Across most of the large-cap tech complex, share buybacks have been pulled back over the last several quarters as AI capital expenditure has crowded out other uses of cash. Nvidia is the conspicuous exception: free cash flow is large enough to fund both. The result is an asymmetric capital-allocation landscape in which the supplier of AI compute can keep buying back stock while the buyers of that compute cannot. That asymmetry has consequences for relative multiples that are easy to miss inside the AI-narrative noise.
Key takeaways
- Hyperscaler buybacks are below recent multi-year trends as capex absorbs free cash flow.
- Nvidia's cash generation supports continued buyback intensity even alongside large reinvestment.
- Net of buybacks, hyperscaler share counts are stabilizing while Nvidia's continues to shrink.
- Multiple-expansion math now embeds this asymmetry — a slow-moving but persistent flywheel.
Why hyperscaler buybacks are compressing
Three forces are converging on the cash-flow statement at the major cloud platforms:
- AI capex. Capacity additions for accelerator-heavy compute fabric are running at multiples of historical capex levels.
- Property, plant, and energy infrastructure. Data-center build-outs require not only compute but also power, cooling, and grid-interconnect investment.
- Working-capital obligations for capacity reservations are higher than they were three years ago.
The mechanical effect on share count
For most of the past decade, hyperscaler share counts trended down as buybacks more than offset stock-based compensation. With buybacks compressed, that net-decrease pattern has slowed or reversed for some names. Equity-based compensation continues to flow at scale, which means dilution is no longer fully absorbed.
Why Nvidia can do both
Nvidia's cash conversion is structurally different. Three operational features stand out:
- Customer prepayments and deposit structures fund a portion of operating cash flow upfront.
- Capital intensity for the design-house model remains modest relative to revenue — most of the heavy capex sits at foundries and customers, not at Nvidia.
- Gross margin and operating margin levels at this scale produce cash flow that easily covers buyback ambition.
How the major tech complex looks on capital allocation
| Company | Buyback intensity | Capex direction | Net share count trend |
|---|---|---|---|
| Nvidia | Elevated | Moderate | Down |
| Microsoft | Moderate, easing | High | Roughly flat |
| Alphabet | Moderate, easing | High | Down slightly |
| Meta | Steady | High | Down |
| Amazon | Limited | High | Flat to up |
Capital allocation is the slow-moving variable that determines whether a multiple compounds or merely persists.
What this asymmetry means for relative valuation
- Nvidia's share-count shrinkage compounds its per-share metrics even at a constant business-level growth rate.
- Hyperscalers face a longer payoff cycle on the capex they are deploying, with no buyback flywheel running in parallel.
- Multiples on the buyers of AI infrastructure should embed the lower cash-return profile, while the supplier multiple embeds the opposite.
Frequently asked questions
Are hyperscalers cutting buybacks because they have to?
Yes. AI capex is competing for the same cash, and prudence dictates funding the productive investment first. Most management teams have been explicit about the tradeoff.
Will Nvidia eventually face the same constraint?
Only if its own capex model shifts materially. As long as foundries shoulder the wafer-cost intensity, the company's cash generation should continue to support buybacks at scale.
Is this a near-term equity tailwind?
It is a slow flywheel rather than a near-term tailwind. The effect compounds quarter by quarter, with the magnitude depending on share-price levels and the consistency of the buyback program.
The bottom line
Capital-allocation patterns inside Big Tech have split into two trajectories. Nvidia keeps the cash-return flywheel running while hyperscalers redirect cash into capex. Multiples are slowly adjusting to reflect that asymmetry, and it will continue to be one of the quieter but more meaningful relative-performance drivers in the AI complex.
