No matter how smart the model is, OpenAI’s compute still can’t get around that global hardware supply chain. Even the most advanced AI ultimately hinges on the chips, memory, and power of a handful of suppliers.
This is a lightweight rundown of OpenAI’s specific supply-chain exposure points. For the full picture of how the entire supply chain works, it’s best to read this alongside The AI hardware supply chain, end to end; to understand its compute footprint and partnership strategy, see OpenAI’s compute map. This piece focuses only on “which links OpenAI is stuck on.”
Nvidia: a dependence it can’t shake in the near term
OpenAI trains its models almost entirely on Nvidia GPUs, and this is its deepest dependence, the hardest to break out of in the short run. Beyond the hardware itself, Nvidia’s software ecosystem (CUDA) acts more like a layer of glue, binding the entire development workflow to its platform.
OpenAI is diversifying too: alongside buying from Nvidia, it has been in talks with AMD and is developing its own chips. But the industry consensus is that in-house chips will initially focus on inference, while the most resource-hungry part, training, can’t get away from Nvidia in the near term. In other words, this chokepoint can be partly dismantled, but not quickly.
In-house chips and TSMC’s process nodes
A key step in OpenAI’s effort to take its lifeline back into its own hands is partnering with Broadcom to develop its own AI chips. But this path has two external exposure points.
The first is the process node. According to supply-chain reports, this chip is fabricated by TSMC on its most advanced process, meaning OpenAI has to compete with Apple and Nvidia for the most capacity-constrained nodes at TSMC. The fight over wafer allocation is fierce, and there have also been reports that mass production of the next, more advanced process node is being delayed. OpenAI has not officially disclosed the exact specs, so outside figures should be treated with caution.
The second is funding. The early production of in-house chips requires enormous upfront capital, and OpenAI is simultaneously carrying hundreds of billions of dollars in compute and chip commitments, which is why outsiders are watching its cash-flow pressure. Some reports dwell on the details of related financing arrangements, but neither OpenAI nor Broadcom has publicly confirmed them, so treat them with caution. It points to one reality: what in-house chips burn through is not just technology, but enormous capital.
HBM: the high-bandwidth memory everyone is fighting over
AI chips can’t do without HBM (high-bandwidth memory), and HBM is currently in severe shortage. OpenAI’s Stargate project has an enormous appetite, its memory needs alone are reported to be set to consume a fairly high share of global DRAM output.
This area is dominated by three companies: SK Hynix, Samsung, and Micron. Some reports say Samsung has won the HBM4 supply for OpenAI’s in-house chips, but accounts differ on “whether it’s exclusive,” so treat that with caution. What is clear is that the HBM shortage is expected to persist for some time, and that’s a strain on any company looking to scale up compute aggressively. For the bottleneck in advanced packaging, see CoWoS.
Power: the most underestimated bottleneck
Beyond chips, the most easily overlooked exposure point is power. OpenAI’s Stargate targets gigawatt-scale compute, with power consumption equivalent to that of a city.
There are two gaps worth watching here. The first is the energy mix: supply currently runs mainly on natural gas, leaving a clear gap with the green-power commitments OpenAI publicly emphasizes. The second is social friction: some US communities oppose large data centers moving in because of environmental and grid burdens. OpenAI has even publicly called for the US to dramatically increase its power generation, treating electricity as a strategic asset in the AI race. Whether power can keep up, and at what cost, is the real bottleneck on its pace of expansion.
The China market and export controls
Finally, there’s the geopolitical thread, stated here from a neutral angle.
OpenAI’s service is available in more than a hundred countries, but mainland China, Hong Kong, and Macau are not officially open, and it has actively blocked access from those regions. Layer on the US-China chip export controls (whose scope has in recent years even extended to renting cloud compute from outside the country), and OpenAI is essentially absent from the China market.
This absence brings two external variables. The first is that the space in this market is filled by homegrown models like DeepSeek and Alibaba’s Qwen; the second is that OpenAI therefore has a harder time obtaining data on Chinese-language usage scenarios, leaving it relatively limited on Chinese-language tasks. This is also seen as one of the reasons it later released open-weight models, hoping to hold on to part of the developer ecosystem.
Penchan’s take
Lining up these chokepoints reveals a contrast: OpenAI is the front-runner in model capability, but on the hardware supply chain it is, like every peer, pulled along by the same chain, Nvidia’s chips, TSMC’s process nodes, the big three’s memory, and whether power can keep up.
It is inching toward “self-sufficiency” through in-house chips, multiple suppliers, and self-built power, but that’s a long and cash-burning road. In the near term, these supply-chain exposure points remain the ceiling on its speed of expansion. Once you grasp this layer, it’s easier to see why the AI giants have in recent years poured so much effort into the chips and power plants that sit so far from the end user.