Over the past week we looked at how AI-related debt issuance is lifting bond yields, and at governments' growing reliance on AI-driven economic growth as a way to repay record debt. So this week we want to check back in on the AI growth scorecard. Are we getting lost in the hype, or is the world really changing under the AI revolution?
We believe life is changing. Google now processes more than 3.2 quadrillion tokens a month — roughly 330 times the level of two years ago. Anthropic and OpenAI have a combined annualised revenue run rate above US$105 billion, up from around US$29 billion at the end of 2025. The four largest hyperscalers are guiding to roughly US$740 billion of capex in 2026, close to double last year. GE Vernova has 116 GW of gas turbines under contract or reserved — around six years of its current output. The prompt for this edition is a presentation Blackstone President Jon Gray gave to investors on 20 September, built around the 1984 advertising line "Where's the beef?" We checked each of his AI claims against independent sources.
Tokens: The Leading Indicator
Tokens are the units of text, code and data that AI models read and write — the closest thing we have to a real-time meter of work being handed to machines. Google disclosed at its May developer conference that it processed 9.7 trillion tokens a month in May 2024, roughly 480 trillion a year later, and more than 3.2 quadrillion in May 2026. That is a sevenfold increase in the past year and roughly 330 times in two years. More than 375 Google Cloud customers each processed over a trillion tokens in the past twelve months.
The better test is whether people pay. Anthropic's annualised revenue run rate rose from about US$9 billion at the end of 2025 to more than US$65 billion at the end of July, with second-quarter revenue above US$11.5 billion against US$787 million a year earlier. OpenAI's run rate passed US$40 billion in August. Businesses do not spend at this pace on tools that fail to earn their keep.
Broad Adoption Has Barely Started
Set against that usage explosion, the most rigorous measure of business adoption looks almost sluggish. The US Census Bureau's fortnightly Business Trends and Outlook Survey, drawn from around 1.2 million firms, found that 19.8% of US businesses used AI in a business function in the fortnight to 3 May 2026. Between December and May the share hovered between 17% and 20%. Use is concentrated: 37% of firms with 250 or more employees and nearly 40% in the information sector, but around 14% in retail, with no significant change among firms employing fewer than twenty people.
Reconciling the two pictures is straightforward. Token growth of 330 times in two years alongside one-in-five firm adoption means usage is being driven by a minority of intensive users: technology firms, financial services, large enterprises and software developers. That is exactly what the early stage of a technology S-curve looks like. The earnings data tells the same story — so far this is largely a picks-and-shovels boom, with the suppliers of the build capturing the gains ahead of the users.
History suggests this is normal. Edison's first commercial power station opened in 1882, yet economic historians found that factory electrification took roughly four decades to show up in productivity statistics, because the gains only arrived once factories were redesigned around the electric motor rather than simply swapping it in for the steam engine. AI is likely to follow the same pattern. Buying the tool is easy. Redesigning the workflow is slow.
What the Productivity Data Actually Says
Blackstone argued that US productivity growth has run at 2.6% over the past two and a half years. The Bureau of Labor Statistics data supports the direction but not quite the magnitude. Nonfarm business productivity grew 1.5% a year across the 2007 to 2019 business cycle, 3.0% in 2024 and 2.1% in 2025, and was up 2.2% in the year to the second quarter of 2026. Across the current cycle from late 2019, it has averaged 2.1% — exactly equal to its long-run average since 1987. The fair reading is a recovery from a weak post-crisis decade rather than a clean break above trend. The electricity precedent says the statistics will lag the reality.
The Barriers Are Real
Social and political opposition has moved sharply. A Gallup survey in March found 71% of Americans oppose a data centre being built in their area — well above the 53% who oppose a local nuclear power plant. Data Center Watch estimates that nearly four dozen projects worth around US$68 billion were blocked or delayed in the second quarter alone. New York's legislature passed a one-year moratorium on permits for large data centres, and Governor Hochul imposed a temporary moratorium by executive order in July. More than 100 local moratoria have been adopted nationally.
The harder constraints are equipment and capital. GE Vernova's gas turbine backlog reached 116 GW against annual output of around 20 GW. Management expects to be sold out through 2030 by year end. Hyperscaler capex is now absorbing close to all of their operating cash flow — Alphabet reported its first negative free cash flow since listing. As we have argued since early in this series, corporate balance sheets cannot carry this alone.
So, What's Around the Corner?
The token data says the sudden phase has already begun for the heaviest users. The next three to five years are about diffusion to the other four in five firms — and that is when the productivity statistics should begin to break above trend. As with electricity, the official data will lag. The bond market will not wait for it, which is why the pace of adoption matters as much to sovereign debt as to technology investors.
Our read is that the regulatory question is shifting from whether AI infrastructure gets built to who pays for it. Resolved in favour of the household ratepayer, that means AI load must bring its own power and fund its own grid upgrades. That moves a large volume of generation, storage and transmission off utility balance sheets and into dedicated, contracted structures — which is the role Monard is built for.
This publication is provided for general informational purposes only and does not constitute financial, investment, legal or tax advice. Views expressed reflect the judgement of Monard's Investment Committee as at the date of publication and are subject to change without notice. Past performance is not indicative of future results.
Sources: Blackstone investor presentation by Jon Gray, 20 September 2026; Google I/O 2026 keynote; Bloomberg, CNBC, TechCrunch reporting on Anthropic and OpenAI revenue; US Census Bureau Business Trends and Outlook Survey May 2026; US Bureau of Labor Statistics Productivity and Costs Q2 2026; GE Vernova Form 8-K filings Q1 and Q2 2026; Edison Electric Institute; Gallup (March 2026); AP-NORC and EPIC (July 2026); Data Center Watch Q2 2026; FactSet Earnings Insight.