'Caution Is Warranted': Ed Dowd Warns Wall Street's AI CapEx Party Is Ending

Authored by Ed Dowd via 'Beyond The Narrative' substack,

The signs are piling up faster than the hype can spin them. AI capex has been the rocket fuel for markets, but the second derivative is turning. Factors ending the party:

• Private credit stalled — flows reversing, redemptions surging, industry effectively paused.

• Enterprise demand cracking — ROI skepticism, token costs biting, data/alpha extraction backlash.

• Power constraints hitting hard — the grid can’t scale without massive, long lead time builds or dystopian reallocation.

• Open-source pressure — Chinese based DeepSeek and now the new open-source frontier model Kimi K3 are rivaling OpenAI and Anthropic frontiers labs at fraction of the price, commoditizing the economics.

Credit markets always end the party. We’re watching it live.

Private Credit: The Silent Pause Button on AI Capex

This is where financing reality bites. Morgan Stanley estimated private credit could fund up to 50% of the external financing needs for the massive AI data center buildout. That channel is now under serious stress.

Flows in private credit are going the wrong way. The industry is effectively paused. Redemption requests are surging, funds are gating, and high-profile bankruptcies plus underwriting scrutiny are flashing warnings. Private credit has become the new junk bond market…except it lacks transparency, liquidity, and is now being stress-tested in real time.

With outflows accelerating, near-term funding from Private credit for AI data center buildouts looks less likely. NVIDIA and others keep popping up in private credit loan books. What happens to repossessed GPUs in a stressed environment?

Continued capex relies on credit markets keeping the spigot open. When that spigot slows or gets expensive, the capex math breaks. Financing could turn prohibitively costly, pausing or dramatically slowing the cycle that has supported the S&P 500 index with roughly 45% of the market cap being AI or AI-adjacent.

Additionally AI infrastructure inflation itself is credit driven. This reflexive credit driven surge in demand has caused cost inflation for chips and data center construction making the past cost projections moot. These inflated costs make ROI hurdles even harder. Current creditors are reassessing their exposure as the Goldman Sachs credit desk has recently highlighted.

Enterprise Demand Is Cracking

The people who are supposed to use this stuff are slowing down. Companies that rushed AI tools into workers’ hands are now reining them in because costs at scale are biting. That’s not theory…it’s the second derivative showing up in real budgets. They are stepping back and trying to assess the ROI from this investment as their AI budgets come in above initial cost projections.

Alex Karp of Palantir laid it out bluntly. Enterprises are livid. They’re paying for tokens that create no reliable value. They’re watching their own workflows, customer data, and competitive alpha get extracted and potentially sold back to competitors. Token pricing itself is the confession…if these models delivered durable, defensible productivity gains at scale, the labs would price on value or take equity cuts, not meter compute.

Open-source pressure

DeepSeek and other cheaper open-source models were already pressuring token pricing for those users who didn’t need the frontier models. To make matters worse, recent newcomer Moonshot AI’s Kimi K3 is a Chinese open-weight model rivaling top US frontier leaders OpenAI and Anthropic at a fraction of the cost. It’s putting direct pressure on closed-source pricing and exposing how over-hyped the token economics have become. Why pay premium rates when open-source alternatives deliver competitive performance? Competitive pricing hurts revenue growth for the two US frontier leaders which is likely to push their IPOs into next year (if at all) and raises the cost of their debt capital. The downstream effect is an eventual capex slowdown for the pick-and-shovel crowd (semiconductors etc.). Simple as that.

Power Constraints: The Physical Wall Nobody Wants to Talk About

Even if the money were flowing freely, the electricity isn’t. AI compute requires tremendous amounts of energy and water. Currently installed electrical capacity can’t handle the aggressive projections without massive new builds, which take years of capital, permitting, and construction or drastic reallocation of existing power away from other uses.

One path is slow and practical: add real capacity. The other gets dystopian fast: reduce human usage or rolling blackouts to free up juice. The reality will be somewhere in between and will play out politically. Local opposition to data centers is already rising over electricity rates, water use, and land. Towns are pushing back. This isn’t abstract futurism…it’s a hard physical constraint on timelines and costs.

Power adds another multiplier to the capex problem. Data centers need gigawatts. Hyperscalers are already forecasting enormous spending just to keep up. When private credit tightens and power infrastructure lags, the combined effect is a slower, more expensive buildout than the bull-case spreadsheets assume. The second derivative doesn’t just slow…it can stall.

Market Concentration and Cyclical Reality

Semiconductor stock valuations have hit record levels recently around 19-20% of the entire S&P 500. That’s not healthy diversification; it’s concentration in a notoriously cyclical industry whose recent boom has been funded by debt to business models that remain unproven at scale.

The biggest AI capex spenders have seen their stocks pull back meaningfully from highs even as broader indices hover near records. ROI concerns are finally showing up in price action. Even the Bank of International Settlements (BIS) has been more sober: ‘AI has boosted confidence via productivity expectations, but it’s also raising job fears, supply bottlenecks, and the risk of overinvestment boom-bust cycles we’ve seen before.’ Power and credit constraints make that overinvestment risk even more acute.

Bottom Line

We’ve seen this movie before. Credit questions profitability first. Physical limits and cheaper open-source alternatives (hello Kimi) force the timeline and pricing reckoning. Circular deals, negative free cash flow, sky-high chip prices, and now power realities all point in the same direction.

The party isn’t over tomorrow but closing time signals are everywhere: semis at peak gross margins, enterprises pausing, private credit tightening, power wall rising, open-source commoditization accelerating. Stock market AI concentration at extremes.

Skepticism isn’t denial of eventual AI value. It’s calling the current valuation and frenzy for what it is…priced for perfection that customers, credit markets, the electrical grid, and open-source competition aren’t delivering.

Watch the flows. Watch power builds. Watch Kimi-style pricing pressure. Watch the second derivative across money, megawatts, and model costs. The unwind in these concentrated, debt-fueled, physically constrained narratives tends to be swift once the marginal equity buyer and/or lender steps away.

Caution is warranted…the bright bar lights are about to be turned on. The distance to a real repricing is shrinking fast. Protect capital. The math doesn’t lie.