Based on supply-chain sourcing (TSMC allocation, HBM contracts, CoWoS packaging), SemiAnalysis argues that JalapeñO isn't trying to out-engineer Nvidia's core GPU architecture but instead targets transformer inference with long context and heavy KV-cache pressure. On that narrow workload, a purpose-built ASIC with the right memory bandwidth and interconnect topology can beat Blackwell on tokens-per-dollar and tokens-per-watt — the metric that actually determines gross margin at ChatGPT scale.
By submitting the SemiAnalysis teardown to HN with the framing 'Better Than Nvidia Blackwell,' bmulholland surfaces the argument that a targeted inference ASIC can outperform a general-purpose GPU on the workloads that matter economically. The 474-point score and 303 comments suggest the framing resonated with the developer audience.
SemiAnalysis frames JalapeñO as the completion of an industry-wide pattern: Google (TPU), Amazon (Trainium/Inferentia), Meta (MTIA), Microsoft (Maia), and Anthropic (co-designing on Trainium 2) have all concluded that paying Nvidia's margin on inference is untenable at billion-query-per-week scale. OpenAI going custom with Broadcom is the last shoe to drop, signaling that vertical integration on silicon is now table stakes for frontier labs.
The teardown emphasizes that JalapeñO follows the exact Broadcom playbook from Google's TPU v1-v5 program: physical design, SerDes IP, and packaging expertise sit at Broadcom, while microarchitecture, compiler stack, and workload targeting sit at OpenAI. This division of labor — not just OpenAI's ambition — is what turns a custom ASIC from a moonshot into a shippable product, and it explains why Broadcom is quietly becoming the second-most-important company in AI infrastructure.
SemiAnalysis published a detailed teardown of what OpenAI and Broadcom have been quietly building: a custom AI accelerator internally nicknamed JalapeñO, positioned as competitive with — and in specific inference workloads, better than — Nvidia's Blackwell generation. The reporting draws on supply-chain sourcing (TSMC allocation, HBM contracts, CoWoS packaging slots) rather than an official OpenAI announcement, which is characteristic of how these chip programs leak before they ship.
The headline claim is not that OpenAI has out-engineered Nvidia's core GPU architecture. The claim is narrower and more defensible: on the specific workload OpenAI actually runs at scale — transformer inference with long context and heavy KV-cache pressure — a purpose-built ASIC with the right memory bandwidth and interconnect topology can beat a general-purpose GPU on tokens-per-dollar and tokens-per-watt. That's the metric that matters when you're serving a billion ChatGPT queries a week and your gross margin is a function of how many tokens each rack produces.
Broadcom is the co-design partner, following the same playbook it ran with Google on TPU v1-v5. The physical design, SerDes IP, and packaging expertise sit at Broadcom; the microarchitecture, compiler stack, and workload targeting sit at OpenAI. TSMC N3 (or a variant) is the process node, and the part reportedly leans hard on HBM3E and Broadcom's in-house optical/electrical interconnect rather than trying to match Nvidia's NVLink domain size directly.
Every major AI lab has now signaled the same conclusion: at scale, renting Nvidia GPUs to serve your own model is a structural margin problem. Google figured this out a decade ago with TPU. Amazon has Trainium and Inferentia. Meta shipped MTIA. Microsoft has Maia. Anthropic is co-designing with Amazon on Trainium 2. OpenAI going custom with Broadcom completes the set — there is no longer a top-tier lab whose long-term inference roadmap is "buy more Hoppers."
The reason is arithmetic. Nvidia's data-center gross margin sits around 75%. When you're OpenAI or Anthropic and inference is your dominant cost of revenue, every dollar you spend on H100s or B200s is roughly 75 cents of margin you're handing to Jensen. A custom ASIC that's even 30% cheaper per served token on your specific workload — even accounting for the multi-billion-dollar NRE and multi-year tape-out cycle — pays for itself in a single generation of scale.
What's genuinely interesting in the SemiAnalysis piece is the argument that inference silicon and training silicon are diverging fast enough that a specialized inference chip can beat a general-purpose GPU without matching its FLOPs. Training wants dense matrix multiply, huge NVLink domains, and flexibility across model architectures you haven't invented yet. Inference — especially long-context transformer inference — is bottlenecked on HBM bandwidth for KV-cache reads and on interconnect latency for tensor parallelism. Optimize for those two things and you can strip out area, power, and cost the GPU spends on flexibility you don't need.
The skeptical counter is worth stating plainly. Custom silicon programs are graveyards. Graphcore, Cerebras, Groq, SambaNova, Tenstorrent — every one has a technically impressive part and a commercial story that's underwhelmed relative to the hype. Google's TPU is the only clear commercial success, and TPU is more than a decade of accumulated compiler, framework, and workload co-design. OpenAI is betting it can compress a decade of TPU-scale software investment into a two-to-three-year window, riding on Broadcom's physical design and its own concentrated inference workload. The upside if it works is enormous. The downside is a multi-billion-dollar write-down and a return to Nvidia dependency on worse terms.
Community reaction on Hacker News (474 points, top of the front page) split predictably. The bulls read it as confirmation that the Nvidia moat is thinner than the stock price implies — inference is the volume workload, inference is where custom silicon wins, and OpenAI is the customer with the most concentrated demand and the deepest pockets to build for it. The bears point out that SemiAnalysis has been wrong before, that "co-designed with Broadcom" is doing a lot of load-bearing work, and that first-silicon perf numbers from any custom program tend to age poorly against Nvidia's next generation on a two-year cadence.
If you're not an AI lab, the immediate practical impact is close to zero. You will not be buying JalapeñO chips. You will not be renting JalapeñO instances on a competitor cloud. OpenAI is building this to serve its own models cheaper, not to sell you silicon.
The second-order effects are what matter. The moment OpenAI's inference cost per token drops meaningfully below what it costs Nvidia-dependent competitors to serve equivalent quality, API pricing follows. You'll see it as another price cut on GPT-class models, another expansion of the free tier, and another squeeze on the margin structure of every startup whose business model assumes today's inference prices as a floor. If you're building on the API, the direction of travel is that per-token costs continue their multi-order-of-magnitude decline — plan capacity and pricing accordingly, don't lock yourself into contracts that assume 2025 prices are stable.
For infrastructure teams running open-weight models on your own GPUs, the signal is slightly different. The reason hyperscalers are all building inference ASICs is the same reason your `vllm` and `sglang` deployments are memory-bandwidth-bound, not compute-bound. If you're sizing hardware for llama-class or Qwen-class inference, the lesson from JalapeñO is the same one TPU has been shouting for years: HBM bandwidth and interconnect matter more than TFLOPs for your workload. An H200 or MI300X often beats an H100 on real throughput for reasons that have nothing to do with peak compute.
The interesting question isn't whether JalapeñO ships — it will, on some timeline, in some form. The question is whether the second generation ships. TPU became a real weapon at v3-v4, not v1. If OpenAI has the organizational patience to iterate on a chip program across three generations while simultaneously shipping models and running a consumer product, the Nvidia hedge becomes a real Nvidia alternative. If it doesn't, JalapeñO joins the long list of first-silicon parts that looked good in the launch deck and quietly disappeared. Either way, the era of "Nvidia is the only game in town" for frontier inference is ending — the only question is how fast.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.