OpenAI just declared war on Nvidia’s monopoly

I’m going to be honest, I didn’t think we’d see this so soon. For years, the story has been that if you want elite AI, you buy Nvidia. But today, OpenAI officially changed the narrative. They just dropped GPT-5.3 Codex Spark, and for the first time ever, it isn isn’t running on H100s or B200s. It’s running on Cerebras hardware.

The Dinner Plate Chip

The hardware powering this isn isn’t a cluster of small cards. It’s the Cerebras Wafer Scale Engine 3—a single, massive chip the size of a dinner plate. Because the entire processor is on one piece of silicon, data doesn doesn’t have to travel across cables or slow motherboards. This partnership, which OpenAI announced back in January, is finally bearing fruit. The result is a staggering 1,000 tokens per second.

I’ve been watching the demos, and it’s unsettling. Code doesn doesn’t stream anymore; it just appears. Because Spark uses a persistent WebSocket connection, the overhead is slashed by 80% per roundtrip. It feels less like a chat and more like the AI is reaching into the future and pulling the completed file back into my IDE.

The data behind the speed

Below is the performance snapshot from the latest benchmarks, showing exactly how Spark stacks up against the standard Codex model in vertical chart form.

Real‑time collaboration is here

If you’re a developer, you’ve probably felt that waiting anxiety—the three-second pause while the LLM thinks. Spark kills that. It features a 128k context window and is optimized specifically for real-time coding. It’s designed as a complementary mode to longer-horizon reasoning. I see it as a tool for real-time collaboration on edits, logic, and interfaces.

Why this matters for you

Spark is rolling out right now to ChatGPT Pro users across the web app, CLI, and the VS Code extension. It is text-only at launch, but the speed makes up for the lack of bells and whistles. More importantly, it proves that the AI industry doesn’t need Nvidia to innovate. OpenAI signed a massive contract with Cerebras to make this happen, and the performance is undeniable.

What about the reasoning

To be clear, Spark is a specialized model. If I need deep architectural planning, I’m still going to use the standard GPT-5.3 or Gemini 3 Deep Think. But for the 90% of our day that is spent writing boilerplate, debugging, and refactoring? This is the new standard. The monopoly is breaking, and the speed is finally catching up to our thoughts.