OpenAI's First Custom AI Chip "Jalapeño":
50% Cheaper Inference, Built to Challenge Nvidia

If you are an AI developer, infra lead, or decision-maker tracking compute costs, wondering whether OpenAI custom silicon can dent Nvidia's grip or when inference bills will actually fall, this article unpacks the June 24, 2026 unveiling of Jalapeño—OpenAI and Broadcom's first purpose-built LLM inference ASIC: ~50% inference cost savings in early tests, TSMC 3nm, a record 9-month tape-out, Microsoft Azure deployment by year-end, and the Blackwell comparison, supply-chain split, and 10 GW target by 2029. Node options are on the NOVAKVM pricing page.

OpenAI is among the world's largest GPU consumers. Every ChatGPT answer triggers server-side inference—generating tokens from input. As GPT-4 and GPT-5 capabilities scale, inference has become the heaviest line item on the path to profitability. For years OpenAI relied almost entirely on Nvidia H100, H200, and Blackwell accelerators—general-purpose chips that waste significant compute in homogeneous LLM serving workloads.

Analogy: Nvidia GPUs are a Swiss Army knife; Jalapeño is a scalpel—LLM inference only, but extraordinarily efficient at that job.

  • Bigger models, bigger bills: Inference is OpenAI's largest opex category at hundreds of millions of daily users.
  • Architectural mismatch: GPUs serve gaming, training, simulation, and inference—flexibility costs efficiency at LLM-only scale.
  • Late but fast: OpenAI is the last major hyperscaler to ship custom silicon, yet claims the fastest advanced ASIC cycle on record.
  • Developer-side drag: Expected API price cuts and long-running local agents both depend on upstream inference efficiency—and on stable 24/7 run environments.
Hyperscaler custom AI silicon (public information)
Company Chip Primary use
Google TPU Training + inference
Amazon Trainium / Inferentia Training + inference
Microsoft Maia 100 Inference
Meta MTIA Inference
OpenAI Jalapeño (2026) Inference

Jalapeño is an ASIC—not a GPU. It does one job: LLM inference. No gaming, no training, no general compute. Richard Ho, who leads OpenAI's hardware program, says the chip was designed from scratch for LLM inference using insights into kernel execution, memory movement, networking, and serving patterns—and early tests run critical workloads near hardware theoretical limits.

  • Blank-slate design: Every decision targets Transformer inference patterns, not general compute retrofitted in software.
  • Minimize data movement: Memory bandwidth—not raw FLOPs—often bottlenecks inference; the architecture cuts unnecessary memory traffic.
  • Balanced compute, memory, and networking: Tuned for real LLM load ratios so utilization stays closer to peak.
  • Broadcom Tomahawk interconnect: High-performance cluster networking for multi-chip serving of frontier models.
  • Celestica system integration: Boards, racks, and server integration for volume production.
  • TSMC 3nm: Same generation as Apple M4 and Nvidia Blackwell—dense, power-efficient.
  • Already in the lab: Engineering samples run ML workloads including GPT-5.3-Codex-Spark at target frequency and power.
Jalapeño early performance metrics (vendor claims; third-party validation pending)
Metric Jalapeño (early tests) Baseline
Inference cost savings ~50% vs. current mainstream AI GPUs
Performance per watt Substantially above SOTA Per OpenAI blog
Absolute performance On par with Blackwell and Google TPU Per Broadcom CEO Hock Tan (Reuters)
Thermal behavior Better than expected OpenAI internal tests

Broadcom CEO Hock Tan (Bloomberg): "So far, Jalapeño has shown cost savings of roughly 50% compared to typical AI GPUs."

OpenAI president Greg Brockman: Jalapeño went from initial design to tape-out in 9 months; parts of the design flow were accelerated using OpenAI's own models (generation not disclosed; VentureBeat cites sources saying prior-generation OpenAI models).

Treat "50%" with skepticism: Still early lab data from Broadcom; a full technical report is promised in coming months; Azure-scale deployment and independent benchmarks are not done yet.

From initial design to tape-out took 9 months—OpenAI and Broadcom claim the fastest advanced high-performance ASIC cycle ever.

  • Deep hardware–software co-development: Model researchers and chip engineers worked together, avoiding guesswork-driven rework.
  • AI-assisted chip design: OpenAI models accelerated parts of the optimization loop.
  • Broadcom IP reuse: Mature silicon implementation and networking IP shortened the path to physical design.
Jalapeño supply chain roles
Role Company Responsibility
Architecture OpenAI LLM inference optimization, full-stack design direction
Silicon & networking Broadcom Implementation, Tomahawk networking, production support
Foundry TSMC 3nm manufacturing
System integration Celestica Boards, racks, server systems, volume integration
First deploy customer Microsoft Azure Data center rollout starting end of 2026

Broadcom is becoming the go-to custom ASIC partner for Google (TPU v5/v6), Meta (MTIA), and now OpenAI (Jalapeño). Public market data: Broadcom shares up ~18% YTD through May 2026 and nearly 7x since late 2022. Like all AI accelerators, Jalapeño needs large HBM pools—SK Hynix and Samsung benefit per Tan's comments.

Near term (end of 2026): Engineering samples in OpenAI labs; commercial rollout to Microsoft and other partners; priority workloads include ChatGPT, Codex, and API inference.

Mid term (2027): Volume production; Broadcom expects deployment above the prior 1.3 GW forecast; chip is "built for current and future LLMs across the industry" with possible external availability.

Long term (through 2029): OpenAI targets 10 GW of compute on custom silicon—roughly ten nuclear plants worth of power; next-gen chip expected 2028 with annual iterations; training silicon may follow (inference only today).

Can Jalapeño replace Nvidia? Short-term matrix
Dimension Jalapeño today Nvidia moat
Workloads Inference only Training + inference
Software Specialized stack CUDA ecosystem, millions of developers
Flexibility ASIC efficiency, architecture-change risk General GPU adaptability
Capital ties Custom silicon + Broadcom $30B Nvidia direct investment in OpenAI (Feb 2026)
Strategy Supply diversification, pricing leverage Vera Rubin platform, ecosystem depth

The real win is leverage—not replacement. Even 20–30% of inference on Jalapeño saves real money, strengthens procurement negotiations, and reduces single-vendor risk. Same playbook as Google, Amazon, and Microsoft: not abandoning Nvidia, but not being fully beholden either.

"Nobody wants to be beholden to Nvidia." — Ben Barringer, global tech research head at Quilter Cheviot (via CNN)

Nvidia's stock reaction was muted; markets see training dominance as safe near term, with custom silicon as a structural long-term pressure.

  • Inference economics reshape business models: If 50% holds in production, ChatGPT and API pricing can fall further, clarifying OpenAI's margin path and lowering the floor of the AI price war.
  • Full-stack AI becomes the bar: OpenAI now designs chip architecture, kernels, memory, networking, scheduling, deployment, and product experience—not just models. Competition shifts from "best model" to "best end-to-end efficiency."
  • Semiconductor winners and losers: Winners include Broadcom, TSMC, and HBM suppliers; pressure on Nvidia (inference share) and AMD (weak inference ASIC story).
Key people (public launch information)
Name Role Launch role
Greg Brockman OpenAI co-founder & president Public launch; full-stack infra framing
Richard Ho OpenAI hardware lead Technical architecture leadership
Hock Tan Broadcom CEO Blackwell-class performance, 50% savings claims
Sam Altman OpenAI CEO Overall strategy; compute independence push

  • Is Jalapeño an Nvidia GPU replacement? Not today. Inference only; training still Nvidia-centric—complementary, not substitutive.
  • Is the 50% savings real? Early Broadcom lab data via Bloomberg; no independent validation yet; full report coming in months.
  • What changes for end users? Cheaper ChatGPT/API and possibly faster responses if savings scale; AI becomes more affordable long term.
  • Why "Jalapeño"? No official explanation; OpenAI has a food-naming tradition—possibly signaling heat or market disruption.
  • Will other AI companies get access? Language suggests industry-wide LLM focus; near-term priority is OpenAI infrastructure.
  • Next generation? Roadmap points to 2028 for gen two, then yearly cadence.
  • Nvidia stock impact? Limited immediate move; training moat intact; custom silicon is a slow structural headwind.

Timeline:

  • Oct 2025: OpenAI and Broadcom announce custom chip partnership
  • Feb 2026: Nvidia $30B direct investment in OpenAI (Vera Rubin compute deal)
  • Jun 24, 2026: Jalapeño public launch; engineering samples running in labs
  • End of 2026: First commercial Azure and partner deployments
  • 2027: Volume production; deployment above 1.3 GW
  • 2028 (expected): Second-generation chip
  • 2029 (target): 10 GW on custom silicon

  1. Split training vs. inference budgets: Jalapeño covers inference only; frontier training stays on Nvidia—account for both separately.
  2. Discount vendor benchmarks: Treat 50% as a hypothesis until OpenAI's technical report, Azure production data, and MLPerf-class third parties weigh in.
  3. Watch the supply chain: Broadcom, TSMC, HBM vendors, and Celestica integration signal production readiness and supply risk.
  4. Reprice API assumptions: Better inference economics may pull ChatGPT/Codex pricing down—design multi-vendor and caching now.
  5. Plan for full-stack competition: Model quality alone won't win; chip-to-product efficiency compounds—match agents and CI to stable, low-latency environments.
  6. Lock in 24/7 dev environments: Run coding agents and long inference jobs on always-on Apple Silicon nodes; see Help Center and order page.
  • Launch date: June 24, 2026 (OpenAI + Broadcom)
  • Inference cost savings: ~50% (Broadcom CEO early lab claim; unverified)
  • Tape-out cycle: 9 months (vendor claim: record for advanced ASIC)
  • Process: TSMC 3nm
  • Long-term compute target: 10 GW by 2029

Running Codex, Cursor agents, and long API soak tests on a laptop hits sleep interrupts, memory ceilings, no true 24/7 uptime, and cross-border network jitter. Shared cloud VMs add Metal overhead and uncontrolled macOS versions. For production iOS/macOS builds, AI agent automation, and always-on coding assistants, NOVAKVM Mac Mini bare-metal cloud rental is usually the better fit: dedicated Apple Silicon, multi-region low-latency nodes, and flexible daily/weekly/monthly terms so engineering spend tracks a predictable execution environment as inference costs fall upstream.

Sources used at writing time; if upstream pages change, treat the links as authoritative.

OpenAI official launch blog

TechCrunch: OpenAI unveils its first custom chip, built by Broadcom

VentureBeat: Jalapeño sped up with OpenAI models

The Next Web: Jalapeño and Nvidia

Bloomberg: OpenAI and Broadcom unveil Jalapeño

Axios: OpenAI moves beyond Nvidia