AI chipmaker Cerebras namedropped by Oracle, alongside Nvidia and AMD

Oracle disclosed on its earnings call that its infrastructure includes Cerebras accelerators, a potential strategic cloud customer for Cerebras. Cerebras still faces concentration risk—G42 accounted for 87% of its H1 2024 revenue—but secured a $1.1B funding round at an $8.1B valuation after withdrawing an IPO filing and says it still intends to go public. Oracle reported better-than-expected results, raised fiscal 2027 guidance and said remaining performance obligations surged to $553B, underscoring strong demand for data-center compute capacity.

Analysis

Oracle’s move to support alternative accelerators is an inflection for incumbents in two ways: it materially shortens the sales cycle and concentration risk for emerging hardware vendors by converting “pilot” projects into cloud-billable capacity, which can shift private valuations more than incremental revenue. If Cerebras or another niche accelerator secures even 1–3 major cloud partners in the next 6–12 months, that will convert concentrated contract value into recurring, multi-region footprint — valuation multiples for niche chips typically re-rate 20–40% once cloud OEMs formalize price lists and SLAs.

For Nvidia and AMD the effect is bifurcated: the former’s software and model-optimization moat still protects high-utilization training economics, but a separate, latency-sensitive layer of demand (real-time inference, embedded LLMs) is being carved out where wafer-scale architectures can offer lower end-to-end latency and total cost per inference. Expect 6–18 months of price competition at the margin in inference workloads and selective displacement in edge or specialized cloud cabinets, which compresses realized GPU ASPs for specific inference classes even as aggregate GPU demand stays strong.

Data center supply chains will feel the tug: a larger installed base for WSE-style parts increases demand for HBM stacks, specialty packaging and TSMC capacity — creating a temporary supply squeeze (lead times 3–9 months) that benefits foundry/pricing power for incumbents and raises replacement capex for cloud operators. The wild cards are software maturity and model portability; without robust compiler/tooling and reference model fits, hardware wins will be limited to bespoke workloads and won’t scale to mainstream LLM fleets.

Key catalysts to monitor over the next 3–12 months are (1) Oracle publishing Cerebras-based SKUs/pricing, (2) OpenAI / other hyperscaler workload migrations onto alternative silicon, and (3) Nvidia’s GTC architecture reveal and any regulatory scrutiny of its acquisitions. Reversals come from fast parity in compiler stacks, large discounts from Nvidia on inference GPUs, or a failure by Cerebras to show materially better latency/TCO for widely used LLMs.

Key Decisions for Investors

• ORCL — Buy a 9–12 month call spread (near‑ATM long, +10–15% OTM short) to capture upside if Oracle monetizes alternative accelerators in pricing and sales. Risk = premium paid; reward = asymmetric if Oracle announces SKU/pricing or sees uplift in high‑margin AI RPOs (target +20–35%).
• NVDA — Avoid directional exposure into GTC; if implied vol spikes, sell a 2–6 week call calendar or short-dated call spread to collect IV premium. Risk: headline product beats; reward: capture mean reversion in post-event IV with limited defined risk.
• Pair trade — Long ORCL / Short NVDA (equal dollar, 3–6 month horizon): play Oracle’s differentiation and near-term monetization vs Nvidia’s already-priced-in growth. Position size small (max drawdown per leg capped at 6–8% of book); target asymmetric payoff of 20–30% if Oracle converts offerings while NVDA faces post-GTC profit-taking.
• AMD — Small tactical long (6–12 months) to play second-source wins in cloud if hyperscalers demand diversity; keep exposure limited and hedge with short-dated puts to cap downside. Reward: capture 25–40% re-rating if AMD secures material design wins; tail risk if it fails to compete on performance/WTP.