The AI Capex Waterfall Doesn't Rain Dividends — It Transfers Them Upstream

The prevailing narrative around hyperscaler capital expenditure is straightforward: Big Tech is building AI infrastructure today that will compound into higher earnings, bigger buybacks, and growing dividends tomorrow. The spending is a bet on the future that returns to shareholders. That framing gets the direction of the value flow wrong.

The AI capex cycle is not a value-creating buildout for the companies writing the checks. It is a value transfer. The four largest hyperscalers — Microsoft, Alphabet, Amazon, and Meta — are collectively committing $725 billion in capital expenditure for 2026, a 77 percent increase from the prior year's $410 billion. Goldman Sachs now projects $5.3 trillion in combined capex for these four companies from fiscal 2025 through fiscal 2030. The money is flowing. The question is who captures the economic rent.

The answer is structural, not aspirational. The companies that sell the tools capture the margins. The companies that buy them capture the depreciation.

The Capital Expenditure and Free Cash Flow Divergence

Table 1 below summarizes the TTM (trailing twelve months) capital expenditure, free cash flow, and free cash flow growth for the four hyperscalers and Nvidia.

Company	Capex (TTM)	Operating Cash Flow (TTM)	Free Cash Flow (TTM)	FCF Growth YoY
Microsoft	$115.9B	$182.9B	$66.99B	-6.5%
Alphabet	$132.4B	$185.7B	$53.27B	-20.2%
Meta	$91.76B	$130.3B	$38.54B	-23.1%
Amazon	$173.0B	$161.4B	-$11.62B	-186.2%
Nvidia	$6.57B	$125.6B	$119.1B	+65.2%

The pattern is not a temporary bump. Amazon has turned negative on free cash flow entirely, with capital expenditure exceeding operating cash flow by $11.6 billion. Meta's free cash flow has contracted by more than a fifth year over year. Alphabet's declined by 20 percent. Microsoft, the most capital-efficient of the four, has still seen free cash flow growth turn negative.

Meanwhile, Nvidia — which spends less than $7 billion on capital expenditure and generates nearly $120 billion in free cash flow — is growing that free cash flow by 65 percent year over year.

That is the waterfall. Hyperscalers absorb capex at scale; chip designers convert the same spending into free cash flow. The dividend narrative assumes the hyperscalers will monetize their infrastructure fast enough to offset the drag. The data does not support that assumption yet.

Where Value Actually Accumulates

The AI value chain splits into two categories based on one structural variable: whether a company controls a supply-constrained chokepoint. The companies that do command gross margins of 53 to 75 percent. The companies that don't operate in single digits. That is the entire argument compressed.

Nvidia's gross margin stands at 74.15 percent with a 64 percent operating margin and 89.4 percent ROIC (return on invested capital). A Blackwell GPU costs approximately $6,400 to manufacture and sells for $30,000 to $40,000. The five- to six-fold markup is not pricing greed — it is the economic rent that accrues to a single dominant supplier when buyers are in a structural bidding war.

Hyperscaler cloud margins tell the contrasting story. AWS operating margin has declined from a peak of roughly 37.5 percent in early 2025 to 35 percent as AI infrastructure depreciation hits the income statement. Microsoft's operating margin stands at 46.8 percent but that includes the high-margin Office and licensing franchise propping up the average. Amazon's operating margin across all segments is 11.5 percent. Meta's is 38.1 percent, but its advertising business carries that figure while its AI infrastructure investment drags.

The hyperscalers are not unprofitable. They are capital-intensive intermediaries. They buy hardware at chokepoint pricing, depreciate it over three to five years, and hope the recurring revenue from cloud subscriptions and AI services outpaces the depreciation charge. That is a margin-compression business model, regardless of whether the underlying demand is real.

The Two-Market Split: Chokepoint Owners Versus Capital Deployers

The AI infrastructure market has bifurcated. One side controls scarce inputs. The other side finances the buildout.

The chokepoint owners include Nvidia at the design layer, TSMC at the foundry and advanced packaging layer, ASML at the lithography layer, and SK Hynix and Samsung at the high-bandwidth memory layer. These companies command gross margins between 53 and 75 percent because they have no viable alternatives at scale. CoWoS packaging capacity is sold out through 2026. EUV lithography has a sole global supplier. HBM contract prices are up approximately 820 percent year over year into 2026.

The capital deployers — Microsoft, Alphabet, Amazon, and Meta — operate in the other market. Their competitive moat is distribution and switching cost, not supply scarcity. They are locked into Nvidia because alternatives like Google's TPUs and custom ASICs are still nascent at scale. They are locked into TSMC because no one else can manufacture at the required nodes. They are absorbing the cost of that lock-in as margin compression.

The implication is fairly straightforward. The companies with the most attractive margins in this cycle are the ones selling to the hyperscalers, not the hyperscalers themselves.

The Monetization Gap

The critical assumption underlying the dividend narrative is that hyperscaler AI revenue will grow fast enough to justify the capex. That assumption is untested at current spending levels. Goldman Sachs projects the combined revenue of pure-play AI model providers — OpenAI at roughly $20 billion in annual recurring revenue and Anthropic at approximately $9 billion — totals less than $35 billion for 2026. That is 3 to 5 percent of the $725 billion in hyperscaler capex.

Hyperscalers build for their own services and enterprise customers as well, not just third-party model vendors. Google has reportedly cracked AI monetization, with its cloud backlog surging 55 percent sequentially to over $240 billion. Microsoft disclosed an $80 billion backlog of Azure orders unfulfilled due to power constraints. Those numbers show demand exists. They do not show that the revenue will cover the incremental depreciation of $725 billion in new assets.

Infrastructure built today may take 18 to 36 months to generate proportional returns. During that period, free cash flow continues to contract. Amazon is already in the hole, spending $173 billion on capital while generating only $161 billion in operating cash flow. If utilization rates fall short or if inference efficiency gains reduce compute demand per workload, the depreciation charge becomes a permanent margin headwind, not a temporary drag.

What Investors Are Actually Pricing

The valuation multiples in Table 2 below show how the market has allocated its confidence.

Company	P/E (TTM)	EV/EBITDA (TTM)	P/S (TTM)	Dividend Yield
Nvidia	34.0x	32.3x	21.4x	0.13%
Microsoft	27.8x	19.0x	11.2x	0.71%
Alphabet	17.7x	24.2x	9.7x	0.24%
Meta	22.2x	13.7x	6.6x	0.35%
Amazon	21.9x	17.6x	3.8x	—

Nvidia trades at the highest multiple in every category, reflecting its position as the primary beneficiary of the capex cycle. Microsoft and Amazon, the two hyperscalers building the most infrastructure, trade at moderate multiples with dividend yields below 1 percent — well below what income-focused investors typically require. Meta trades at the lowest EV/EBITDAmultiple, a reflection of the market's skepticism about whether its aggressive capex raise to $125-$145 billion will generate commensurate returns.

Alphabet at 17.7 times earnings looks cheap by comparison, but that multiple already prices in the assumption that Google Cloud monetization will accelerate enough to offset $132 billion in annual capex. It is a reasonable valuation only if that acceleration materializes.

Investor Takeaway

The AI capex cycle is real. The spending is real. The assumption that it will "rain future dividends" to hyperscaler shareholders is a misattribution of who captures value.

The key issue is not whether AI demand remains strong. Demand appears sufficient to sustain capex levels for the foreseeable period. The more important question is whether hyperscalers can raise the revenue yield on their AI infrastructure fast enough to offset the structural margin compression that comes from buying into chokepoint pricing. Nvidia's 89 percent ROIC and 74 percent gross margin set a very high bar. AWS operating margins declining from 37.5 percent to 35 percent set a trajectory that does not yet justify a dividend expansion narrative.

The forward condition to watch is simple: hyperscaler free cash flow growth must turn positive at current capex levels before dividend expansion becomes credible. Until then, the waterfall flows upstream. Investors who want exposure to the AI buildout with durable margins should look at the chokepoint owners. Investors who own the hyperscalers for dividends should understand that the dividends are being delayed, not accumulated.

The constraint migration in this cycle has not moved toward the hyperscalers. It has moved toward the companies that control supply. The capex waterfall is a toll road, and the toll collectors are not the ones writing the checks.