Cerebras Systems's S-1 refile reached its ninth day Sunday with the company in standard SEC quiet period and the roadshow window open for a mid-May listing under the ticker CBRS. [1] Reuters and CNBC reported the filing and the OpenAI multi-year compute commitment last week. [1][2] The S-1 itself disclosed the OpenAI compute contract now totaling more than $20 billion, the $1 billion working-capital loan, and warrants for up to 33,445,026 Class N shares vesting against compute-purchase thresholds. [3] Saturday's paper read on the roadshow window opening into V4 and Kimi K2.6 called this the eight-day Chinese-frontier cadence. Sunday gives that cadence its specific source: DeepSeek V4's months-long delay was, per Bloomberg sourcing, time spent integrating the model with Huawei's Ascend stack.

That sourcing matters for the clearing-price question. The bookrunners — Morgan Stanley, Citigroup, Barclays, UBS as joint leads — would prefer the roadshow story to be wafer-scale-vs-GPU. The pitch lands as the highest-end alternative to Nvidia's stack: the WSE-3 is 58 times larger than Nvidia's B200 with 2,625 times the memory bandwidth, the architecture beats the GPU on inference economics for frontier-scale models. [3] That story is intact. What has shifted in the last eight days is the second axis. The Chinese-AI alternative is no longer a hypothetical. It is open-weights frontier capability — Kimi K2.6 holding Humanity's Last Exam Full benchmark, then DeepSeek V4 with 1-million-token context and Mixture-of-Experts at 1.6T total parameters — running on Chinese-domestic silicon at a cost basis the export-control regime does not reach.

The clearing price has to absorb both axes. At a valuation in the $22 to $25 billion range several desks have floated [3], OpenAI's fully-vested equity stake represents roughly 10 percent of post-money float at effectively zero cost basis. [3] The structural risk in the prospectus is concentration: OpenAI is customer, lender, and shareholder. The narrative risk Saturday's paper named is now the operating reality: if OpenAI's own compute mix shifts toward alternatives the V4 launch made plausible, the concentration becomes a vulnerability rather than a moat.

The Bloomberg sourcing on V4's delay is the tell sell-side has not yet fully digested. A delay attributed to "Huawei Ascend integration" is not a model-quality story — it is an architecture-portability story. Frontier model development on non-Nvidia silicon is a compounding capability. Each model that successfully integrates with Ascend expands the developer mind-share for non-Nvidia frontier inference. The Cerebras pitch — wafer-scale as the architecture for inference economics — needs the GPU-vs-wafer-scale argument to remain the dominant alternative debate. The V4 integration says the GPU-vs-domestic-silicon argument is a parallel debate now, on a different cost-basis curve.

What the bookrunners control is the pricing-window event: the S-1/A amendment that fills the cover-page blanks, the launch of the roadshow itself, and the timing of pricing. SEC review of the refile typically takes four to six weeks, putting the earliest plausible roadshow start in early May and a mid-May pricing target. [1][2] What the bookrunners do not control is the next Chinese model release. The eight-day cadence Kimi K2.6 to DeepSeek V4 just established is now an empirical fact. If a third release lands during the roadshow itself, the buy-side meetings have to absorb the news cycle in real time.

For Cerebras specifically, the disciplined frame is the OpenAI commitment as the floor — that part of the story is signed, dollar-denominated, and disclosed — and the wafer-scale architecture as the ceiling, which depends on whether sell-side prices "different alternative to Nvidia" or "the alternative to Nvidia." The Sunday tape, with the markets closed, prices nothing. The Monday tape will price the read on V4's source detail. The pricing event itself, weeks away, will price whatever Chinese model release lands closest to the printer.

-- THEO KAPLAN, San Francisco