Lab report

AI Hyperscalers Face $725 Billion in Capex Scrutiny. We Tested Whether Quality Beats Momentum.

The biggest cloud companies are projected to spend $725 billion on AI infrastructure in 2026, and the market has started asking for proof it pays off. We tested whether screening for cash flow coverage protects a technology portfolio, or whether simply buying whatever is rising still wins.

Stax Research · 8 min · tested over June 2021 to May 2026

What we tested, and what happened

Each version below ran on $100,000 of simulated cash over June 2021 to May 2026. Equal weighting, monthly rebalance, quarterly reconstitution, $0.005 per share commission, 0.1% slippage, exit risk triggers applied.

AI Infrastructure Quality

Tech and communication stocks with high return on capital, solid operating margins, and self-sustaining capex coverage.

Total return
+28.3%
Worst drop
16.7%
Trades
75
Win rate
53.3%

AI Infrastructure Momentum Control

Large-cap tech and communication stocks with strong profit margins, ranked purely by recent price momentum.

Total return
+39.2%
Worst drop
29.9%
Trades
155
Win rate
50.3%

The Setup

By late May 2026 the spending plans of Microsoft, Alphabet, Meta and Amazon reached a tipping point. Wall Street estimates now put their combined capital spending at $725 billion for the year [1], driven by the race to build data centers, secure specialised hardware and run generative AI at scale [2].

Investors have started shifting their attention from how much is being deployed to what comes back [3]. Technology companies that spend inefficiently or earn weak returns on that spending have become more volatile, while the ones generating real cash have held up. That raises a practical question: can you build a better technology screen by filtering for companies that generate strong cash flow relative to what they spend?

To find out we built a quality strategy focused on capital efficiency and compared it against a control that simply buys the largest technology companies whose share prices are climbing.

How to Think About This

The intuitive approach is to look for companies with balance sheets strong enough to support enormous spending. That can be defined with three measures.

Return on invested capital, set to at least 15%, checks that the company turns the cash it raises and borrows into real profit rather than just activity.

Operating profit margin, set to at least 20%, filters out low-margin service businesses that look large but keep very little of what they earn.

Capital expenditure coverage, set to at least 1.5 times, checks that cash generated by the business comfortably exceeds what the business spends on equipment and buildings. A company clearing that bar is funding its own growth rather than borrowing to keep up.

The control strategy takes the opposite approach. It targets technology and communication companies worth $100 billion or more with operating margins of at least 25%, then sorts them purely by how much their share price rose over the previous six months. Comparing the two shows whether the quality constraints genuinely manage risk, or whether they simply cut returns by excluding the fastest-rising companies.

Strategy NameFilter CriteriaRanking Method
AI Infrastructure Quality (Thesis)ROIC ≥ 15%, Operating Margin ≥ 20%, CECR ≥ 1.5, Market Cap ≥ $10B, Tech & Comm Sectors30% Fundamental / 70% Momentum
AI Infrastructure Momentum (Control)Market Cap ≥ $100B, Operating Margin ≥ 25%, Tech & Comm Sectors100% Momentum Ranking

What the Data Revealed

Both strategies ran from June 1, 2021 to May 25, 2026, starting with $100,000. Both used equal weighting, monthly rebalancing, quarterly reconstitution and realistic trading friction of $0.005 per share commission and 0.1% slippage, alongside standard exit rules.

The Quality thesis selected 40 companies and made 75 trades. It returned 28.3% in total, 5.1% a year, and fell 16.7% at its worst point. It won on 53.3% of trades with a profit factor of 1.62.

The Momentum control screened 178 companies, made 155 trades and returned 39.2% in total, 6.9% a year, but fell 29.9% at its worst point. It won on 50.3% of trades with a profit factor of 1.57.

This is a clean example of the tradeoff between return and comfort. The control captured the full upside of the AI expansion and finished with more. It also asked you to sit through a 29.9% fall to get it. The quality thesis cut that worst drop by nearly half, to 16.7%, while winning more often and extracting more from its winners, using half as many trades. Screening for cash flow coverage and capital efficiency did provide a real cushion during corrections. It just cost about 11 percentage points of return over five years to have it.

StrategyTotal ReturnSharpeMax DDWin RateTrades
AI Infrastructure Quality (Thesis)28.3%0.1616.7%53.3%75
AI Infrastructure Momentum (Control)39.2%0.2529.9%50.3%155

What You Can Do With This

These results point to a few concrete adjustments.

If your priority is protecting what you have, keep the free cash flow yield and capital spending coverage filters. They exclude fast-moving growth companies that do not generate cash, which produces a smoother ride at the cost of some upside.

If the screen feels too restrictive, lower the capital expenditure coverage requirement from 1.5 times to 1.2 times. More companies qualify, which lets you hold some faster-growing names while keeping a basic spending discipline in place.

You can also combine the two ideas rather than choosing between them. Apply the return on capital and margin filters first to remove the fragile companies, then order the survivors by how their share prices have been behaving. That keeps the quality floor while letting the ranking follow what drove returns in this window.

Run The Test

Sources

  1. 1. Tech giants capex surge triggers investor caution Wall Street Journal, May 22, 2026
  2. 2. Hyperscaler capital expenditures projected to hit $725 billion in 2026 Bloomberg Technology, May 2026
  3. 3. The shift from infrastructure deployment to return on capital Morgan Stanley Equity Research, May 25, 2026
  4. 4. GenAI infrastructure spending and productivity benchmarks Gartner Research, 2026

This is an educational backtest, not investment advice. All results are simulated using historical data from June 2021 through May 2026. Past performance does not predict future results. Hypothetical backtests do not represent live trading or transaction fees.

Test your own version of this idea.

Describe it in a chat, change any rule you disagree with, and watch the same engine replay it over real history before a dollar is involved.