Nvidia Now Takes 47 Cents of Every Capex Dollar • The Weighted Average

Nvidia’s data center segment booked $89.0 billion in the quarter ended July 26, up 117% year over year, and management guided the whole company to $108 billion next quarter while assuming no China data center revenue at all, according to Nvidia’s fiscal Q2 2027 results. Annualize the data center line at that run rate and it reaches roughly $356 billion. S&P Global expects Alphabet, Amazon, Meta, Microsoft, and Oracle to spend about $750 billion on capital expenditure in 2026, as summarized in coverage of the backup-power supply chain.

Divide one by the other and you get the number that should govern every infrastructure negotiation this quarter: Nvidia’s annualized data center revenue equals about 47 cents for every dollar the five largest US hyperscalers will spend on capex in 2026. Neither company published that ratio. It is imperfect — Nvidia sells to neoclouds, sovereigns, and labs well beyond those five buyers, and hyperscaler capex covers land, shells, and power Nvidia never touches — but the magnitude is the point. One vendor’s revenue is now the same order of size as the entire capital budget of the industry’s biggest builders.

The moat moved off the die

The strategic shift is that Nvidia’s advantage is no longer only the GPU. The same results release announces Vera, which Nvidia calls the first CPU built for AI agents, alongside the Groq 3 LPX inference accelerator in full production, Spectrum-6 switching, and BlueField-4 STX storage processing. TechCrunch’s reporting on how Nvidia’s advantage is moving beyond the GPU quotes Nvidia storage VP Jason Hardy claiming “upwards of 3x improvement in these operations” where the Vera CPU accelerates data movement into flash.

That framing explains the capex share better than any GPU market-share chart. As racks scale, the bottleneck migrates from FLOPs to orchestration — getting bytes to the accelerator without stalling it. Buyers who assumed second-source silicon would cap Nvidia’s take are discovering the take was never only silicon. OpenAI’s own answer, the Jalapeño inference chip designed to minimize data movement, attacks the same problem from the other direction, which is itself confirmation that the problem is real. This paper’s earlier arithmetic on what a Rubin deployment costs per megawatt priced the hardware; this quarter prices the vendor’s grip on the whole rack.

Nvidia also moved to fund the demand it books, announcing partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize over $500 billion of third-party capital for AI infrastructure, subject to definitive agreements. A supplier organizing the financing for its own customers is an efficient market-maker and a concentrated risk, depending entirely on who holds the paper when utilization disappoints.

Jensen Huang’s framing in the release is worth reading as a demand claim rather than a slogan: “This time last year, one lab alone was driving the buildout; today, we have a golden age of new AI labs and startups, multiple frontier labs scaling in parallel, a thriving open-model ecosystem and physical AI coming online.” Concentration risk falling is the bull case for the ratio holding. It is also unverifiable from outside, since Nvidia does not break out revenue by customer class, and the company returned about $26.0 billion to shareholders in the quarter — a capital-allocation posture that assumes the demand curve is real.

What breaks the 47-cent read

Three things, in order of likelihood. The denominator is the weakest link: $750 billion is an S&P Global estimate for five companies in a spending environment that has repeatedly revised upward, and Dell’Oro’s forecast that data center capex surpasses $3 trillion by 2030 — nearly double its own January 2026 view — shows how fast these baselines move. A bigger denominator shrinks the ratio without weakening Nvidia at all.

The numerator is also a snapshot annualized, and Nvidia is guiding sequentially higher, which cuts the other way. Gross margin guidance ticking down to 74.0% from 75.0% hints at mix shift toward systems with more bought-in content — memory especially, where the paper has tracked a 3.6x memory bill riding inside a 15% server price increase.

The real operator question is not whether Nvidia wins the quarter but whether the ratio is durable. Watch two indicators: whether the top four US hyperscalers hold at roughly half of global capex, as Dell’Oro’s Baron Fung projects, and whether the AI-specialized cloud segment — model builders and neoclouds, forecast at nearly a 60% CAGR — keeps buying full Nvidia racks rather than mixing accelerators. If custom silicon absorbs inference while Nvidia keeps training plus orchestration, the 47 cents drifts down without any single deal being lost.

Two further tests sit inside the disclosure. Nvidia secured land, power, and shell capacity through an SB Energy partnership in Ohio to host its own compute, and it is establishing sovereign AI factories with Korea’s SK Telecom and NAVER and with the Japanese government. Vendor-owned capacity and state-backed buyers both change who bears utilization risk, and both dilute the simple story that hyperscaler budgets set the ceiling. If sovereign programs become a durable third demand pool, the 47-cent ratio understates Nvidia’s position rather than overstating it.

For buyers negotiating capacity now: price the system, not the chip. Ask what the CPU, networking, and storage layers cost as a share of the rack, because that is where the incremental margin is moving, and it is the part second-sourcing does not solve. And treat China’s exclusion from guidance as an option, not a loss — export-control geography is still unsettled, and a policy change would land on top of a forecast that already assumes zero.

The governance corollary belongs on the same page as the procurement one. Racks bought this quarter run agents next quarter, and today’s lead finds that the interval between a public hint about a bug and the first automated probe has collapsed to about ten minutes. Nvidia’s own answer is partly institutional — it formed an Open Secure AI Alliance and put confidential-computing GPUs inside Apple’s Private Cloud Compute — but the buyer still owns the operational clock. Capacity contracts that specify throughput and say nothing about patch cadence are half a purchase order.