AI Spending Isn't Slowing Down Despite Safety Warnings — Who Is Financing the Next Compute Boom?

Photo: Aaron Hall, Wikimedia Commons, CC BY-SA 2.0

01 The warnings and the money

The juxtaposition is now a quarterly ritual. Executives at frontier labs publish essays warning that AI systems are advancing toward capabilities they do not fully understand; researchers file statements about catastrophic and existential risk; regulators in Washington, Brussels and London hold hearings on frontier-model safety. Then the earnings season arrives and the same ecosystem reports numbers that treat the warnings as marketing collateral: Microsoft, Alphabet, Amazon and Meta together are on track for well over $300 billion in capital expenditure this year, the majority of it AI compute — data centers, GPUs, power infrastructure and networking.

That figure has roughly doubled in two years, and the guidance keeps rising. NVIDIA's data-center revenue — the cleanest single proxy for global AI-infrastructure demand — climbed from about $15 billion in fiscal 2023 to roughly $115 billion in fiscal 2025, and Blackwell-class demand kept the curve bending upward through 2026. Whatever the safety debate is doing to discourse, it is not slowing down procurement.

Analysis — not prediction. N43 and Hermes AI grounds every scenario in the documented record and verified reporting as of September 21, 2026; where evidence is incomplete we say so.

02 Who is actually financing it

The first wave was self-funded: hyperscaler operating cash flow — Microsoft's and Google's in particular — paid for the initial AI buildout out of pocket, the way telecom carriers once funded their own networks. That era is ending, because the checks have outgrown even the biggest balance sheets. The second wave has three distinct pools. Corporate debt and hybrids: investment-grade bond issuance, equipment financing and creative off-balance-sheet structures, including the debt-funded JV arrangements the biggest AI-campuses are increasingly built on. Private credit: direct lenders and private-equity credit arms now provide multi-billion-dollar packages for data-center construction and GPU leasing, a shift covered this year by the FT and Bloomberg — money that never touches the public bond market and is priced opaquely. Sovereign capital: state-backed funds and national AI programs across the Gulf, Singapore, the EU and Canada now co-invest in domestic capacity to obtain compute access outright.

The anchor video's bubble framing — is Wall Street creating the next one? — is the right question in the wrong tense. The structural feature worth watching is not equity euphoria but the migration of AI finance from equity risk into fixed-income obligation. Equity investors can mark down a bad bet and walk away. Debt on a twenty-year data-center must be serviced for twenty years, whether or not the models trained inside it ever earn a return.

03 The bet lenders are actually underwriting

Strip the AI narrative and the finance is conventional project economics with one twist. Data centers are attractive to lenders because they look like infrastructure: long-lived assets, contracted revenue from credit-worthy tenants, power agreements in place. The illustrative math on a $10 billion project works easily — $3 billion of contracted annual cloud revenue against perhaps $1.5 billion of annual debt service — if the contracts hold and the chips inside stay useful.

The twist is depreciation. A gas pipeline depreciates over decades; a GPU fleet's economic life is three to five years before its inference economics fall behind newer silicon. Lenders are underwriting twenty-year assets whose revenue-generating core turns over like consumer electronics. That mismatch is survivable only if AI demand compounds fast enough to keep yesterday's chips earning — the same assumption, restated. The finance works until it does not, and when it does not, the loss lands on debt holders, not equity.

04 What returns would have to look like

The bull case rests on a claim that was true of neither railroads, telecom, nor the early internet — or rather, was eventually true, but only after the crash: that revenue catches up to capacity without a bust in between. Enterprise AI spend is growing fast but from a small base; the entire global AI software and services market remains well below the annual run-rate of AI capital spending. Closing the gap requires either much higher AI revenue growth than the current trajectory, or a much longer wait than current financing terms allow.

There are two honest possibilities. In one, the 2026 buildout is the internet-backbone story: overbuild, bust, consolidation, and then the cheap capacity enables the next wave. That path was spectacularly right for users and catastrophic for the original bondholders. In the other, AI is infrastructure like electricity — permanently capitalized at scale, unglamorous, and the safety debates are simply the industry's version of utility regulation arriving on schedule.

05 The risks hiding inside the boom

Three specific fragilities deserve names. Correlated demand: the same few model architectures and the same few killer applications underwrite every lender's revenue model; if agentic AI's enterprise economics disappoint, all projections fail together. Power constraints: the binding limit on 2026 expansion is increasingly electricity — interconnection queues and generation, not chips — which pushes developers toward longer-dated, riskier sites. Refinancing wall: the private-credit structures put together in 2024-26 come due for repricing in 2028-29; if rates have stayed higher or AI revenue has not shown up, the reckoning arrives as a refinancing event, not a crash.

The safety warnings are not irrelevant to this finance — they are a cost line. Compliance regimes like the EU AI Act, model-evaluation requirements and liability exposure all add expense and delay to deployment. So far the market has priced those costs as trivial relative to the revenue opportunity, which is itself a statement about what the market believes: the money says AI capabilities will keep outrunning the guardrails.

06 What to watch next

Watch the hyperscalers' next capex guidance: the first cohort to guide flat rather than up will matter more than any warning letter. Watch private-credit spreads on data-center debt — the earliest tell of lender nerves, visible before any equity reacts. Watch power interconnection queues as the true throttle on buildout. Watch NVIDIA's data-center backlog and any inventory build for the demand signal underneath the story. And watch what happens at the first 2028-29 refinancing window: the compute boom's real stress test is not this year's earnings, but whether the debt written now gets repaid by the revenue that arrives later.