The AI inference chip market saw two closely matched blockbuster moves between late 2025 and April 2026. Nvidia agreed to acquire Groq for $20 billion, while OpenAI said it would buy more than $20 billion worth of chips from Cerebras. On the same day OpenAI’s expanded deal surfaced, Cerebras filed for a Nasdaq IPO seeking a $35 billion valuation and aiming to raise $3 billion. Taken together, the transactions point to one shift: competition in AI compute is moving from training toward inference.
Inference is taking a larger share of AI spending
According to the source material, global AI compute spending in 2023 was still led by training workloads, with inference playing a secondary role. That mix is changing fast. Inference accounted for 50% of all AI compute spending in 2025, and the figure is expected to reach two-thirds in 2026. Lenovo Chairman and CEO Yang Yuanqing said at CES 2026 that the spending mix could flip from “80% training and 20% inference” to “20% training and 80% inference.”
The reason is straightforward. Training is usually a one-time or periodic expense, while inference is tied to ongoing user demand. Once a model is built, every chatbot response, coding request, or other live query creates another inference call, and those costs accumulate at scale.
Why Nvidia moved on Groq
Nvidia’s H100 and H200 GPUs were built with training in mind, where raw compute throughput matters most. The source argues that inference bottlenecks often come from memory bandwidth instead. For each user request, model weights must be moved from memory into compute units before an answer is generated. In high-volume deployments, that transfer delay becomes a practical performance limit.
Groq and Cerebras both focus on SRAM-based architectures optimized for inference. On December 24, 2025, Nvidia announced its acquisition of Groq, its largest deal since the $7 billion Mellanox purchase in 2019. The source describes Groq’s LPU as the fastest inference chip service in public benchmarks at the time, and says the deal also brought in founder Jonathan Ross and several senior engineers with Google TPU backgrounds.
Nvidia has not stood still on product development. The source says its Blackwell B200 architecture delivers 4x the inference performance of the H100 and is already being deployed at scale. Even so, the company chose to buy technology rather than rely only on internal upgrades.
OpenAI deepens its ties with Cerebras
OpenAI’s strategy looks different. Instead of buying a company outright, it is building a tighter supplier relationship. In January 2026, OpenAI signed a $10 billion, three-year compute procurement deal with Cerebras. By April 17, the disclosed terms had changed materially, with the chip purchase commitment rising to more than $20 billion.
The updated arrangement went beyond procurement. The source says OpenAI will receive Cerebras warrants, allowing its stake to rise to as much as 10% of the company’s total equity as purchase volume increases. OpenAI will also provide $1 billion in data center construction funding. That combination suggests a supplier-building effort, not just a purchase order.
At the same time, the source says OpenAI is working with Broadcom on its own ASIC and expects production by the end of 2026. One track secures external inference capacity; the other aims at in-house chip development.
Cerebras IPO brings valuation gains and concentration risk
Cerebras filed its Nasdaq IPO paperwork on April 17 with a target valuation of $35 billion. That compares with a valuation of $8.1 billion in September 2025 and $23 billion in its February 2026 funding round. The IPO target implies another premium of roughly 52% over the February level.
This is not its first attempt to list. The source says Cerebras withdrew its 2024 IPO effort after the Committee on Foreign Investment in the United States intervened on national security grounds because key customer G42 contributed 83% to 97% of annual revenue. G42 is now gone from the shareholder list, and OpenAI has become the major customer drawing attention. That leaves investors with a familiar question: whether customer concentration has really improved or merely shifted to a different name.
Cerebras enters the market with clear upside and equally clear uncertainty. A growing inference market could create substantial room even for a smaller share player. But OpenAI’s own ASIC effort could become a direct substitute if it reaches volume production on schedule. The fight over who controls inference capacity, chip design, and the surrounding ecosystem is no longer abstract.