Dear readers,
As I send this newsletter, the question of the US election outcome is, of course, top of mind. After months of suspense, we will know today if we get our first woman president or a second Donald Trump term.
Of course, the election will have a defining impact on the global AI race, with contrasting views on Biden’s executive order: will regulation be reinforced or loosened? On the other hand, AI has already had a significant impact on the U.S. elections, with AI deepfakes and scams being a top concern for election officials. As always, the actual impact will take weeks to unveil, but this election may be remembered not just for its outcome but as the moment when AI's influence on democracy - and democracy's influence on AI - became undeniable.
In the meantime, recent earnings reports from Meta, Amazon, and Alphabet reveal an uncomfortable truth: The genAI revolution collides with physical reality, creating ever-increasing infrastructure needs (and asset-heavy balance sheets) that could reshape tech economics and test investor patience.
Consider Meta's latest earnings: While genAI initiatives drive engagement, they also pushed capital expenditure projections to $35-40 billion for 2024 (up from $28B in 2023), a significant increase that initially rattled investors. Amazon's earnings showed similar strains, with internal documents revealing "zombie" data centers - facilities built but non-operational due to power constraints that could last through 2030. Meanwhile, Alphabet’s finance chief flagged an “increase” in next year’s investments, evidencing the difficulty in predicting the spending at stake.
This infrastructure challenge extends beyond public tech giants. OpenAI, despite its $157 billion valuation and projected $11 billion in 2025 revenue, expects to lose $5 billion this year on operations alone. GenAI is creating a new class of tech companies that look more like utilities or industrial firms in their economic profile - marking a fundamental shift from tech's traditional asset-light model.
While most software companies will remain technically asset-light, many are partner-dependent on cloud and LLM providers with heavy infrastructure. They could face both increased costs and capacity constraints shortly, putting pressure on gross margins. 60-70% gross margins should be the new normal in an AI-enhanced or AI-native world. A handful of software companies also want more strategic control and are testing a hybrid approach in which they own a part of the infrastructure (notably inference GPUs). This shift is creating new strategic vulnerabilities and forcing a fundamental rethinking of software economics. We explore why genAI’s next phase is about atoms, not bits in this month’s OpEd.
Two converging trends explain the change and the strategic risks involved:
First, the sheer scale of infrastructure needs is staggering. Global data center investment reached $22 billion in the first five months of 2024. According to Goldman Sachs, power demand from AI is projected to increase 160% by 2030. A single ChatGPT query consumes 10 times the power of a traditional Google search, while AI training requirements double every 6-12 months.
Second, physical constraints are becoming binding. Tech giants are making unprecedented moves to secure power: Microsoft is restarting the Three Mile Island nuclear plant, Alphabet is contracting multiple small modular reactors, and Amazon is investing in wind farms. Yet these efforts may not be enough. As Amazon's internal documents reveal, the company faces "headwinds in power, zoning and permitting, water, and workforce/labor that are providing challenges to our long-term capacity growth."
While private markets continue funding genAI's potential, public market reactions to growing capital expenditure needs suggest increasing scrutiny of the massive investments required to deliver on AI's promise. I could not agree more, but with a nuance. GenAI is a long-term play; innovation is still at full speed – Where is ChatGPT-5, has LLM hit its ceiling? was my favorite panel at last month’s TEDAI in SF - and results on genAI native application and AI-enabled tech and tech-enabled companies will take time to unfold. A fundamental approach is needed but with a multi-timeframe perspective.
OpenAI's trajectory illustrates these challenges. Despite its consumer success, the company must pivot toward enterprise customers where economics are more favorable, and infrastructure investments can be better monetized. The latest release of its Strawberry model, tailored for complex enterprise needs, goes in that direction. Its consumer business, while impressive, should be viewed as a customer acquisition channel for enterprise offerings rather than a standalone profit center. Other LLM providers, including Mistral and Cohere, have already taken this route. We take a look at OpenAI’s economics here.
For tech investors, this requires a fundamental rethinking of frameworks:
Traditional software metrics are challenged, starting with gross margins and, to some extent, fixed asset turnover (an accounting ratio that we had forgotten in tech and that is included with the “Balance Sheet Strength” pillar of our AGI methodology)
In a world of higher capex, Free cash flow supplants operating cash flow (an evolution already existing in the software world where FCF is now what is used when computing the Rule of 40)
Infrastructure and hardware choices become critical
Risk assessment must incorporate strategic vulnerabilities and physical constraints
We're entering a phase where AI's success depends less on algorithmic breakthroughs and more on solving fundamental business model challenges.
The implications extend beyond software to the entire tech ecosystem. The massive infrastructure buildout required for AI - estimated to double U.S. data center capacity in six years - creates new investment opportunities in real estate, energy, and infrastructure. Traditional technology investors must now understand power markets, real estate development, and utility regulations. It’s a fascinating and welcome evolution breaking down investment silos.
The AI revolution isn't slowing; atoms, not bits, will define its next phase. Success requires bridging the growing gap between digital ambition and physical reality and rethinking strategic frameworks, a transformation that will rewrite the rules of tech investment and innovation.
The winners will be those who recognize this reality first and act decisively.
Have a great read and a good election night!