OpenAI's massive chip bet highlight aggressive strategy

OpenAI has committed to acquiring 26 gigawatts of advanced data processors from Nvidia, AMD, and Broadcom in less than a month — a move that Gil Luria of D.A. Davidson says will require “hundreds of billions of dollars.”

The San Francisco-based AI firm behind ChatGPT, which does not expect to be profitable until 2029, is forecasting billions in losses this year despite generating about US$13 billion in revenue.

The chip orders would use as much electricity as 20 standard nuclear reactors, according to industry estimates.

Nvidia plans to invest up to US$100 billion in OpenAI over several years, with OpenAI using the funds to buy Nvidia’s chips and Nvidia taking a stake in the startup.

AMD has also offered OpenAI the option to acquire equity in the company, an unusual move in financial circles.

Experts note that OpenAI’s aggressive spending contrasts with rivals like Google and Meta, which fund AI through profitable businesses.

Some analysts warn this could fuel a speculative bubble, while others see lasting demand for AI technology.

🔗 Source: Agence France-Presse

🧠 Food for thought

Implications, context, and why it matters.

Chip-on-Wafer-on-Substrate (CoWoS) capacity expansion cannot keep pace with OpenAI’s 26 GW chip commitment timeline

Buying over 10 million data processors would strain limited capacity. Global CoWoS plus similar packaging (the step that connects multiple chips and memory into one module) is projected at about 1.313 million 12-inch wafers in 2026 1. Nvidia alone is predicted to book roughly 595,000 CoWoS wafers that year 2.
Taiwan Semiconductor Manufacturing Company (TSMC), the world’s largest contract chipmaker, could reach 88,000 to 93,000 CoWoS wafers per month by late 2026 with 26 to 33 percent yearly growth 1 3.
TSMC’s 3nm and 5nm production lines (advanced manufacturing nodes for leading-edge chips) are expected to be 100% booked in the first half of 2026 4. Most new 3nm output already goes to Apple and Qualcomm. MediaTek, Nvidia, and AMD take much of the rest. That leaves little room for extra orders tied to OpenAI’s buildout.

GPU financing and collateralization infrastructure remains underdeveloped for chip-backed lending at scale

High-end accelerators (GPUs and specialized AI chips) are customized and often tied to specific platforms, which complicates repossession or resale after a default.
Lenders can build valuation methods, custody programs, insurance for AI chip inventories. TSMC is expected to raise 3nm and 5nm wafer prices by 5 to 10 percent in 2026, with CoWoS packaging up about 15 to 20 percent 4.
Nvidia invests in OpenAI while OpenAI buys Nvidia systems, so chip values hinge partly on the health of the same firms supplying them 5.