Capital Markets Update #26

Let’s assume hyperscalers et. al. spent around $425B in AI related CapEx in 2025 (Goldman Sachs).  Estimates on future capital spend by this group vary, but taking the mean of the Goldman range would suppose another $6,000B of “AI CapEx” is yet to be spent by this cohort through 2031.  These figures roughly account for anticipated spending on compute and data centers, alone, with a relatively insignificant investment in “power” which must reference some sort of behind-the-meter generation on-site. More broadly, JPMorgan expects another $1.0T to be invested in US grid infrastructure by 2035, while you can pick your number as to how much domestic manufacturing will be onshored - partially as a result of the premium we place on supply chain security for AI and its related upstream inputs.

The point of this thought piece is to set aside the profitability critique of these investments and instead look at the AI movement’s broader impact on the US economy.  We can slice and dice monetization outcomes for owners of frontier models to eternity, with wanton disregard for the fact that no comp truly exists to provide an underwritable base and downside case for future AI-attributable revenue.  Reasons for that are self-evident; it seems to us all revenue projections for this category are closer to heuristics than underwriting.  On an ROE basis, you can almost rationalize the spend given the amount of leverage in the system.  On a total revenue basis, it’s hard to imagine where the orders for AI tokens keep coming from in an increasingly price-competitive universe of relatively interchangeable models.  And by the way, fighting over price assumes you have a customer.  Its still up for grabs as to how far generative AI’s “efficiency service” can reach into the corporate sphere.  So, implicit within this dichotomy is the fact that debt is probably more at risk than anything else, having lent quite heavily against a credit rating as opposed to rational EBITDA coverage levels.  Ultimately, hyperscalers will hang a giant AI sign around the neck of each incremental dollar of revenue growth from a 2025 base into the future.  Stock prices for these AI-related enterprises, along with the massive supply chain pumping goods and services into these entities, will gyrate wildly into the future.  Volatility is probably the name of the price game here, at least until we’ve seen two or three years of public reporting from OpenAI and Anthropic.      

What’s more interesting to us is the fact that all of this capital spending is actually incredibly healthy for the American economy for a myriad of reasons beyond the narrow objective that is solving to the fair market value price for Meta, Alphabet, OpenAI or Anthropic.  Think about what the AI boom has forced America to reconcile, in practice.  Our energy grid is woefully inefficient and insufficient.  AI has supplied the future demand to privately finance the upgrade of the grid system.  What was probably going to become some sort of massive public-works fumble is being pulled forward and executed by the private sector.  We’ll give some credit here to the Inflation Reduction Act and its associated tax credit regime for clean energy projects, but not that much.  Without future AI-related demand, most projects would never pencil.   Undoubtedly, many functions beyond AI are becoming digitized and life (including goods and services production) will continue to expand into the digital realm going forward.  AI has catalyzed the private financing of a previously unimaginable increase in compute capacity – and that compute capacity can do anything, not just run AI models. If we over-supply the market with compute, perhaps the result is other technologies and digital functions have room to grow fueled by the theoretical price washout of exceptionally high-performance compute over-supply.  It’s not our realm of expertise, but seems rational. 

A few others:  Annualized quarterly productivity growth averaged a paltry 1.0% from 2010 – 2020.  AI has provided the initial inertia or catalyst for a labor productivity boom commensurate with levels seen during the late 1990’s tech boom – if not even better.  From 1998 - 2000, the economy generated 3.5% - 4.0% productivity growth, up from about 0% in 1995.  We even saw intermittent readings in excess of 4% productivity growth through YE 2005.  For reference, YOY productivity growth as of Q1 2026 was 2.8%, which is noteworthy compared labor productivity growth of ~1% during the 2010 – 2020 decade. You cant attribute much, if any, of our recent increase in productivity directly to AI. But what AI has successfully done so far is help pull the focus of C-Suites economy-wide towards efficiency, following a complete lack thereof in the over-stimulated, slightly gluttonous post-covid years.  Finally, AI momentum has fostered the proliferation of data centers throughout the United States, sometimes (and increasingly more often) in rural parts of our country.  Data centers often demand continuous base load from the grid system, which in certain communities has caused prices for ratepayers to spike dramatically.  The 2025 capacity price spike in the PJM Interconnection increased electricity bills by about 30% for 65 million people across the region (NRDC).  According to Avanza energy, ratepayers saw an average $10 - $30 increase in monthly bills throughout the year. However, according to the WSJ, a rural Louisiana school district is awarding teachers a $70k bonus in 2026 due to the increase in sales tax received by the parish from a Meta data center.  One doesn’t cancel out the other, but its not all bad.

Again, all of these externalities associated with the AI spend dynamic have little relationship to the overall price at IPO of OpenAI and Anthropic.  However, regardless of where equities price AI risk, the impact of AI momentum on the US economy extends far and wide – and is overwhelmingly positive in many, if not nearly all respects…so far.