Layer 3 of 8 · 7 companies
Chips: accelerators and custom silicon
How long can one layer keep most of the profit?
Chip companies kept 72% of the $280B of revenue they added in two years: most of the build-out’s profit has collected here, overwhelmingly at Nvidia. The question for investors is not whether this layer is profitable but how long one company can keep most of the profit while every customer works to design around it.
- combined revenue, latest four quarters
- $544B
- combined revenue, latest four quarters
- Nvidia data-center revenue, +90% on the year before
- $278B
- Nvidia data-center revenue, +90% on the year before
- operating margin across the layer
- 48%
- operating margin across the layer
- of the two-year revenue gain kept as gross profit
- 72%
- of the two-year revenue gain kept as gross profit
The toll booth
Every AI data center is built around accelerators, and for most of the build-out that has meant Nvidia’s. Its data-center revenue over its latest four quarters was $278B, +90% on the year before, and its company-wide gross margin was 75% with an operating margin of 65%. Those are software-like margins on hardware, and they are the clearest sign of scarcity in the whole stack.
Nvidia now splits its data-center revenue by customer type. In its latest quarter hyperscalers accounted for $48.7B and AI clouds, industrial and enterprise customers for $40.3B. The second group matters for the circular-financing story: it includes the GPU clouds Nvidia invests in and buys capacity back from. Two direct customers accounted for 22% and 14% of Nvidia’s fiscal 2026 revenue, so concentration is high even before counting the companies they resell to.
The scale of the bet runs the other way too. Nvidia reports $279 billion of supply and capacity commitments, up from $119 billion a quarter earlier, mostly to TSMC, memory makers and packaging suppliers. That is confidence in demand, contracted years ahead, and a fixed obligation if demand disappoints.
Quarterly revenue: chips
The layer’s 6 US-listed companies, five largest shown separately; Other is Marvell
- Nvidia
- Broadcom
- Intel
- Qualcomm
- AMD
- Other
Source: company 10-K and 10-Q filings (SEC XBRL); fiscal quarters aligned to the nearest calendar quarter. Companies appear from their first reported quarter, so newly listed ones (such as CoreWeave, Sandisk or Astera Labs) add to the total partway through.
Show the numbers
| Quarter | Nvidia | Broadcom | Intel | Qualcomm | AMD | Other |
|---|---|---|---|---|---|---|
| Q1 2022 | $8.3B | $8.1B | $18.4B | $11.2B | $5.9B | $1.4B |
| Q2 2022 | $6.7B | $8.5B | $15.3B | $10.9B | $6.5B | $1.5B |
| Q3 2022 | $5.9B | $8.9B | $15.3B | $11.4B | $5.6B | $1.5B |
| Q4 2022 | $6.1B | $8.9B | $14.0B | $9.5B | $5.6B | $1.4B |
| Q1 2023 | $7.2B | $8.7B | $11.7B | $9.3B | $5.4B | $1.3B |
| Q2 2023 | $13.5B | $8.9B | $12.9B | $8.5B | $5.4B | $1.3B |
| Q3 2023 | $18.1B | $9.3B | $14.2B | $8.6B | $5.8B | $1.4B |
| Q4 2023 | $22.1B | $12.0B | $15.4B | $9.9B | $6.2B | $1.4B |
| Q1 2024 | $26.0B | $12.5B | $12.7B | $9.4B | $5.5B | $1.2B |
| Q2 2024 | $30.0B | $13.1B | $12.8B | $9.4B | $5.8B | $1.3B |
| Q3 2024 | $35.1B | $14.1B | $13.3B | $10.2B | $6.8B | $1.5B |
| Q4 2024 | $39.3B | $14.9B | $14.3B | $11.7B | $7.7B | $1.8B |
| Q1 2025 | $44.1B | $15.0B | $12.7B | $11.0B | $7.4B | $1.9B |
| Q2 2025 | $46.7B | $16.0B | $12.9B | $10.4B | $7.7B | $2.0B |
| Q3 2025 | $57.0B | $18.0B | $13.7B | $11.3B | $9.2B | $2.1B |
| Q4 2025 | $68.1B | $19.3B | $13.7B | $12.3B | $10.3B | $2.2B |
| Q1 2026 | $81.6B | $22.2B | $13.6B | $10.6B | $10.3B | $2.4B |
| Q2 2026 | $96.2B | $29.6B | $16.1B | $9.9B | $11.5B | $2.7B |
Custom silicon: the way around the toll booth
Every large buyer now designs its own accelerator, many of them with a design partner such as Broadcom or Marvell. Broadcom’s semiconductor segment, which includes its custom accelerators and Ethernet switch chips, earned $59.4B over its latest four quarters, +75%. Broadcom reports that one distributor accounted for half its revenue in the latest quarter and that its top five end customers accounted for about 55%: custom silicon is a business of a few enormous customers.
Marvell’s data-center revenue was $7.2B over its latest four quarters, +33%. Arm, which licenses the CPU designs inside most custom cloud chips, reported $4.9B of revenue for its fiscal year at a gross margin of 98%: a small revenue line with a toll on nearly every custom chip.
Custom chips don’t have to beat Nvidia’s to matter. They only have to be good enough for a buyer’s own most common workloads at a lower cost. Each one that succeeds moves a slice of spending from the highest-margin company in the stack to a design partner earning less, and to the foundry, which earns the same either way.
Nvidia’s data-center revenue, by quarter
Revenue from Nvidia’s Data Center market platform
Source: Nvidia 10-K and 10-Q filings (XBRL, Data Center market platform). Fiscal quarters aligned to the nearest calendar quarter.
Show the numbers
| Quarter | Revenue |
|---|---|
| Q1 2023 | $4.3B |
| Q2 2023 | $10.3B |
| Q3 2023 | $14.5B |
| Q4 2023 | $18.4B |
| Q1 2024 | $22.6B |
| Q2 2024 | $26.3B |
| Q3 2024 | $30.8B |
| Q4 2024 | $35.6B |
| Q1 2025 | $39.1B |
| Q2 2025 | $41.1B |
| Q3 2025 | $51.2B |
| Q4 2025 | $62.3B |
| Q1 2026 | $75.2B |
| Q2 2026 | $89.0B |
AMD and Intel: the challengers
AMD’s data-center segment, which includes its Instinct GPUs and EPYC server CPUs, brought in $6.7B in its latest quarter. AMD has used an unusual tool to win share: warrants for up to 160 million of its shares each to OpenAI and Meta at a penny a share, vesting in tranches as they hit purchase milestones for its Instinct GPUs and as AMD’s share price reaches set targets. That is a discount paid in stock, and it shows how hard it is to dislodge an incumbent with a software ecosystem as entrenched as Nvidia’s.
Intel’s Data Center and AI segment earned $20.2B, +22%. Server CPUs remain essential in AI racks, but at a far smaller share of the spend than accelerators. Intel’s larger AI question is its foundry: whether it can win outside customers that would give the industry an alternative to TSMC. Qualcomm, growing +2%, is the bet on AI moving to phones and PCs, where it already sells the processors.
What keeps the margins up, and what could bring them down
Nvidia’s margins rest on three things: the best-performing chips, a software platform developers have built on for well over a decade, and complete systems (chips, networking, racks) that are faster to deploy than assembling parts. As long as buyers are racing for capacity, they pay for time-to-deployment, and that favors the incumbent.
Margins in this layer would come under pressure from three directions: custom chips taking the most common workloads, buyers slowing their capex so that time-to-deployment matters less, and supply catching up so customers can negotiate. Gross margin is the early warning. A falling margin at Nvidia while its revenue still grows would say more about the cycle than any revenue figure.
The companies, side by side
Latest four reported quarters for US filers; latest fiscal year for 20-F filers (marked). Underlined names have their own company page; the others link to the company’s filings on SEC EDGAR. Margins are left blank where a company doesn’t report gross profit or cost of revenue.
| Company | Revenue | vs year before | Gross margin | Operating margin | Free cash flow |
|---|---|---|---|---|---|
| Nvidia NVDAGPUs, systems and networking; the largest seller in the stack | $303B | +83% | 75% | 65% | $127B |
| Broadcom AVGOCustom accelerators (XPUs) for hyperscalers, plus Ethernet switch silicon | $89.1B | +49% | 69% | 48% | $39.4B |
| Intel INTCXeon server CPUs and a foundry trying to win outside customers | $57.0B | +7% | 39% | 0% | $2.8B |
| Qualcomm QCOMEdge AI on phones and PCs; early data-center accelerators | $44.1B | +2% | 54% | 23% | $10.4B |
| AMD AMDInstinct GPUs and EPYC server CPUs | $41.3B | +40% | 53% | 16% | $8.4B |
| Marvell MRVLCustom silicon and optical DSPs for cloud data centers | $9.5B | +31% | 52% | 17% | $1.7B |
| Arm ARMLicenses the CPU designs inside most custom cloud chipsFiscal year to March 31, 2026 (20-F) | $4.9B | +23% | 98% | 18% | $979M |
Revenue and growth. Total revenue over the four quarters, and its change from the four quarters before. Fast growth with a flat margin usually means volume; fast growth with a rising margin usually means pricing power.
Gross margin. The share of revenue left after the direct cost of what was sold. It is the best single measure of how much of each dollar a company keeps rather than passes on to its own suppliers.
Operating margin. What is left after research, sales and overhead too. The gap between gross and operating margin shows how much the company spends to stay competitive.
Free cash flow. Operating cash flow minus capital spending: the cash the business generated after paying to maintain and grow itself. Negative free cash flow during fast growth means the growth is consuming cash.
Where the money comes from and goes
This layer is paid by AI clouds and servers and networking. Its own spending flows to foundry & tools, memory. For the dollar amounts and who keeps them, see following the money.
Financing links that involve this layer
Quoted from the filings. Numbers match the money-loop map.
11Nvidia → OpenAICredit or guarantee
Lease guarantees for OpenAI, capped at $105B
“The campus will exclusively host our compute under 20-year leases to OpenAI, subject to limited exceptions, with our obligation capped at $105 billion in the aggregate, subject to certain conditions including SB Energy, the lessor, satisfying applicable ready-for-service conditions.”
Nvidia 10-Q for the period ended July 26, 2026, filed August 26, 2026
12Nvidia → Nvidia’s private holdingsEquity investment
$99B invested, $25B more committed
“These include equity investments of $99 billion and equity investment commitments of $25 billion as of July 26, 2026.”
Nvidia 10-Q for the period ended July 26, 2026, filed August 26, 2026
13Nvidia → AI cloud partnersPurchase commitment
$36B of cloud capacity it has agreed to buy back
“Our commitments, which are typically six years in duration, totaled $36 billion as of July 26, 2026, and decrease as capacity is used by third-party customers or by us for our research and development efforts.”
Nvidia 10-Q for the period ended July 26, 2026, filed August 26, 2026
14Nvidia → Nvidia’s suppliersPurchase commitment
$279B of supply and capacity commitments
“We have significantly increased our supply and capacity commitments from $119 billion last quarter to $279 billion as of July 26, 2026 to meet future demand.”
Nvidia 10-Q for the period ended July 26, 2026, filed August 26, 2026
15Nvidia → CoreWeaveEquity investment
$2.0B share purchase, January 2026
“In January 2026, we entered into a securities purchase agreement with NVIDIA Corporation for a private placement of approximately 23 million shares of our Class A common stock at a purchase price of $87.20 per share, for aggregate gross proceeds of $2.0 billion.”
CoreWeave 10-Q for the period ended June 30, 2026, filed August 12, 2026
19AMD → OpenAICustomer warrant
Warrant for up to 160M AMD shares at $0.01
“In October 2025 and February 2026, the Company issued warrants to OpenAI OpCo, LLC (OpenAI) and Meta Platforms, Inc. (Meta), each entitling the holder the right to purchase up to 160 million shares of the Company’s common stock at an exercise price of $0.01 per share (the OpenAI Warrant and the Meta Warrant, respectively).”
AMD 10-Q for the period ended June 27, 2026, filed August 5, 2026
20AMD → MetaCustomer warrant
Warrant for up to 160M AMD shares at $0.01
“In October 2025 and February 2026, the Company issued warrants to OpenAI OpCo, LLC (OpenAI) and Meta Platforms, Inc. (Meta), each entitling the holder the right to purchase up to 160 million shares of the Company’s common stock at an exercise price of $0.01 per share (the OpenAI Warrant and the Meta Warrant, respectively).”
AMD 10-Q for the period ended June 27, 2026, filed August 5, 2026
What to watch in the next filings
- Nvidia’s gross margin and the hyperscale vs AI-cloud split of its data-center revenue.
- Broadcom’s semiconductor segment growth and customer concentration: the read on custom accelerators.
- Equity investments, capacity buy-backs and guarantees in Nvidia’s 10-Q, and the warrants AMD has issued.
- Nvidia’s supply and capacity commitments: confidence when they rise, risk if demand slows.
- Any customer moving a major workload to its own chips, announced on earnings calls or visible in the design partners’ results.
Free · quarterly · from SEC filings
Get the AI Stack Report every quarter
After each Big Tech earnings season we update the map from the new filings and publish what changed: who spent, who got paid, who kept it and what the filings reveal about the loop. Free, in Rob May’s Investing in AI newsletter.
Form not showing? Subscribe on Substack. No spam; unsubscribe any time.
Questions about chips
How much of Nvidia’s revenue comes from data centers?
Nvidia reported $278B of data-center revenue over its latest four quarters, the large majority of its total revenue of $303B. It now splits data-center revenue between hyperscalers and AI clouds, industrial and enterprise customers.
Why are Nvidia’s margins so high?
Because demand for its accelerators has exceeded supply, and its software platform and complete systems make it costly for buyers to switch. Its gross margin over its latest four quarters was 75% and its operating margin 65%.
What are custom AI chips?
Accelerators designed by the cloud companies themselves for their own workloads, usually with a design partner such as Broadcom or Marvell, and manufactured by TSMC. Google’s TPUs and Amazon’s Trainium are examples. They aim to cut the cost of common workloads rather than to beat Nvidia outright.
Why did AMD give OpenAI and Meta warrants?
To win large orders for its accelerators. Each holder can buy up to 160 million AMD shares at a penny each as purchase milestones are met, which is in effect a discount paid in stock rather than cash.
Research and education, not investment advice: nothing here is a recommendation to buy or sell any security. Figures come from companies’ SEC filings as reported; companies named were chosen because they show how the stack works, not as a view on their shares.