OpenAI Says Its New Chip Outperforms Nvidia's Blackwell As Nvidia Prepares Earnings Release

OpenAI unveiled its first in-house accelerator, Jalapeño, claiming 1.5x–1.9x more inference throughput and 1.7x–3.6x lower end-to-end latency (with 2.1x–4.1x faster ultra-low-latency interactive inference) versus Nvidia Blackwell GB200/GB300 racks. The article stresses caveats—Jalapeño is inference-only, excludes speculative decoding, and does not disclose system power; deployment starts later in 2026 with volume in 2027. Into Nvidia’s fiscal Q2 close, consensus clusters around revenue of ~$91–$92B (about 2x YoY) and EPS ~$2.09, with prior guidance ~$91B ±2%; Wedbush expects a beat, but investors are watching whether custom silicon from hyperscalers becomes a structural drag.

Analysis

This is less a near-term NVDA earnings shock than a gradual deterioration in its pricing power. The immediate pool of capital is likely to treat the reveal as headline risk, but the economic effect is deferred: the first real threat is not lost 2025 revenue, it is a slower 2026-2027 replacement cycle as hyperscalers use custom inference silicon to negotiate lower rack prices and richer networking attach. That favors AVGO most directly, with CLS as a quieter beneficiary of the systems buildout; the real loser is the implied terminal multiple on NVDA if the market starts capitalizing a smaller share of future AI capex.

The key timing distinction is that tonight’s print should still reflect orders booked months ago, so the stock can rally on a beat and still underperform over the next 1-3 months if management sounds defensive on hyperscaler concentration, margin, or 2026 visibility. The falsifier is simple: if Nvidia re-anchors the street with a stronger-than-expected data-center guide and explicitly frames custom silicon as complementary rather than substitutive, the threat gets pushed out another cycle. If instead the call leans on China exclusions or avoids commenting on customer self-supply, that’s a sign the market will start paying for growth less generously.

Consensus is probably missing that this is a multiple-risk story before it is a unit-volume story. Even if NVDA continues to ship, the mix of inference workloads migrating in-house can compress gross margin and reduce the scarcity premium, while AVGO monetizes the design wins and networking content around those deployments. The cleanest expression is to own the picks-and-shovels layer and be patient on the downside in NVDA; the risk is that investors overreact tonight to a chip that is not yet at scale and built for inference only, which would create a better short entry after the post-earnings bounce fades.

Key Decisions for Investors

• Long AVGO / short NVDA as a 3-6 month pair trade, entered on any post-earnings rally in NVDA; thesis is that custom silicon compresses Nvidia's multiple before it meaningfully dents 2025 revenue, while Broadcom captures the design-win monetization. Stop if NVDA re-accelerates 2026 data-center guide or AVGO booking commentary softens.
• Do not short NVDA into the print aggressively; wait for the call. If the stock gaps up on a beat but management is vague on hyperscaler concentration or custom-silicon pressure, initiate a tactical short or put spread targeting the next 4-8 weeks of multiple compression.
• Overweight AVGO on weakness rather than strength-chasing; the 6-18 month setup is cleaner because it is the beneficiary of customer self-supply, not the company facing substitution risk. Use any broad semis selloff to add with a tighter stop than NVDA because expectations are rising quickly.
• Add CLS as a smaller, higher-beta beneficiary of systems integration buildout, but only as a watchlist trade until 2027 production timing becomes firmer; the thesis depends on conversion of design wins into volume, not the press-release language.
• For hyperscaler exposure, prefer GOOGL or AMZN only if the market starts rewarding internal silicon as margin defense; otherwise treat them as the medium-term winners from lower compute cost, not as immediate catalysts.