AI Weekly Malaysia

OpenAI claims its 700W 'Jalapeño' ASIC, co-developed with Broadcom, delivers up to 1.9x throughput per kilowatt and 3.6x lower latency compared to Nvidia's 1,400W GB300 flagship GPU, based on first-published benchmarks. The chip runs at half the power envelope of Nvidia's part.

Why: If these numbers hold under independent testing, inference cost-per-query could shift significantly toward custom ASICs over Nvidia GPUs — builders pricing AI features should watch whether OpenAI passes efficiency gains downstream via API pricing, and whether Broadcom-backed custom silicon accelerates the trend of large AI labs going in-house on chips rather than buying Nvidia.

Apple announced a new Mac Mini starting at $899 with an M6 chip (12-core CPU, 12-core GPU, 16GB RAM, 256GB storage), shipping September 22 with macOS 27 and Siri AI. Apple claims 4x AI performance over the M4 model, and notes the Mac Mini has grown popular for running local AI agents like OpenClaw and Hermes. An M5 Pro variant starts at $1,699 with 24GB RAM and 512GB storage; the previous $599 base model has been discontinued.

Why: If you're evaluating hardware for local AI agent workloads, the new base Mac Mini now ships with 16GB RAM standard (up from 8GB on older base models) at $899, but the cheapest entry point has risen from $599 to $899. The M5 Pro variant at $1,699 with 24GB RAM may be the better value for serious local inference work. Decide whether to pre-order now or wait for benchmarks against the M4 before committing.

OpenAI revealed benchmark results for its custom inference chip, Jalapeño, at Hot Chips, showing higher tokens-per-user and throughput-per-kilowatt than Nvidia Blackwell on Semianalysis's InferenceX benchmark. Developed with Broadcom, Jalapeño targets prefill and communication bottlenecks by keeping KV cache local. OpenAI's Richard Ho said small-volume deployment arrives end of 2026, with meaningful scale in 2027.

Why: If you build on OpenAI's API, Jalapeño could eventually translate into lower latency and lower per-token costs once it scales in 2027, but nothing changes today. Teams heavily dependent on OpenAI inference costs should watch whether promised efficiency gains pass through to API pricing, rather than assuming Nvidia-based alternatives will remain the default.

OpenAI revealed its custom 'Jalapeño' inference accelerator at Hot Chips, developed with Broadcom: 128 chips per system delivering 1.7 exaFLOPS with 27 TB of HBM. On SemiAnalysis' InferenceX benchmark suite, Jalapeño showed 1.5x-1.9x higher peak throughput and 1.7x-3.6x lower end-to-end latency versus unnamed competitors across GPT-OSS-120B, DeepSeek R1, and Kimi K2.5. Volume production targets 2027, and the chip is inference-only—OpenAI still plans to use Nvidia and AMD GPUs for training.

Why: These are pre-production, vendor-selected benchmarks for chips that won't ship at volume until 2027, so nothing changes operationally today. But if the 1.5-3.6x latency advantage holds, builders heavily dependent on OpenAI's API for real-time agent workloads could see meaningful cost-per-token and latency improvements downstream—worth tracking but not worth re-architecting around yet.

Apple announced M6 (2nm, 12-core CPU, 12-core GPU, Dual 16-core Neural Engine, 170GB/s bandwidth) in the new Mac mini, and M5 Ultra (quad-die, up to 36-core CPU, 80-core GPU, 1.2TB/s bandwidth) in the new Mac Studio. M5 Ultra targets large local AI model workloads with 50% more unified memory bandwidth than M3 Ultra.

Why: If you're evaluating local AI inference hardware, the M5 Ultra's 1.2TB/s unified memory bandwidth and quad-die architecture are the specs to compare against GPU alternatives for running large models on a desktop. For everyone else, M6's 2x Neural Engine uplift is incremental and doesn't require any action today.

Eigen Labs' Darkbloom project has become a paid inference provider on OpenRouter, letting Apple Silicon Mac owners earn an estimated $120-200/month per machine by contributing idle compute. The network currently has 250 machines online, has served ~4.5B tokens, and reports $102K ARR. The inference engine runs in a single hardened Swift process using mlx-swift-lm, with macOS kernel-level protections (PT_DENY_ATTACH, Hardened Runtime) to prevent prompt/response data leakage, routed through a Go coordinator in an AMD SEV-SNP confidential VM.

Why: If you have idle Apple Silicon Macs (M1 MacBook Pro or better, Mac mini preferred), you can sign up at darkbloom.dev to monetize them as inference nodes—but cloud-rental arbitrage is explicitly banned. The privacy architecture is worth studying if you build distributed inference or edge AI systems: the single-process hardened Swift model with kernel-level debugger denial is a concrete pattern for running untrusted inference workloads safely.

OpenAI reports first benchmark results for 'Jalapeño,' its custom AI inference chip, claiming it sits on the Pareto frontier for both GPT-OSS 120B and DeepSeek R1 670B across multiple operating points. The post highlights improvements in tokens per user, throughput per kilowatt, and time-between-tokens, and notes the chip was designed using AI and architected so AI could program it.

Why: These are OpenAI's own self-reported benchmarks on OpenAI's own chip running OpenAI's own models—treat the claims as marketing until independent third-party measurements appear. If even partially true, cheaper and faster inference would lower API costs and latency for anyone building AI agents or SaaS products on OpenAI, but no pricing change or API access detail is announced here, so there is nothing to switch or decide yet.

OpenAI published first benchmark results for Jalapeño, its first custom inference chip, claiming it widens the lead over previous-best TBT (time between tokens) on the InferenceX public benchmark. Written by Sarah Friar, the piece frames OpenAI's strategy as a vertically integrated stack spanning data centers, chips, frontier models, developer platform, and consumer/enterprise products, where each layer reinforces the others.

Why: This is a vendor announcement about proprietary silicon that most builders cannot directly adopt or change because of. The only practical signal is that if Jalapeño's TBT gains hold up under independent testing, OpenAI API latency for token streaming could improve—relevant if you ship latency-sensitive agent or chat UIs on OpenAI. Until independent benchmarks confirm, no action is required.

Kyoto University demonstrated a silicon carbide (SiC) transistor that operates at 600°C, using standard ion implantation and a bottom-gate design to address leakage and voltage drift problems. The fabrication approach is compatible with existing semiconductor manufacturing processes, avoiding the need for specialized equipment.

Why: For builders working on industrial IoT, automotive, or deep-well/geothermal sensing hardware in Southeast Asia's harsh environments, this signals that extreme-temperature-rated SiC transistors may become cheaper and more accessible as they can be produced on standard fab lines. However, this is early-stage university research with no commercial timeline, so no near-term design decisions should change based on it.

Apple announced new M6 and M5 Ultra Apple silicon chips, debuting in refreshed Mac Mini and Mac Studio form factors. The article contains no specifications, benchmarks, pricing, or release dates beyond the headline announcement.

Why: No actionable detail is provided—no specs, performance numbers, or pricing to evaluate against current Intel/AMD/Ryzen AI developer workstations. Developers considering a Mac for local ML inference or on-device agent development should wait for benchmark data before deciding; this announcement alone gives nothing to act on.