OpenAI Broadcom Custom Chip Challenges NVIDIA AI Dominance

On June 24, 2026, OpenAI and Broadcom unveiled a custom AI chip design—codenamed 'Jalapeño'—that will power the next generation of ChatGPT inference. The New York Times reported the chip design will consume 10 gigawatts of electricity at scale, an amount that could power millions of households, signaling a radical rethinking of AI infrastructure.

OpenAI and Broadcom revealed a custom AI chip design on June 24, 2026, targeting inference workloads for ChatGPT and future models.
The chip is designed to run on 10 gigawatts of electricity—enough to power millions of homes—highlighting the massive energy demands of AI at scale.
This partnership directly challenges NVIDIA’s dominance in AI chips, potentially reshaping the hardware supply chain.
Energy infrastructure emerges as the next critical bottleneck for AI scaling, with implications for data center operators and utilities.

Why Is OpenAI Building Its Own Chip Now?

According to the New York Times, the chip design—internally called 'Jalapeño'—is optimized for inference, the process of running trained AI models to generate responses. This is a strategic pivot: while NVIDIA’s GPUs excel at training, inference at ChatGPT’s scale requires different trade-offs. OpenAI said the custom design could reduce inference costs by up to 40% compared to NVIDIA’s H100 clusters, a figure The Register confirmed based on leaked internal documents. The timing is critical: OpenAI is burning cash on inference compute, and owning the silicon gives it margin control.

How Does the Broadcom-OpenAI Chip Compare to NVIDIA’s Offerings?

OpenAI and Broadcom Cut NVIDIA Out of the Loop

Broadcom brings its expertise in custom ASIC design, previously used for Google’s TPUs. The Jalapeño chip uses a systolic array architecture tailored for transformer models, unlike NVIDIA’s general-purpose tensor cores. The Register reported that early benchmarks show 2x throughput per watt on GPT-4 class models. NVIDIA’s upcoming Blackwell architecture, however, promises 4x improvement over H100, so the gap may narrow. The key advantage for OpenAI is vertical integration: no margin stack, no supply chain dependency.

Feature	OpenAI/Broadcom 'Jalapeño'	NVIDIA H100	NVIDIA Blackwell (projected)
Architecture	Systolic array (custom)	Tensor core (general)	Tensor core (enhanced)
Target workload	Inference-optimized	Training + inference	Training + inference
Throughput per watt (GPT-4 inference)	2x H100 (estimated)	1x (baseline)	4x H100 (projected)
Power per chip	700W (estimated)	700W	1000W (estimated)
Supply chain	TSMC 3nm, Broadcom-designed	TSMC 4nm, NVIDIA-controlled	TSMC 3nm, NVIDIA-controlled
Verdict	Winner in inference cost per query	Still dominant for training	TBD—depends on real-world power efficiency

What Does 10 Gigawatts Mean for AI Data Centers?

The 10-gigawatt figure is staggering. According to the New York Times, this is enough to power 3 million U.S. homes—more than the entire city of Chicago. The Register noted that no single data center today can handle that load; it would require multiple gigawatt-scale facilities, likely co-located near hydroelectric or nuclear plants. This creates a new bottleneck: energy infrastructure. OpenAI will need to partner with utility companies and possibly build its own renewable energy farms. The cost of power alone could exceed $1 billion per year at current industrial rates.

Who Wins and Who Loses From This Chip Design?

Winners: Broadcom gains a marquee customer and validates its custom ASIC business against NVIDIA. OpenAI reduces dependency on NVIDIA and gains control over its cost structure. TSMC wins more advanced packaging orders. Utility companies like NextEra Energy and Dominion Energy could see massive new demand.

Losers: NVIDIA loses a high-volume inference customer, though it still dominates training. AMD and Intel, which are still competing for AI workloads, face an even tougher market as custom chips proliferate. Hyperscalers like Microsoft and Google, who rely on NVIDIA, may feel pressure to accelerate their own custom chip efforts.

Is This the End of NVIDIA’s AI Chip Monopoly?

Not yet. NVIDIA still commands 80%+ of the AI chip market, and its CUDA ecosystem is deeply entrenched. But this move by OpenAI signals that the largest AI companies are willing to invest billions to escape vendor lock-in. According to The Register, OpenAI’s chip will enter production in late 2027, meaning NVIDIA has at least 18 months to respond with better pricing or a dedicated inference chip. The real test will be whether OpenAI can scale production and deployment without delays.

My thesis: The Jalapeño chip is a credible threat to NVIDIA’s inference monopoly, but the 10-gigawatt power requirement reveals that energy, not silicon, is the new moat in AI.

In the short term, this announcement puts pressure on NVIDIA to lower prices or risk losing more inference business. Broadcom’s stock will rise on the news, while NVIDIA’s may dip modestly. In the long term, the real winner is the energy sector: AI companies will become the largest new electricity consumers since the 1990s dot-com boom. OpenAI’s chip design is smart engineering, but its success hinges on solving a grid-scale problem that no chip can fix. I believe we will see OpenAI announce a major renewable energy partnership within 12 months to address the power gap.

By Q2 2027, OpenAI will announce a 5-gigawatt renewable energy deal with a major utility like NextEra Energy to power its Jalapeño data centers.
By Q4 2027, NVIDIA will release a dedicated inference chip, potentially called 'InferX,' with 3x the throughput per watt of H100, specifically targeting custom ASIC competitors.
By 2028, at least two more hyperscalers—likely Microsoft and Google—will announce custom AI chip partnerships with Broadcom or Marvell, further eroding NVIDIA’s market share in inference.

June 2026
Jalapeño chip design unveiled
OpenAI and Broadcom announce custom AI chip with 10-gigawatt power requirement.
Late 2027
Production start
Jalapeño chips enter production at TSMC 3nm.
2028
Full deployment
OpenAI deploys Jalapeño chips across data centers, targeting 40% cost reduction.

Timeline of Events

June 2026: OpenAI and Broadcom unveil Jalapeño chip design; New York Times reports 10-gigawatt power requirement.
Late 2027: Expected production start for Jalapeño chips at TSMC 3nm.
2028: Full deployment in OpenAI data centers, potentially reducing inference costs by 40%.

Inference Cost per Million Tokens (Estimated)

Estimated inference cost per million tokens: H100 ($0.60), Jalapeño ($0.36), Blackwell ($0.30 projected). Chart shows Jalapeño undercutting H100 but still trailing Blackwell on pure cost—though vertical integration gives OpenAI other advantages.

Article Summary

OpenAI and Broadcom’s custom chip is a direct assault on NVIDIA’s inference stronghold, but energy infrastructure is the hidden constraint.
The 10-gigawatt power figure is a wake-up call: AI scaling is now an energy problem, not just a compute problem.
NVIDIA still dominates training, but the inference market is fragmenting—and custom ASICs will capture 30% of it by 2028.
Broadcom emerges as a key player in AI hardware, challenging NVIDIA’s ecosystem lock-in.
Investors should watch utility partnerships and energy announcements as closely as chip benchmarks.