Cerebras Systems will report its first earnings as a public company on June 23, and Wall Street is watching for one critical signal: whether the wafer-scale chip maker has diversified beyond its two Abu Dhabi customers, who accounted for roughly 86% of 2025 revenue. The per-share loss that will likely dominate headlines is almost meaningless; the real story sits in the backlog conversion rate and the customer roster.

Why Does Customer Concentration Matter for an AI Chip Maker?

Cerebras priced its initial public offering at $185 per share in May 2026 and raised approximately $5.55 billion, making it the largest US technology offering of the year. The stock opened near $350 on its first day and closed at $311, up roughly 68%, before settling near $235 by mid-June. That volatility reflects investor uncertainty about whether the company's business model can scale beyond a handful of sovereign buyers.

The concentration problem is stark. In 2024, G42, an Abu Dhabi AI group, represented 85% of Cerebras's revenue. The IPO narrative promised diversification, and G42's share did fall to 24% in 2025. But the gap did not fill with a diversified customer base. Instead, Mohamed bin Zayed University of Artificial Intelligence, or MBZUAI, became the largest customer at 62% of revenue. Combined, these two UAE-affiliated entities accounted for roughly 86% of the company's 2025 revenue. For a US-listed chipmaker whose customers' AI ambitions track a single government's strategy, that represents a single-country dependency with significant geopolitical sensitivity.

What's Inside the $24.6 Billion Backlog?

Cerebras reported $24.6 billion in remaining performance obligations as of December 31, 2025, a figure that anchors every bull case on the stock. Against $510 million in actual 2025 revenue, that is nearly 48 times promised-to-delivered. The overwhelming majority of this backlog traces to one counterparty: OpenAI. An agreement announced in January and formalized in late 2025 commits OpenAI to deploy 750 megawatts of Cerebras inference capacity through 2028, in a deal Cerebras values at more than $20 billion.

The catch is timing. Secondary estimates suggest only about 15% of the backlog will convert to revenue by the end of 2027, with the bulk arriving in 2028 and later. This means the figure that matters on June 23 is not the backlog itself, but the pace at which it becomes recognized revenue. A backlog anchored to one well-funded AI leader is both the bull case and the risk: it represents real, contracted demand, but its conversion runs on a single customer's buildout schedule.

How to Evaluate Cerebras's Growth Trajectory

• Revenue Acceleration: Cerebras grew from $99.5 million in Q1 2025 to $171.4 million in Q4, a 72% climb across the year that accelerated into the finish, with Q4 up 26% over Q3. Full-year revenue landed at $510 million, up 76% year-over-year.
• Margin Pressure: Gross margin eased from about 42% in 2024 to 39% in 2025, the wrong direction for a company asking the market to pay up for scale. The Q1 guide and any comment on where margin goes next will reveal whether the ramp is healthy or merely busy.
• Customer Diversification: One piece of genuine diversification worth tracking is Cerebras's partnership with Amazon Web Services to offer its inference inside the AWS ecosystem. A hard number attached to AWS on the earnings call would matter more than another OpenAI headline, as it would put Cerebras silicon in front of mainstream cloud customers rather than a handful of sovereign buyers.

The reported 2025 "profitability" was an artifact of accounting, not a sign of operating leverage. Cerebras booked GAAP net income of roughly $238 million for 2025, but almost all of it came from a single non-cash gain of about $363 million tied to the change in fair value of a forward-contract liability, recognized entirely in the second quarter. Strip that out and the picture inverts: Cerebras lost money on a GAAP basis in the other three quarters, and its non-GAAP net loss for the year was about $75.7 million, wider than 2024.

Consensus expects Cerebras to report a loss of roughly 14 cents per share for Q1 2026, and that is the right thing to expect from a company still scaling into its backlog. The trap is mistaking the prior year's accounting gain for operational profitability.

What Makes Cerebras's Hardware Distinctive?

The Wafer-Scale Engine, or WSE-3, is the most distinctive piece of AI hardware in production. It is a single chip the size of a dinner plate, etched with 4 trillion transistors and 900,000 cores on one continuous slab of silicon, where a conventional design would be sliced into dozens of separate graphics processing units, or GPUs. Keeping an entire model on one wafer removes the networking bottleneck that throttles clusters of discrete chips.

That architecture has found its moment in inference, the process of running already-trained models, just as the industry's spending tilts from training toward serving. Cerebras markets eye-watering speed, claiming its systems generate output for large open models like Llama at well over 2,000 tokens per second, multiples faster than GPU-based services. Those are company benchmarks and deserve the usual skepticism, but the direction holds: as AI shifts from building models to running them at scale, raw inference throughput becomes the product, and it is the field where Cerebras's architecture shines.

The earnings report on June 23 will reveal whether Cerebras's technical advantage can translate into a diversified customer base. Investors who want this story to mature need to see a recognizable American or European logo on the customer list. That is the real test of whether Cerebras is a generational AI infrastructure play or a single-country bet masquerading as one.