Premium: What If…We’re In An AI Bubble? (Part 2)

Last week I laid out a handful of “what‑if” scenarios that illustrate how tightly coupled the AI ecosystem is to a handful of megaprojects—massive data‑center builds, GPU supply chains, and venture‑capital financing. The response was overwhelming, making that piece the most‑read article I’ve ever published. If you missed it, here’s a quick reminder of the questions we already explored:

• What if the AI industry moves to entirely token‑based billing?
• What if organizations can’t afford to keep spending on AI?
• What if the AI capacity crunch never ends and data centers aren’t being built?
• What if CoreWeave can’t keep up with its capacity demands?
• What if hyperscalers can’t build data centers fast enough?
• What if hyperscalers have warehouses of uninstalled GPUs?
• What if hyperscalers write off a large chunk of GPUs?
• What if data‑center construction demand collapses?

The missing gestalt

Most commentary treats each of those facts in isolation. A single data‑center delay is a hiccup; a high‑priced GPU is a market‑signal; a venture‑capital round is a vote of confidence. What I’m trying to surface is the cascade that follows when the underlying assumptions break down.

A chain reaction of defaults

If construction of new AI‑focused data centers slows to a crawl—as early indicators from permitting pipelines and supply‑chain bottlenecks suggest—several things happen simultaneously:

1. Capacity ceiling for OpenAI and Anthropic – Both firms already account for roughly half of the compute backlog for Amazon, Google, and Microsoft. Without new megawatts, their ability to expand is capped.
2. Revenue shortfall for hyperscalers – The promised AI‑related revenue streams that underpin shareholder guidance evaporate, putting pressure on earnings guidance and stock valuations.
3. Unpaid data‑center debt – The United States faces about $178.5 billion in data‑center‑related project financing slated for 2025. If the facilities never become operational, the debt becomes non‑performing, threatening banks that financed the builds.
4. GPU inventory glut – NVIDIA shipped over 3 million Blackwell GPUs in 2025. With no new racks to host them, a sizable portion will sit in warehouses, forcing write‑offs or steep discounting. The upcoming Vera Rubin GPUs risk becoming stranded assets as well.
5. Financing freeze – Credit institutions, already nervous about “choking” on data‑center debt, will tighten lending, making it even harder for any player to raise the $44 million per megawatt capital required for a new AI‑grade facility.

In an optimistic timeline where a 2024‑started build finishes only in 2027‑2028, the “latest” GPUs will be two to three years old by the time they’re installed. Projects will either launch with outdated hardware or be forced to overpay for newer silicon that no one can actually use.

The demand illusion

Proponents of the AI boom point to “insatiable” compute demand. That claim holds up when you look at days‑to‑months of order backlogs, but it unravels over a year‑plus horizon. The bulk of current capacity is already locked into OpenAI and Anthropic. When you strip those two out, the remaining performance obligations from Google, Amazon, and Microsoft plateau.

If genuine, distributed demand existed, we would see exploding RPOs (remaining performance obligations) across the board. Instead, the market is dominated by a few large contracts, and the rest of the ecosystem—mid‑size AI startups, niche model providers, and even internal R&D teams—shows little appetite for additional megawatts.

Revenue vs. utilization

Some argue that the revenue flowing to NVIDIA, AMD, Samsung, and SanDisk proves a healthy downstream market. While it’s true that these companies have seen billions of dollars in sales, the money is largely forward‑looking—pre‑orders, inventory purchases, and speculative stock‑buy‑backs—not a reflection of sustained utilization.

The consumer side adds another layer of confusion. Higher prices for consoles or smartphones are often blamed on AI‑driven component shortages, but those price hikes are also driven by inflation, supply‑chain disruptions, and macro‑economic factors unrelated to AI compute.

Financing under a frothy debt market

The AI data‑center boom is being funded by the same private‑credit firms that chased SaaS growth from 2018‑2022, assuming perpetual expansion. When due‑diligence is thin—evidenced by Apollo’s John Zito calling many valuations “all wrong”—the risk of systemic under‑funding becomes acute.

Venture‑capital returns have slumped to a TVPI of 0.8‑1.2× since 2018, meaning investors are barely breaking even. Yet the same investors continue to pour money into AI startups, betting on a future where every model becomes profitable. The reality is that most AI‑focused companies are still cash‑negative, relying on continual VC inflows to stay afloat.

What if the bubble bursts?

Imagine a scenario where:

• VC funding dries up because limited partners see diminishing returns.
• AI startups go to zero as they cannot achieve profitability without cheap compute.
• OpenAI and Anthropic become the de‑facto lenders of last resort, bailing out smaller labs in exchange for equity or exclusive API access.
• Inference services remain unprofitable, forcing hyperscalers to subsidize compute at a loss.

The result would be a feedback loop: less compute demand → lower GPU sales → tighter financing → fewer data‑center builds → even less demand.

Counter‑perspectives

1. Tech optimism – Some analysts argue that new architectural breakthroughs (e.g., optical interconnects, specialized ASICs) could dramatically lower the cost per megawatt, reviving the build‑out.
2. Government stimulus – Strategic investments in AI infrastructure from national governments could fill the financing gap, especially in regions looking to become AI hubs.
3. Alternative revenue models – Companies might shift from token‑based billing to subscription‑plus‑usage hybrids, smoothing cash flow and reducing the need for massive upfront spend.

These arguments have merit, but they often underestimate the timing lag between R&D breakthroughs and large‑scale deployment. Until the hardware pipeline catches up, the financial pressure remains.

Bottom line

The AI bubble narrative frequently ignores the interdependence of three pillars:

1. Physical capacity – data‑center construction and GPU availability.
2. Financial scaffolding – debt markets, private credit, and venture capital.
3. Demand concentration – the fact that a handful of labs consume the majority of compute.

If any one of these pillars wobbles, the entire structure can wobble with it. The next few years will reveal whether the market can re‑align these forces or whether we’ll witness a hard landing that forces a recalibration of AI’s growth expectations.

---

If you’d like to read the full premium story, including the speculative scenarios about venture‑capital collapse and AI’s lender‑of‑last‑resort dynamics, subscribe to Where’s Your Ed At.