OpenAI has confirmed a alliance with Broadcom to design and manufacture its own artificial intelligence chips, a move that seeks to secure computational muscle and adjust costs in the midst of the expansion of its generative models.
As both companies have explained, OpenAI will take over the design of the accelerators and Broadcom will be responsible for development, manufacturing and deployment, with a schedule that points to the second half of 2026 and an infrastructure sized for around 10 gigawatts of electricity.
What the agreement includes and how the functions are distributed
The plan contemplates systems based on the Broadcom's Ethernet stack and network technology, integrated into both OpenAI's own data centers and external partner infrastructures, with networking, memory, and computing integration tailored to your workloads.
The company maintains that with specific hardware for your models You'll be able to gain efficiency, accelerate training and inference, and ultimately deliver faster, more cost-effective models to businesses and users.
Sam Altman has described the move as key to building the necessary infrastructure that will bring the benefits of AI to scale, while Broadcom CEO Hock Tan highlights that OpenAI is driving some of the more advanced border models.
Timing, magnitude, and consumption: the 10 GW bar
The deployment would start at the end of 2026 with a aggregate capacity equivalent to 10 GW, a figure that, for comparative purposes, far exceeds the electricity generated by the Hoover dam and could supply more than eight million homes U.S..
This leap in scale puts the electricity suppliers and the data center supply chain, which will have to meet demands for power, cooling, and components at a pace that is unusual even for the technology industry.
Industry references estimate that a 1 GW data center could be around 50.000 billion dollars of investment, of which about 35.000 million are typically allocated to chips and accelerators; extrapolated, the financial effort for 10 GW would be enormous.
Less dependence on Nvidia and side deals
The alliance with Broadcom fits into the strategy of diversify suppliers and reduce dependence on Nvidia, whose supply and prices have strained the market. In parallel, OpenAI maintains a contract with AMD for 6 GW and collaborates with Nvidia in another 10 GW infrastructure, which places the total commitment around 26 GW of capacity compromised.
In addition to these partners, the company has advanced agreements with Oracle for cloud infrastructure and conversations with Samsung and SK Hynix linked to the supply of memory and critical components for their future systems.
The economic terms of the agreement with Broadcom have not been made public, and the funding scheme has not been detailed either. In information released by financial media, AMD has reportedly granted OpenAI a warrant on up to 160 million shares subject to capacity deployment conditions, in addition to the supply of its next series MI450.
Broadcom's market and position in the ecosystem
After the announcement, Broadcom rebounded nearly 10% on the stock market., in line with the movements already seen at AMD and Oracle following their respective pacts with OpenAI. For Broadcom, this is one of its most relevant contracts and reinforces its transit from networks and telecommunications to the large-scale AI infrastructure.
The company already supplies XPU accelerators and solutions to giants like Google, Meta or ByteDance, and has seen its revenues and capitalization grow strongly since 2022 in the heat of the Generative AI.
Operational challenges: energy, costs and efficiency
The AI ​​business is still far from being fully profitable for many players, and build data centers of this caliber It involves complex permits, massive investment, and energy consumption that can strain local and regional grids.
OpenAI is confident that the custom chips help contain computing costs and improve energy efficiency per task, reducing bottlenecks compared to the purchase of commercial processors in a market increasingly competitive.
The joint venture between OpenAI and Broadcom marks a turning point in the race for AI infrastructure: a program of its own accelerators with a 2026 horizon, a 10 GW energy threshold, and parallel alliances with AMD, Nvidia, and other partners that seek to secure supply, reduce costs, and scale capabilities without relying on a single supplier.
