It’s big. Artificial intelligence is changing our work and our lives in ways we see already and in others we can’t predict. At Magnolia, we’ve learned to make predictions sparingly. In that sense, this moment is made for us.

The cost of building an AI-enabled future is big. How big? Compared to past infrastructure buildouts, it’s on track to be the biggest.‍

Note: Rail figures adjusted to 2026 dollars capture relative economic capacity (GDP-equivalent weight); standard adjustment yields ~$350B. AI figures based on 2023–2030+ U.S. projections, with post-2026 dotted/shaded values representing industry consensus estimates.
‍
^{‍Sources: Bureau of Economic Analysis; Federal Highway Administration; Historical rail databases; Industry estimates (Dell'Oro, Synergy, Gartner).}

And the comparison is even more striking when we consider the compressed timeline.

A (big) caveat: A projection is a kind of prediction, and the estimates rely on a number of unknowns.

The chips need data centers to run, the data centers need permitting, wires to deliver electricity, and energy to generate power. To stay on its current course, the buildout will also require demand for all this intelligence and pricing power to cover the cost, and the capital to pay for it.

None of this is guaranteed. Consider:

• The frontier models from Anthropic and OpenAI are producing extraordinary results for cutting edge projects, but for most people AI adoption is using a chatbot instead of a search engine–a commodity service.
• Even business-friendly states and municipalities are pushing back on data center development because they don’t want to live near them and they don’t want them driving up their electricity bills.
• AI needs better PR: an executive who predicts massive job displacement or possible human extinction shouldn’t be surprised at society’s lack of enthusiasm.
• Electrical utilities are less willing and able to provide guarantees of power to new projects.
• If cost and availability of power are the rate-limiting factors for AI development in the US, China–with its cheaper electricity–can compete even more effectively on price than they already do, with open models.
• Enterprises (and others, like publishers and even mathematicians!) have good reason to consider doing inference in-house and even owning their own hardware: a large language model trained on their data (as we’ve seen) can be an existential threat.
• The capital markets are facing a big test: the parent companies of Claude and ChatGPT seem likely to come public at lower valuations and raise less capital than predicted just months ago; bonds of some of large AI players (e.g., Oracle, CoreWeave) are trading at discounts; the appetite for massive private loans for infrastructure is stretching investor demand.

Still, there’s a strong case for the buildout to continue or even accelerate. If the frontier models prove able to meet some of society’s deepest needs–cures for cancer, for example–we should expect stronger support from investors and even government backstops. If AI is truly the battlefield on which great powers will compete, national support of every element that it requires–power, transmission, permitting–will follow. And let’s not discount innovation: to take just one example, modern data centers require about as much water as a few almond trees.

Given the importance of its development to financial markets, we intend to revisit these ideas in future letters and blog posts, but we’ll also report the impact on our services and the financial industry, as well as our perspectives on its broader impact. Artificial intelligence is a big subject.