The AI silicon ledger: GPUs, custom chips, foundry and memory, October 2026
Nvidia's data center revenue hit $89.0B and Broadcom's AI chips $16.7B, while Micron's sales rose 4.8 times as memory became the AI bottleneck.
Key findings
- Nvidia’s Data Center revenue reached $89.0B in its fiscal Q2 2027 (quarter ended July 26, 2026), up 117% from $41.1B a year earlier and 93% of total revenue of $96.2B. The company guided to $108.0B, plus or minus 2%, for the following quarter.
- Custom silicon is growing faster from a smaller base: Broadcom’s AI semiconductor revenue rose 221% to $16.7B in its fiscal Q3 2026 and is guided to $21.7B next quarter. Nvidia’s Data Center line was still 5.3 times that figure.
- Memory has become the swing factor. Micron’s revenue was $54.23B in the quarter ended September 3, 2026, 4.8 times the $11.32B of a year earlier and larger than TSMC’s $40.20B in calendar Q2. Micron’s GAAP gross margin was 86.8%.
- Nvidia’s supply and capacity commitments rose to $279B at July 26, 2026 from $119B one quarter earlier, an increase the company attributed primarily to memory procurement. Total commitments were $366B.
- TSMC’s high-performance computing platform was 66% of Q2 2026 revenue. The foundry raised 2026 capex to $60B to $64B and reported September revenue up 54.6% year over year. It reports Q3 results on October 15.
- Demand beyond the hyperscalers is growing faster than hyperscaler demand: Nvidia’s AI clouds, industrial and enterprise revenue rose 138% to $40.3B, against 102% growth to $48.7B for hyperscale.
The ledger at the latest quarter
This report follows the money one layer below the hyperscaler spending tracked in our AI capex ledger: the chip designers, the foundry and the memory makers that turn that capex into revenue. The lines below are not like for like. Fiscal quarters end on different dates, AMD’s Data Center segment includes server CPUs, Marvell’s includes non-AI products, and TSMC’s figure is a derived platform share. They are shown together to give a sense of scale, not to be summed.
| Company | Line | Quarter ended | Revenue | Year over year | Next-quarter guide |
|---|---|---|---|---|---|
| Nvidia | Data Center | July 26, 2026 | $89.0B | +117% | $108.0B total revenue, plus or minus 2% |
| Micron | Cloud Memory and Core Data Center units | Sept. 3, 2026 | $34.3B | n/a | $61.5B total revenue, plus or minus $1.5B |
| TSMC | HPC platform (66% of revenue, derived) | June 30, 2026 | about $26.5B | n/a | $44.6B to $45.8B total revenue |
| Broadcom | AI semiconductors | Aug. 2, 2026 | $16.7B | +221% | $21.7B AI semiconductors |
| AMD | Data Center segment | June 27, 2026 | $6.7B | +107% | about $13B total revenue, plus or minus $300M |
| Marvell | Data center end market | Aug. 1, 2026 | $2.2B | +46% | $3.15B total revenue, plus or minus 5% |
Merchant GPUs: Nvidia and AMD
Nvidia’s revenue has roughly doubled in a year while GAAP gross margin rose from 72.4% to 75.0%. Data Center revenue grew 18% from the prior quarter alone. From fiscal Q1 2027 the company reports Data Center in two parts. Hyperscale was $48.7B in the latest quarter, 55% of the segment, and AI clouds, industrial and enterprise (ACIE) was $40.3B. ACIE grew 138% year over year against 102% for hyperscale, so the customer base is broadening even as the largest buyers keep growing.
| Nvidia fiscal quarter (ended) | Total revenue | Data Center | GAAP gross margin |
|---|---|---|---|
| Q2 FY26 (July 27, 2025) | $46.7B | $41.1B | 72.4% |
| Q3 FY26 (Oct. 26, 2025) | $57.0B | $51.2B | 73.4% |
| Q4 FY26 (Jan. 25, 2026) | $68.1B | $62.3B | 75.0% |
| Q1 FY27 (Apr. 26, 2026) | $81.6B | $75.2B | 74.9% |
| Q2 FY27 (July 26, 2026) | $96.2B | $89.0B | 75.0% |
| Q3 FY27 outlook | $108.0B, plus or minus 2% | not guided | 74.0%, plus or minus 50 bps |
Concentration remains a feature of the ledger. Nvidia’s 10-Q shows five direct customers accounting for 22%, 14%, 13%, 11% and 10% of accounts receivable at July 26, 2026, or 70% combined. Receivables stood at $63.1B and inventory at $31.6B.
AMD is the second merchant supplier, at a different scale. Its Data Center segment, which combines EPYC server CPUs and Instinct GPUs, grew 107% to $6.72B in Q2 2026 and was 58% of company revenue. AMD does not break out AI GPU revenue. Its release names OpenAI among the customers deploying its Helios rack-scale systems, and AMD has separately announced agreements for up to 6 GW of Instinct GPUs with each of OpenAI and Meta and up to 2 GW with Anthropic. Guidance for Q3 2026 is revenue of about $13B.
| AMD quarter | Data Center segment | Total revenue |
|---|---|---|
| Q2 2025 | $3.24B | $7.69B |
| Q3 2025 | $4.34B | $9.25B |
| Q4 2025 | $5.38B | $10.27B |
| Q1 2026 | $5.78B | $10.25B |
| Q2 2026 | $6.72B | $11.54B |
Custom silicon: Broadcom, Marvell and the in-house programs
Broadcom is the main route by which hyperscalers and labs build their own accelerators, and its AI line is now most of the company’s growth. AI semiconductor revenue was 56% of total revenue of $29.6B in the latest quarter.
| Broadcom fiscal quarter (ended) | Total revenue | AI semiconductor revenue | AI growth, year over year |
|---|---|---|---|
| Q1 FY26 (Feb. 1, 2026) | $19.3B | $8.4B | +106% |
| Q2 FY26 (May 3, 2026) | $22.2B | $10.8B | +143% |
| Q3 FY26 (Aug. 2, 2026) | $29.6B | $16.7B | +221% |
| Q4 FY26 guidance (Nov. 1, 2026) | about $34.8B | $21.7B | +236% |
On its September 2 call, according to a third-party transcript, management put fiscal 2026 AI revenue at about $58B and described a path to about $115B in fiscal 2027, said custom accelerators (XPUs) were 73% of AI revenue in the quarter, and named Google, Anthropic, OpenAI and Meta among six XPU customers. Those forward figures are management commentary taken from a transcript, not filed guidance, and should be read with that caveat.
Marvell, the second custom-silicon supplier, reported data center revenue of $2.17B in the quarter ended August 1, 2026, up 46% and 79% of company revenue. It does not disclose custom AI revenue in dollars.
| Program | Design partner | Latest disclosed status | Source type |
|---|---|---|---|
| Google TPU | Broadcom | Anthropic announced an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity from 2027 (April 6, 2026) | Company release |
| Amazon Trainium, Graviton, Nitro | In-house (Annapurna) | Amazon said its chips revenue run rate is over $25B, up from over $20B a quarter earlier; a Trainium-only figure is not disclosed | Company statement |
| Meta MTIA | Broadcom | Multi-generation partnership announced April 14, 2026, with more than 1 GW in the first phase | Company release |
| OpenAI custom accelerator | Broadcom | 10 GW framework agreement dated October 13, 2025; an inference chip was unveiled June 24, 2026 with first deployment targeted by the end of 2026 | Company release; framework agreement per trade press |
| Microsoft Maia | In-house | Maia 200 on TSMC 3nm reported entering production deployments | Trade press |
Foundry: TSMC
Most of the accelerators above are manufactured by TSMC, which makes its results the cleanest read on aggregate AI silicon demand. The HPC platform’s share of revenue rose from 55% in Q4 2025 to 66% in Q2 2026, and gross margin reached 67.7%.
| TSMC quarter | Revenue (US$) | HPC share of revenue | Gross margin |
|---|---|---|---|
| Q2 2025 | $30.07B | 60% | 58.6% |
| Q3 2025 | $33.10B | 57% | 59.5% |
| Q4 2025 | $33.73B | 55% | 62.3% |
| Q1 2026 | $35.90B | 61% | 66.2% |
| Q2 2026 | $40.20B | 66% | 67.7% |
Monthly revenue for July, August and September 2026 sums to about NT$1,494B, and September alone was up 54.6% year over year. Management guided Q3 revenue to $44.6B to $45.8B, raised 2026 capex to $60B to $64B from $52B to $56B, and lifted its 2026 revenue growth outlook to slightly above 40% in US dollar terms. On the July call management also said advanced packaging capacity is limiting how fast customers can grow, which makes packaging, alongside memory, a gating input for accelerator supply.
Memory: the new bottleneck
The largest change in the ledger over the past year is in memory. Micron’s fiscal 2026 revenue was $133.19B, 3.6 times fiscal 2025, and its two data center business units together were 63% of fourth-quarter revenue. The Core Data Center unit alone grew to $18.0B from $1.6B a year earlier.
| Micron business unit | FQ4 2026 (ended Sept. 3, 2026) | FQ3 2026 | FQ4 2025 |
|---|---|---|---|
| Cloud Memory | $16.28B | $13.77B | $4.54B |
| Core Data Center | $18.00B | $11.52B | $1.58B |
| Mobile and Client | $13.11B | $11.52B | $3.76B |
| Automotive and Embedded | $6.82B | $4.63B | $1.43B |
| Total revenue | $54.23B | $41.46B | $11.32B |
| GAAP gross margin | 86.8% | 84.6% | 44.7% |
The Korean suppliers show the same pattern. SK hynix reported Q2 2026 revenue of KRW 79.3 trillion and operating profit of KRW 60.5 trillion, a 76% operating margin, and said HBM4 mass shipments began in the quarter. Samsung’s semiconductor division reported operating profit of KRW 89.2 trillion on revenue of KRW 127.5 trillion. None of the three disclosed high-bandwidth memory (HBM) revenue in these releases, so the AI share of memory sales cannot be measured directly from filings.
Commitments: the supply chain is being prepaid
The buyers’ side of the memory squeeze is visible in Nvidia’s commitments. Supply and capacity obligations rose by $160B in a single quarter, to nearly three times the quarter’s revenue.
| Nvidia commitments at July 26, 2026 | Amount |
|---|---|
| Supply and capacity | $279B |
| Cloud service agreements | $29B |
| Data center leases not yet commenced | $25B |
| Equity investments | $25B |
| Capital expenditures | $8B |
| Total | $366B |
Of the supply and capacity total, $92B falls due in the remainder of fiscal 2027, $87B in fiscal 2028 and $88B in fiscal 2029. Nvidia also disclosed guarantees with a maximum exposure of $108.5B, of which $105.0B is a cap on guarantees entered into in August 2026 for an SB Energy campus in Ohio, to be leased to OpenAI for 20 years. The guarantees take effect in phases, with the first expected in fiscal 2029. Chip suppliers are no longer only selling into the buildout; they are underwriting parts of it, a theme we cover in the AI lab funding tracker.
China and export controls
China is close to absent from the current numbers. Nvidia said shipments of its older Hopper-generation products to China were under 1% of Data Center revenue in the latest quarter, and its $108.0B outlook assumes no Data Center compute revenue from China. Under a Bureau of Industry and Security rule dated January 15, 2026, as summarised by export-control counsel, sales of chips such as Nvidia’s H200 to approved Chinese buyers are reviewed case by case, with exports limited to 50% of the aggregate processing performance shipped to US customers. A separate presidential action imposes a 25% tariff on such chips. AMD said around its Q4 2025 results, as reported by trade press, that about $390M of MI308 revenue came from China, and it has not given a figure since. Any reopening would be additive to the guidance above; none is assumed in it.
What to watch
- October 15, 2026: TSMC reports Q3 2026 results, including the HPC share, 2026 capex and any update on packaging capacity.
- November 3, 2026: AMD reports Q3 2026, after saying in August that Helios would begin to ramp in the second half.
- Mid-November 2026: Nvidia’s fiscal Q3 report, against the $108.0B outlook. The date had not been confirmed by press release as of October 11.
- December 2026: Broadcom and Micron report; neither date has been announced.
- Gross margin: whether memory costs show up in accelerator margins, given the scale of Nvidia’s memory commitments.
Methodology
Scope: the largest listed suppliers of AI accelerators, custom silicon, foundry capacity and memory. Data covers each company’s most recent reported quarter as of October 11, 2026, with prior quarters for trend. Fiscal calendars differ: Nvidia’s fiscal Q2 2027 ended July 26, 2026; Broadcom’s fiscal Q3 2026 ended August 2, 2026; Marvell’s fiscal Q2 2027 ended August 1, 2026; Micron’s fiscal Q4 2026 ended September 3, 2026; AMD and TSMC report calendar quarters.
Sources and definitions: figures are taken from company earnings releases, SEC filings, Nvidia’s CFO commentary and TSMC’s management reports and monthly revenue releases. Nvidia margins are GAAP because the company changed its non-GAAP definition in fiscal 2027. Broadcom’s AI semiconductor revenue is as defined by the company. AMD’s Data Center segment and Marvell’s data center end market include non-AI products. TSMC’s HPC revenue is derived by applying the reported 66% platform share to Q2 2026 net revenue. Micron’s data center figure is the sum of its Cloud Memory and Core Data Center business units.
Computations: year-over-year growth, shares, ratios and sums were computed in Python from the sourced figures. The summary table and chart place lines with different definitions and period ends side by side for scale only.
Limitations: Broadcom’s forward AI revenue figures, its customer names and XPU mix come from a third-party call transcript that carries an accuracy disclaimer. Several in-house chip program details come from trade press and are labelled as such. Nvidia’s receivables concentration is from its 10-Q; revenue concentration by customer was not verified against the filing and is not shown. HBM revenue is not disclosed by any memory supplier. This report is not investment advice.
Sources
- NVIDIA second quarter fiscal 2027 results, Form 8-K exhibit (as of August 26, 2026)
- NVIDIA Q2 fiscal 2027 CFO commentary (as of August 26, 2026)
- NVIDIA Form 10-Q for the quarter ended July 26, 2026 (as of July 26, 2026)
- AMD second quarter 2026 financial results (as of August 4, 2026)
- Broadcom third quarter fiscal 2026 results, Form 8-K exhibit (as of September 2, 2026)
- Broadcom Q3 fiscal 2026 earnings call transcript, third party (as of September 2, 2026)
- Marvell second quarter fiscal 2027 results (as of August 27, 2026)
- TSMC 2Q26 management report (as of July 16, 2026)
- TSMC 2Q26 earnings conference transcript (as of July 16, 2026)
- TSMC September 2026 revenue report (as of October 8, 2026)
- Micron fiscal Q4 and full-year 2026 results, Form 8-K exhibit (as of September 30, 2026)
- Samsung Electronics second quarter 2026 results (as of July 30, 2026)
- SK hynix Q2 2026 results, reprint of company release (as of July 30, 2026)
- Amazon on its chips business and run rate (as of July 31, 2026)
- Anthropic on its Google and Broadcom compute agreement (as of April 6, 2026)
- ArentFox Schiff on the BIS rule for H200 exports (as of January 16, 2026)
- NVIDIA first quarter fiscal 2027 results, Form 8-K exhibit (as of May 20, 2026)
- NVIDIA fourth quarter fiscal 2026 results, Form 8-K exhibit (as of February 25, 2026)
- NVIDIA third quarter fiscal 2026 results, Form 8-K exhibit (as of November 19, 2025)
- AMD first quarter 2026 financial results (as of May 5, 2026)
- AMD fourth quarter 2025 financial results (as of February 3, 2026)
- AMD and OpenAI strategic partnership announcement (as of October 6, 2025)
- DatacenterDynamics on AMD’s Q4 2025 results and China revenue (as of February 2026)
- Broadcom second quarter fiscal 2026 results, Form 8-K exhibit (as of June 3, 2026)
- Broadcom first quarter fiscal 2026 results, Form 8-K exhibit (as of March 4, 2026)
- TSMC 1Q26 management report (as of April 16, 2026)
- TSMC 3Q25 management report (as of October 16, 2025)
- Amazon first quarter 2026 results (as of April 2026)
- Meta on its custom silicon partnership with Broadcom (as of April 14, 2026)
- OpenAI on its inference chip with Broadcom (as of June 24, 2026)
- Redmondmag on Microsoft’s Maia chips (as of September 30, 2026)
New reports, trackers and scorecards on markets, AI and digital assets, sent when they publish.
Subscribe to the newsletterQuestions on this report
How much AI chip revenue does Nvidia report?
Nvidia's Data Center segment generated $89.0B in the quarter ended July 26, 2026, up 117% year over year and 93% of total revenue of $96.2B. The company guided to total revenue of $108.0B, plus or minus 2%, for the following quarter.
How big is custom AI silicon compared with Nvidia?
Broadcom, the largest custom accelerator supplier, reported $16.7B of AI semiconductor revenue in the quarter ended August 2, 2026, up 221%, and guided to $21.7B. Nvidia's Data Center revenue was about 5.3 times Broadcom's AI revenue in the latest quarters.
Why does memory matter for AI chips?
Accelerators need large amounts of high-bandwidth memory, and supply is tight. Micron's revenue rose to $54.23B in the quarter ended September 3, 2026 from $11.32B a year earlier, and Nvidia's supply and capacity commitments rose to $279B from $119B in one quarter, mainly for memory.
BlockWest Research reports are built from public filings, company disclosures and third-party data providers named in each report. Figures are point-in-time and are not updated after publication unless noted. This is research and commentary, not investment advice. Spotted an error? Email the desk via the contact page.