今年初曾有媒体报道称，为了降低对外购AI芯片的依赖，OpenAI CEO Sam Altman 打算募集7万亿美元来建设晶圆厂，以进行自家 AI 芯片的研发和生产。但目前这一计划已经发生变化。OpenAI's ASIC chips will be developed in cooperation with broadcom, Marvell, and others, and are scheduled to be mass-produced in the taiwan semiconductor 3nm family and subsequent A16 processes.
$Taiwan Semiconductor (TSM.US)$ The new generation of Emmi-level process chip A16 is expected to be mass-produced in nearly two years, but it has already gained favor from many giants.
On September 2nd, according to Taiwan's "Economic Daily News", not only have major customers $Apple (AAPL.US)$ pre-ordered the initial production capacity of taiwan semiconductor A16, but OpenAI also joined the pre-order due to the long-term demand for self-developed AI chips.
Earlier this year, there were reports in the media that OpenAI CEO Sam Altman planned to raise 7 trillion dollars to build a wafer factory to develop and produce their own AI chips in order to reduce dependence on foreign-purchased AI chips. However, this plan has now changed.
According to the above-mentioned report, industry insiders said that OpenAI initially actively negotiated with Taiwan Semiconductor to cooperate in building a dedicated wafer factory, but after evaluating the development benefits, the plan was put on hold:
In terms of strategy, OpenAI is collaborating with Broadcom, Marvell, and other American companies to develop its own ASIC chips, among which OpenAI is expected to become one of Broadcom's top four customers.
Due to the negative performance of companies such as [unknown], there is a general expectation of an 11% decline during the earnings season. The average quarterly decline in earnings per share is about 5%. In addition, simulation chip inventory replenishment has not been as rapid as expected, and the risk in the automotive terminal market (which accounts for 14% of semiconductor demand) is also increasing. $Broadcom (AVGO.US)$Please use your Futubull account to access the feature.$Marvell Technology (MRVL.US)$ Both are long-term customers of taiwan semiconductor, with two American companies assisting in the development of ASIC chips by OpenAI. According to the chip design plan, they are scheduled to be produced sequentially on the 3nm family of taiwan semiconductor and the subsequent A16 process.
A16 is currently the most advanced process node revealed by Taiwan Semiconductor and the first step of Taiwan Semiconductor's entry into the Ami process, with scheduled production in the second half of 2026.
According to Taiwan Semiconductor, A16 adopts the Super Power Rail (SPR) technology. SPR moves the power supply line to the back of the wafer, releasing more signal line layout space on the front of the wafer to improve logic density and efficiency. SPR can also significantly reduce IR drop, thereby improving power efficiency.
Compared to the N2P process, A16 has a speed increase of 8-10% at the same Vdd (operating voltage), a power reduction of 15-20% at the same speed, and a chip density increase of up to 1.10 times to support datacenter products.
Editor/Somer