Cerebras Systems began trading on Nasdaq under ticker CBRS on May 14, pricing at $24 per share and opening at $31.50, a 31% first-day pop that delivered an $8.2 billion post-money valuation. The offering raised $740 million in primary capital, with underwriters led by Morgan Stanley and Goldman Sachs. No secondary shares were sold by insiders in the base deal, a structural choice that sidesteps the usual post-lockup volatility and signals founder Andrew Feldman's confidence in the next twelve months.

Cerebras manufactures wafer-scale AI training chips—single silicon dies the size of a dinner plate, each containing 850,000 cores and 40 gigabytes of on-chip SRAM. The architecture is the inverse of Nvidia's modular GPU clusters: one monolithic piece of silicon instead of thousands of smaller chips stitched together with high-speed interconnects. The company disclosed $340 million in trailing twelve-month revenue in its S-1, up 68% year-over-year, with 42% gross margins and a path to EBITDA breakeven by Q4 2026. The customer concentration is tight—73% of revenue came from four hyperscale AI labs in 2025, including G42 in Abu Dhabi and an unnamed U.S. foundation model developer.

The offering matters because it confirms that AI infrastructure capex is now structural, not speculative. Nvidia still commands 92% of the accelerator market, but Cerebras proves there is room for architectural heterogeneity when training runs exceed 10^25 FLOPs and interconnect latency becomes the bottleneck. Allocators should note the gross margin profile: Cerebras ships appliances, not raw chips, which means the company controls the full stack from silicon to cooling to orchestration software. That vertical integration supports 42% gross margins even at low scale, a figure that expands as wafer yields improve and TSMC's 5nm production costs decline. The IPO also validates a new class of compute buyer—sovereign AI funds and national labs willing to pay a 30-40% premium over Nvidia for supply-chain independence and architectural differentiation.

Operators should watch three follow-on events. First, the 180-day lockup expires on November 10, but the no-secondary structure means insider selling pressure will be muted unless the stock doubles from here. Second, Cerebras will likely announce a second-generation wafer with 1.2 million cores by Q3 2026, which would allow training of 2-trillion-parameter models on a single CS-3 appliance instead of a cluster. Third, watch for Amazon Web Services or Microsoft Azure to announce availability of Cerebras instances by year-end—cloud distribution would de-risk the customer concentration and open access to the $18 billion market for on-demand AI compute.

The fact worth holding: Cerebras shipped 42 systems in Q1 2026, each priced between $2 million and $3 million, and the backlog now sits at 89 units with deposits already paid.