OpenAI's Compute Bottleneck: Supply-Chain Exposure from Nvidia and TSMC to Electricity

No matter how smart its models are, OpenAI’s compute still depends on the global hardware supply chain. Chips, manufacturing, memory, and electricity are each controlled by just a few companies.

This article looks only at where OpenAI gets stuck. For the whole chain, read The AI hardware supply chain; for its data center plans, read What is Stargate?.

Nvidia: a dependence that cannot be removed soon

Nvidia is the core of OpenAI’s training infrastructure. This is its deepest dependence. Beyond hardware, Nvidia’s software ecosystem (CUDA) also ties the development process to its platform.

In 2026, the tie grew tighter:

Shareholder: Nvidia took part in the funding round OpenAI completed in March 2026, investing $30 billion.
Next-generation chips: OpenAI plans to use 3 GW of inference and 2 GW of training compute on Nvidia’s Vera Rubin systems.

At the same time, OpenAI is spreading its risk:

AMD: It signed a 6 GW GPU agreement. The first 1 GW of MI450 is planned for deployment starting in the second half of 2026.
Custom chips: Jalapeño, an inference chip built with Broadcom, is covered in the next section.

There is currently no evidence that custom chips will replace training GPUs.

The custom Jalapeño chip and TSMC manufacturing

OpenAI’s first custom chip is called Jalapeño. Broadcom helped design it. It is made only for inference (answering user questions), not for training models. The two companies aim to deploy 10 GW of custom accelerators, starting in the second half of 2026 and finishing by the end of 2029.

The Wall Street Journal reported in October 2026 that Broadcom is arranging more than $50 billion in financing for OpenAI’s custom chips, with no deal reached yet.

The outside exposure is in manufacturing. Jalapeño reportedly uses TSMC’s 3-nanometer process, so it must compete with Nvidia and other AI chips for TSMC’s advanced manufacturing and packaging capacity.

For the product side, read What products does OpenAI make?.

HBM: memory everyone is fighting for

AI chips need HBM (high-bandwidth memory), and HBM is in short supply. Samsung and SK Hynix have reportedly signed preliminary supply agreements with OpenAI, and Stargate’s demand is rumored to reach up to 900,000 DRAM wafers per month.

Jalapeño also uses HBM, reportedly HBM4 supplied by Samsung. The more chips OpenAI builds, the bigger an HBM buyer it becomes, competing directly with Nvidia for the same supply. For the advanced-packaging bottleneck, also read about CoWoS.

Electricity: the most underrated bottleneck

Buying chips is not enough. Data centers also need enough power. Stargate has more than 9 GW of planned capacity, but research group Epoch AI estimated in April 2026 that only about 0.3 GW at Abilene, Texas was running at that time, while the other sites were still under construction. OpenAI has not published real-time capacity.

New sites also take years. In August 2026, OpenAI signed a deal for about 8 GW of capacity at the PORTS-Pike campus in Ohio, and Nvidia is providing credit support for building the first 4.25 GW. The first 800 MW is not expected until 2028.

Power prices and regulation are limits too. According to CNBC, OpenAI paused Stargate UK in April 2026 over energy costs and regulatory uncertainty.

The Chinese market

According to ChinaTalk, mainland China and Hong Kong are not among OpenAI’s supported regions, and OpenAI also blocks API traffic from unsupported regions. OpenAI is therefore absent from the Chinese market, which local models such as DeepSeek and Alibaba’s Qwen fill instead.

Penchan’s summary

OpenAI leads in model capability, but in hardware it is pulled along the same chain as every rival: Nvidia’s chips, TSMC’s manufacturing, HBM from the memory makers, and power that may or may not keep up.

AMD, Jalapeño, and its own data centers are three steps toward independence, but each takes years. In the near term, these supply-chain links still set the ceiling on how fast OpenAI can expand.

FAQ

Does OpenAI rely entirely on Nvidia chips?

Nvidia remains central to OpenAI’s training infrastructure, and in March 2026 it invested $30 billion in OpenAI. OpenAI signed a 6 GW GPU agreement with AMD to spread its capacity, and Jalapeño, its custom chip built with Broadcom, is made for inference. There is currently no evidence that Jalapeño will replace training GPUs.

Who manufactures OpenAI's custom chip?

Broadcom helped design the first-generation Jalapeño inference chip. It reportedly uses TSMC’s 3-nanometer process, with Samsung supplying HBM4 memory. The companies have not officially disclosed every specification or supplier.

Why is data center electricity a problem?

Stargate has more than 9 GW of planned capacity, but Epoch AI estimated in April 2026 that only about 0.3 GW was operating at that time, at Abilene, Texas. OpenAI has not published real-time capacity. According to CNBC, Stargate UK was also paused in April 2026 over energy costs and regulatory uncertainty. Whether power can keep up is a key bottleneck for expansion.

Is OpenAI available in China?

Not officially. According to ChinaTalk, mainland China and Hong Kong are not among OpenAI’s supported regions, and OpenAI also blocks API traffic from unsupported regions. Local models such as DeepSeek and Alibaba’s Qwen fill this market instead.