A single stone creates a thousand waves! In just nine months, OpenAI's self-developed ASIC chips have already surpassed Nvidia's Blackwell?
```
OpenAI's first self-developed chip, Jalapeño, has burst onto the scene, overturning the existing landscape of the AI chip industry with an astonishing speed—this is not only a product launch, but a signal: AI is reshaping chip design itself.
According to a WallstreetCN article and Bloomberg on August 25, OpenAI stated that Jalapeño leads Nvidia’s GB300 in two key metrics: AI workload per unit power and response speed. This chip was co-developed by OpenAI and Broadcom, designed specifically for the AI inference stage, and is expected to be put into practical use later this year. OpenAI chip lead Richard Ho says Jalapeño can achieve strong performance at a low power draw of 700 watts, which helps greatly reduce data center power costs.
According to semiconductor research firm SemiAnalysis, this chip took about only 16 months from design initiation to tape-out. The critical CoWoS packaging and tape-out node was completed in November 2025, just 9 months ago—which is well below the industry's usual 18 to 36 month cycle.
SemiAnalysis researchers personally visited OpenAI’s lab and used its proprietary benchmark suite InferenceX for practical testing of Jalapeño. The conclusion was straightforward: This chip outperformed all Nvidia, AMD, and Google chips they previously tested in performance-per-watt (perf/W).
Even more noteworthy is that this result was achieved when Jalapeño had not even enabled speculative decoding, prefill/decoding split deployment, or other optimizations, while the competing chips were running with their best configurations enabled.
No wonder even famed semiconductor analyst Dylan Patel remarked bluntly, "It's not just Blackwell that's toppled, even Nvidia’s Rubin chip has been surpassed!"
Analysts believe this result is a direct challenge to Nvidia’s market position, causing the market to reconsider the competitive landscape of AI chips.
Test Data: Jalapeño Crushes Blackwell on Multiple Key Metrics
SemiAnalysis’s test results reveal that Jalapeño’s advantage in inference efficiency is significant.
For the GPT-OSS 120B model, Jalapeño outputs about 1,459 tokens per second, while Nvidia’s GB200 yields only 535; in terms of end-to-end latency, Jalapeño completes a task in just 1.65 seconds, while GB300 takes nearly 6 seconds—a gap of 3.6x. In high-interaction scenarios—with GB300 pushed to its fastest decode speed (169 tokens/sec)—Jalapeño’s throughput is 104.3 times that of GB300.
On the core data center efficiency metric—tokens per megawatt per second—Jalapeño achieves about 53 million tokens/MW/s on the GPT-OSS model, while the GB200 NVL72 reaches only about 10 million.
SemiAnalysis points out that this metric is essentially equivalent to the number of tokens produced per joule, directly determining the revenue ceiling of a data center—under today’s compute environments constrained by power, this advantage is particularly critical.
At the system cost level, SemiAnalysis calculated power, cooling, networking, etc., and found Jalapeño’s total ownership cost per chip per hour is about $1.56, almost the same as the H100’s $1.55, while Nvidia’s Vera Rubin is as high as $3.61.
It’s worth noting that SemiAnalysis also raised several key caveats.
First, the model used for testing is not the most cutting-edge; Nvidia and AMD have already published AgentX-based results on much larger models (e.g., DeepSeek V4 Pro, Kimi K3), while Jalapeño has not undergone AgentX testing—which better reflects multi-turn, long-context production scenarios.
Second, a more appropriate comparison would be Nvidia’s Vera Rubin, which also uses HBM4, rather than Blackwell; Vera Rubin’s perf-per-watt is about 5.4x higher than GB200 NVL72, and its tokens-per-TCO (total cost of ownership) is almost equal to Jalapeño.
Third, Jalapeño is still in the engineering prototype stage, with volume production expected to ramp up starting in 2027.
AI-designed Chips: The GPT-Astra and Codex "Flywheel Effect"
A key reason why Jalapeño could be developed so rapidly is the deep involvement of AI tools. According to SemiAnalysis, OpenAI made extensive use of internal AI models during the chip design process, including the widely discussed GPT-Astra.
Andrew Curran, a user on social platform X, cited OpenAI information that GPT-Astra was deeply involved throughout the Jalapeño R&D process:
"Through Codex and GPT-Astra, the team tuned three open-source models that were not initially planned for Jalapeño mass production to high-performance status—within two months."
This means AI is not only accelerating chip design itself, but quickly expanding the range of models that the chip can support.
SemiAnalysis's data further quantifies this contribution:
AI-assisted design reduced SIMD unit area by 8%, matrix engine area by 10%, and improved timing and power consumption over the original version.
At the software level, OpenAI used an internal extended version of Codex to write Jalapeño’s kernel, with some kernel code up to 3,000 lines; for the most performance-critical attention and MoE modules, AI-generated code was 1.5 to 1.8 times faster than code written by top human engineers.
SemiAnalysis commented sharply: OpenAI models (e.g., GPT-5.6 Sol), running on Nvidia GPUs, are being used to design a chip that poses a real threat to CUDA’s moat—"Nvidia's own GPUs are now actively nurturing their potential successor."
Architecture Analysis: Why Is "General Purpose" Actually Faster?
It’s widely assumed that Jalapeño is a chip deeply customized for OpenAI's own models—but SemiAnalysis concludes the opposite: Jalapeño is a general-purpose AI inference chip capable of running various models and workloads—including the game "Doom" ported with Codex.
Architecturally, Jalapeño’s core design philosophy is "eliminate fixed latency." Unlike GPUs, which depend on complex memory hierarchies, Jalapeño directly binds compute cores to HBM slices, and synchronizes cores via a dedicated high-bandwidth collective network, greatly reducing memory access latency.
In addition, Jalapeño uses out-of-order (OoO) cores with L1 cache, rather than the commonly used software-managed scratchpads of other accelerators, which allows it to approach hardware theoretical peak performance in small-batch, low-latency scenarios.
In terms of memory bandwidth, Jalapeño uses HBM4, achieving 15.4TB/s of in-package memory bandwidth and 22 GB/W HBM bandwidth, compared to Nvidia Rubin’s 11.1 and GB300’s 5.71.
SemiAnalysis points out, Jalapeño’s HBM4 pin speed hits 10Gbps, slightly higher than Nvidia Rubin’s 9.6Gbps, and its HBM supplier may be Samsung.
Notably, Jalapeño chose not to do prefill-decoding split deployment (PDD). SemiAnalysis offers a detailed explanation:
In real production environments, parameters such as input-output ratio, concurrency, and cache hit rate are constantly changing. Fixed resource partitioning can lead to global utilization decline, while a homogeneous resource pool can dynamically respond to traffic changes, flexibly allocating between latency-sensitive requests and high-throughput batch processing.
CUDA Moat: Cracks Emerging?
SemiAnalysis’s report makes a weighty assertion: CUDA’s moat may already be dead.
The justification is the software ramp-up speed comparison. Nvidia Rubin's CoWoS tape-out was completed a month before Jalapeño’s, but so far, the only public Rubin benchmark data comes from CoreWeave's engineering sample—unlike OpenAI, Nvidia hasn’t allowed third-party lab testing. SemiAnalysis believes this disparity is not due to hardware, but software maturity.
Starting from scratch with a software stack instead let OpenAI shed historical baggage and make cleaner architectural decisions. OpenAI used its self-developed kernel programming language Gluon (built on Triton) and internal inference engine "Teacup," and completed kernel tuning with Codex. SemiAnalysis observed that in under two weeks, Jalapeño’s throughput at specific interaction speeds more than doubled; in eight days, the team extended tensor parallelism from TP8 to TP32, enabling full-cabinet scale deployment across racks.
Yet, SemiAnalysis also notes current testing only covers the relatively simple 8k1k workload, and has not completed AgentX multi-round long-context testing. For more complex agent workloads, components like routers and prefix caches will become new challenges.
Next Steps: B0 in Fab, 10GW Blueprint Unfolds
The Jalapeño story is far from over.
SemiAnalysis discloses that all public testing used A0 stepping samples; B0 stepping has now entered the fab, expected to improve performance per watt by another 25% over A0—the MXFP4 compute of a single B0 die will reach 13.4 PFLOPs, with TDP maintained at 700W.
At the system level, OpenAI, together with Celestica, has designed a full rack solution: each ASIC cabinet houses 128 Jalapeño chips, while a dual-cabinet system draws roughly 160kW, about the same as Nvidia’s GB300 double-width cabinet.
In terms of scalable deployment, a single expansion network domain can connect up to 2,048 Jalapeño XPUs across 16 racks. For mass production, SemiAnalysis expects ramp up to start in 2027, with the next milestone being a 100MW deployment.
On a grander scale, OpenAI and Broadcom have signed a 10GW custom accelerator partnership agreement—roughly equivalent to the full output of ten nuclear power units.
WallstreetCN article writes that OpenAI chip lead Richard Ho said the company has achieved real reductions in infrastructure power and cost—"this is only the first step." He also stressed that Nvidia remains an important partner, and that OpenAI’s demand for compute is so huge it will not abandon current suppliers in the short term.
Analysts point out that Jalapeño’s significance may not lie in how much it can replace Nvidia chips today, but in that it proves something that was widely doubted: an AI company, using AI tools, was able to produce a truly competitive chip in record time. This flywheel has already started turning.
Risk Warning and DisclaimerThe market has risks; investments must be made cautiously. This article does not constitute personal investment advice, nor does it take into account the special investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are appropriate to their own circumstances. Investment decisions based on this article are at your own risk. ```