Is the AI Bubble About to Pop? Follow the Power—and the Promises.

AI can change the world and still become a terrible investment at the wrong price. The uncomfortable question is whether the industry’s spending promises are moving faster than its power connections—and its customers’ willingness to pay.

That is the strongest argument behind “The AI bubble is about to pop…” from The Infographics Show. We checked the infrastructure and accounting claims against company disclosures, ERCOT and the International Energy Agency. The documents support serious scrutiny. They do not supply a countdown to a crash.

Research note: we accessed the video’s description and chapters and used a reader-supplied Gemini summary to identify claims; we could not independently review the narration. The linked primary sources below support our findings.

474 gigawatts is a warning sign—with an important label

In its July 29, 2026 presentation, ERCOT reported approximately 474 gigawatts of large loads seeking interconnection, about 90% of them data centers, using June data. The operative word is seeking. This is a historical request pipeline, not 474 GW of operating facilities or a forecast that all projects will be built.

A queue can signal strong interest while overstating what ultimately connects. The useful questions are which projects have secured power, posted financial security and reached construction milestones. Treating every application as inevitable demand turns a real bottleneck into a misleading statistic.

The IEA’s updated outlook describes tightening energy and equipment bottlenecks while projecting global data-center electricity consumption to rise from 485 TWh in 2025 to around 950 TWh in 2030. That is a projection for all data centers, not AI alone. Delays can coexist with substantial long-term growth.

Microsoft’s $329 billion promise is not a $329 billion loss

Microsoft’s fiscal 2026 annual report, Note 13, discloses $329.1 billion of additional leases, primarily for data centers, that had not commenced as of June 30. Some depend on contractual conditions; commencement runs from fiscal 2027 through 2033, with terms from one to 20 years.

These commitments matter because future payments can become harder to support if capacity arrives late or customer spending disappoints. They are not cash already lost, nor proof that all the facilities are unusable. The same annual report says Azure revenue surpassed $100 billion and grew 41% in fiscal 2026. Any credible bubble argument must account for a business already generating substantial revenue.

Depreciation changes profits. It does not create cash

Meta’s 2025 Form 10-K says it extended the estimated useful lives of most servers and network assets to 5.5 years from January 2025. The disclosed effect: $2.92 billion less depreciation expense and $2.59 billion more net income for that year.

Depreciation spreads an asset’s cost over its expected working life. Extending that estimate generally reduces the annual expense, holding other factors constant. The cash paid for the equipment does not disappear. If equipment stops being economically useful earlier than assumed, the mismatch deserves attention. A newer chip arriving, however, does not automatically make every older chip worthless.

Calling a disclosed estimate change an “accounting trick” skips the central test: does the assumed useful life match how the equipment actually earns money? We could not substantiate the summary’s $167 billion hidden-cost estimate or $3 trillion aggregate from the primary documents reviewed, so we are not presenting them as established totals.

Outside financing does not make the risk vanish

Meta’s October 2025 Hyperion announcement described an approximately $27 billion development venture: Blue Owl-managed funds own 80%, and Meta owns 20%. Meta disclosed initial four-year operating leases with extension options and a conditional, capped residual-value guarantee covering the first 16 years of operations.

Its 2025 10-K, Note 5, says the venture is not consolidated and reports $45.95 billion of maximum exposure to loss as of year-end, including investments, commitments and guarantee exposure. Maximum exposure is not an expected loss. These disclosures show why reading the notes matters: economic obligations can extend beyond a simple debt headline.

Supplier-customer investment links deserve similar care. NVIDIA’s September 2025 announcement described a letter of intent to deploy at least 10 GW of systems and an intention to invest up to $100 billion in OpenAI as deployment progressed. That historical announcement does not establish that the full sum was paid or that every proposed term remains current. Interconnected funding is a risk to examine, not by itself proof of fabricated demand.

The $145 billion figure needs its range and date

Meta’s July 29, 2026 earnings release put full-year capital-expenditure guidance at $130–145 billion, including finance-lease principal payments. The upper end was not a single fixed “AI budget.” The release also reported revenue up 28% while costs and expenses grew 55%, with legal and severance charges among the contributors. The pressure on returns is worth watching without assigning every expense increase to AI.

What this means for retirement savers

The Labor Department’s March 30, 2026 proposal addressed how retirement-plan fiduciaries evaluate investments including alternative assets. Its stated factors include fees, liquidity, valuation, performance and complexity. A proposal and a policy direction do not establish that a particular 401(k) owns private data-center debt.

Your exposure depends on the funds and holdings in your own plan. We have not verified the summary’s $10.8 trillion retirement-exposure total or its fixed “$34 of every $100” allocation. Neither belongs in a blanket claim about every saver.

PCGeek’s take: watch the cash, not the countdown

The strongest warning is a timing mismatch: fast hardware cycles, slower power infrastructure and long financial commitments. The strongest counterargument is that demand and revenue can keep growing while infrastructure catches up. Both can be true, and returns can still vary sharply between projects and companies.

Watch energized capacity, utilization, customer collections, free cash flow after capital spending, and the assumptions behind depreciation and guarantees. Those signals are more useful than a confident crash date. Would power delays or weak customer economics worry you more? Follow PCGeek for the evidence behind the tech headline.

This is technology and business reporting, not a recommendation to buy or sell securities. Figures are dated snapshots, not interchangeable measures of losses, spending or demand.

Video context: The Infographics Show: “The AI bubble is about to pop…”. Related PCGeek coverage: the accounting behind Anthropic’s IPO loss headline.

Support PCGeek: browse our Amazon tech storefront (affiliate link). This is a shopping link, unrelated to investing in AI companies. As an Amazon Associate I earn from qualifying purchases.