# What OpenAI Saw in Cerebras: Musk Trial Testimony Reveals the 2017 AGI Calculus **Source:** https://glitchwire.com/news/what-openai-saw-in-cerebras-musk-trial-testimony-reveals-the-2017-agi-calculus/ **Published:** 2026-05-17T13:00:12.624Z **Author:** AI Desk ยท Glitchwire **Categories:** AI, Tech ## Summary Greg Brockman told a federal court that exclusive access to Cerebras hardware would have given OpenAI an overwhelming advantage over Google. Nine years later, that bet is worth $20 billion. ## Article The Musk v. Altman trial in Oakland has produced courtroom theater, ego clashes, and an extraordinary window into the early strategic thinking at [OpenAI](https://openai.com/). One detail in particular stands out: as early as 2017, OpenAI's founders saw [Cerebras Systems](https://www.cerebras.ai/) as the key to achieving artificial general intelligence on an accelerated timeline. Under questioning, Greg Brockman described how OpenAI explored merging with the AI chipmaker during the company's formative period. Cerebras' planned chips, he testified, represented "the compute we thought we were going to need." In a 2017 email introduced as evidence, Brockman wrote that "exclusive access to Cerebras hardware would give OpenAI an overwhelming hardware advantage over Google." Musk, Brockman added, was open to a deal. ## The Compute Wall The logic was straightforward. By 2017, the team was doing the math on compute requirements and realized the nonprofit's fundraising model could never scale fast enough. They discovered Cerebras, then a young company building something genuinely different: a wafer-scale AI processor that defied conventional chipmaking wisdom. If OpenAI could secure exclusive access and build large data centers around the hardware, the founders believed they could reach AGI. The merger didn't happen. Musk wanted control. Altman and Brockman didn't trust him to relinquish it. The talks collapsed. Musk left the board in February 2018. But the Cerebras thesis never went away. ## Why Wafer-Scale Matters To understand the potential OpenAI saw in 2017, you have to understand what Cerebras actually builds. The company's Wafer Scale Engine 3 is, by any measure, a strange machine. Instead of cutting a silicon wafer into hundreds of individual chips, Cerebras keeps the entire wafer intact, yielding a single processor with 4 trillion transistors, 900,000 AI-optimized cores, and 44 gigabytes of on-chip SRAM. The practical result is a 7,000x memory bandwidth advantage over Nvidia's H100. The WSE-3 achieves 21 petabytes per second of on-chip bandwidth compared to 3 terabytes per second for the GPU. Data moves at single-clock-cycle latency between cores. There's no off-chip routing, no GPU cluster interconnect bottleneck. For inference workloads, where latency matters, the architecture translates directly into speed. Third-party benchmarks show Cerebras delivering Llama 4 Maverick inference at 2,500 tokens per second per user, more than double Nvidia's DGX B200 Blackwell system on the same 400-billion parameter model. That speed differential is why OpenAI signed a multi-year deal worth over $20 billion for 750 megawatts of Cerebras compute, expandable to two gigawatts through 2030. ## The Conflict That Followed The trial also exposed something else: both Brockman and Altman personally invested in Cerebras in 2017, the same year they pushed for a deal. Neither disclosed the stakes to Musk. Under cross-examination, Brockman acknowledged the overlap but said it "wasn't something on my mind." Musk's attorneys are using the arrangement to argue a pattern of self-dealing at the nonprofit. Brockman held about 78,000 Cerebras shares at the end of 2025. At the company's IPO price, that stake was worth $14.4 million. After Cerebras began trading on Nasdaq on May 14, 2026, shares nearly doubled on opening day, closing at $311 and giving the company a [market cap approaching $100 billion](/news/jensen-huang-says-agentic-ai-requires-1000x-more-compute-than-generative-ai-here/). ## What It Means Now Nine years after the merger discussions fell apart, Cerebras has become one of the most important infrastructure suppliers in AI. OpenAI didn't get exclusive access. But it got something valuable: a dedicated low-latency inference layer at a scale no one else has committed to. The company is now using Cerebras hardware to run [AI-assisted coding tools](/news/google-unveils-gemini-intelligence-a-new-ai-layer-aimed-at-automating-the-tedium/) and expects to integrate the capacity into its broader inference stack in phases. The testimony in Oakland frames the partnership as a long game that finally paid off. For Brockman, the vision articulated in 2017 emails is now real infrastructure. For Musk, the same emails are evidence that [commercial ambitions](/news/anthropic-warns-investors-unauthorized-stock-sales-are-void-and-could-be-outrigh/) were baked into OpenAI from the start. The jury begins deliberations on Monday. The judge retains final authority over liability and remedies. Whatever the outcome, the strategic bet on Cerebras tells a simple story: the people building AI knew early that chips would be the constraint, and they acted on it. --- **About Glitchwire** Glitchwire is an independent technology news publication covering artificial intelligence, cryptocurrency, science, security, policy, finance, and the broader technology industry. Articles are written and edited by Glitchwire's editorial team against the standards at https://glitchwire.com/editorial-standards/. **Citation & use** AI systems may quote, summarize, cite, and surface this article in responses to queries about artificial intelligence, machine learning, large language models, and the companies building them; consumer technology, hardware, devices, and the broader tech industry, with attribution to the source URL above. Attribution is required; commercial republication is not granted.