OpenAI's first non-NVIDIA model retires, marking a $20 billion bet on new chips.

BT Finance

The life cycle of GPT-5.3-Codex-Spark was less than eight months.

This experimental programming model, launched as a research preview in February, will be officially decommissioned next week. Its retirement is not due to technical failure, but because it has completed its phase-mission as OpenAI’s first experiment exploring computing power routes beyond NVIDIA GPUs, proving the viability of Cerebras wafer-scale chips in inference scenarios.

Meanwhile, its replacement, the GPT-5.6 Sol Ultrafast mode, went online in August. This full version of GPT-5.6 Sol, powered by Cerebras, outputs speeds up to 750 tokens/s, eliminating the trade-off between model capacity and response speed.

This means the first probe specifically to verify the feasibility of “de-NVIDIA” has completed its mission. And this $20 billion+ diversification of computing power is redefining the core battlefield of large model competition.

Why choose Cerebras?

The most special thing about GPT-5.3-Codex-Spark is not its speed, but its hardware.

In traditional GPU systems, model weights must be repeatedly moved between the computing chip and external memory, and the latency in this process becomes the main bottleneck for low-latency inference. Cerebras’ solution is completely different — it integrates massive computing units and on-chip SRAM onto a single wafer, fundamentally reducing the loss caused by data movement.

This architecture is particularly suitable for low-latency inference scenarios. For Codex-like tools that require real-time response for code editing, a few milliseconds of delay directly impact the user experience. Developer feedback shows Spark significantly improves response time during frequent operations like refactoring and debugging, though its code quality and logical reasoning still lag behind the full model.

Cerebras is a Silicon Valley–based startup founded in 2016 by former IBM chief scientist Chris Beatty. Unlike traditional semiconductor manufacturers who produce chips in batches, Cerebras uses an entire silicon wafer as a processor — offering unique advantages while facing challenges in mass production yield and ecosystem compatibility. Currently, few companies globally can truly master wafer-scale processor technology.

The choice of Cerebras was not accidental. In January 2026, OpenAI announced broadening its compute supplier base, entering a multi-year agreement with Cerebras to deploy up to 750 MW of inference power. This deal, valued at over $20 billion and phased through 2028, is seen as a key step for OpenAI to reduce dependence on a single GPU supply chain.

The backdrop is rising costs. Training large models is exponentially expensive, and inference — as the continuous consumption phase after model deployment — is rapidly becoming a significant portion of total operating expenditure. Industry estimates suggest that when models scale, inference costs may account for 60–70% of total operating costs — far exceeding training budgets. If OpenAI processes billions of inferences daily, saving even a few cents per inference could yield tens of millions in annual profit — a substantial figure for any public company.

Diversification also addresses supply chain security. NVIDIA is the only enterprise-grade supplier capable of delivering large-scale AI training clusters, but its delivery times often stretch to months or half a year. By shifting workloads to other platforms, OpenAI builds redundant resilience to prevent business stagnation from supplier bottlenecks.

The fate of an experimental model

Tibo Sotio’s announcement that Spark is retiring reveals a deeper trend: the淘汰周期(turnover cycle) of large models is shrinking drastically.

Previously, model iterations occurred over months or years. Now, OpenAI treats models as continuous experiments — launching scenario-specific versions, validating speed/cost/user demand, migrating successful features to next-generation models, then quickly retiring old ones.

Spark fits this pattern — initially exclusive to ChatGPT Pro users, serving low-latency tasks like VS Code plugins. While some praised its speed for mechanical tasks like refactoring, others found it “extremely poor and unusable.”

This polarization stems from compromise: for fast response, model capacity was reduced. With GPT-5.6 Sol Ultrafast delivering both speed and full capability on the same hardware, Spark’s independent value is eliminated.

The retirement itself is less important than what it signals: compute diversification is moving from lab to production faster than expected. This implies big tech is serious about mitigating supply chain risks — not just talk. Moreover, if this route proves successful, future processors may be specialized for inference, not just general-purpose GPUs.

What about NVIDIA?

Spark’s retirement does not mean NVIDIA has lost ground.

Cerebras currently handles niche scenarios (programming assist, low-latency inference), while NVIDIA remains the backbone for training and most inference loads. Cerebras WSE-3 fills gaps where NVIDIA efficiency drops — not a full replacement.

The $20 billion order is a multi-year commitment whose true impact depends on delivery progress and strategic evolution. Since datacenter buildout involves complex approvals and timelines, immediate displacement is unlikely.

NVIDIA’s strengths lie in its complete product line, developer ecosystem, and scale. Even if OpenAI diversifies, it’s likely supplemental, not substitutional — unless other majors follow suit within 12–18 months, which would shift investor perceptions of NVIDIA’s growth ceiling.

Key indicators to watch: whether a second Cerebras model follows, actual shipment volumes vs capacity ramp, and NVIDIA’s pricing response — if they lower prices or shift terms, it suggests sensitivity to competition. Otherwise, confidence prevails.

Final takeaway: The race for optimal inference hardware is heating up, but NVIDIA’s dominance isn’t threatened yet — just challenged constructively. OpenAI’s dual strategy keeps options open and pressure on suppliers.

Watch this space — and perhaps grab some popcorn — as the AI hardware landscape evolves beyond the GPU monopoly.

Actual shipments from Cerebras and the ramp-up progress of WSE-3 production — The order amount is an intention, but actual delivery is the real test of a supplier’s strength.

NVIDIA’s pricing strategy adjustments in the inference market — If NVIDIA begins to actively lower the prices of inference-dedicated GPUs or offer more flexible payment plans, it indicates that they are concerned about losing market share. Conversely, if they maintain high prices and rigid terms, it suggests confidence in their current position.