The AI boom is leading to a massive computing crunch… electricity bills are skyrocketing… but aren’t the hyperscalers supposed to pay?… invest in the “AI demand shocks”… where senior analyst Brian Hunt is looking today
Everyone’s talking about oil, but I think what the world is mainly short of is tokens.
That line comes from Ben Pouladian, an engineer and tech investor quoted in The Wall Street Journal on Sunday.
To make sure we’re all on the same page, a token is simply a unit of measurement tracking how much computing power an AI task consumes.
Think of it as the basic currency of the AI economy – every query, every generated document, every autonomous agent action draws from the supply. And right now, that currency is running short.
This has significant investment implications that we’ll get to shortly. But first, let’s look at the scope of the problem…
The computing crunch is already reshaping how AI companies operate
The WSJ ran a detailed look this week at what’s happening inside the AI infrastructure stack, and the numbers are striking.
Over the past several months, demand has exploded for “agentic” AI, the latest evolution of AI. “Agents” don’t just answer questions but autonomously perform tasks: writing code, scheduling appointments, managing complex multistep workflows.
The shift from conversational AI to agentic AI is causing a dramatic spike in computing consumption that existing supply chains weren’t built to absorb.
Here’s the WSJ with an example of the astonishing demand:
Token use in OpenAI’s API—a platform where mostly enterprise users access its software—rose from six billion a minute in October to 15 billion a minute in late March.
The supply side can’t match it.
According to the Ornn Compute Price Index, hourly rental prices for the most advanced Blackwell-generation GPUs from Nvidia Corp. (NVDA) – the ones that power modern AI – have risen to $4.08 per hour, up 48% from just two months ago.
The strain is already showing in ways that are hitting end users directly
Anthropic announced in late March that it would begin rationing computing access during peak weekday hours. Enterprise clients have started switching to competing providers. OpenAI scrapped its Sora video-generation app in part to redirect computing resources toward higher-priority products.
The WSJ captured just how acute the pressure has become. From the article, quoting J.J. Kardwell, CEO of cloud infrastructure company Vultr:
There’s a massive capacity crunch that’s unlike anything I’ve seen in more than five years running this business.
The question is, why don’t we just deploy more gear?
The lead times are too long. Data center build times are long. The power available through 2026 is already all spoken for.
Reread that last sentence….
The power is already spoken for.
So, what’s the significance of that?
Well, let’s take it one step further – what really is a token shortage?
Basically, it’s an electricity shortage wearing a tech hat.
Power: the constraint behind the constraint
Pull back one layer from the computing crunch and you find an energy problem that runs deeper and wider than most people realize.
On Monday, Bloomberg published an in-depth look at the growing size of America’s electricity bills, and the data behind it raises an eyebrow.
From Bloomberg:
The North American Electric Reliability Corp., the country’s grid security regulator, forecasts that US power demand in summer will rise 224 gigawatts over the next decade — roughly the equivalent of adding 180 million homes.
One analyst said the last comparable surge came during World War II.
The AI buildout is a central driver of this increase.
Let’s look at an example to drive this home…
The Pennsylvania-New Jersey-Maryland (PJM) Interconnection is the nation’s largest power grid, stretching from the Midwest to the East Coast. Over just the three years ending in May 2028, data centers are projected to add at least $23 billion to customer bills on the PJM grid alone. That’s an increase of more than 50%.
Meanwhile, in parts of eastern Pennsylvania, electricity prices have already risen 200% since 2020.
Now, to offset this, in his February 24 State of the Union address, President Donald Trump told America’s largest technology companies they would “have the obligation to provide for their own power needs.”
Ten days later, Amazon.com Inc. (AMZN), Alphabet Inc. (GOOG), Meta Platforms Inc. (META), Microsoft Corp. (MSFT), OpenAI, Oracle Corp. (ORCL), and Elon Musk’s xAI gathered at the White House and signed what the administration calls the Ratepayer Protection Pledge – committing to “build, bring, or buy” all the power and grid infrastructure required for their data centers, with none of those costs passed to American households.
Trump was candid about the politics driving it. At the signing ceremony, he told the assembled tech executives:
They need some PR help because people think that if a data center goes in there, electricity prices are going to go up.
So, that should settle it, right?
Here’s the part that doesn’t make the press releases
The pledge is basically forward-looking, but the price increases Bloomberg documented have been building for years.
The PJM grid’s capacity prices – what utilities must pay generators for electricity – exploded from $28.92 per megawatt-day in the 2024-’25 delivery year to $329.17 in the 2026-’27 delivery year.
That’s an 11X increase already baked into utility rate structures long before anyone signed anything at the White House. But it’s still not enough – the most recent PJM capacity auction fell 6.6 gigawatts short of available supply.
Plus, as I just noted, the pledge addresses future data center buildouts, not existing ones.
So, today’s power bills reflect:
- The power infrastructure already built to serve current hyperscaler facilities
- The rate increases already approved by utility commissions to pay for the related grid upgrades
These costs are very real, very big, and baked in – and the pledge doesn’t unwind them.
Plus, even for new projects, signing a pledge to build your own power plant doesn’t make one appear. Permitting a new generation facility takes two to four years. Construction takes more years on top of that. And by the time the concrete is poured, the demand it was designed to meet has often already doubled.
Which brings us to the investment implications…
When demand outruns supply, the winners aren’t always who you expect
Every major tech boom in history has eventually run into massive demand for the critical components related to that technology’s buildout.
Brian Hunt, editor of Money & Megatrends, has built an entire investment framework around this dynamic. He calls it the “AI demand shock.”
Here’s Brian explaining the core concept:
For the past three years, the best way to make money quickly in stocks has been to locate an industry where an AI “demand shock” is about to strike… and then invest there before the shock arrives.
Not a supply shock, mind you, where a war or a pandemic abruptly cuts off the supply of a resource like oil.
Instead, I’m talking about a “demand shock,” where demand for a specific resource or manufactured product suddenly skyrockets… and sends its price hundreds of percent higher.
The historical examples Brian cites drives home his point.
In 2023, the AI demand shock for advanced semiconductors sent Nvidia up 525% in under two years. Around the same time, the sudden need for data center cooling systems drove Comfort Systems USA Inc. (FIX) up 1,000%. Meanwhile, demand for advanced optical systems – the components that allow fast data transfer between AI servers – drove Lumentum Holdings Inc. (LITE) up 1,164% in two years.
None of these companies are AI companies in the headline sense like OpenAI or Anthropic. Instead, they’re the picks-and-shovels suppliers to the AI buildout – the firms sitting upstream of the technology, making the physical things the technology couldn’t exist without.
Brian explains why these moves tend to be so large and so fast:
AI is advancing at such a rapid pace that AI-driven demand shocks are now happening every year… and creating the fastest – and most lucrative – stock market moves we’ve ever seen.
The typical manufacturing industry needs five-to-10 years to build operations capable of meeting increasing demand. Same with mining industries that supply critical raw materials.
But our new, lightning-fast technological cycles now move way, way faster…
We now have crazy mismatches in the economy’s interlocking and interdependent parts.
It’s like we have a rocket engine attached to the drivetrain of a Toyota Corolla.
This mismatch between the rocket’s engine and the Corolla’s drivetrain is where the investment opportunity lives.
Brian’s current focus: the chemistry of AI
While opportunities are all over the place, Brian’s Tuesday issue of Money & Megatrends highlighted a sector that might surprise you…
Chemicals.
Here’s Brian:
The chemicals sector is commonly thought of as an “old economy” industry that produces products such as plastics, paints, solvents, cleaners, and pesticides.
However, some chemical firms are involved in “brand new economy” activities related to AI.
The chemicals industry sits upstream of almost every physical component in the AI stack: from the specialty gases and materials used to manufacture semiconductors to the high‑purity solvents, coatings, coolants, flame retardants, and advanced polymers that make modern data centers possible.
Every AI server relies on a long chain of ultra-specific and ultra-pure chemicals. As AI usage explodes, the need for more chemicals and more sophisticated, higher-purity ones increases.
Brian’s March 6 recommendation to add chemicals to your portfolio is already paying off…
Chemours Co. (CC) is up 37% since that note and just hit a new one-year high. And Element Solutions Inc. (ESI) has added 15% and also just hit a new one-year high.
For readers who want Brian’s full analysis – including the specific names he’s watching for connected to all the demand shocks he’s tracking beyond chemicals – his research is available for free in Money & Megatrends
Circling back to the token shortage
Ben Pouladian’s observation that opened our Digest – the world is mainly short of tokens – is true. The shortage is real, and it’s disrupting AI companies’ ability to serve their users right now.
But pull back far enough and you’ll see the broader sequence…
- The token shortage is downstream of a GPU shortage…
- Which is downstream of a data center shortage…
- Which is downstream of a power shortage…
- Which is downstream of an infrastructure buildout that the physical world simply hasn’t had time to complete.
This is what a genuine technology transition looks like from the inside. The demand arrives faster than the supply chain can respond.
The companies positioned in the middle of these constraint points – the ones making the chemicals, cooling the servers, supplying the power infrastructure – are where the big money is being made today…and where we want to be.
We’ll keep tracking this with you here in the Digest.
Have a good evening,
Jeff Remsburg
(Disclaimer: I own LITE)
