Cerebras Systems builds AI infrastructure. It reported earnings for the second time as a public company on Wednesday, August 12th, posting a net loss of about $450 million and missing revenue estimates. Shares fell sharply.
What the company makes
Cerebras builds computer chips the size of a dinner plate and supercomputers made to train and run AI models fast. It runs two business segments.
Hardware made up 30% of total revenue. It covers designing, making, and selling large wafer-scale processors and on-site AI systems. These are built to speed up AI work without needing supply-limited parts like high-bandwidth memory.
Cloud and services was just under 70% of total GAAP revenue. It runs AI data centers that sell high-speed inference as a service. Inference means a trained AI model uses what it learned to handle new information and give new answers. This segment also supports enterprise agentic workflows, coding tools, and large language model deployments.
The chip and how it is made
Most chipmakers cut chips out of a silicon wafer, a thin flat round slice of very pure crystalline silicon that acts as the base for chips and other circuits. Cerebras instead makes each chip from the whole wafer, giving it the largest processors ever built. There are no direct competitors making 8 and 1/2 inch square chips.
A Cerebras Wafer Scale Engine 3 chip is about 58 times larger than an Nvidia (NVDA) Blackwell B200 chip. It holds 4 trillion transistors versus the B200's 208 billion. The large size gives faster processing and leaves room to put RAM memory directly on the chip. Other companies use off-chip memory, which is slower because data travels farther. Cerebras says this makes its systems much faster than normal supercomputer clusters.
Competition
Its main hardware rival is Nvidia (NVDA), the market leader in AI training and inference. Other rivals are AI chip startup Groq and AMD (AMD). In cloud and hyperscaler work, its main competition is Google's (GOOGL) custom tensor processing unit systems for internal cloud work. Other players are Amazon's (AMZN) AWS, which makes custom training and inference chips, plus Microsoft (MSFT) and Meta (META), which build custom internal accelerators called Maya and MTIA.
The numbers
Total GAAP revenue for the quarter was $180 million, below the street's $194 million estimate. Cloud and services revenue was $126 million, up 281% year-over-year. Hardware sales fell 23% year-over-year to $51.1 million. Core revenue more than doubled and the cloud business nearly quadrupled. GAAP gross margin was 14%. Core gross margin improved 940 basis points to 41%.
CEO and co-founder Andrew Feldman said demand for fast inference is enormous and the company aims to meet it. He said Cerebras wants to triple its revenue in 2027.
Positives
Cerebras raised $6.4 billion in gross proceeds through its IPO and closed a revolving credit line of up to $850 million to speed up buying data centers. It had $25.4 billion in remaining performance obligations as of June 30th, 2026.
It has separate partnerships with AMD (AMD) and Amazon's (AMZN) AWS to build faster AI inference. The AMD partnership aims to keep Cerebras's high speeds while raising throughput by up to five times, and is set to go into production in the fourth quarter of 2026.
Cerebras expects core revenue of $214 to $216 million for the third quarter. It raised full-year guidance to $880 to $890 million, up from the prior $855 to $865 million.
Concerns
Execution risk is the main worry. Like many chip companies, Cerebras does not make its own chips. It designs them and outsources making them to firms like Taiwan Semiconductor (TSM). Those makers have limited production capacity, which is a bottleneck. The CEO also said the main limit on AI growth is data centers - both their scarcity and slow buildout, plus the power infrastructure needed to scale AI.
Customer concentration is another concern. Cerebras has a multi-year deal with OpenAI worth over $20 billion, about 80% of its performance obligations. The risk grows because much of its revenue comes from groups tied to the United Arab Emirates. That creates exposure to foreign policy shifts or tariff and export news.
The chart
Cerebras began trading on May 15th, so there is limited price history. Shares bottomed near 160.81 and formed a triple-bottom pattern, hitting that level and bouncing several times. Trading has stayed range-bound between a floor near 160 and a ceiling near 243. Price trended up into earnings along a short-term white trend line, but a pre-report pop faded, leaving price below that line. The high point was about 266, matching a high from late May soon after trading began.
RSI, a momentum gauge, needs at least 14 trading days of data to read the pace of gains or losses. Its green upward trend line broke, like the price trend line, but RSI stays above the 50 midline, giving a slightly bullish momentum read.
For moving averages, simple ones are used here because data is thin: a 20-day simple moving average in yellow and a 50-day in blue, both bunched between about 208 and 210, offering possible support.
Volume profile shows a smaller node from 181 to 191 near the extreme lows. Most trading since the start sat between 200 and 233, with the point of control - the heaviest trading area, marked by a thick red line - at 221.50. Activity thins above 233, picks up again between about 275 and 307, and stays light above that.
On expected move, the September 18th monthly expiration, 36 days out, shows the market does not expect a big move past the lows, bottoming near 170. That area could be major support or a breakdown point if price drops. To the upside, even out to November the market does not expect a return to the highs near 386.
Summary
Earnings showed large year-over-year growth but missed Wall Street estimates by a wide margin. Investors focused on guidance coming in weaker than expected, even though the full-year forecast was raised. The open question is how big Cerebras can become in AI hardware and hyperscaler markets against giants like Nvidia (NVDA), and how much its partnerships can drive future growth.
