AI Capex, DeepSeek and Nvidia's Monster on the Horizon ??

This is a guest post by the folk at the Pragmatic Optimist.

• To get access to my best work, consider a Paid subscription for less than $2 a week.
• Support independent curation and journalism without Billionaire ownership or influence.
• To read the full article as it's intended to be viewed read the original here.

Go Paid

Hyperscalers are raising the stakes on their AI capex. DeepSeek’s R1 innovation is one more reason why.

I can't stop reading DeepSeek and AI Capex type content. Today Amrita Roy and Uttam Dey help us analyze all of this. Their Newsletter is also evolving: (all that comes next are their words, I hope you enjoy the article!)

Introducing The Pragmatic Optimist 2.0

Up until now, the content strategy at The Pragmatic Optimist (which we started in September 2023) was about connecting the dots in macroeconomics, technology, and culture in a jargon-free language to help readers understand the big picture.

Looking forward to 2025 and ahead, we have narrowed down our focus to the AI value chain, where our objective is to help readers and investors navigate the evolving innovation landscape, identify rock-solid businesses with strong growth trajectories, and make long-term investments in the space with a high probability of success.

As a result, our content strategy will involve understanding how different industries and companies fit into the overall AI ecosystem, how their product innovation roadmap impacts their revenue and profit cycles, and assessing industry crosscurrents and how that impacts forward valuations.

The Pragmatic Optimist 2.0 will be a paid publication, where a paid subscriber gets the following:

?? 1 Monday Macro per quarter

?? 9-12 industry research posts per quarter with a focus on relevant stocks ?? in the AI ecosystem.

?? Subscriber chat that contains all live trades with commentary on key events that signal an acceleration/deceleration in the companies’ profit cycles

?? Post comments and/or DM us with any questions.

?? Access to The Pragmatic Optimist’s portfolio (coming up in H2)

??As an extension of our gratitude, all existing and new paid members until the end of March will be able to lock in our annual rate of $80 for a lifetime.

(Go Premium for $80 annually FOREVER)

I also rank in the Top 2% of all analysts, with a 71% success rate and an average return of 29% per rating, with some of my most successful long calls that include Palantir $PLTR, Twilio $TWLO, Reddit $RDDT, Cloudflare $NET, Atlassian $TEAM and Wix $WIX.

Here are some of our most popular posts from our archive.

1. AMD’s Q4 Review: Data Center Business Hits Scale Feb ‘25
2. How Nvidia Gets Impacted By The AI Chips Ban Jan ‘25
3. Our S&P 500 Target For 2025 Jan ‘25
4. The internet is about to be flooded with AI Agents. May ‘24
5. Yes, AI is in a bubble. No, it won't pop just yet. Mar ‘24

The Q1 AI Capex Roundup: Further Loosening The Purse Strings

One word to summarize the AI developments from the last few 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 $META, Apple $AAPL, Amazon $AMZN, Netflix $NFLX, and Alphabet’s Google $GOOG 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, wiping out over $1.2T from the US markets, led by Nvidia $NVDA 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, global AI startups 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.

And Big Tech isn’t backing down. In fact, they’re raising the stakes and upping their AI infrastructure investments per their latest conference calls and earnings reports published in the past two weeks. DeepSeek, according to hyperscalers, is just another mega catalyst for AI, and these large tech companies are seeking to shore up their strategic advantages by doubling down their spending plans. The question remains: At what cost?

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 the hyperscalers’ recent earnings conference calls to understand how they see the demand landscape of AI products and services and their forward capex plans.

A Quick Primer On Hyperscalers’ AI Journey & Goals

Up until now, the Magnificent 7 companies have driven a lion’s share of the S&P 500 returns, thanks to their profit margin expansion from the AI technology supercycle.

During this technology supercycle, the goal for the hyperscalers so far has been to position themselves as vertically integrated compute providers that generally consist of three layers that include 1) Infrastructure, 2) Models, and 3) Applications to support widespread AI adoption and innovation across enterprises and consumers.

Take Amazon, for example. At the infrastructure layer, its EC2 instances offer both Nvidia GPUs as well as the company’s own custom silicon chips, such as Trainium & Inferentia. Then, Amazon also offers large language models via Amazon Bedrock. Finally, at the third layer, or the Application layer, which is what most of us can see, Amazon pushes its suite of applications, such as its GenAI-powered assistant, Amazon Q.

And so far, this strategy for hyperscalers to be vertically integrated compute providers seems to have worked. As GenAI catalyzed a new technology supercycle, infrastructure companies have been the early beneficiaries, via their capital investments to support new AI workloads.

Take a look at the graph below, where cloud providers saw their revenue growth accelerate from Q2 FY23 onwards as a result of growing AI workloads after a period of enterprise spend optimization.

In FY24, Big Tech spent $217.3B in capex, growing 55% YoY. For 2025, capex is likely to follow a similar strong growth trajectory, as companies have already signaled a willingness to invest substantially more this year, predominantly for AI, with estimates ranging up to $300B.

Battery Ventures, a global technology-focused VC, estimates a cumulative revenue uplift of $2T for public cloud providers by 2030 from maturing use cases of AI.

Despite global macro uncertainty remaining elevated, Battery Ventures believes “most companies are still in the early stages of deployment,” according to their 2024 State of Enterprise Tech Spending Survey. The VC firm’s findings noted that “a tough macro is the new normal,” but capex would continue to scale from here—a belief system that I resonate with given the high growth environment we currently are in and the tangible demand being seen in GenAI.?

So it only makes sense for hyperscalers to continue to deploy their volumes of cash that they have accumulated over the past few years.

In case you didn't notice, these companies are sitting on massive war chests of cash and highly profitable core businesses that can fund their AI-related capex without them needing to secure the equity and debt financing that is typical of such investment booms.

Enter DeepSeek Into The Capex Equation

Until now, the fierce competition for dominance in GenAI, which includes the capex investment race, has centered around the idea of “scaling laws.”

The basic idea behind scaling laws in AI are that the larger your model, the more “capable” it is. Large models have large volumes of model parameters and are typically trained on larger amounts of data, measured in tokens, which often require copious amounts of compute power (more computing chips and power), thus enabling the model to get more “capable”, a.k.a. perform more complex tasks and score better on benchmarks.

Training these models costs millions of dollars, funded by AI capex, which is often spread or amortized over years.

But, last September, OpenAI kicked off a paradigm shift in perspectives around scaling laws with the launch of their o1 reasoning model. The paradigm shift here is that the o1 model’s performance consistently improves not just with training but also via “inference test-time compute,” a technique that enhances existing AI models during the inference phase, or when the model is being used. The direct implications from o1’s reasoning model meant a change in efficiently utilizing the underlying resources to develop the model, resulting in that paradigm shift in thinking about scaling laws.

Here is how Noam Brown, an OpenAI researcher, framed the paradigm shift at a TED AI conference in San Francisco at the end of last year:

“It turned out that having a bot think for just 20 seconds in a hand of poker got the same boosting performance as scaling up the model by 100,000x and training it for 100,000 times longer.”

Then, Chinese startup DeepSeek comes along and releases their own reasoning model, R1, which shows how efficient its R1 model is. Not only did DeepSeek demonstrate R1’s outperformance vs. OpenAI’s o1 reasoning model across benchmarks, but it also revealed R1 was developed with a budget of just $6M, 13 times cheaper than OpenAI’s o1. Plus, DeepSeek’s R1 model is fully open-source.

You see, DeepSeek’s model is not an innovation in terms of performance. It’s an innovation in efficiency. Therefore, with cheaper models like R1, we should see an explosion in building with AI, which in turn should lead to more inference test-time compute, thus ultimately translating to significantly higher compute demand from where we stand now. In fact, Nvidia’s CEO Jensen Huang believes that Inference will be a billion times larger than Training in this video.

However, this also changes the math for all the players in the AI ecosystem. I am talking about model providers, semiconductors, cloud providers, and more.

Hyperscalers Defend Capex Plans; Praise the “DeepSeek Catalyst”

While markets initially freaked out on the launch of R1, wiping $1.2T from the US markets on Monday, January 27, investors poured back $4.3B into the tech-heavy fund $QQQ the following day, its biggest one-day haul since 2021.

Microsoft and Meta Platforms were the first two hyperscalers to report their earnings and began by defending their capex plans, saying it was crucial to stay competitive in the field.

Microsoft’s chief, Satya Nadella, emphasized that AI isn’t just getting cheaper; it’s getting better. As a result, this could drive entirely new use cases, leading to more businesses, developers, and industries integrating AI into products, services, and operations. Result? Accelerating demand for cloud compute and infrastructure.

Microsoft has earmarked $80B for AI in its current fiscal year, while Meta has pledged as much as $65B. However, investors remained cautious, particularly with Microsoft, which saw its shares decline 6% right after its earnings call.

Part of that has to do with Azure, which saw a sequential slowdown in revenue growth to 31%, missing the high end of expectations, while their margins came under pressure from higher AI-related costs.?

Nevertheless, the bright spot was that it exceeded its projections of $10B for the annual run rate of its AI businesses, delivering an annual run rate of $13B, up 175% YoY, as enterprises begin to move from “proof of concepts” to enterprise-wide deployment, while scaling laws continue to compound across both pre-training and inference time compute. Plus, it also saw 60% quarter-on-quarter growth in usage intensity of Copilot driven by seat expansion, while they simultaneously expanded their total addressable market with new offerings like Copilot Chat and Copilot Studio to embed AI agents into workflows.

For context, this puts AI at 5% of Microsoft’s overall revenue.

Looking forward, the management expects double-digit revenue and operating income growth in FY25, with Azure growth stabilizing at 31-32% and Copilot driving M365 adoption. At the same time, they also reiterated that their AI-driven demand continues to outpace supply, which they believe should ease by the end of FY25. What is also important to keep in mind is that their CFO, Amy Hood, discussed that capex growth will likely moderate in FY26, especially as data centers catch up with demand.

Meta Platforms, on the other hand, is aggressively investing in compute and data center capacity as it continues developing its Llama family of models. Meta noted improving engagement as well as monetization, with its daily active users growing 5% YoY to 3.35B users and Average Revenue Per User accelerating by 15% YoY to $14.25 in Q4 FY24.

But unlike Microsoft, Meta’s operating margin expanded on both a sequential and year-over-year basis to 48% despite expanding losses from their Reality Labs business division, as they streamlined their operating expenses and improved monetization efficiency through their Core and GenAI initiatives.

Meanwhile on GenAI, Zuckerberg has big ambitions where he expects Meta AI to become the leading AI assistant in 2025 as they launch their multimodal Llama 4 during the year with agentic capabilities. Plus, he also pointed out that 2025 might be the year where their Meta AI glasses become the next computing platform, as they have been talking about for some time.

All of this requires large, long-term capital investments, and Zuckerberg is sticking to his long-term view, where he said the following during the earnings call:

“And at this point, I would bet that the ability to build out that kind of infrastructure is going to be a major advantage for both the quality of the service and being able to serve the scale that we want to.”

Finally, Google & Amazon echoed Meta’s goals for increasing investments in their respective data center infrastructure to scale capacity so that they could meet the strong demand they are seeing.?

While Google’s revenue growth accelerated by 14% to $350B in 2024 and operating margins expanded by 470 bp to 32%, it was really the 63% jump in Google’s capex that stirred up some concern. The concerns became more pronounced when markets saw Google's cloud unit was reporting slower year-on-year growth at 30% y/y in Q4, down from the previous quarter’s 35% y/y and underwhelming Q4 expectations of 32.5%.

In contrast to Google’s Cloud growth, Amazon’s AWS reported slightly better figures, where AWS grew at 19% y/y in Q4, at the bottom range of the market’s expectations of 19-20%. For the full year, Amazon reported consolidated sales of $637B, up 11% y/y, while capex for the year grew 48% to $83B.

Particularly, when it came to questions about DeepSeek’s business implications for Amazon AWS, this is what Andy Jassy, CEO of Amazon, said:

“First of all, I think like many others, we were impressed with what DeepSeek has done. And I think if you run a business like AWS and you have a core belief like we do that virtually all the big generative AI apps are going to use multiple model types and different customers are going to use different models for different types of workloads. You're going to provide as many leading frontier models as possible for customers to choose from. And that's what we've done with services like Amazon Bedrock. And it's why we moved so quickly to make sure that DeepSeek was available both in Bedrock and in SageMaker, faster than you saw from others and we already have customers starting to experiment with that.

I believe the cost of inference will meaningfully come down. I think it will make it much easier for companies to be able to infuse all their applications with inference and with generative AI. And I think it's going to -- if you run a business like we do where we want to make it as easy as possible for customers to be successful building customer experiences on top of our various infrastructure services, the cost of inference coming down is going to be very positive for customers and for our business.”

For FY25, Amazon sees its capex rising by ~20% to cross $100B, a much more reasonable increase in capex versus the planned capex budgets of its hyperscaler peers. Also, unlike other hyperscalers, Amazon uses a portion of its capex to continue the buildout of its logistics network, warehousing, and robotics projects to support its e-commerce goals.?

But when it comes to AI and the capital investments needed to support the capacity requirements for addressing the AI demand, Google & Amazon’s capex budgets move in a slightly different trajectory as compared to Microsoft and Meta.

Both Google & Amazon have been increasing their investments & resources over the past twelve months in building out their own custom-made AI chips. While Amazon offers Trainium1/Inferentia2 chips for model training and inference purposes, Google has been ramping up the production of their in-house custom-silicon TPU chips in order to improve the cost-efficiency of running AI workloads for their customers. Both these companies rely on partnering with semiconductor vendors such as Marvell $MRVL, Broadcom $AVGO, etc., to build their custom semi chips.

Connecting the dots together, 2025 should witness another 45-50% increase in capex from hyperscalers. While many investors are increasingly questioning the ROI of AI capex, especially after the launch of DeepSeek’s R1 model, we believe that we are about to enter a period of time where we see an explosion of companies building with AI (better economics), leading to widespread AI adoption across use cases among enterprises and consumers, which ultimately results in a significant rise in the demand for compute.?

Having said that, the demand curve is not likely to be a straight line. In the short term, some of these companies may face headwinds to their margins from capex depreciation.

But that doesn’t change the long-term outlook, where model-agnostic hyperscalers should stand to win from growing compute demand, as Satya Nadella and Andy Jassy repeatedly emphasized. As for Meta, it wants a future where open-source wins. Therefore, as it implements some of R1’s innovations into Llama, it should drive growing GenAI adoption through its advertising and hardware business lines at lower infrastructure costs, thus justifying its capex spend.

References

Guest post guide for AI Supremacy - by Michael Spencer