AI Is Not a Chip Trade

ASML, the only manufacturer of extreme-ultraviolet lithography systems, reported a gross margin of 54.0% for the quarter ended in June 2026.

Super Micro's fiscal 2026 (year ended 30 June 2026) net sales were $39.063 billion, up 78% from $21.972 billion, while GAAP gross margin fell to 10.8% from 11.1%; in fiscal 2019 its gross margin was 14.2%.

Hon Hai Technology Group (Foxconn), the largest contract manufacturer of AI servers, reported a gross margin of 6.12% and an operating margin of 3.75% for the quarter ended 30 June 2026.

Fabrinet's fiscal 2026 (year ended 26 June 2026) revenue was $4.641 billion, up 36%, while GAAP gross margin was 12.0%, down from 12.2%.

An NVIDIA GB300 NVL72 rack-scale system is specified by NVIDIA as a liquid-cooled rack of 72 GPUs requiring up to 142 kW.

Microsoft's chief financial officer told investors on 29 April 2026: 'For calendar year 2026, we expect to invest roughly $190 billion in capital expenditures which includes approximately $25 billion from the impact of higher component pricing.'

On 30 July 2026 Amazon's chief executive said the company now expects to spend approximately $220 billion in cash capital expenditure in 2026, up from a prior estimate of about $200 billion, attributing the increase to the higher cost of memory, and said that even at that amount 'we will still not have enough capacity to meet all the demand we have in 2026.'

SK hynix reported an operating margin of 76% for the quarter ended 30 June 2026 — its all-time high — and reported an operating LOSS of KRW 7.73 trillion for the full year 2023, an operating margin of roughly minus 24%.

Micron reported a GAAP gross margin of 84.6% for the quarter ended 28 May 2026, against 39.8% for the fiscal year ended 28 August 2025.

NVIDIA disclosed that in the quarter ended 26 July 2026 one direct customer represented 16% of total revenue, and that for the first half three direct customers represented 16%, 15% and 13%.

For fiscal year 2026, sales to a single customer represented 28.1% of Super Micro's total net sales.

NVIDIA disclosed that its supply commitments increased from $119 billion at the end of the prior quarter to $279 billion as of 26 July 2026.

Lawrence Berkeley National Laboratory's 2026 'Queued Up' study finds that the median United States generation project reaching commercial operation in 2025 took 61 months from its interconnection request, up from 36 months for projects completed in 2015 and 22 months in 2005.

ERCOT reported tracking approximately 474 GW of large loads seeking interconnection as of June 2026, of which about 90% are data centres — and NERC documents that data-centre demand projections in Texas 'are reduced by almost 50% of their original requested load level to reflect the observed consumption behavior of existing data centers.'

TSMC began high-volume production on its N4 process at its Arizona site in the fourth quarter of 2024, having announced the project in May 2020.

NVIDIA's fiscal 2026 revenue was $215.938 billion, up 65% year on year, while GAAP gross margin FELL 3.9 percentage points to 71.1% from 75.0%. The company attributes the decline to the transition from Hopper HGX systems to full-scale Blackwell datacenter solutions.

TSMC's chief executive said on the company's second-quarter 2026 earnings call that 'our packaging capacity is so tight that now it limits my customers' growth', and separately that developing a new process technology, building the capacity and ramping it 'now takes five to seven years. There are no shortcuts.'

Asked on the same call why TSMC is no longer the most profitable semiconductor manufacturer and what its pricing approach is, its chief executive answered that TSMC is a partner and its customers have to be successful — 'I don't want to squeeze them out from the market' — and that it does not suddenly raise prices by large multiples.

CoreWeave's annual report states that approximately 67% of its 2025 revenue came from its top customer, Microsoft, named in the filing.

NVIDIA disclosed that for fiscal year 2026 sales to one direct customer represented 22% of total revenue and sales to another direct customer represented 14%.

Fabrinet's annual report discloses four customers each at 10% or more of revenue, totalling 57.4%: Cisco 19.9%, NVIDIA 16.3%, Nokia 10.7%, Amazon 10.5%. NVIDIA's share fell from 35.1% in fiscal 2024 and 27.6% in fiscal 2025, while total revenue rose 35.7% — so the falling percentage does not mean falling dollars.

TIME TO ADD SUPPLY, eight layers, each from a primary or agency source. Another data-centre building on a campus that already has power: 12-23 months (TeraWulf's 401 MW lease at its Kentucky campus). A large power transformer, order to delivery: 18-48 months (Wood Mackenzie, quoted in an SEC-filed investor presentation; the US Department of Energy separately reports 36-month lead times commonly quoted and maxima reaching 60). A new advanced-packaging plant, construction start to first manufacturing: 24-42 months (Amkor, Arizona). A new US high-voltage transmission line, CONSTRUCTION ONLY: 24-48 months (US Department of Energy, Grid Deployment Office). A leading-edge logic fab, announcement to volume production: 54 months (TSMC Arizona — announced May 2020, N4 volume production Q4 2024). A new DRAM fab, groundbreaking to first wafer output: 45-63 months (Micron, Boise). A heavy-duty gas turbine plant, order to commercial operation at current backlogs: 36-84 months. A new copper mine, discovery to first production: 16-20 years (IEA, average of major mines that came online 2010-2019).

Microchip Technology's revenue for the fiscal year ended 31 March 2026 was $4.713 billion — up 7% year on year, but 44% BELOW its fiscal-2023 peak of $8.439 billion. Three years into the AI buildout, this semiconductor company's revenue is far below where it stood before it began.