Yesterday Nvidia put Groq 3 LPX into production. Today the other side of that decode fight posted numbers, and IBM dropped an Apache agent you can pull on a box you already own.

OpenAI took Jalapeño to Hot Chips on Tuesday. First custom inference ASIC. Broadcom silicon. Not a training chip. Not a GPU replacement speech. TechCrunch has Richard Ho, OpenAI’s hardware lead, saying the results are a “very, very significant performance advance over state of the art.” The comparison set is Nvidia Blackwell GB200/GB300 on SemiAnalysis InferenceX. The deployment calendar is the actual product: small volumes at the end of 2026, real volume in 2027.

The number OpenAI wants you to remember is 1.5 to 1.9x more AI work per watt. The Verge has the rest of the vendor range: 1.7x to 3.6x lower end-to-end latency across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T. GPT-OSS is quoted around 1,400 tokens per second per user. DeepSeek R1 around 700. Those are OpenAI-run benches that SemiAnalysis verified in the lab. They are not an independent AA chart. Treat them like any first-silicon slide until someone else rents the rack.

The package is 700 watts on the datasheet and at or below 550 watts on the workloads they showed. Data Center Dynamics has the rack math: 128 chips per deployment, 1.7 exaflops of 4-bit compute, 27.5 TB of HBM4, 15.4 TB/s per package. A full pod is 2,048 ASICs. Inference only. Prefill and communication delays were the design target — keep the KV cache local, move less data, skip the usual latency-versus-throughput trade.

Ho was careful about the rest of the stack. OpenAI is not walking away from Nvidia. Jalapeño is the in-house inference SKU next to the GPUs they already buy. Yesterday’s Groq LPX story was Nvidia answering the same decode problem with silicon it acquired. Today’s story is a lab that still writes the biggest checks in the GPU market also shipping its own ASIC. Those two sentences can both be true. The interesting fight is 2027 volume, not a Tuesday chart.

If you are going to wait a year for any of this to show up as an endpoint you can call, a used workstation and 64GB of DDR5 still beat a reserved rack you cannot order.

IBM shipped the local agent, Apache 2.0, three sizes

While the silicon people argued about watts, IBM released Granite 4.2. 3B, 8B, and 30B. Decoder-only. Native 128K context, 512K on the 30B with the long-context extension. Apache 2.0. Hugging Face, Ollama, GitHub, LM Studio. No license theater.

The pitch is not “beats Claude.” The pitch is a reasoning-focused Granite you can host. Native chain-of-thought in think tags. Switchable full-thinking, non-thinking, and low-effort modes on one set of weights. The 8B and 30B went through an extra agentic RL block: terminal, web search, tools, software-engineering sandboxes. The 3B can call tools. It did not get the same specialized pass.

IBM’s own numbers, from the research post and the model cards: AIME25 at 78.33 on the 3B and 89.17 on the 30B. SWE-Bench Verified at 47.67 on the 8B and 57.00 on the 30B. Those are vendor benches. Ars Technica is the clean independent read, and it is mostly about why this family exists: predictable on-prem deployments, not a frontier leaderboard. IBM also shipped Granite Speech 5.0 Turbo CTC at 470 million parameters, no LLM backbone, aimed at laptops and call-center transcription. They claim it can transcribe three hours of audio in a second on a single H200. That is a speech SKU, not a coding agent.

Same-day, this is the stack you can actually run. Pull the 8B on a machine you already have. Route the hard jobs to the 30B. Leave the 3B on the edge box. If you need a second screen while Ollama finishes the pull, a cheap 24-inch monitor remains undefeated.

Apple sold the desk version of the same bet

Apple put M6 in the Mac mini and M5 Ultra in the Mac Studio. M6 is Apple’s first 2 nm chip: 12-core CPU, 12-core GPU with Neural Accelerators, dual 16-core Neural Engine, up to 32 GB unified, 170 GB/s. Mini starts at $899. M5 Ultra is the first quad-die M-series: up to 512 GB unified and 1.2 TB/s of bandwidth. Studio Ultra starts at $5,499. Preorder now, ships September 22.

Ars notes the quiet product: Thunderbolt 5 clustering is no longer a hobbyist hack. Apple is selling daisy-chained Studios as a way to run models that do not fit in one chassis. That is the same local-inference story as Granite, with a nicer case and a worse RAM tax. If you are actually going to cluster minis, budget Thunderbolt 5 cables before you budget another GPU rumor.

Nvidia reports fiscal Q2 after the close. Do not trade a Tuesday ASIC slide against a number that has not printed. The useful split for today is simpler:

The takeaway is not that Nvidia is finished. It is that inference silicon is no longer one company’s SKU list, and the local agent stack is no longer waiting on a cloud reservation. Jalapeño has benches. Granite has weights. The earnings print tonight will tell you whether the GPU factory still believes both of those sentences.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.