The $10 Trillion Gap in Hyperscaler AI Spending

The four largest US hyperscalers (Microsoft, Alphabet, Amazon, and Meta) are on track to spend about $725 billion in combined capital expenditure in 2026, up from roughly $410 billion in 2025, according to figures attributed to Goldman Sachs credit strategist Amanda Lynam and reported by Yahoo Finance in June 2026. Most of that money is going into AI data centers and the chips that fill them. That single number is the starting point for an argument that technology consultant Dr. Jeffrey Funk, author of “Unicorns, Hype and Bubbles,” has been pressing on LinkedIn: the spending only makes sense if AI produces revenue at a scale the world has never seen from any software category.

Key Takeaways:

• Big Four hyperscaler capex is around $725 billion in 2026 and analysts at Morgan Stanley project it could reach $1.2 trillion to $1.4 trillion by 2028.
• The AI model builders that are supposed to fill those data centers earn tens of billions in annual run-rate revenue today, not the trillions the buildout implies.
• Bridging that gap by the late 2020s would require revenue growth faster than any comparable technology has produced, and productivity gains larger than the 1870 to 1970 boom documented by economist Robert Gordon.
• The bull case argues the comparison is wrong: AI targets the multi-trillion-dollar labor market, not the $1.4 trillion software market, and Goldman Sachs estimates $8 trillion in present value from AI productivity gains.
• The debate is now about timing and triggers (data center cancellations, chip shortages, rising rates), not whether the math is tight.

How the Capex Forecasts Keep Rising

The defining feature of hyperscaler capex forecasts is that each new vintage revises the previous one upward. BCA Research (2026) made this point directly with a chart plotting successive forecast vintages: estimates that sat near $150 billion to $200 billion around 2020 now run to roughly $700 billion for 2026, and the forward projections keep climbing. The table below reconstructs that trajectory from the underlying company and analyst figures rather than reproducing the chart.

The individual guidance behind these totals is public. Amazon guided to roughly $200 billion in 2026 capex and later signaled about $220 billion, per CNBC. Alphabet raised its 2026 range to $195 billion to $205 billion, reported by MLQ.ai. Meta reported $72.2 billion of capex for 2025 and guided higher for 2026 in its Q4 2025 results. Note that “$700 billion” fits the 2026 Big Four figure well, but it does not describe 2025, when combined capex was closer to $410 billion. Funk’s framing of roughly $700 billion “now” is accurate for the current year and overstates the recent past.

The $10 Trillion Revenue Gap

Capex on this scale needs to earn a return. Sequoia Capital’s David Cahn framed an early version of this problem in 2024 as the “$600 billion question”: roughly twice Nvidia’s data center revenue for total cost of ownership, doubled again for a target margin, gave an annual AI revenue figure the industry needed to justify its spend. Scale the buildout up to today’s numbers and the required revenue rises with it. Some analysts now put the revenue that would justify a $1 trillion-plus annual buildout in the range of $10 trillion per year by the late 2020s. That is the figure Funk anchors on.

Set that against what the model builders actually earn. OpenAI’s annualized revenue run-rate topped $20 billion at the end of 2025 and crossed $40 billion by mid-2026, according to Bloomberg reporting. Anthropic began 2025 near a $1 billion run-rate and passed $5 billion within eight months, per the company’s own disclosures, and told investors its annualized run-rate reached about $65 billion by July 2026, reported by CNBC. DeepSeek’s annualized revenue was nearing $500 million in mid-2026, per PYMNTS citing The Information. These are large, fast-growing businesses. They are also one to two orders of magnitude below the revenue the capex implies.

The 1000x Jump From Tens of Billions to Trillions

The gap between roughly $40 billion to $65 billion of run-rate at the leading builders and a $10 trillion target is close to 1000-fold. Funk’s point about the pace is arithmetic: revenue that doubles every year still takes about ten years to grow 1000x, because 2 to the tenth power is 1024. No major software business has sustained annual doubling for a decade. OpenAI roughly tripled its run-rate over the past year and Anthropic grew faster than that, so the early slope is steep. Sustaining that slope long enough to close a 1000x gap, across the whole sector rather than one breakout company, is the part with no precedent.

The comparison to existing markets is where Gary Marcus sharpened the argument. Global healthcare spending was about $9.8 trillion in 2022, per the World Health Organization, and global food retail spending is on the order of $9.7 trillion, per Statista, with restaurant and foodservice spending adding roughly $3.2 trillion on top, per Euromonitor. The entire global software market, by contrast, is forecast at about $1.4 trillion for 2026, per Gartner. A $10 trillion AI revenue pool would be roughly seven times the size of all software sold today and would sit alongside food and healthcare as one of the largest categories of human spending.

The Productivity-Growth Assumption

Revenue at that scale has to come from somewhere. The companies buying AI must pay the model builders, which means AI has to deliver enough measurable productivity gain for buyers to hand over trillions. Northwestern economist Robert Gordon documented in “The Rise and Fall of American Growth” (Princeton University Press, 2016) that the century from 1870 to 1970 was a unique burst of productivity driven by electricity, the internal combustion engine, indoor plumbing, and modern chemistry, and that these gains “cannot be repeated.” Gordon’s data show post-1970 productivity growth from computing has been comparatively modest. For AI to justify the current buildout on Funk’s math, it would have to drive productivity growth beyond even that special century.

The early enterprise evidence is mixed. An MIT study from Project NANDA in 2025 found that about 95 percent of enterprise generative-AI pilots delivered no measurable profit-and-loss impact, as covered by Fortune. Goldman Sachs economists, reported in early 2026, estimated AI added close to zero to measured US GDP growth in 2025, per Tom’s Hardware. Those are early readings on a technology still being deployed, not verdicts, but they sit uncomfortably next to the productivity assumption the spending requires.

The Case For the Buildout

The strongest counter-argument attacks the software comparison itself. Andreessen Horowitz argues that AI’s addressable market is labor, not software: global software-as-a-service revenue is a few hundred billion dollars a year, while the US labor market alone runs to roughly $13 trillion, and the firm has pointed out that nurses’ wages alone exceed all SaaS revenue, in its analysis of AI and labor markets. On that view, measuring AI revenue against the $1.4 trillion software market understates the target, because AI can be paid for automating work rather than for selling seats of software.

Goldman Sachs Research put a number on the upside. Its economists estimated in October 2025 that AI could generate about $8 trillion in present-discounted value from productivity gains, within a range of roughly $5 trillion to $19 trillion, and argued that AI investment as a share of GDP remains modest compared with past technology cycles, reported via Yahoo Finance. Nvidia CEO Jensen Huang has made the durable-asset case, calling the current period the largest infrastructure buildout in history and arguing the chips and systems are revenue-generating assets rather than speculative overbuild, as covered by the IEEE ComSoc blog. Data centers, power connections, and networking gear built now have useful lives measured in years, so even a revenue shortfall does not erase the asset base the way a pure software bust would.

Goldman Sachs is worth watching because the same firm holds both views. Its Head of Global Equity Research, Jim Covello, published a widely read 2024 note titled “Gen AI: Too Much Spend, Too Little Benefit?” arguing that to earn a return on roughly $1 trillion of cost, AI would have to solve complex problems it was not built to solve. One research house producing an $8 trillion upside estimate and a “too much spend” warning captures how unsettled the question is.

What Could Set Off a Correction

The math being tight does not schedule a collapse. Menlo Ventures reported that spending on model APIs reached $8.4 billion in the first half of 2025, more than double the prior six months, in its mid-year market update, so demand is real and growing even if it is far below the eventual target. The debate among people who take the bubble framing seriously is about timing and triggers rather than direction. Three candidates recur: cancellations or local opposition that stall data center construction, chip supply shortages that break the buildout schedule, and rising interest rates that make financing a trillion-dollar annual capex line far more expensive.

Funk’s core arithmetic holds up against the sources. Hyperscaler capex is around $725 billion in 2026 and credibly heading toward $1.2 trillion or more later this decade. The revenue that would justify it, on the analyst framing he cites, is close to $10 trillion a year, while the model builders that would earn it are at tens of billions today. Closing that gap requires growth and productivity gains with no historical precedent. The bull case does not dispute the size of the gap so much as argue the target market is labor rather than software, which is a claim about a market that does not yet exist at scale. Both sides are betting on numbers that have not happened. That is what makes it a bubble debate rather than a settled forecast.