to revisit the story again. At the time, AWS injected $1.25 billion into Anthropic, and the two agreed to begin porting LLM’s Claude family to the cloud maker’s Trainium AI training chips and Inferentia AI inference chips . We delved deeper into Trainium and Inferentia chips in December 2023 and explained how AWS could undercut Nvidia GPUs for AI training and inference with its strategy. In March of this year, Amazon injected another $2.75 billion into Anthropic, and last week it invested another $4 billion. At the time, using Nvidia “Hopper” H100 accelerators on the cloud, we calculated that $4 billion only covered the cost of training about three dozen 2 trillion parameter LLMs over a period of time. of 90 days.
With credible local AI chips and sufficient volumes to reduce unit cost, AWS can provide better value on AI clusters than with very expensive Nvidia GPUs. And with the Claude models developed for Trainium and Inferentia, Anthropic can become their biggest customer by evolving their models to make them better and more accurate. AWS can continue to iterate on hardware to meet the needs of Anthropic software, creating a virtuous loop that can then be extended to the AWS-operated Bedrock AI platform service, released from beta a year ago, which supports supports Claude as well as a multitude of other LLMs, and which already has tens of thousands of customers paying for Claude on the cloud.
At some point, the revenue stream from Claude models running on Bedrock is large enough to generate enough profit to actually be able to cover the AI training costs and inference needs of Amazon, the retailer, and the company. entertainment. This crossover point likely occurred several years ago with generic data center infrastructures, although it is difficult to calculate precisely.
It’s the genius of managing a very large cloud while carrying out another activity: in the case of Microsoft, it distributes software and tracks its use, and with Amazon, it sells products online and sends them to us. either on the roads or on the Internet. Interestingly, the IT infrastructure needed for Google’s search, video streaming and advertising businesses is still vast, and Google Cloud is still not profitable enough for this effect to take hold at the Chocolate Factory. But we believe that day will inevitably come.
And, as we’ve said, we believe big cloud investments in big LLM providers will continue at a steady pace, even as the latter continue to try to raise money independently and keep their clouds sugar daddies as minority shareholders.
That’s the trick, and it’s probably also one of the reasons why Elon Musk decided not to build the 100,000 GPU “Colossus” machine at xAI in collaboration with Oracle Cloud and decided to take over a former Electrolux vacuum cleaner factory outside of Memphis, Tennessee and have Supermicro and Dell build the iron for Colossus. Musk knows very well, thanks to Tesla and SpaceX, that cloud AI is much more expensive than on-premises AI. And in the long term, we wouldn’t be surprised to see the “Dojo” AI engines and associated systems created by Tesla used at all four companies controlled by Musk. (Including X, formerly known as Twitter, in addition to xAI, Tesla and SpaceX.) Why not? Musk clearly wants to control the fate of these companies and is rich enough to invest in building his own platform for these companies.
It would be even funnier to see Dojo break away from Tesla and sell its technology to other Musk companies. Why not?
For OpenAI and Anthropic, their independence depends on their ability to raise successively larger “up round” financing, which increases their valuation and thus dilutes the stakes of their cloud sugar daddies.
Scouring the internet, we see that Anthropic has a valuation of just $18 billion, which seems low enough for AWS to still have a minority stake in the company. By our calculations, that’s 44.4 percent of Amazon when considering investment versus valuation. Comparing AWS’ investment to the $13.7 billion in total funding, that represents a 58.4% stake, which would not make AWS a minority shareholder. (And therefore this can’t be how companies calculate their stakes, we think.)
OpenAI just raised $6.6 billion last month and has a valuation of $157 billion; Microsoft, Nvidia, SoftBank and a number of venture capital firms have invested a total of $21.9 billion for OpenAI, and it’s on track to generate perhaps $3.7 billion in revenue this yearbut is also expected to lose $5 billion. If Microsoft’s investments were counted as financing – and we are not suggesting that they are – Microsoft would own at least 59.3% of OpenAI. This clearly didn’t happen, because if it had, Microsoft would own OpenAI and behave accordingly. But somewhere north of $13 billion, against a valuation of $157 billion, there’s no less than an 8.3 percent stake.
With fundraising underway at xAI, the Musk AI company is rumored to be worth between $40 billion and $45 billion. xAI has raised $11.4 billion in four funding rounds, including a $5 billion round last week. This funding round covers the cost of the Colossus machine and its data center, we believe, based on back-of-the-envelope calculations. Just purchasing systems with 100,000 H100 GPUs would cost on the order of $4.7 billion, including networking.
Maybe Musk will build a cloud? This seems inevitable, once you see how the math works. Then you could all use the xCloud (as we might call it) and subsidize the data processing needs of Tesla, SpaceX, xAI and X.
At some point, we also think it’s inevitable that Nvidia will build a cloud. The profits are simply too great to avoid.
Maybe Jensen Huang, co-founder and CEO of Nvidia, and Musk can do it together? Made you laugh. But stranger things have happened in this computer racket. This one probably won’t do it. But imagine if they built competing clouds. . . .
