Last updated: August 6, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
Governments are making a dangerous bet on the AI boom: they're subsidizing the power and data-center buildout, becoming a direct customer of the same handful of AI labs, and building growth forecasts around a technology that the U.S. Treasury's own internal analysts now compare to the dot-com crash. The bet isn't hypothetical anymore — it shows up in a central bank warning, a leaked federal report, and utility bills in a couple dozen states.
Short answer: Governments are making a dangerous bet on the AI boom: Treasury's own draft report compares it to the dot-com crash, the BIS flags $1 trillion in hyperscaler spending as a recession risk, and states are already reversing data-center tax breaks as power bills rise. In my testing, treat any "government-backed" AI vendor as unproven, not automatically safe.
I test AI tools for a living, which means I read vendor pricing pages and government filings in the same week, and lately they've stopped matching up. A state legislature votes to end a tax break it handed a hyperscaler two years ago. A federal agency signs a nine-figure AI contract, then quietly excludes the vendor that pushed back on how its model could be used. None of this shows up in a product review, but it shapes whether the tool you're paying for still exists, at the price you're paying, in eighteen months.
What's actually happening
Three separate warnings landed within six weeks of each other in mid-2026, and none of them came from AI skeptics. The Bank for International Settlements — the institution central banks themselves answer to — used its June 29, 2026 annual report to flag roughly $1 trillion in AI-related capital spending by the five largest hyperscalers across 2025 and 2026, warning that "these episodes ended with an eventual reversal in investment, inducing economy-wide recessions" historically, and that a correction here could unwind "much faster than previous banking crisis episodes," according to Fortune’s coverage of the report.
Days later, a draft U.S. Treasury report — written for Secretary Scott Bessent and Federal Reserve Chair Kevin Warsh, and still awaiting formal sign-off — reached a similar conclusion from inside the government itself: AI firms are now so embedded across the economy that a downturn in the sector could ripple into stock markets, private credit, and utilities, in a pattern the analysts likened to the dot-com bubble, per NOTUS’s July 6, 2026 reporting. That's a striking gap from the administration's public line — Bessent has publicly praised tech firms for investing $750 billion into AI buildout in 2026 and called AI a "key driver of America's new Golden Age" — but his own department's analysts were modeling the downside at the same time.
The growth math backs up why governments are leaning in so hard despite the warning. Apollo Global Management's chief economist, Torsten Slok, estimated in the firm's own June 2026 mid-year outlook that AI-related spending on data centers, chips, and power infrastructure is contributing roughly one full percentage point of 2026 U.S. GDP growth — in an economy expected to grow only a little above 2%, according to Apollo’s published outlook. Strip AI spending out of that forecast, and the growth story shrinks by close to half. That's the bet: a huge share of measured economic health now rests on one sector continuing to spend at a pace even its own regulators call a recession risk.
What you'll need
Nothing technical — just fifteen minutes and a short list. Pull up your household or business's AI subscriptions (ChatGPT, Claude, Gemini, or any specialized tool billed monthly), your last electricity bill, and — if you're a business owner — any AI vendor whose product a government contract, grant, or tax break helped make cheap. You're not trying to predict when a correction happens. You're checking how much of what you rely on depends on subsidies and customer concentration that could change with one budget cycle or one lawsuit.
Step-by-step: check your own exposure to the government's AI bet
1. Check whether your state is one of the roughly two dozen pulling back incentives
AI-driven electricity demand pushed residential power bills up enough that a subsidy nobody used to notice became a political liability. Illinois, Arizona, and Ohio have already paused data-center tax incentives, New Jersey froze its program, and Virginia lawmakers are weighing an end to an exemption that costs the state about $1.6 billion a year — Georgia is projecting $2.5 billion in losses from its own program, and Texas about $1 billion, according to Stateline’s February 2026 analysis. If your utility rate has jumped and there's a data center nearby, that's not a coincidence — check your state legislature's site for pending bills before you assume the rate is permanent.
2. Don't treat "government contract" as a safety stamp
The Pentagon awarded identical $200 million contract ceilings to OpenAI, Anthropic, Google, and xAI in July 2025 for agentic AI development, treating all four as equally vetted, per CNBC’s reporting at the time. Less than a year later, the Defense Department formally designated Anthropic a "supply chain risk" and excluded it from new classified-network deals, after the company refused to let its models be used for autonomous weapons or mass domestic surveillance — a designation Anthropic is still fighting in court, per CNBC’s April 2026 coverage. A big government deal tells you a vendor won a procurement process, not that its business, its guardrails, or its long-term viability have been independently checked.
3. Ask how much of your vendor's revenue is one customer, one deal, or one investor away from changing
If a chunk of a vendor's growth story depends on a single hyperscaler partnership, a single circular financing deal, or a single government contract, that's concentration risk you're inheriting the moment you build your workflow around it. In my testing, I now ask this directly in a sales call or support chat, and a vendor that can't answer plainly is telling you something.
4. Separate "AI helped GDP" from "AI helped me"
A macro number like Apollo's one-point GDP contribution measures data-center construction and chip sales, not whether the specific tool on your desk is worth what you pay for it. Judge your own AI spend on your own results — hours saved, output you'd actually use — not on how large the industry's spending number sounds.
5. Keep a second vendor warm for anything business-critical
If an AI tool touches billing, customer support, or a workflow you can't run manually anymore, don't single-source it the way some of these deals single-source a government agency to one lab. Keep a second option evaluated, even if you're not paying for it, so a pricing change or an outage isn't a scramble.
6. Recheck this list quarterly, not once
Contracts get canceled, tax breaks get repealed, and companies get excluded from deals faster than most people update their vendor list. Set a quarterly reminder to redo steps one and two, the same way you'd recheck a mortgage rate or an insurance policy.
Example prompts you can copy
- Test vendor transparency: "If I'm paying for your product, what percentage of your company's revenue comes from a single customer, government contract, or partner? If you can't share an exact number, give me a rough range."
- Stress-test your dependence: "Here's how my team uses [tool] day to day: [describe it]. If this vendor raised prices 3x or shut down with 30 days' notice, what would break first?"
- Check the macro-vs-mine gap: "Explain the difference between 'AI is boosting GDP' and 'AI is boosting my specific business,' using my situation: [describe your use case]."
- Track the policy angle: "Search for any pending state legislation in [your state] about data-center tax incentives or utility rate changes tied to AI infrastructure."
Common mistakes to avoid
The mistake I see most is assuming a large government contract or a subsidized power rate is evidence a technology is safe, proven, or permanent — Anthropic's Pentagon exclusion above happened to a company that had an identical contract to its three biggest rivals a year earlier. Second is confusing macro AI spending with your own results; the two move independently, and a strong GDP contribution from data-center construction says nothing about whether your $30-a-month tool is worth renewing. Third is single-sourcing a business-critical workflow to one AI vendor the way some agencies have single-sourced entire missions to one lab, which leaves you with no fallback if a contract, lawsuit, or price hike hits. Fourth is treating this as a reason to avoid AI tools altogether, which throws out real, working tools because of a macro risk that mostly sits with hyperscalers and their balance sheets, not with your ChatGPT subscription.
Signs of a durable AI investment vs. bubble-era warning signs
| Signal | Durable investment | Bubble-era warning sign |
|---|---|---|
| Financing | Disclosed, traceable revenue and contracts | BIS flags deals with "poorly disclosed" terms and assets pledged more than once |
| Customer base | Diversified across many paying customers | Growth leans on one hyperscaler deal, one government contract, or circular investor financing |
| Government relationship | Contract won on merit, reviewed periodically | Contract awarded, then reversed within a year (Anthropic's Pentagon status) |
| Local cost | Subsidy tied to measurable public benefit | Tax break survives after residents' power bills visibly rise |
| Official tone | Public and internal assessments roughly agree | Public "Golden Age" messaging vs. an internal report modeling a dot-com-style downturn |
Tools that make this easier
None of this requires guessing. My AI mania and decision-making guide covers the verification habit — a second opinion, a small pilot — that catches a confident-sounding wrong answer before it becomes a decision. If you're weighing a vendor's pitch specifically, Silicon Valley sees AI as the solution – for everyone else walks through separating a company's growth story from what actually solves your task. For the utility-bill side of this piece, my map of where AI data centers are driving up power bills shows whether your state is one of the ones absorbing the cost. If you want the broader bubble picture, Apple will watch everything burn when the AI bubble bursts and why corporate America pulled back on AI spending both cover the corporate side of the same risk. And if the concern here is cost creep in your own AI stack, AI is getting way too expensive covers what's actually driving your bill up. Once you've narrowed down which tools are worth keeping, my AI tool ratings hub and free AI tools roundup are where to compare or pilot something new without adding another subscription you'd have to unwind later.
My take
I don't think the fix here is avoiding AI tools — I use several every day, and so does most of my audience. The fix is not confusing a government's bet with your own. A state subsidizing a data center, a federal agency signing a nine-figure contract, or a central bank flagging trillion-dollar exposure are all decisions made at a scale and with incentives that have nothing to do with whether the tool on your desk is worth its price. Run the checklist above once now, keep a second vendor warm for anything critical, and recheck it every quarter. That costs less than the subscription you're deciding whether to renew, and it's the same distance between "government-backed" and "government-vetted" that Anthropic just found out the hard way.
Frequently Asked Questions
Is the "AI bubble" risk from governments actually measurable, or just a narrative?
It's measurable. The BIS's June 2026 annual report flagged roughly $1 trillion in hyperscaler AI capital spending as a recession risk, and a U.S. Treasury draft report separately compared the AI boom's structure to the dot-com bubble — two independent institutions, inside and outside government, pointing at the same pattern.
Does this mean governments should stop investing in or contracting with AI companies?
No. The warnings are about concentration and disclosure, not about AI being worthless — Apollo estimates AI-related spending is adding roughly a percentage point to 2026 U.S. GDP growth. The risk is treating any single AI vendor's government relationship as proof of stability.
How long does it take to check my own exposure?
About fifteen minutes for a personal or small-business check: list your AI subscriptions, check your state's data-center tax-incentive news, and ask one vendor how concentrated their revenue is. A team decision takes longer only because more people need to see the answer.
What's the single fastest check if I only do one thing?
Ask your AI vendor directly what share of their revenue comes from one customer, contract, or investor. In my testing, a vendor that dodges that question is telling you as much as one that answers it.
Did Anthropic actually lose a Pentagon contract over this?
Yes. The Defense Department formally designated Anthropic a "supply chain risk" and excluded it from new classified-network deals after a dispute over how its models could be used, even though Anthropic had received an identical $200 million contract to OpenAI, Google, and xAI a year earlier. Anthropic is contesting the designation in court.