The development of AI technology is the stock market issue of 2026, 2027, 2028, and likely beyond. AI has created trillions in stock market value, and set off an epic arms race among technology giants looking to dominate the space.
It's also stoked fear of a bubble that threatens to take down stocks and the U.S. economy. A Barron's analysis recently suggested that the U.S. economy can handle spending 25% of economic output before a new-technology bubble pops. That's roughly $7.5 trillion, a threshold that won't be reached for years.
That's the one number investors worried about AI should know. Here are five more.
1 Gigawatt
AI computing is often expressed in gigawatts, which is the power required to keep the computers running. A gigawatt of AI compute might need 500,000 to 700,000 Nvidia GPUs and cost $40 billion. Several firms have building estimates. Investors can also look at SpaceX's reported results.
Elon Musk's rocket and AI company has spent roughly $42 billion for 1.4 gigawatts of AI compute. That's closer to $30 billion a gigawatt, but SpaceX's initial data center, Colossus I, went into an existing factory in Tennessee. It didn't have all the typical infrastructure costs. What's more, Musk has a knack for doing things fast with a focus on costs. Colossus I's initial phase was ready in 122 days.
SpaceX is renting about 330,000 Nvidia GPUs to Anthropic and Alphabet's Google for about $2.2 billion a month, or $26 billion a year, which means SpaceX is earning more than $50 billion annually for each gigawatt of AI compute. Spending $40 billion to generate $50 billion in revenue isn't a bad business at all.
$1 Trillion
The hyperscalers, Meta Platforms, Microsoft, Amazon.com, and Alphabet, are leaning into that business, and plan to spend about $1 trillion on new plants and equipment in 2027, up from $700 billion in 2026. That's roughly half of the total capital spending for the entire S&P 500, whose companies spent less than $600 billion on new plants and equipment in 2021, the year before OpenAI's ChatGPT launched. Back then, the hyperscalers accounted for roughly 20% of total capital spending.
Hyperscaler capital spending didn't pick up materially until 2025, with the monster jump in 2026. The AI arms race isn't all that old, with prior data center spending focused on the cloud transition, not AI computing.
1,300 Terawatt-Hours
The four hyperscalers are expected to spend an additional $3.7 trillion through 2029. Adding in Oracle, SpaceX, and others means spending could easily reach $6 trillion, cumulatively, on AI by the end of the decade. That might leave the U.S. with 150 gigawatts of AI computing capacity. Keeping that on requires about 1,300 terawatt-hours of electricity. That's a roughly 30% increase over current U.S. electricity demand.
That could be a problem. Electricity demand just doesn't grow that fast. Not all of the electricity will fall onto the existing grid, though. Data center operators are starting to put in their own power sources. That's prudent, according to Melius Research analyst Ben Reitzes. He believes AI data center operators should ensure that residential power bills go down, not up, which will lessen the impact of growing not-in-my-back-yard sentiment.
22 billion
While the scale of spending is incredible, it isn't a Field of Dreams situation. Demand for AI services is exploding. Alphabet reported generating 22 billion tokens a minute in the second quarter, up from 16 billion in the first quarter.
Investors should know what tokens are. It's how AI is priced. A token is essentially a chunk of text, or four characters. User prompts are broken into input tokens. The answer generates output tokens. The industry standard is dollars per million tokens. ChatGPT 4 was originally priced at $30 per million input tokens and $40 per million output tokens.
Over time, model and hardware capabilities have gone up while costs have come down. Nvidia's new Vera Rubin processors promise 10 times the AI throughput of its Blackwell chips. And pricing for the best models might be $2 per million on the input side and $10 on the output side. New, open-source, and mini models can cost pennies per million.
Falling token costs enable Jevon's paradox, when falling costs for something mean people use more of it, resulting in an increase in total market size. What Alphabet's token numbers illustrate is that AI demand is growing. Its cloud services backlog was $514 billion at the end of the second quarter, up from just over $100 billion a year ago.
0.08
It's no wonder Alphabet is spending. AI spending has eaten into its and other tech firms' free cash flow, but they remain what Taiwan Semiconductor Manufacturing CFO Wendell Huang calls "wealthy and healthy."
The four hyperscalers generated roughly $200 billion in 2025 free cash flow. By 2027, that number is expected to be negative $60 billion. That's a concern, but 2027 earnings before interest, taxes, depreciation, and amortization, or Ebitda, are expected to total about $1 trillion. Those businesses can support $2 trillion in debt without breaking a sweat.
Combined, the four have essentially no debt. Net debt, which is debt less cash, totaled about $67 billion at the end of the second quarter, according to FactSet. That gives them a current debt-to-Ebitda ratio, a common measure of balance sheet strength, of 0.08. A ratio between one and two times isn't considered alarming for most businesses.
There are numbers inside the five numbers. They are only a guide to help investors understand the AI trade. For now, all the numbers add up.
Comments