Roughly 35-40% of every dollar of hyperscaler AI capex — about $110-130B of the $300B+ being spent in 2025 — ends up on Nvidia's income statement. That's the short answer. The longer answer is more interesting.
No single company has ever captured this much of an entire industry's capital budget. Understanding exactly where the money goes — and where it doesn't — tells you more about the durability of the AI trade than any earnings headline.
Nvidia's Share of Hyperscaler AI Capex, Explained
Nvidia's share of hyperscaler AI capex is approximately 35-40%, equal to about $110-130B of the $300B+ that Microsoft, Google, Meta, and Amazon are spending on infrastructure in 2025. The exact figure depends on definition: counting only the compute portion of capex pushes Nvidia's share above 50%, while including land, buildings, and power — which Nvidia does not supply — pulls it below 40%.
That concentration is unprecedented. In the cloud-buildout era of 2015-2019, no single supplier captured more than about 10-15% of hyperscaler capex; spend was spread across Intel, Dell, server ODMs, networking vendors, and construction. AI flipped that. One vendor now sits at the center of nearly every cluster being built.
How Much of the $300B Goes to Nvidia, by Buyer
The table below estimates 2025 capex by company and the slice flowing to Nvidia. These are analyst-consensus ranges, not reported segment figures — Nvidia does not disclose revenue by customer, and hyperscalers do not disclose vendor breakdowns. Treat them as directional.
| Buyer | 2025 Capex | Est. to Nvidia | Nvidia % | Own Silicon |
|---|---|---|---|---|
| Microsoft | ~$80B | $30-40B | ~45% | Maia (early) |
| Amazon (AWS) | ~$100B | $20-30B | ~25% | Trainium/Inferentia |
| ~$75B | $15-25B | ~28% | TPU v6 (mature) | |
| Meta | ~$65B | $25-35B | ~48% | MTIA (early) |
| Oracle / CoreWeave / others | ~$40B+ | $20-30B | ~60% | None |
| Total | $360B+ | $110-130B | ~35-40% | — |
The pattern is clear: buyers with mature in-house silicon (Google's TPU, Amazon's Trainium) send a smaller share to Nvidia, while neoclouds like CoreWeave and Oracle — which have no custom chip — send the most. Track the underlying capex on the Big Tech Earnings dashboard.
The Data-Center Math Behind Nvidia's Capex Share
You can sanity-check the share from the seller's side. Nvidia's data-center segment exited fiscal 2025 above a $115B annualized run-rate, up from roughly $47B the year prior — a jump of about 145%. Data center is now nearly 90% of Nvidia's total revenue. The largest cloud customers — by Nvidia's own filings, a small handful of buyers — account for an estimated 40%+ of revenue, with two customers each above 10%.
Run the unit math and it holds up. A single GB200 NVL72 rack lists in the $3M+ range. A 100,000-GPU training cluster — now table stakes for a frontier lab — runs $3-4B in Nvidia hardware alone, before networking, power, or the building. Build a dozen of those across the hyperscalers in a year and you are at $40-50B without trying. Stack in inference fleets, and $110-130B is conservative.
Nvidia's gross margin tells the rest of the story: around 73-75%. The hyperscalers are spending depreciating capital at a rate that compresses their own free cash flow, while Nvidia converts the same dollars into the highest-margin hardware business at scale in history.
What the Other 60% of AI Capex Buys
If Nvidia takes 35-40%, where does the rest of the $300B+ go? Understanding the non-Nvidia slice matters because it's where the share-shift battle is being fought.
Data-center construction & land
~20-25%Shells, fit-out, and real estate — zero Nvidia content
Power & cooling infrastructure
~10-15%Substations, generators, liquid cooling; the new bottleneck
Networking (non-Nvidia)
~5-8%Arista, Broadcom, optics — though Nvidia also competes here
Custom AI silicon (ASICs)
~8-12%TPU, Trainium, MTIA, Maia — the direct Nvidia substitute
Storage & memory
~5-8%HBM is constrained; SK Hynix and Micron benefit
AMD & other GPUs
~3-5%MI300/MI350 gaining in inference workloads
The interesting line is custom silicon. Google's TPU program is a decade old and now handles the bulk of its internal training; Amazon's Trainium2 is winning real Anthropic workloads. Every dollar that moves to in-house ASICs is a dollar that leaves Nvidia's share — which is why Nvidia's slice is high but slowly contracting at the margin.
Why Nvidia's Share of AI Capex Stays This High
Three structural moats keep the concentration sticky even when buyers actively want to diversify:
Why Spend Concentrates
- ✓ CUDA: 15+ years of software lock-in for training
- ✓ Networking: InfiniBand/NVLink scale clusters rivals can't match
- ✓ One generation ahead on training perf-per-watt
- ✓ Supply allocation rewards the biggest committed buyers
Why the Share Erodes Over Time
- ✕ Inference (the majority of future compute) is less CUDA-dependent
- ✕ Custom ASICs are 30-50% cheaper per token at scale
- ✕ Buyers hate single-vendor dependence on principle
- ✕ AMD and Broadcom-built chips are now production-grade
What This Means for Investors and Founders
The single most important number to watch is not Nvidia's revenue — it's Nvidia's share of compute capex. Revenue can keep climbing while share falls, because the pie is growing 30-40% a year. If total AI capex hits $500B+ by 2027 and Nvidia's share drops from ~38% to ~30%, that's still ~$150B in Nvidia revenue — more dollars, smaller slice.
For founders building in the AI infrastructure layer, the opportunity is in the other 60%: cooling, power, networking, inference optimization, and the software that makes non-Nvidia silicon usable. The most defensible bets are the ones that help hyperscalers reduce their Nvidia dependence — that's a $100B+ structural tailwind. Compare valuations across the stack on the AI Valuations dashboard.
One company captures 35-40% of an entire industry's capital budget.
The AI trade is, for now, a bet that Nvidia's ~$120B share keeps growing faster than buyers can build their way out of it.
Track hyperscaler spending and Nvidia's share on the Big Tech Earnings Dashboard at Value Add VC. Originally published in the Trace Cohen newsletter.