The AI Capex Roundup
Insights from Microsoft and Meta's earnings call on capex outlook and how DeepSeek's R1 impact AI spending landscape for cloud providers?
«The 2-minute version»
The AI developments from the previous two weeks created quite a market whiplash. While investors cheered the announcement of the $500B AI infrastructure investment called The Stargate Project, sentiment quickly turned south with the launch of DeepSeek’s R1 model, with roughly $1.2T wiped out of the US markets on Monday, January 27.
It started with: OpenAI’s ChatGPT launch in November 2022, which soon led to a renewed revenue acceleration cycle in cloud providers and semiconductors as AI unlocked new workloads. Investor optimism started rising, and very soon the Magnificent 7 companies were driving the majority of the S&P 500 gains.
At the same time: Capital expenditure also started ticking higher as hyperscalers invested in growing their data center capacity (GPUs, storage, networking, land, power, etc.) to support the buildout of large frontier models.
But: By the end of November 2024, headlines started screaming of “scaling laws ending.”. But OpenAI soon released its o1 reasoning model, ushering in a paradigm shift in scaling with a technique known as “inference test-time compute.”
And then came DeepSeek’s R1: A new reasoning model from China that matched the performance of OpenAI’s o1 model. Except, it is 13 times cheaper to run and is fully open source. And one more thing. This just put the conversation of ROI on AI Capex front and center.
Did all the AI spending just go to waste? While investors will be asking CEOs of the Big Tech companies tougher questions on the utilization rate of their data centers, cheaper models like the R1 should not be feared. In fact, we should see an acceleration of companies building with AI as models become cheaper, leading to significantly higher compute demand. That’s right, I am talking about widespread AI adoption and autonomous agents and apps running in parallel.
Plus: Microsoft and Meta also defended their capex plans, earmarking $80B and $65B, respectively, for 2025. While Microsoft is seeing an acceleration in its AI business with an annual run rate of $13B, Satya Nadella is optimistic about DeepSeek’s implications for its business. In the meantime, Zuckerberg has ambitious plans for 2025, where he sees Meta AI becoming a leading AI assistant, but watch for capex depreciation headwinds.
⏰ We are running a limited time offer for all new annual subscribers to lock in $80 for a lifetime. Paid subscribers will get full access to 9-12 Research posts like this one every quarter with select stock picks in the AI ecosystem. Prices will go up higher at the end of March.
Let’s set the stage- Making sense of it all
One word to summarize the AI developments from the last two weeks - Whirlwind
On January 21, President Trump announced a $500B private sector AI infrastructure investment called The Stargate Project. Its goal: improve AI infrastructure in the US at a time when AI gains mainstream adoption and competition from China is staring right in the face.
For context, $500B is roughly:
2x what the US government spent on the entire Apollo program (adjusted for inflation).
2x more than Meta, Apple, Amazon, Netflix, and Google's combined R&D spending in 2023.
Just about the same number of billions as the entire GDP of the UAE.
As markets cheered the announcement of The Stargate Project on January 21, sentiment quickly turned south when DeepSeek, the Chinese AI firm, launched its reasoning model R1 four days later, wiping out over $1.2T from the US markets, led by Nvidia NVDA 0.00%↑ on Monday, January 27.
The thing is that DeepSeek’s R1 model rivals OpenAI’s o1 in performance at just 10% of the cost. Plus, its choice to open-source R1 will have a profound effect on the AI ecosystem.
Up until now, US model makers have been locked into a single paradigm of building ever-larger and more compute-hungry models. This has led to Big Tech capex (capital expenditure) reaching a record-breaking level as hyperscalers invest in AI infrastructure to capitalize on seemingly insatiable demand.
But, with growing chatter around AI scaling laws plateauing and the shift from pre-training to inference as a possible way to scale LLMs (large language models), DeepSeek’s R1 marks a massive catalyst towards widespread AI adoption.
At the same time, it has also brought the topic of ROI (return on investment) on AI infrastructure under massive scrutiny.
Some are thinking if AI computing gets cheaper, shouldn’t demand for expensive AI chips and cloud infrastructure decline?
Not necessarily.
In fact, a more pragmatically optimistic way to look at it would be that cheaper models like R1 should lead to an explosion of building with AI. The idea is that when something becomes cheaper and more accessible, it generally leads to more consumption. And in this case, more consumption of cheaper models should lead to more inference compute, which in turn translates to higher compute demand.
Think of autonomous agents making decisions, chatbots solving complex queries, or systems interacting in real time. The more agents or apps running in parallel, the more steps the agents take, the more compute resources are needed.
That’s right, I am referring to Jevons Paradox that the whole AI community seems to be talking about after Satya Nadella, CEO of Microsoft MSFT 0.00%↑ tweeted about it post DeepSeek’s R1 launch, easing fears that more efficient AI would hurt demand for high-performance chips and cloud computing.
Having said that, I do not believe that the shift from pre-training to inference via smaller reasoning models is going to be a straight line. Hint: Market Volatility.
In this post, I will untangle the latest developments in AI to understand its implications on AI infrastructure capex moving forward. I will also dissect management commentary from Microsoft and Meta META 0.00%↑ latest earnings call to understand how they see the demand landscape of AI products and services and their forward capex plans.
In order to unlock the remainder the post (and similar ones in the future) and get access to the the thread on Subscriber chat next week where I will highlight my observations on the hyperscaler complex with commentary on valuations and price targets, consider upgrading to paid for a limited time offer of $80 annually.