OpenAI has just announced its first series of custom chips for training AI models, developed in collaboration with Broadcom. This is not just another innovation — it is a declaration of independence from Nvidia, which has dominated the artificial intelligence market for years. What are the benefits and risks of this move? Does OpenAI have a chance to outperform Nvidia, or will it only deepen the industry divide?
In June 2026, OpenAI unveiled more than just a new language model — the first-ever chip designed exclusively for the training requirements of large AI models. The "OpenAI Chip v1" project, realized in partnership with Broadcom, is a response to rising costs, reliance on a single supplier, and growing geopolitical tensions in the semiconductor supply chain. Is this a game-changing breakthrough, or a risky investment that could delay AI development?
A new king of AI chips? "OpenAI Chip v1" specifications
According to information gathered by TechCrunch, OpenAI's first custom chip is a 3-nanometer processor, manufactured in Broadcom's facilities using advanced lithography. Key technical parameters include:
- Computational performance: approximately 2.2 TFLOPS in matrix operations (similar to the Nvidia H200), but with 30% lower power consumption — approx. 500W at full load.
- Memory: 64 GB of HBM3E memory, allowing large models to be stored directly on the chip and reducing latency associated with data transfer.
- Architecture: optimized for parallel processing and scalability — each chip can work with others in a cluster, which is crucial for training large-scale models.
- Release date: the first units are expected to reach OpenAI data centers in Q4 2026, with mass production planned for 2027.
This is not a modification of an existing GPU — it is a ground-up design aimed at maximum energy efficiency while maintaining computational power comparable to Nvidia's top-tier chips. As reported by SemiAnalysis, only about half of 3nm chips pass production tests in the first batch, meaning OpenAI and Broadcom had to overcome significant technological challenges.
Why did OpenAI decide on custom hardware?
The decision to create custom chips was not made lightly. According to Bloomberg, in 2025, OpenAI relied on 90% Nvidia infrastructure — primarily A100 and H100 GPUs. This dependency became a problem for several reasons:
1. Rising costs and supply constraints
- Nvidia raised chip prices by 40% in 2025, significantly increasing model training costs — for instance, GPT-4 required an investment of over $100 million in infrastructure.
- The company restricted chip access for non-business customers, which slowed down OpenAI's model development.
2. Supply chain control
Custom chips allow OpenAI to have full control over production, security, and development. This is particularly important in the context of geopolitical tensions — many Nvidia chips are produced in factories in Taiwan or China, which carries the risk of supply disruptions or industrial espionage.
3. Competitive advantage
OpenAI wants to become self-sufficient in terms of infrastructure, which is crucial for the development of next-generation models — such as GPT-5, which is expected to require up to 10 times more compute power than GPT-4. Custom chips also offer the opportunity for faster iterations and better hardware optimization for the specific needs of language models.
Reuters reports that the agreement with Broadcom involves multi-billion dollar investments in production, demonstrating the seriousness of the commitment. The open question, however, is whether OpenAI will manage to deploy its own chips before the release of new models — especially since Nvidia is preparing the Blackwell series for 2027.
Broadcom: technology partner or strategic player?
OpenAI's collaboration with Broadcom is not just a production contract — it is an alliance of two tech giants. Broadcom, known primarily for networking chips and high-frequency semiconductors, is entering the AI chip market with significant momentum.
According to the Financial Times, the collaboration model assumes that OpenAI owns the intellectual property rights, while Broadcom is responsible for production and a portion of the financing. This approach allows OpenAI to avoid the risks associated with building its own fabs, but simultaneously makes it dependent on a single manufacturer.
Broadcom's long-term goal, however, is to expand its product portfolio beyond networking chips and become a key AI chip supplier for other companies — such as startups or data centers. This could lead to new dynamic competition in the market, especially as companies like AMD, Intel, and Qualcomm are also investing in AI chips.
Market reaction: what is happening with Nvidia and the competition?
OpenAI's announcement had an immediate impact on the markets. According to CNBC, Nvidia's stock fell by 8% within 48 hours of the news about OpenAI's custom chips. This is a signal that investors fear competition that could threaten Nvidia's position as the leader in the AI GPU market.
What are the reactions of other giants?
- Google: Since 2016, the company has used its own TPU (Tensor Processing Units) to train models like PaLM 2. The TPU v5e is cheaper than Nvidia's offerings but less flexible. Can OpenAI outperform Google in energy efficiency?
- Amazon: It is developing Trainium chips for AWS, but their deployment is mainly limited to its own cloud. Amazon does not plan a broad external market release.
- Meta: It is testing its own AI chips, such as MTIA, but currently on a limited scale. The company is focusing primarily on optimizing existing models.
- Microsoft: It relies mainly on Nvidia but is investing in alternative solutions, for example through collaboration with AMD.
According to The Information, Nvidia responded by cutting H200 prices by 15%, which may be an attempt to maintain competitiveness. In 2027, the company plans to introduce the Blackwell series, which is expected to be twice as efficient as current chips. This means the battle for the AI chip market is just heating up.
Pros and cons of custom chips: what does OpenAI gain?
Owning custom chips is a huge opportunity, but also a risk. Here are the most important benefits and threats:
| Benefits | Drawbacks and risks |
|---|---|
|
|
According to Analog Devices, current Nvidia chips achieve about 1.9 TFLOPS at 700W, while the "OpenAI Chip v1" is expected to reach 2.2 TFLOPS at 500W. This is a real saving that could translate into lower model training costs. However, as SemiAnalysis emphasizes, 3nm chip production remains a challenge — even the biggest players, like Intel, have trouble scaling this process.
Technological and business challenges: what could go wrong?
Designing and producing custom chips is no small feat. Here are the biggest threats that OpenAI and Broadcom must face:
1. Production issues
Producing 3nm chips is a costly and risky endeavor. According to SemiAnalysis, only about 50% of 3nm chips pass production tests in the first batch. This means OpenAI and Broadcom had to account for the necessity of several design iterations before achieving stable production.
2. Software and compatibility
OpenAI must adapt its frameworks, such as PyTorch or Triton, to the new chips. There is a risk that lack of compatibility with existing models will slow down the chip deployment process. This is especially important as many enterprises use off-the-shelf solutions based on Nvidia.
3. Finance and ROI
Investment in chip design and production is a cost in the tens of billions of dollars. OpenAI is counting on a return on investment through long-term savings, but this is a long-term strategy. In the meantime, the company must find funds for further AI model development, which may prove difficult in the face of growing competition.
According to the Financial Times, OpenAI must also account for rising R&D costs, which could reach billions of dollars annually. This means the chip project is not just a technological innovation, but also a financial challenge.
What does this mean for the future of AI? The new chip war
OpenAI's decision is not just a step toward independence from Nvidia — it is the beginning of a new chip war. The AI industry faces a major choice: continue using off-the-shelf solutions like Nvidia chips, or invest in custom hardware.
If OpenAI succeeds, other companies will be forced to follow suit. This could lead to a market split among several key players, each offering their own solutions. However, this approach has its drawbacks:
- Increased costs for the entire industry: Competition between chip manufacturers could lead to higher prices for customers.
- Ecosystem fragmentation: Different chips will require different optimizations, which could hinder collaboration between models.
- Risk of technological waste: Investments in many different solutions could lead to a fragmentation of research efforts.
On the other hand, supplier diversification is an opportunity for greater flexibility and innovation. Companies like Google, Amazon, or Meta will be able to choose solutions best suited to their needs, which could accelerate AI development.
Ultimately, there doesn't have to be one winner in this war. As Dario Amodei emphasizes in his recent interview, "The human compass in the eye of the AI cyclone. A fascinating vision of the future according to Dario Amodei," the future of AI depends not only on technology but also on human capacity for adaptation and collaboration. Custom chips are just one of many tools that will shape this future.
Summary: revolution or a risky move?
The "OpenAI Chip v1" project is undoubtedly a breakthrough step that could change the rules of the game in the AI industry. Custom chips allow for greater control, lower costs, and faster model development, but they also come with high risk and massive investments.
Does OpenAI have a chance to outperform Nvidia? It is hard to predict for now. Nvidia remains the leader in the AI GPU market, and its new Blackwell series is expected to be even more efficient. However, OpenAI's decision is a signal that the AI industry is entering a new phase — a phase where control over hardware becomes critical.
One thing is certain: the chip war has only just begun. And its outcome may decide not only the future of OpenAI, but also who will shape the face of artificial intelligence in the next decade.
Comments