OpenAI's Jalapeño Chip Beats Blackwell; Meta Halts Layoffs

OpenAI’s first custom silicon drew serious investor attention today after benchmark results put it ahead of Nvidia’s Blackwell on inference efficiency, arriving at a pointed moment: hours before Nvidia’s Q2 earnings report.

Key points

• OpenAI’s Jalapeño chip beat Nvidia Blackwell on key inference-efficiency tests in SemiAnalysis benchmarks, built in partnership with Broadcom.
• The result lands directly ahead of Nvidia’s Q2 earnings, making it one of the most closely watched reports of the year.
• Meta halted further workforce reductions that had been planned for November, following a 10% headcount cut in May.
• One analyst argues Meta’s cloud strategy remains unclear despite $130-$145 billion in planned 2026 CapEx, and suggests the company should consider renting out spare compute.
• AMI has ordered 9,000 Nvidia Vera Rubin chips, making it one of Asia’s earliest adopters of the next-generation platform.

Does Jalapeño change the Nvidia calculus?

OpenAI’s Jalapeño chip, co-developed with Broadcom, posted inference benchmarks that beat Nvidia’s Blackwell systems according to SemiAnalysis testing. The chip represents OpenAI’s first serious move into custom silicon, following a path already taken by Google (TPUs) and Amazon (Trainium/Inferentia).

The timing is deliberate or at minimum convenient. Publishing benchmark results the same day Nvidia reports earnings sends a clear signal to the market: hyperscalers and large AI labs are no longer content to be purely Nvidia customers. Custom silicon built for specific workloads, particularly inference rather than training, can undercut Nvidia on efficiency metrics that matter for per-query economics at scale.

For investors, the key question is margin trajectory. CNBC noted that custom AI silicon gaining ground across major tech companies represents a “new threat” to Nvidia’s margins. That framing is worth taking seriously, even if Nvidia’s near-term order book remains packed. Broadcom, which co-developed Jalapeño, is the quieter beneficiary here: each custom chip win for a major AI lab adds to its custom ASIC business, which has become a significant growth driver.

What Meta’s layoff pause and CapEx pile signal

Meta’s decision to stop further cuts is a notable shift. The company trimmed roughly 10% of its workforce in May, and additional reductions had been planned for November. Halting those suggests either that the efficiency push has achieved its targets, or that the company wants to retain talent as AI competition for engineers intensifies.

The more pressing strategic question comes from a Seeking Alpha analysis published today, which points out that Meta’s $130-$145 billion 2026 CapEx commitment is enormous relative to its current cloud monetization posture. The argument: Meta is building data center capacity at a scale that rivals hyperscalers, but unlike AWS, Azure, or Google Cloud, it has no public mechanism for renting that compute to third parties. If a meaningful share of that infrastructure sits underutilized, the return on capital from that spending becomes harder to justify.

Meta did not provide clarity on its cloud strategy in Q2, according to the piece. Whether the company eventually opens compute access to external customers, or simply consumes everything internally for AI model development and inference on its own platforms, has real implications for how investors should value the CapEx. GF Value currently pegs META as 32.4% undervalued, though that figure reflects a model-based estimate rather than a consensus target.

Nvidia’s earnings and the memory angle

With Nvidia’s Q2 results due imminently, one upstream story is worth watching. Server manufacturers have notified major data center customers including Microsoft, Google, and Oracle that prices for Nvidia-based AI systems will rise more than 15% for early 2027 shipments, driven in part by rising memory costs. Samsung and SK Hynix are reportedly gaining bargaining power as a result, with the increases affecting both Vera Rubin and Grace Blackwell-based systems. (We covered the initial server price hike story on August 23; today’s reporting adds the specific memory supplier dynamic.)

Meanwhile, AMI’s order for 9,000 Vera Rubin chips illustrates that demand for next-generation Nvidia hardware remains strong across Asia, even as the Jalapeño benchmark results circulate. Both things can be true: custom silicon is advancing, and Nvidia’s order pipeline is not about to empty.

Nvidia’s actual earnings numbers, and more importantly its forward guidance on data center revenue, will do more to set near-term direction for the broader AI chip sector than any single benchmark result. That report is the event to watch for the rest of today.

This article is for informational purposes only and does not constitute investment advice.