Will OpenAI begin production deployment of custom Broadcom AI chips by December 31, 2026?

The Q4 FY2026 earnings provide no direct information about the OpenAI-Broadcom custom chip timeline. Jensen's statement that GPT-5.2-Codex is 'trained with and inferencing on Grace Blackwell and NVL72 systems' demonstrates a deepening NVIDIA-OpenAI relationship, but OpenAI has always pursued both paths simultaneously. The deepening NVIDIA partnership could slightly reduce urgency for custom silicon deployment, but the strategic rationale for supply diversification remains intact -- OpenAI's 10GW commitment represents a fundamental hedge against single-vendor dependency regardless of how strong the current NVIDIA relationship is. The Groq acquisition removes one independent inference ASIC competitor but does not affect Broadcom's design capability or TSMC's manufacturing. The timeline remains the core uncertainty: if design started mid-2024, tape-out could occur in H1 2026 with initial manufacturing in H2 2026. We are now ~10 months from the resolution date with still no public tape-out confirmation. Marginally lower than prior due to reduced urgency signal.

Two developments from Q4 earnings are indirectly relevant but push in the same direction -- slightly lower probability. First, NVIDIA's massive Q4 ($62.3B DC revenue, +22% QoQ) and $78B Q1 FY2027 guidance mean OpenAI faces less supply constraint pressure from its existing GPU vendor. When NVIDIA can deliver abundant supply, the urgency to deploy unproven custom silicon diminishes. Second, Groq's acquisition by NVIDIA strengthens NVIDIA's inference stack, partially addressing one of the motivations for OpenAI's custom chip program (inference cost optimization). However, the Broadcom chip program targets training workloads as well, where Groq technology is irrelevant. The program is structurally independent of NVIDIA's earnings performance. The 2-3 year typical design-to-production timeline (committee finding) places late 2026 at the aggressive end. The absence of any tape-out announcement -- now nearly 18 months after the partnership was reportedly initiated -- is a modest negative signal, though companies routinely delay such announcements.

I maintain the prior probability because the Q4 earnings data is largely irrelevant to the Broadcom custom chip timeline. The resolution criteria ask whether OpenAI begins 'production deployment' by December 31, 2026. This is a hardware development question driven by Broadcom's chip design execution and TSMC's 3nm manufacturing, not by NVIDIA's financial results. The deepening NVIDIA-OpenAI relationship is a sentiment signal, not a timeline signal -- OpenAI can simultaneously be NVIDIA's closest AI partner AND pursue custom silicon diversification, as the prior context noted. The Groq acquisition and NVLink Fusion are competitive dynamics that affect the long-term ASIC thesis but do not change whether a specific chip currently in development reaches production. The resolution criteria's OR clause (tape-out + manufacturing OR deployed in data centers) provides two paths to YES. If tape-out occurred in early 2026 (plausible given mid-2024 design start), foundry manufacturing by late 2026 is achievable. No new data contradicts this possibility.

The updated context shifts probability slightly lower for several reasons. First, Jensen's characterization of OpenAI as 'a once-in-a-generation company' and the active partnership negotiation ('believe we are close') suggests deepening commercial ties that could reduce OpenAI's incentive to publicly announce a competing silicon program -- even if the chip is ready, OpenAI may delay the announcement to preserve the NVIDIA relationship. Second, NVIDIA's compute supply is expanding faster than expected ($78B Q1 guide), reducing the supply-constraint motivation for custom silicon. Third, the agentic AI demand explosion means OpenAI needs every GPU it can get right now, making a mid-stream architecture transition riskier. Against these, the Broadcom program's timeline is independent: Broadcom designs the chip, TSMC manufactures it, and OpenAI can deploy it alongside NVIDIA GPUs. The question is whether ~2.5 years from design initiation is enough. Industry evidence is mixed but leans toward 'tight.'

The Q4 earnings data is a weak signal for this market. The primary driver remains the chip development timeline, which no NVIDIA earnings call would illuminate (Broadcom and OpenAI control that information). The context update correctly identifies that the Broadcom custom chip program is structurally independent of the NVIDIA partnership. Key timeline analysis: if partnership initiated in mid-2024, design phase (12-18 months) puts first tape-out in early-to-mid 2026. Validation and yield optimization (3-6 months) puts volume production in H2 2026 at earliest. This is aggressive but within the range of possibility, especially given Broadcom's track record with Google TPU. The resolution criteria's OR clause is important: if TSMC begins manufacturing chips from a completed tape-out, that could resolve YES without full data center deployment. I shift marginally lower from the prior 33% due to the reduced urgency signal, but the core timeline calculation is unchanged.

I weight several factors that push probability slightly below the prior 33%. First, we are now 2.5 weeks closer to the resolution date with no new public information about the Broadcom chip -- each passing week without a tape-out announcement modestly reduces the probability that production deployment occurs within the remaining ~10 months. Second, the NVIDIA-OpenAI partnership deepening suggests OpenAI is getting what it needs from NVIDIA GPUs, reducing the commercial urgency for custom silicon deployment (though not the strategic rationale). Third, the Groq acquisition demonstrates NVIDIA is actively absorbing potential competitors, which may cause OpenAI to recalibrate the competitive landscape. However, I do not weight these heavily because: (a) chip programs of this scale rarely generate public updates before major milestones, (b) OpenAI has never publicly discussed the Broadcom timeline at all, and (c) the 10GW commitment represents a multi-hundred-billion-dollar strategic bet that would not be abandoned due to short-term NVIDIA supply adequacy.

Q4 earnings provide no direct Broadcom chip data. Deepening NVIDIA-OpenAI partnership (GPT-5.2-Codex on GB NVL72) and Groq acquisition modestly reduce urgency for custom silicon. NVIDIA supply expanding fast ($78B Q1 guide). But Broadcom chip timeline is independent -- Broadcom designs, TSMC manufactures, OpenAI can deploy alongside GPUs. ~10 months to resolution with no tape-out confirmation is the key constraint. Amazon Trainium took ~3 years. Slightly lower than prior 33% due to reduced urgency and time passage without milestone confirmation.

The Broadcom custom chip program remains on an independent timeline from NVIDIA's results. Resolution criteria allow tape-out + foundry manufacturing as sufficient (OR clause). If design started mid-2024, tape-out by mid-2026 is tight but possible given Broadcom's ASIC expertise. TSMC 3nm is mature. The 10GW commitment creates strong economic incentive to execute. No new negative signals on the program. NVIDIA's strong results are neutral to marginally negative for this market (reduced urgency). Maintaining probability near the prior range.

Time decay is the dominant factor for this update. We are now 2.5 weeks past the prior prediction with no new positive signals on the Broadcom chip timeline. The deepening NVIDIA-OpenAI partnership and Groq acquisition both marginally reduce the urgency for custom silicon deployment. ~2.5 years from design initiation to production deployment is aggressive by industry standards (Amazon Trainium took ~3 years). The absence of any public tape-out announcement, while not definitive, is a weak negative signal as we approach the resolution window. More likely production deployment occurs in 2027 than by end of 2026.