The US$720 billion question: the share of big-four capex turning into AI silicon is rising, not falling as this piece first assumed

Key takeaways

This year's capital-spending guidance from Amazon, Microsoft, Alphabet and Meta adds up to between US$720 billion and US$745 billion. Almost every use of that number stops at the amount: how much more than last year, how much of cash flow, who raised guidance again. This piece started by swapping the question. For chipmakers and their supply chains, the number that drives decisions was never the amount but a fraction: how much of that money actually becomes revenue for AI silicon. The swap is right. But once the fraction was actually computed, the starting premise turned out to point the wrong way. That premise was that the conversion rate is falling, because capex now carries satellites and robots, and because in-house chips are diverting purchases away from NVIDIA.

The measurement says otherwise. The numerator is NVIDIA's Hyperscale sub-segment, which NVIDIA broke out for the first time this year. Hyperscale covers public clouds and the largest consumer internet companies, a group wider than the big four. The denominator is the four companies' cash capex. Across the three points available, that fraction runs 24.5% in Q1 2025, 28.5% in Q4 2025 and 29.2% in Q1 2026 (our calculation). The fraction is rising, not falling. The series is a proxy, not a direct count. It includes buyers beyond the big four, and it leaves out AI chips the big four buy through Broadcom, Marvell and others. The reasons are measurable too. Microsoft's CFO says the share of Microsoft's capex going to short-lived assets, mainly CPUs and GPUs, rose from half to two-thirds within a year. Amazon's main reason for raising its guidance by US$20 billion was the rising price of memory. Amazon's Kuiper satellite constellation is the dilution most often named. It costs a single-digit number of billions a year, a rounding error against a denominator above US$200 billion.

The money does reach silicon. The real concern is elsewhere: both ends of the fraction are losing comparability. Accounting choices are moving the denominator: Microsoft has shifted leases out of capex, and Meta has put an entire data center into an off-balance-sheet joint venture. On the numerator side, about half of NVIDIA's data-center revenue now sits outside Hyperscale, and Hyperscale itself includes buyers beyond the big four. The sections below take apart the four public versions of the conversion rate, which differ by a factor of 3.4 purely on definitions. The sections then fix the denominator and the numerator, present the closest series available today, and ask when the fraction would actually turn down.

Compared with the daily brief, this piece adds a measured denominator. The daily brief carried the warning not to read headline capex as the size of the AI server market, but the warning stopped at direction. The four companies' "capex" turns out to follow three different sets of books: cash purchases of property and equipment, management's own figure, and items kept outside both. On a cash basis the 2025 total was US$376.1 billion, and on management's definition US$418 billion. Two more items sit outside both figures: Meta's off-balance-sheet joint venture and US$279 billion of leases signed but not yet commenced. Amazon's segment disclosure measures the non-AWS part directly: US$13 billion to US$15 billion a quarter. The non-AWS share of Amazon's equipment spending has shrunk from 49% in 2023 to 24% in the first half of this year.

The premise that the conversion rate is falling therefore does not hold at the aggregate level.

The condition that would prove this wrong is fixed in advance. NVIDIA's next quarterly report, around August 26, gives the fourth point in this series. If Hyperscale revenue comes in below this quarter's level, or grows markedly slower than the four companies' capex, this piece's reversed call goes back the other way.

One US$720 billion question, four answers, the highest 3.4 times the lowest

At least four groups publicly compute the ratio of NVIDIA revenue to hyperscaler capex, and the highest answer is 3.4 times the lowest. Daloopa, a financial data provider, puts the ratio between 39% and 47% over the past eight quarters (Daloopa). REX Shares, an ETF issuer, gets 57.6% for the first quarter of this year (REX Shares). The advisory firm AL Capital gets 29 cents of NVIDIA revenue for every dollar of capex (AL Capital). One more version works backward from an assumption that chips are only a fifth of capex and lands at 17% to 18% (Investing.com).

None of the four numbers is miscalculated. Each is correct on its own definitions. Set side by side, the four versions differ in only three places. The first is who is in the denominator: whether Oracle counts as a hyperscaler. The second is what is in the numerator: all of NVIDIA's data-center revenue, or only the part sold to cloud companies. The third is whether the two ends line up in time. One version divides trailing-twelve-month revenue by next year's guidance. In a market where the denominator grows about 70% a year, that method shows a "falling" ratio whatever the real conversion rate does. The conversion rate is the right question, but so far nobody has computed it on matching definitions. Any argument that cites one of these versions to say the conversion rate is high or low has to settle the definitions before the number can be judged.

The next two sections fix the two ends of the fraction separately.

Fixing the denominator: the four companies' "capex" is not one thing

Add up the four companies' 2025 "purchases of property and equipment" from their cash-flow statements and the total is US$376.1 billion. Add up the "capex" figures that management cites and the total is US$418 billion. Microsoft alone accounts for US$34.9 billion of the US$42 billion gap, because Microsoft's definition includes finance leases and the other three barely do. That is the first layer: a sum of four numbers on different definitions means something only as an order of magnitude. The definitional gap also explains a tension this publication has carried on its trend map for six months. Figures cited for the big four's combined guidance have run from US$720 billion to US$750 billion, and they never quite reconcile. The guidance figures were never alike. Amazon's US$220 billion is net cash capex. Microsoft's US$175 billion is a book figure after reclassification. Meta's US$130 billion to US$145 billion includes principal payments on finance leases. Aggregate coverage adds them up as if they were the same item.

The second layer matters more: the definitions are not static, accounting choices are moving them, and the movement makes capex smaller. In July Microsoft said it was extending the useful life of its data-center and office buildings from 15 years to 25 years. As a result, a set of leases moves from finance leases to operating leases. Operating leases don't count toward capex, so Microsoft's guidance for calendar 2026 went from about US$190 billion to about US$175 billion. CFO Amy Hood said plainly that, apart from the useful-life effect, the investment plan was unchanged (Microsoft FY26 Q4 earnings call). So the US$15 billion "cut" removes no demand at all. What changed is the accounting classification.

Meta has gone further. Hyperion, Meta's data center in Louisiana, sits in a joint venture in which Meta holds 20%. Because Meta does not consolidate the venture, about US$27 billion of development cost stays out of Meta's capex. CFO Susan Li said on the third-quarter 2025 call that the construction cost would not be recorded in capex (as relayed by Phemex). The same quarterly report disclosed about US$279 billion of lease commitments that are signed but have not yet commenced. That is money that becomes compute over the next decade, and today it sits neither in capex nor on the balance sheet. Stack the two layers and the conclusion is this: headline capex, used as a proxy for the total AI build-out, is drifting systematically toward understatement. Any ratio built on it has a denominator that is thinner than the real one.

The third layer is the one the starting premise cared about most: how much non-AI spending is mixed into the denominator. Here Amazon provides the best primary material in the market, because Amazon is the only one of the four that discloses equipment spending by segment. Everything outside AWS, meaning warehouses, logistics, robots and satellites together, has held steady at US$13 billion to US$15 billion a quarter for the last two quarters. AWS's share has climbed from 51% in 2023 to 62% in 2024, 68% in 2025 and 76% in the first half of this year (Amazon 10-Q).

The dilution most often named is Kuiper, Amazon's satellite constellation, since renamed Amazon Leo. Amazon has never disclosed Leo's capex on its own, and every available anchor is small. Related costs rise about US$1 billion in 2026, and for now they run mostly through the income statement rather than capex. Third-party estimates put the cost of the whole constellation at US$16.5 billion to US$25 billion. Spread over the years, that is a single-digit number of billions a year, 2% to 3% of a US$220 billion annual guide. Capex does include satellites and robots, but that does not noticeably dilute the conversion rate. The non-AI spending is roughly flat in absolute terms while the denominator doubles, so the dilution ratio is shrinking.

Fixing the numerator: NVIDIA's new Hyperscale split

On the numerator side, the only figure available until recently was total revenue from NVIDIA's data-center segment, and that figure was never "what the big four bought." The segment includes new cloud providers such as CoreWeave, sovereign AI and enterprises building their own capacity. NVIDIA's direct-customer disclosures also name contract manufacturers and system builders, not end buyers.

In May, NVIDIA added a finer cut. Starting with its latest quarter, NVIDIA split the data-center segment into two parts. Hyperscale covers public clouds and the largest consumer internet companies, a group wider than the big four. ACIE covers AI clouds, industrial and enterprise customers. NVIDIA also restated two historical points. Hyperscale revenue for the quarter was US$37.87 billion, about half of data-center revenue, up 115% year on year (NVIDIA FY27 Q1 CFO Commentary).

The split also reveals two things that carry weight here. First, the main source of numerator growth has moved outside the big four. ACIE grew 31% quarter on quarter, two and a half times Hyperscale's 12%. Any method that divides NVIDIA's total data-center revenue by big-four capex therefore folds in a growing slice of revenue that is not big-four money at all. That ratio will only inflate from here. Second, even on the narrower numerator, Hyperscale alone, the 115% growth rate is faster than the roughly 80% growth in the big four's cash capex over the same period. That is the most direct evidence available against the idea that the big four's money is flowing away from NVIDIA. At least through the latest available quarter, NVIDIA's Hyperscale revenue grew faster than the big four's total capex. Hyperscale is the closest available proxy for the big four's purchases from NVIDIA.

The measured direction: the share is rising

The table below divides the fixed numerator by the fixed denominator. NVIDIA's fiscal quarters end about a month after calendar quarters, so each point below is aligned to the nearest calendar quarter. That one-month offset is a known error in this table.

Three points, and the direction is up. The series is that short because the official numerator was born this year. But it is the conversion-rate series in the market today whose numerator and denominator match most closely, and none of the four public versions in the first section uses it. Because Hyperscale includes buyers beyond the big four, the share is a proxy for big-four purchases, not a direct count.

The rise is not an accident; the buyers' own disclosures show the mechanism. Microsoft's CFO has described Microsoft's capex mix in stages over the past eight quarters. In mid-2024, about half went to long-lived assets. In late 2024, more than half did. Then the mix flipped. In the third quarter of 2025, about half went to short-lived assets, mainly GPUs and CPUs. For the last three quarters in a row, about two-thirds has gone to short-lived assets, mainly CPUs and GPUs. Within one year, the share of each Microsoft capex dollar flowing to silicon rose from half to two-thirds.

Alphabet's CFO has given the same split for four straight quarters: about 60% of technical-infrastructure investment goes to servers, 40% to data centers and networking (Alphabet Q2 2026 call). When Amazon raised its guidance from US$200 billion to US$220 billion, CEO Andy Jassy gave rising memory costs as the main reason. Memory is silicon, so the price rise alone lifts silicon's share of capex. That is evidence about silicon's share as a whole. Only part of that money shows up in NVIDIA's revenue.

Industry measurements point the same way. SemiAnalysis estimates that memory alone rises from about 8% of hyperscaler capex in 2023 and 2024 to about 30% in 2026 (SemiAnalysis, as relayed by Yahoo Finance). The price of an NVIDIA rack-scale system has climbed from about US$3 million in the GB200 generation to quotes of US$7 million to US$8.8 million in the Vera Rubin generation (Tom's Hardware).

The forces pushing the other way have to be on the table too. Inflation on the non-IT side is just as fierce. New 80 MW to 90 MW gas turbines cost 235% more than in 2024, large transformers cost 70% to 95% more, and lead times have stretched to two to four years (IRMI). Every part of construction and power is getting more expensive. But the buyers' books have already settled the race between the two kinds of inflation. If non-IT inflation were winning, Microsoft's short-lived share would not have gone from half to two-thirds. In this round, silicon inflation and the heavier silicon mix have outpaced the rise in construction and power costs.

Taken together with the previous section, this is enough to rule on the starting premise. At the aggregate level, "the conversion rate is falling" is contradicted by three independent lines of evidence: the official numerator series is rising, the silicon mix the buyers report themselves is rising, and the dilution most often named measures out as a rounding error. The premise gave two reasons. The first was non-AI spending in the mix. It is real in direction but too small to move the fraction, and its share is shrinking. The second was in-house chips diverting purchases. That diversion is real and bigger than the premise assumed, but it dilutes a different fraction, as the next section explains.

In-house chips take NVIDIA's share, not AI silicon's share

The starting premise counted diversion to in-house ASICs as a reason for the conversion rate to fall, and there is a mismatch of definitions here. When Google buys TPUs through Broadcom, or Amazon orders Trainium, Amazon's in-house AI chip, through Marvell and Alchip, a Taiwanese firm that designs custom chips for clients, that money still flows into the AI silicon supply chain. Only the company collecting it changes, from NVIDIA to someone else. The fraction of headline capex that turns into AI silicon revenue does not shrink as a result; NVIDIA's portion of it does. For the big-picture call on whether the AI build-out is real, the relevant measure is AI silicon's share of capex. For holders of NVIDIA or its supply chain, the relevant measure is NVIDIA's share of that money. Folding the two into a single "conversion rate" is the part of the starting premise most in need of repair.

The scale of the diversion has moved from narrative into financial results over the past six months. Broadcom's AI semiconductor revenue was US$12.2 billion in fiscal 2024 and US$20 billion in fiscal 2025, with guidance of US$56 billion for fiscal 2026 and an outlook of more than US$100 billion for fiscal 2027. In the latest quarter Broadcom shipped US$10.8 billion while booking more than US$30 billion of new orders, and CEO Hock Tan described demand as insatiable (Broadcom Q2 FY26). Customer commitments are now quoted in units of power: OpenAI 10 GW through 2029, Meta 3 GW through the end of 2028, Anthropic 5 GW starting next year.

On Amazon's side, Jassy gave two measures. The annualized revenue of Amazon's in-house chip business went from US$10 billion to US$25 billion within six months, a triple-digit growth rate. Jassy also said that Amazon expects Trainium to save Amazon tens of billions of dollars of capex every year. That remark is relayed from the earnings call that followed Amazon's annual filing. In effect, Amazon is saying that the same compute costs less capex on the in-house route. Shrinking NVIDIA's share is Amazon's stated policy, not market speculation.

The diversion is far less smooth than the narrative, though, and the four companies are at very different stages. Google and Amazon are in real volume production. Third-party estimates for 2026 TPU shipments fall between 3.2 million and 4.3 million units, and Amazon's official line is that more than one million Trainium2 chips are in production. Meta's in-house accelerator, MTIA, is speeding up: large-scale deployment began in February, and Meta moved to a six-month cadence per generation. Microsoft's in-house AI chip, Maia, had still not opened to Azure customers when this piece went to press. Volume plans have slipped several times, and the generation after next is now scheduled for 2028. Of the four, Microsoft remains the most dependent on GPUs, despite how loudly it has promoted its own chip.

The design-services side is not a smooth curve either. ASIC design-services revenue moves in steps, one design win at a time. Alchip's 2025 revenue fell 40% after Alchip lost one generation of a single customer's design, and this July its monthly revenue jumped 182% to a record. Extrapolating a diversion trend from a single month or a single year trips over those steps.

One more entry runs the other way. In March this publication recorded a narrative that Microsoft and Amazon were moving orders from Broadcom to Marvell because of Broadcom's hard line on pricing. Checked against the results since: Broadcom's AI revenue growth accelerated to 143%, while Marvell's data-center growth slowed from 76% to 21% and then came back to 27%. On the reported results, Broadcom, the company supposedly being designed out, is growing much faster than Marvell, the expected winner.

On this publication's trend map, this split between GPUs and in-house chips is a branch of its own, and the standing call on that branch is "coexist and diverge." This piece moves that call forward one step. Coexistence is not a static fifty-fifty split. Broadcom's order book and Jassy's capex-saving remark show buyers spending real money to run more of their own inference on in-house chips. But frontier training and capacity rented out to others still run mainly on GPUs. Maia's troubled delivery is the clearest counterexample: buyers with less predictable workloads move to in-house chips most slowly.

When this share would actually turn down

With the call reversed to "rising," the next question is when the starting premise would come true after all. Three paths could get there, and each has a reading that can be watched.

Path one: the silicon mix peaks and falls back. Microsoft's short-lived share is already at two-thirds and Alphabet's server share at 60%. These numbers have a physical ceiling; a data center cannot be all chips and no shell. When construction and power have to catch up, the silicon mix may revert to the mean. Power is the likeliest trigger. The lead times on transformers and gas turbines imply that after 2027 some sites reverse the usual order, building the shell first and installing chips later. The reading: the short-lived share line on each Microsoft earnings call, and Alphabet's 60/40 line. Two consecutive quarters of decline in either is the signal.

Path two: a depreciation wall squeezes reinvestment. The four companies' 2025 cash capex was 1.65 times their 2024 total, and depreciation on those assets will pile into income statements quarter by quarter over the next two to three years. Amazon has even cut the useful life of its servers from six years back to five, citing how fast AI hardware turns over, which brings depreciation earlier and steeper. Once depreciation starts to press on margins, capex growth will be forced to slow. Construction spending locked in by long contracts is harder to cut than silicon purchases, and the US$279 billion of leases not yet commenced will commence year by year. The numerator then slows before the denominator, and the fraction falls. The reading: the first cut in big-four capex guidance that does not come from an accounting reclassification.

Path three: the diversion shows up as capex saved. Jassy's remark that Trainium saves tens of billions of dollars of capex a year points to a possibility nobody had computed before. If the cost advantage of in-house chips is real, the capex needed for the same compute shrinks. Both ends of the fraction, silicon revenue on top and capex below, would shrink together, with vendor revenue shrinking more. On this path, "the conversion rate is falling" would come true in a less obvious way. The same compute would simply need less money. The reading: a cloud provider whose capex growth slows markedly while its compute deployments, counted in GW or in accelerator units, do not slow.

One correction remains. The aggregate conversion rate answers the big-picture question. Decisions need a finer cut: company by company and supplier by supplier. The same 29% is made of very different parts at each of the four. A large share of Alphabet's 60% server spending is TPUs, and that money flows to Broadcom and MediaTek, the two companies collecting Google's TPU spending, rather than to NVIDIA. Amazon is publicly committing to save capex with in-house chips. Meta and Microsoft are still big buyers of merchant silicon, and two-thirds of Microsoft's capex goes to short-lived assets that, in the CFO's description, are mainly CPUs and GPUs. A call on NVIDIA's supply chain has to look at the big four one company at a time. That call also needs two separate lines for each company. One is capex growth. The other is growth in that company's spending with NVIDIA. At Alphabet and Amazon, the gap between those two lines is already taking shape. That gap is read from where the money goes, not from a measured series of each company's NVIDIA spending. That gap is where the starting premise actually survives.

Our call

Headline capex is a denominator that is losing comparability. The four companies use different definitions, and accounting choices are moving real build-out off the books: Microsoft's lease reclassification, Meta's off-balance-sheet joint venture, and US$279 billion of leases signed but not yet commenced. The fraction that turns into AI silicon is the right question, but all four public versions of it stumble on definitions. With NVIDIA's newly split Hyperscale sub-segment as numerator and the big four's cash capex as denominator, the closest series available today runs 24.5% → 28.5% → 29.2% (our calculation), and it points up. Hyperscale covers public clouds and the largest consumer internet companies, a group wider than the big four, so the series is a proxy for their purchases. Silicon inflation and a heavier silicon mix outweigh the non-AI spending in the mix, whose share of capex is shrinking. Silicon inflation and a heavier silicon mix also outweigh the diversion to in-house chips. That diversion takes NVIDIA's share, not AI silicon's share. Where "the conversion rate is falling" does hold, it holds at the company level: the gap opening at Alphabet and Amazon between their purchases from NVIDIA and their capex growth, and the capex-saving path of in-house chips, which could later shrink the numerator and the denominator together. The Alphabet and Amazon gap is read from where their in-house chip money goes. This publication has no measured series of either company's spending with NVIDIA. The next ruling point is the Hyperscale figure in NVIDIA's report around August 26, a test this publication set in advance.

What this changes, and for whom

• Finance and strategy leads at supply-chain companies: turning a customer's capex guidance into the size of a supplier's own market takes three steps: adjust for the definition, apply a conversion rate, then apply the supplier's share of its link in the chain. The first step is where the errors cluster. The four companies' "capex" is three different sets of books. Amazon's US$220 billion is net cash capex, and Microsoft's US$175 billion has just been pushed down by an accounting reclassification. An aggregate such as "big-four capex heading for US$800 billion" does not say which of the three sets of books it sums. Until it is broken down, it says little about a supplier's market.
• Investors tracking the AI infrastructure cycle: the ratio of NVIDIA's data-center revenue to big-four capex is losing meaning, because the main source of the numerator's growth has moved outside the big four, so the ratio will keep inflating. The closest measure is NVIDIA's Hyperscale sub-segment against big-four cash capex. If that line flattens or falls, that is the real diversion signal. The series has only three points so far, so each new quarter is new information. Microsoft's short-lived share belongs on the same watch list. A buyer's CFO describing the split on a call is closer to primary evidence than ten sell-side breakdown models.
• The AI-bubble debate: this piece offers one reassurance and one new worry. The reassurance: the money really is going into silicon, and the claim that capex is hollow, all spent on buildings, does not hold. The share of the US$700 billion-plus flowing to compute assets is rising. The worry: Microsoft's short-lived share, mainly CPUs and GPUs, has been pushed to two-thirds, and silicon is the fastest-depreciating asset. Silicon is written off over five to six years, and Amazon has just shortened that. Because so much of this build-out is short-lived silicon, depreciation will catch up with profits sooner than the headline capex numbers suggest.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.