Why the AI Bubble Could Burst: Power, Chips, and China

The AI Boom Has a Structural Problem Nobody Wants to Talk About

The AI bubble debate has moved well past theory. It now has numbers, timelines, and a widening gap between what the market is pricing in and what the infrastructure can actually support. In a recent interview with The Economist, Elon Musk laid out a three-part thesis that maps the fault lines of America's AI dominance with uncomfortable precision — and it lines up closely with warnings from Steve Eisman, the investor who famously shorted the subprime mortgage market before the 2008 collapse.

This isn't a story about whether AI is real or transformative. It clearly is. This is a story about whether the companies currently capturing most of the market's AI-driven returns can sustain the valuations baked into their share prices when the structural constraints start to bite. And right now, three of those constraints are becoming impossible to ignore: electricity, semiconductors, and model pricing.

---

The Power Gap Is Wider Than Most Investors Realise

Start with the one input nobody in a server farm can negotiate around: electricity.

Musk's framing is blunt — AI development is essentially a function of two inputs, chips and power. And on power, the United States is losing ground fast. China already generates more electricity than the US, Europe, and India combined. According to Bloomberg data, China added the equivalent of 40% of the entire US grid's capacity in a single year and has built more generation capacity in the last four years than the whole US grid has accumulated in total.

Musk's projection is that China reaches four times US electricity production — roughly proportional to its population. That is not a short-term anomaly. That is a structural advantage compounding over decades.

The US is not ignoring the problem. The Trump administration's 2025 ratepayer protection pledge pushed tech companies to self-power their data centers rather than leaning on public grid infrastructure. But the execution gap is brutal:

• ~2,000 GW of proposed US power generation is currently sitting in interconnection queues
• The median wait time to get new generation switched on is approximately 5 years
• In Texas alone, ERCOT is managing a 474 GW queue of large loads, with data centers accounting for 90% of it
• Some projects are staring down 12-year delays

The gap between ambition and delivery is why hyperscalers and AI labs are turning to gas turbines — not as backup power, but as primary energy sources. XAI's Colossus Supercomputer in Memphis runs on approximately 35 gas turbines. A typical hospital uses one or two as backup. That gap tells you everything about how overstretched the US grid is relative to AI's energy appetite.

The downstream effect of this turbine rush is a broken supply chain. Wait times for large gas turbines have stretched to five years. GE Vernova's backlog sits at 116 GW; Siemens Energy's at 69 GW. Turbine prices for plants coming online in 2030–2031 are up approximately 75%. The companies manufacturing this infrastructure — GE Vernova, Siemens Energy, Mitsubishi Heavy Industries — are running hot on analyst expectations, with GE Vernova up roughly 68% and Siemens Energy up 65% over the past year.

For investors, this is a classic picks-and-shovels dynamic: the energy infrastructure enabling AI may prove more durable than the AI companies themselves.

---

China's Chip Shortage Is Temporary. America's Power Shortage Is Not.

The second structural pressure point is semiconductors — specifically, the US export restriction strategy and whether it is actually working.

The logic of the restrictions is sound: Nvidia makes the world's best AI chips, the US controls Nvidia's export policy, so restricting China's access to cutting-edge silicon slows its AI development. The Biden-era rules blocked advanced chip exports; the Trump administration loosened things marginally, approving limited H20 sales in late 2024 with a 25% tariff attached. Nvidia's Blackwell architecture and the upcoming Rubin chips remain completely off-limits to China.

But here is the problem: the restrictions have worked too well in one narrow sense. They have given China every incentive to build a fully independent semiconductor ecosystem — and it is making progress.

SMIC, China's answer to TSMC, has demonstrated 7-nanometre production using older deep ultraviolet (DUV) lithography machines — technology that predates the advanced EUV systems ASML supplies to Western fabs. Five-nanometre production is now reportedly within reach. China's homegrown DUV machines still lag ASML's gold standard on performance and reliability, but the trajectory is clear: the gap is narrowing, not widening.

China's chip shortage is a years-scale problem. America's power shortage is a decades-scale problem. That asymmetry is what makes Musk's analysis worth taking seriously — because the US is on the wrong side of the longer timeline.

---

The Price War That Could Crack the AI Narrative

The third pressure point is the one most directly connected to stock market valuations: model pricing.

Even with significant chip constraints, Chinese AI labs are producing models that are genuinely competitive at a fraction of the cost. Musk himself — who has every commercial incentive to talk up US AI — acknowledged that China's Kimi model is "getting quite close" to Anthropic's Claude in capability terms.

The cost data is striking. Research from AI analytics firm Artificial Analysis shows the following per-query cost comparisons:

Model	Cost Per Query
DeepSeek V4 Flash	$0.03
Kimi K3 (Moonshot AI)	$0.86
OpenAI ChatGPT 5.6 Soul	$186.00
Claude Sonnet 5 (Anthropic)	$3.15

These are not marginal differences. DeepSeek's pricing is orders of magnitude cheaper than the leading US models. And while top-tier capability still commands a premium, the distance between "good enough" and "best in class" is compressing fast.

This is where Steve Eisman's concern becomes very concrete. Eisman — who shorted subprime mortgage securities before the 2008 crisis — has been warning that a price war in AI model services will squeeze OpenAI and Anthropic's margins severely. His estimate is that those two companies together account for roughly 70% of all AI-related revenue flowing into the major hyperscalers: Microsoft, Amazon, and Google.

Follow the chain:

1. Cheaper Chinese models pull cost-sensitive customers away from OpenAI and Anthropic
2. OpenAI and Anthropic face margin compression and slow revenue growth
3. Both companies have committed to hundreds of billions in future hyperscaler spending — commitments already baked into Microsoft, Amazon, and Google share prices
4. If those commitments get revised downward, the hyperscalers' AI growth projections deflate
5. The AI premium in equity valuations starts to unwind

This is not a theoretical cascade. It is a plausible sequence of events with documented precedent — whenever a disruptive price entrant commoditises a market that the market is valuing as a premium product.

---

What the Current AI Standoff Actually Looks Like

Step back from the individual data points and the structural picture becomes clear.

America's current position: Controls the best chips, leads on frontier model capability, but is constrained by power infrastructure that takes years to build and a grid already under severe strain.

China's current position: Chip-constrained by US export restrictions, but generating far more electricity, building more grid capacity every year, and producing AI models that are competitive and dramatically cheaper.

The critical asymmetry: China's chip problem is solvable on a 3–7 year timeline through domestic semiconductor development, manufacturing scale, and — potentially — continued negotiated access to some US components. America's power problem is solvable on a 15–30 year timeline, because physical infrastructure — transmission lines, power plants, grid interconnections — cannot be accelerated by software or capital alone.

The entity solving the shorter-duration problem wins the long game, assuming model quality continues to converge. And on current trends, that is China.

---

What Investors Should Be Watching

None of this means the AI trade is over. Infrastructure buildout continues. Demand for compute is real. But the risk-reward calculus for the highest-valued AI beneficiaries looks increasingly asymmetric. Here are the indicators worth tracking:

• Model pricing trends: If DeepSeek, Kimi, and other Chinese models continue to undercut US pricing while narrowing the quality gap, margin compression at OpenAI and Anthropic becomes a near-term earnings risk for Microsoft and Amazon
• Hyperscaler capex guidance: Any revision to committed AI infrastructure spending from major cloud providers would be an early signal that the revenue assumptions underpinning those investments are softening
• China's semiconductor progress: SMIC's yield rates and nanometre milestones are more important geopolitically than most equity analysts currently price in
• US grid interconnection clearances: Watch for any policy acceleration — or further delays — in getting new generation capacity online. The 5-year median wait time is the single biggest structural drag on US AI scaling
• Energy infrastructure earnings: GE Vernova, Siemens Energy, and Mitsubishi Heavy are executing in a market with long order backlogs and rising prices — a more durable position than many pure-play AI software names

---

The Bottom Line: Three Risks, One System

The AI bubble, if it bursts, will not burst because AI stopped being useful. It will burst because the market priced in a version of the future — US dominance, unchallenged hyperscaler revenue growth, and unlimited compute scaling — that the physical world cannot deliver on schedule.

Power constraints in the US are structural and slow to fix. China's chip constraints are temporary and self-correcting under competitive pressure. And the pricing dynamics in AI model services are moving in exactly the direction that compresses the margins of the companies whose revenue growth is baked into the largest equity valuations in the world.

That is not a reason to panic. It is a reason to be precise about what you own, why you own it, and what assumptions need to hold for those positions to perform.

---

This article is for informational purposes only and does not constitute financial advice. Always consult a qualified financial professional before making investment decisions.

---

Frequently Asked Questions

Why does electricity matter so much for AI development?

Modern AI training and inference — the two core processes that make AI models work — are extraordinarily energy-intensive. Large data centres running thousands of high-performance GPUs can consume as much electricity as a small city. The more compute you run, the more power you need. This makes electricity generation capacity a direct ceiling on how fast any country can scale its AI infrastructure. Elon Musk has described electricity and chips as the only two real constraints on AI progress — and right now, US electricity generation cannot keep pace with demand.

How significant is China's lead in electricity production?

Very significant. China currently generates more electricity than the United States, Europe, and India combined. In one recent year alone, China added generation capacity equivalent to 40% of the entire US grid. Over the last four years, it has built more new generation capacity than the total accumulated capacity of the US grid. Projections suggest China could reach four times US electricity production — a gap that would make the energy advantage in AI development structurally decisive over the long term.

Do US chip export restrictions actually slow China's AI progress?

In the short term, yes. China cannot access Nvidia's most advanced chips — the Blackwell architecture and upcoming Rubin chips remain completely off-limits — which limits the raw compute available to Chinese AI labs. But the restrictions have also accelerated China's domestic semiconductor ambitions. SMIC has demonstrated 7-nanometre chip production using older DUV lithography equipment, and 5-nanometre production is reportedly in progress. The restrictions create a temporary bottleneck but a permanent incentive to build independence. Most analysts expect China's chip gap to narrow significantly within a decade.

How would a Chinese AI price war affect US tech stocks?

The mechanism is indirect but powerful. Chinese AI models like DeepSeek are already priced at a fraction of the cost of US equivalents — in some comparisons, 98% cheaper per query. If Chinese models reach "good enough" quality at these prices, enterprise customers will face strong incentives to switch. That puts margin pressure on OpenAI and Anthropic, which investor Steve Eisman estimates account for roughly 70% of AI-related revenue flowing to major hyperscalers like Microsoft, Amazon, and Google. If OpenAI and Anthropic slow their spending commitments — commitments already factored into hyperscaler valuations — the AI growth premium embedded in those share prices could deflate meaningfully.