Why Cerebras’ Mind-Boggling LLM Raw Speed Is Still Falling Into Nvidia’s Massive Software Trap

Alex Sirois

Fri, June 26, 2026 astatine 9:21 AM CDT 3 min read

Quick Read

• CBRS secured a $20B+ OpenAI woody yet guided to antagonistic operating margins, portion NVDA's 75% gross margins uncover the level moat's fiscal power.

• Cerebras' wafer-scale spot delivers a 21x velocity vantage implicit Nvidia hardware, but each large LLM model natively optimizes for CUDA, requiring costly customized engineering.

• Don't wait: the expert who called NVIDIA successful 2010 conscionable revealed his apical 10 AI stocks. See the afloat database FREE now.

NVIDIA (NASDAQ: NVDA) and Cerebras Systems (NASDAQ: CBRS) conscionable delivered net that framework the aforesaid question from other ends. Nvidia posted different blowout 4th built connected its CUDA bundle stack. Cerebras, caller disconnected its May IPO, showed jaw-dropping inference velocity yet guided full-year operating margins negative. The moat is developer gravity.

One Sells Platforms. The Other Sells Speed.

Nvidia's Q1 FY27 deed $81.61 cardinal successful revenue, up 85.2% YoY, with Data Center unsocial reaching $75.25 cardinal connected 92% growth. Networking soared 199% arsenic InfiniBand, NVLink and Spectrum-X locked customers deeper into the stack. Jensen Huang told investors NVIDIA is "the lone level that runs successful each cloud, powers each frontier and unfastened root model, and scales everyplace AI is produced", and the numbers backmost the claim.

Cerebras' archetypal study arsenic a nationalist institution landed differently. Q1 GAAP gross reached $193.4 million, up 94% YoY, with unreality services increasing 178%. A multi-year, $20 billion-plus OpenAI inference woody covering 750 megawatts anchors near-term growth. Yet absorption guided full-year operating margins to antagonistic 28% to antagonistic 32%. Speed sells. Scaling it economically is harder.

Software Gravity Beats Wafer-Scale Throughput

Independent benchmarks amusement Cerebras' wafer-scale plan delivering a 21x velocity vantage implicit Nvidia hardware for latency-sensitive, low-batch inference. The drawback is that each large LLM model and endeavor developer stack is natively optimized for Nvidia architecture retired of the box, portion Cerebras requires specialized compilation and customized engineering enactment for thing disconnected the well-trodden path.

Don't wait: the expert who called NVIDIA successful 2010 conscionable revealed his apical 10 AI stocks. See the afloat database FREE now.

Nvidia's $119 cardinal successful proviso commitments and $80 cardinal added to its buyback authorization awesome absorption is doubling down connected the platform. Cerebras raised $5.6 cardinal astatine IPO and is funneling it into information halfway capableness for OpenAI's decode workloads portion AWS Trainium 3 handles prefill. That is simply a focused inference stake riding connected 1 customer's roadmap.