Cerebras Drops 19% in 3 Months: Buy, Sell or Hold the Stock?

Cerebras shares have fallen 19% over three months as Q2 2026 GAAP net loss reached $450.5 million and core gross margin fell 590bps sequentially to 40.6%; Q3 margin is guided lower at 38-40% with operating margin of negative 25% to negative 23%. The company nevertheless holds $25.4 billion in RPO, expects core revenue to more than triple in 2027, and plans to expand manufacturing capacity more than 10-fold in 2026. Liquidity of more than $8.6 billion and improving loss estimates support the longer-term case, but OpenAI-related customer concentration, premium valuation, heavy infrastructure spending and AI-chip competition underpin the Hold stance.

Analysis

The central underwriting issue is not demand visibility but cash conversion. A hardware-plus-cloud model that temporarily substitutes owned/rented capacity for customer-delivered systems can turn booked demand into a capital-intensity problem: revenue may scale before depreciation, power, lease and service costs normalize. That makes the next two quarters a test of unit economics rather than of backlog; a gross-margin recovery without a corresponding improvement in operating cash burn would not validate the long thesis.

Customer concentration creates an asymmetric multiple risk. If a dominant customer defers deployments, renegotiates pricing, or shifts workloads to internally sourced GPUs, CBRS would face both utilization deleveraging and a reduced credibility of its contracted-demand narrative. By contrast, NVDA's installed software ecosystem and broad customer base let it monetize the same inference spend with less single-customer and asset-utilization exposure; AMD is the more direct beneficiary if buyers seek a second-source accelerator rather than a specialized architecture.

The contrarian upside is that avoiding constrained memory and advanced-packaging inputs could permit faster physical deployment than GPU competitors during supply tightness. But that advantage is only valuable if performance-per-dollar converts into repeatable third-party workloads, not customer-specific capacity commitments. Watch disclosed non-anchor-customer revenue, owned-versus-rented capacity mix, and operating cash flow over the next 1-3 months; over 6-18 months, the key falsifier is failure to demonstrate sustainable gross-margin expansion as new capacity enters service.

Key Decisions for Investors

• Maintain no outright CBRS long ahead of the next results. Upgrade only if gross margin turns sequentially higher after the guided trough, operating cash burn improves, and non-anchor customers contribute enough to reduce concentration; failure on any two metrics supports a short/watchlist bias rather than a dip-buy.
• Use a 3-6 month relative-value expression: long NVDA / short CBRS in equal beta-adjusted dollars. The trade captures diversified inference monetization versus execution, utilization and concentration risk; cover the short if CBRS reports material new hyperscaler revenue with improving cash conversion, not merely additional RPO.
• For second-source AI exposure, prefer AMD over CBRS on a 6-12 month horizon only after MI-series supply and hyperscaler adoption data confirm share gains. AMD offers a cleaner upside path if buyers diversify away from NVDA, while CBRS requires simultaneous proof of technology, deployment and financing execution.
• Monitor TSM for incremental wafer-volume upside, but do not treat CBRS capacity expansion as a meaningful standalone TSM catalyst until shipment volumes are independently visible. The relevant alert is a material revision to TSM advanced-node utilization or capex commentary, which would validate broader accelerator demand rather than company-specific bookings.