> ## Content Index > Fetch the complete content index at: https://www.sotkn.com/llms.txt > Use this file to discover other available public pages before exploring further. # Who gets paid when the U.S. AI chip is three cheques? - URL: https://www.sotkn.com/markets/us/us-hardware-acceleration-october-2026/ - Published: 2026-10-04T02:56:40.000Z - Updated: 2026-10-04T02:56:51.000Z - Description: In the United States, the accelerator is not one market. Nvidia has already been paid for chips it shipped, CoreWeave only for the hours it delivered, and Amazon is still in talks on about $8 billion of installed chips. - Author: Sean - Tags: #market-us In the United States, the accelerator is not one market. An accelerator is the chip that runs AI training, and the later answers, faster than a normal processor. Nvidia has already been paid for chips it shipped. CoreWeave has been paid only for the hours it delivered. Amazon is still in talks to move about $8 billion of installed Nvidia chips into a side company and lease them back. Ready is not paid. ## The sale Nvidia sells the graphics chip, the GPU, that most of these halls still run. In the quarter ended 26 July 2026, Nvidia's data-center revenue was $89.0 billion. That was up 117% from a year earlier. Nvidia said so on 26 August 2026\. Gross margin was 75%. That money is from selling the platform. It is not rent for hours. The $89.0 billion is the whole data-center line. It is not one chip. Nvidia's guide for the next quarter, $108.0 billion, is a forecast. It is not cash. Broadcom sells a different chip. An ASIC is a chip built for one customer's job, not a chip sold to everyone from a catalogue. On 2 September 2026 Broadcom said AI semiconductor revenue was $16.7 billion in the quarter ended 2 August 2026\. That was up 221% from a year earlier. The same sentence covers custom accelerators and the networking that ties them together. Do not call all $16.7 billion an ASIC sale. The $21.7 billion Broadcom expects in the next quarter is a guide, not a receipt. AMD sells server processors and its own AI GPU, Instinct, in one reported bucket. Data-center revenue was $6.7 billion in the quarter ended 27 June 2026, up 107% from a year earlier. AMD said both products drove that number. It did not split them. In the same release, AMD said Anthropic would deploy up to 2 gigawatts of a later Instinct chip in AMD racks. "Up to" is a plan. It is not an invoice. Amazon also sells chips of its own. On 30 July 2026 Andy Jassy said Amazon's AI business and its chips business had each passed a run rate of more than $25 billion. That is Amazon as a seller. It is a different fact from the lease talk below. ## The rent CoreWeave does not sell the chip. It rents the hours on chips it has put in halls. Revenue was $2.6 billion in the quarter ended 30 June 2026, against $1.2 billion a year earlier. CoreWeave reported that on 11 August 2026\. It still lost $626 million in the quarter. Interest was $640 million. The operating loss was $49 million. The hours can be real, and the profit can still be missing, because the buildings and the debt sit on CoreWeave, not on Nvidia. CoreWeave also said revenue backlog was about $104 billion on 30 June 2026\. That figure leaves out more than $25 billion of newer promises from early in the next quarter. Backlog is work the company hopes to deliver if the halls are ready. It is not cash. ## The talk On 2 October 2026 the Financial Times reported that Amazon is seeking to shift about $8 billion of Nvidia Grace Blackwell chips into a special-purpose vehicle and lease them back. The chips are already in more than a dozen U.S. data centers in five states, including Nevada and Virginia. The talks were still going. Amazon declined to comment. Nothing has closed. That case is already on the desk: [Ask isn't closed: who gets paid when Amazon seeks an $8B chip leaseback](https://www.sotkn.com/markets/us/us-amazon-nvidia-leaseback-october-2026/). This note does not retell it. It only uses the line. Installed is not financed. On 30 July 2026 Jassy said Amazon's capital spending this year would reach $220 billion, lifted by the price of memory. CNBC reported the remark that day. That is Amazon spending its own money on buildings, chips, and the rest of the company. It is not a cheque from a customer. It is not a combined total for Alphabet, Meta, and Microsoft. ## What the other headlines are not On 29 September 2026 Bain wrote that hardware and semiconductor stocks grew at a 24% compound annual rate from 2020 to 2026, against 6% for software stocks. Bain said software had been the fastest-growing tech sector for nearly two decades, and that hardware is ahead for the first time in years. That is share prices. It is not chip sales. On 18 November 2025 Deloitte predicted that inference, the work of answering with a trained model, would be about two-thirds of all AI compute in 2026\. A year earlier it had been about half, Deloitte said, and a third in 2023\. That is a forecast of compute. It is not a count of dollars, of maths operations, or of GPU hours. Deloitte has not published a measured split for 2026. On 9 October 2025 TrendForce expected cloud companies' own ASICs to grow 44.6% in 2026, against 16.1% for GPUs. The count was AI server shipments, not revenue. A forecast from last autumn is not a 2026 result. On 12 February 2026 Samsung said it had started mass production of HBM4 and had shipped it to customers. HBM is the stacked memory that sits next to the accelerator. CoWoS is TSMC's way of packing that chip and that memory on one base. Samsung's ship date is a product fact. It is not a price. It is not a claim that packing stays short through 2027\. Bain wrote on 29 September 2026 that TSMC plans to raise that packing capacity by 80% and has still said it cannot meet demand. Bain gave no end-date of 2027. On 28 January 2026 Google said its LiteRT software can use the phone neural chip, the NPU, on MediaTek and Qualcomm parts. A phone chip does not pay for a U.S. data hall. Three clocks, and they do not add up. Nvidia's $89.0 billion is a sale that cleared. CoreWeave's $2.6 billion is rent that cleared, and its $104 billion backlog has not. Amazon's $8 billion move is still a talk. ## Sources - Nvidia, quarterly results for the quarter ended 26 July 2026, filed 26 August 2026\. Data-center revenue $89.0 billion. [https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/q2fy27pr.htm](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/q2fy27pr.htm?ref=sotkn.com) - Broadcom, results for the quarter ended 2 August 2026, 2 September 2026\. AI semiconductor revenue $16.7 billion. [https://www.sec.gov/Archives/edgar/data/1730168/000173016826000076/avgo-08022026x8kxex99.htm](https://www.sec.gov/Archives/edgar/data/1730168/000173016826000076/avgo-08022026x8kxex99.htm?ref=sotkn.com) - AMD, results for the quarter ended 27 June 2026, 4 August 2026\. Data-center revenue $6.7 billion. [https://www.sec.gov/Archives/edgar/data/2488/000000248826000121/q22026991.htm](https://www.sec.gov/Archives/edgar/data/2488/000000248826000121/q22026991.htm?ref=sotkn.com) - CoreWeave, results for the quarter ended 30 June 2026, 11 August 2026\. Revenue $2.575 billion. Backlog about $104 billion. [https://www.sec.gov/Archives/edgar/data/1769628/000176962826000362/coreweave2q26earningspress.htm](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000362/coreweave2q26earningspress.htm?ref=sotkn.com) - Amazon, second-quarter results, 30 July 2026\. AI and chips run rates. [https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Second-Quarter-Results/](https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Second-Quarter-Results/?ref=sotkn.com) - CNBC, 30 July 2026\. Jassy on $220 billion of 2026 capital spending. [https://www.cnbc.com/2026/07/30/amazon-amzn-q2-earnings-report-2026.html](https://www.cnbc.com/2026/07/30/amazon-amzn-q2-earnings-report-2026.html?ref=sotkn.com) - Financial Times, Michelle Chan, 2 October 2026, as carried by the Financial Post. About $8 billion Grace Blackwell talks. [https://financialpost.com/technology/amazon-offload-8bn-nvidia-chips-investors](https://financialpost.com/technology/amazon-offload-8bn-nvidia-chips-investors?ref=sotkn.com) - Second Order, U.S. Markets, 4 October 2026\. The leaseback note this draft does not replace. - Bain & Company, Global Technology Report 2026, 29 September 2026\. Stock CAGR, not chip revenue. [https://www.bain.com/about/media-center/press-releases/2026/global-ai-market-could-hit-$6-trillion-annually-by-2031-through-unlocking-value-and-innovation--bain--cos-7th-global-technology-report/](https://www.bain.com/about/media-center/press-releases/2026/global-ai-market-could-hit-$6-trillion-annually-by-2031-through-unlocking-value-and-innovation--bain--cos-7th-global-technology-report/?ref=sotkn.com) and [https://www.bain.com/insights/hardware-strikes-back-in-the-ai-era-technology-report-2026/](https://www.bain.com/insights/hardware-strikes-back-in-the-ai-era-technology-report-2026/?ref=sotkn.com) - Deloitte Insights, 18 November 2025\. Forecast that inference would be about two-thirds of AI compute in 2026\. [https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2026/compute-power-ai.html](https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2026/compute-power-ai.html?ref=sotkn.com) - TrendForce, 9 October 2025\. Forecast of 2026 AI server shipment growth, ASIC 44.6% versus GPU 16.1%. [https://www.trendforce.com/insights/nvidia-scale-up-technology](https://www.trendforce.com/insights/nvidia-scale-up-technology?ref=sotkn.com) - Samsung Newsroom, 12 February 2026\. HBM4 mass production and commercial shipments. [https://news.samsung.com/global/samsung-ships-industry-first-commercial-hbm4-with-ultimate-performance-for-ai-computing](https://news.samsung.com/global/samsung-ships-industry-first-commercial-hbm4-with-ultimate-performance-for-ai-computing?ref=sotkn.com) - Google Developers Blog, 28 January 2026\. LiteRT NPU support. [https://developers.googleblog.com/litert-the-universal-framework-for-on-device-ai/](https://developers.googleblog.com/litert-the-universal-framework-for-on-device-ai/?ref=sotkn.com)