Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Bottom line: The September 2025 report was substantially correct, but “starting next year” is now outdated. OpenAI has publicly confirmed a collaboration with Broadcom for 10 gigawatts of custom AI accelerators and networking systems, targeted to begin deployment in the second half of 2026. Reuters later reported that engineering samples were running in OpenAI’s labs and had been tested with GPT-5.3-Codex-Spark. This is a custom infrastructure project—not OpenAI manufacturing chips in its own factory or preparing a retail Nvidia replacement.
What OpenAI actually announced
On October 13, 2025, OpenAI and Broadcom announced a collaboration to design and deploy 10 gigawatts of custom AI accelerators and networking systems. OpenAI said deployment was targeted to begin in the second half of 2026 and continue through the end of 2029. The announcement is available at OpenAI’s announcement.
The 10-gigawatt figure describes planned data-center system power capacity, not a stated number of chips. Converting it into chip count would require assumptions about accelerator power, rack design, networking, cooling and utilization.
What “OpenAI’s own chip” means
OpenAI designs the accelerator
OpenAI is contributing the workload requirements and accelerator design. An AI accelerator is specialized for machine-learning calculations; it is not necessarily a general-purpose graphics processor such as an Nvidia GPU.
Free tools Windows power users keep installed
One-click scans. No signup required.
#1 Best Overall
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
Broadcom develops and integrates the systems
Broadcom is more than a contract manufacturer in this arrangement. OpenAI says Broadcom will help develop and deploy the accelerators and network systems, which can include custom-silicon engineering, high-speed networking, rack integration and supply-chain coordination.
TSMC reportedly fabricates the silicon
Reuters reported that OpenAI sent its completed design to TSMC for fabrication. OpenAI therefore is pursuing custom chip design, not vertically integrated semiconductor manufacturing. The available reporting does not establish the process node, packaging technology, wafer volume or production yield.
The initial customer is likely OpenAI itself
Early reports described the chip as intended for OpenAI’s own operations. Neither the OpenAI announcement nor the reviewed reporting establishes a retail product, public customer-access program, pricing or a complete specification sheet.
Rank #2
- High-Performance AI Processing: The MX3 is designed to handle the most demanding AI computer vision workloads, delivering exceptional performance and efficiency.
- Flexible Integration: The MX3 can be easily integrated into your existing systems via its M.2 M-key form factor and support for Linux operating systems.
- Energy Efficient: The MX3 is designed to provide high performance while minimizing power consumption.
- Comprehensive Software Development Kit (SDK): The MX3 is supported by a comprehensive SDK that simplifies development and deployment.
- Hardware compatability: The MX3 is compatible with the PCI-SIG M.2 M-key 2280 Specification. It can be used with the Raspberry Pi 5 with a M-key 2280 HAT.
How the project moved from report to testing
| Date | Milestone | Status and source |
|---|---|---|
| 2023 | Exploration of custom AI chips | Reported by Reuters; an early investigation rather than a confirmed product. |
| September 2025 | First chip planned for 2026 with Broadcom and TSMC | Financial Times reporting cited by Reuters and Bloomberg: Reuters report and Bloomberg report. |
| October 13, 2025 | 10-gigawatt Broadcom collaboration announced | OpenAI targeted deployment from the second half of 2026 through the end of 2029. |
| June 24, 2026 | Engineering samples in OpenAI labs | Reuters reported samples running and tested with GPT-5.3-Codex-Spark: Reuters report. |
| Second half of 2026 | Targeted initial system deployment | Public target announced by OpenAI and Broadcom; samples are not the same as volume deployment. |
| End of 2029 | Targeted completion of announced deployment | Schedule stated in OpenAI’s collaboration announcement. |
Axios separately reported that OpenAI planned to use its first homegrown chip for customer queries later in 2026, but that remains reported timing rather than a fully documented operational rollout: Axios.
Why OpenAI wants custom silicon
- Lower serving costs: A small improvement in performance per watt or cost per token matters when serving enormous volumes of model interactions.
- Supply security: Owning part of the design can reduce exposure to shortages, allocation limits and pricing from a dominant supplier.
- Hardware-software co-design: OpenAI can tune memory movement, networking and scheduling for its models and software stack.
- Capacity planning: A dedicated platform may give OpenAI more predictable infrastructure expansion.
- Workload specialization: Stable, high-volume inference workloads can justify an accelerator that would be less useful for broad experimentation.
These goals imply complementing commercial GPUs, not immediately replacing them.
Why this is not an Nvidia replacement—at least initially
Nvidia GPUs bring mature CUDA software, libraries, developer tools, high-bandwidth memory configurations and extensive distributed-training support. Frontier-model training and research also require flexibility as architectures change.
Rank #3
- ✅Powered by 26 Tera-Operations Per Second (TOPS) Hailo-8 AI Processor. 2.5W typical power consumption
- ✅Scalable, enabling simultaneous processing of multi-streams & multi-models
- ✅Enabling real-time, low latency and high-efficiency AI inferencing on the edge devices
- ✅Supports TensorFlow, TensorFlow Lite, ONNX, Keras, Pytorch frameworks
- ✅Supports Linux and Windows. Supports the temperature range of -40°C to 85°C
A custom accelerator may perform well on selected inference jobs while being less suitable for new kernels, training experiments or workloads outside OpenAI’s design assumptions. Before it can displace substantial Nvidia capacity, OpenAI would need reliable yields, production volume, compiler and library support, networking at scale, and evidence that the economics hold under real traffic. Nothing in the available sources says OpenAI has stopped buying or using Nvidia hardware.
What the first chip is known to do—and what is not known
Reuters reported laboratory samples tested with GPT-5.3-Codex-Spark. Other coverage has characterized the first design as inference-focused, but OpenAI has not published a full architecture or benchmark package.
- Not publicly established: exact inference-versus-training workload split.
- Not publicly established: memory capacity, bandwidth, interconnect design or accelerator wattage.
- Not publicly established: process node, advanced packaging, wafer volume or yield.
- Not publicly established: whether the design is a standalone ASIC, a system-on-chip or one component of a rack-scale system.
- Not publicly established: cost per token, performance relative to Nvidia GPUs or the share of OpenAI traffic it will serve.
Has OpenAI started production?
The careful answer is that working samples were reported in OpenAI’s labs by June 24, 2026, while the companies publicly targeted initial deployment in the second half of 2026. Those are different milestones:
Rank #4
- 48GB AI graphics accelerator
- Design completion and foundry submission.
- Engineering samples.
- Qualification, reliability and software testing.
- Initial data-center deployment.
- Volume production and migration of meaningful traffic.
The reviewed sources do not establish exact mass-production volume, commercial yield, deployed chip count or whether the systems were already serving a significant share of customer requests as of August 18, 2026.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Potential effects on OpenAI customers
If deployment succeeds, OpenAI could gain lower serving costs, improved capacity planning, or better latency for models that fit the accelerator. Those gains could support pricing, margins or throughput, but no customer-facing price reduction or universal speed improvement has been verified. Performance may vary by model, region and workload while the platform is being introduced.
ChatGPT users and API customers cannot currently select or purchase this chip. It is an internal infrastructure component rather than a publicly offered accelerator.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Best Value
- Professional AI & Creator Workstation: AMD Radeon AI PRO R9700 GPU with 32GB GDDR6 is engineered for AI development, professional content creation, and compute-intensive workloads.
- Massive 32GB Memory Capacity: 32GB of GDDR6 memory on a 256-bit bus provides ample bandwidth for large AI models, 8K video editing, and complex 3D rendering.
- Advanced RDNA 4 with AI Accelerators: 64 Compute Units with 3rd Gen Ray Tracing and dedicated 2nd Gen AI Accelerators for groundbreaking AI performance and visual computing.
- Professional Blower Cooling: Efficient single blower design exhausts heat directly out of the chassis, ideal for multi-GPU workstation and server configurations.
- Enterprise-Grade Thermal Solution: Vapor chamber heatsink with industrial Honeywell PTM7950 thermal interface material ensures reliable cooling under sustained professional loads.
What it means for the AI-chip market
The project reinforces a broader shift toward custom silicon among companies with enormous and predictable AI workloads. Broadcom could gain a major custom-accelerator and networking engagement, while TSMC’s role illustrates that chip design and chip fabrication remain separate businesses. The deployment also increases demand for memory, advanced packaging, networking equipment, cooling and data-center power.
For Nvidia, the risk is selective workload displacement and greater bargaining pressure—not an immediate loss of its entire market. Nvidia’s software ecosystem and training hardware remain valuable even if OpenAI routes some inference through its own accelerator.
What to watch next
- Evidence that production systems, rather than samples, are operating in data centers.
- Public benchmarks showing performance, power and cost on defined OpenAI workloads.
- Disclosure of how much customer-query traffic moves to the accelerator.
- Details on memory, networking, packaging and manufacturing scale.
- Additional chip generations and whether they support training as well as inference.
- Any change in OpenAI’s Nvidia purchases, cloud capacity agreements or infrastructure disclosures.
Common misunderstandings
“OpenAI is manufacturing chips itself.”
Not according to the reported arrangement. OpenAI designs the accelerator, Broadcom helps develop and deploy the systems, and TSMC is the reported fabricator.
“Ten gigawatts means a known number of chips.”
It does not. The number refers to planned system power capacity and cannot be converted responsibly without undisclosed system specifications.
“Samples prove commercial success.”
Samples demonstrate technical progress, not production yield, reliability, operating cost or large-scale deployment.
“The chip will be sold to consumers.”
No reviewed source supports a retail product or general cloud-access program.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
