Cerebras surges 15% as OpenAI locks in 750MW of AI compute

Cerebras shares jumped 15.3% after Wedbush tied the chip designer to OpenAI's GPT-5.6 Sol Ultrafast tier, backed by a 750-megawatt commitment.

Cerebras shares surged 15.3 percent Monday after a Wedbush note cast the chip designer as the exclusive compute backbone for OpenAI's GPT-5.6 Sol Ultrafast mode, delivering 750 tokens per second — 14 times faster than standard.

"The market is realizing that speed is not a benchmark item. Speed changes user engagement, it changes agentic performance, and it changes AI productivity. Fast inference unlocks new applications and new markets," Andrew Feldman, chief executive of Cerebras, said on the company's Q2 call.

The infrastructure commitment behind that tier is what changes the math. OpenAI has committed to purchasing 750 megawatts of computing capacity from Cerebras by 2028, underpinning the $25.4 billion in remaining performance obligations Cerebras reported at quarter-end. Q2 core revenue came in at $209.9 million, up 103 percent year over year, with core cloud and services revenue up 287 percent. Cerebras raised full-year 2026 core revenue guidance to $880 million to $890 million and reiterated plans to more than triple core revenues in 2027. The headline revenue figure of $180.11 million missed the $193.55 million consensus, but Wedbush and Craig-Hallum maintained bullish ratings after that report.

Monday's move is also a reassessment of last Wednesday's selloff, when shares tumbled about 17 percent on the revenue miss, leaving the stock well below its 52-week high of $386.34. Analyst consensus stands at $291.64 with 10 buy and 1 hold ratings. The next event is close: Cerebras plans to unveil its fourth-generation system at the Supernova Conference the week of Aug. 19, alongside targets to double speed each year and increase throughput by more than 20 times.

Wafer-Scale Engine Breaks the Memory Bandwidth Bottleneck

The speed advantage traces to Cerebras's wafer-scale architecture, which differs from the GPU clusters that dominate AI inference. Traditional GPUs run autoregressive decoding by shuttling model weights between off-chip HBM and compute cores, a process limited by memory bandwidth and cross-chip communication latency. Cerebras's WSE-3 integrates 4 trillion transistors, 125 petaflops of AI compute, and up to 44 gigabytes of on-chip SRAM, keeping model parameters resident on the die. For models larger than a single chip, a pipelined-across-wafers mechanism distributes each network layer across multiple wafers, with parameters held in each chip's SRAM.

The result, per Cerebras's blog, is that GPT-5.6 Sol completed all 2,500 questions on the Humanity's Last Exam benchmark in 11 hours and 11 minutes, versus 78 hours and 27 minutes for Anthropic's Claude Fable 5 — nearly seven times faster at similar accuracy. On GDP-Val, a benchmark measuring economically valuable knowledge work, Ultrafast delivered a 5.6 times end-to-end speedup without quality loss. OpenAI said the tier is 11 times faster than Fable 5 and five times faster than Opus 4.8 in Fast mode.

Intel and AMD Move on 13F Filings and a Cerebras Tie-Up

Intel rose about 1.4 percent Monday, a modest move following Q2 results filed July 23 showing revenue of $16.13 billion, up 25.4 percent year over year, with Data Center and AI revenue up 59 percent to $6.26 billion. The bigger story is the Aug. 14 13F cycle: NVIDIA disclosed 214,776,632 Intel shares valued at $29.99 billion, equal to 47.27 percent of its 13F portfolio, and SoftBank Group disclosed 86,956,522 shares valued at $12.14 billion, equal to 66.81 percent of its portfolio. Coatue Management opened a new position of 12,084,027 shares valued at $1.69 billion. These are point-in-time disclosures as of June 30, filed Aug. 14.

AMD traded essentially flat to slightly lower Monday. The company posted Q2 revenue of $11.54 billion, up 50.1 percent year over year, with Data Center revenue up 107 percent to $6.72 billion. AMD is also a Cerebras partner: the two are combining AMD Helios racks for prefill with Cerebras CS systems for decode, a configuration that maintains Cerebras speed while increasing throughput by five times, with availability in Q4 2026. Tiger Global Management opened a new position of 674,727 AMD shares valued at $391.96 million, and Duquesne Family Office opened a new position of 72,900 shares valued at $42.35 million.

Cerebras drew its own high-profile disclosure on the Aug. 14 cycle. Atreides Management, the fund run by Gavin Baker, disclosed 3,110,086 shares valued at $687.33 million, its fourth-largest disclosed position by value.

For investors, the OpenAI commitment gives Cerebras a contracted revenue base that most AI chip startups lack, but the stock still trades well below its 52-week high of $386.34 after the Aug. 13 selloff. Wedbush raised its price target to $290 from $280 last week, and the consensus of $291.64 implies roughly 20 percent upside from Monday's levels. The question is whether Cerebras can convert its speed advantage into durable share against NVIDIA, whose CUDA software stack remains the industry default — and whether the 750-megawatt buildout, which requires substations, cooling loops, and networking gear, comes online on schedule.

This article is for informational purposes only and does not constitute investment advice.