Artificial-intelligence chip maker Cerebras Systems will report its second-quarter earnings on Wednesday afternoon, and it might be another volatile day for the stock on Thursday.

The company went public in May with an offering price of $185. The stock opened at $350, rose to $386, but fell straight from there, hitting a low of $161 by the end of June. Since then it has fluctuated between $162 and $244. Shares have had moves in either direction of greater than 3% on 42 of the 61 days they have traded publicly.

Wall Street analysts expect "core revenue" of $191 million, about the same as last quarter, but up 84% from last year.

The company uses a non-standard revenue metric because sales figures are impacted by two idiosyncrasies that are not strictly part of operations. First, it gets some pass-through revenue from customers that comes with no margin, so the company subtracts that. Second, the company also issued warrants to two customers, and the amortization of the warrants gets subtracted from revenue under standard accounting, so the company adds that back in for core revenue.

Last quarter, the company had $193.4 million in sales, but it subtracted $4.1 million of pass-through revenue and added back $2.1 million in warrant amortization to arrive at core revenue of $191.3 million.

Analysts expect adjusted operating loss to be $62 million and adjusted loss per share of 17 cents. Wall Street does not expect profits until next year.

Cerebras is operating in a crowded field of AI chip producers trying to steal some business from Nvidia, the dominant player. But the company's Wafer-Scale Engine chip, known as WSE, stands out. While a 300 millimeter silicon wafer typically produces many chips, the WSE is the size of the entire wafer. It has computing and memory on the same slab of silicon, blowing through one of the most restrictive bottlenecks in AI servers: the link between computing and memory. Small and medium-sized AI models run very quickly on the WSE chip and the servers built around it.

There is a shift happening in AI workloads that is opening up an opportunity for Cerebras and other non-Nvidia chip makers. At first, new AI data centers were used mostly for training new models, a laborious and expensive process. But this year we are starting to see inference -- the process of running these models for chatbots or agents -- beginning to take over. Agents are software that can use AI models to accomplish a complex series of tasks from simple conversational commands, and they chew through inference at a rate that no human can match.

Nvidia won the first round of the data-center boom, but its customers see the rise of inference as an opportunity to diversify their supply chains. Cerebras is now fielding orders from more customers. In the past, most of its sales came from two entities closely tied to the United Arab Emirates royal family, but last quarter its $25 billion backlog was dominated by a three-to-five year OpenAI contract to rent Cerebras servers in the cloud. OpenAI also has the option of increasing the deal significantly.

Cerebras also has newer deals with Amazon.com and Advanced Micro Devices. The structure of these contracts provides for WSE to work alongside Amazon and AMD chips for inference tasks that require high speeds. Amazon has also received stock warrants as part of that deal.