OpenAI just announced Jalapeño — its first custom inference chip, built in partnership with Broadcom. That single move tells you more about the future of AI infrastructure than any benchmark comparison or funding round.
For years, Nvidia has been the undisputed king of AI hardware. Every major AI company — OpenAI, Google, Meta, Anthropic — builds on Nvidia’s GPUs. But that dominance is cracking, and the companies building the most powerful AI systems in the world are starting to build their own chips to break free.
Why Build Custom Silicon at All?
The motivations come down to three converging pressures that every scaling AI company eventually hits.
Cost. Nvidia’s GPUs are expensive. When you’re running inference for hundreds of millions of ChatGPT users, the price per token directly impacts your bottom line. Purpose-built chips can slash inference costs by optimizing for the specific math your models actually do, rather than carrying the overhead of a general-purpose GPU designed for everything from gaming to scientific simulation.
Control. When your entire business depends on hardware you don’t make, you’re at the mercy of someone else’s release schedule, pricing decisions, and supply constraints. Owning your silicon roadmap means you plan around your needs, not Nvidia’s priorities.
Performance. General-purpose GPUs are phenomenal at training — the parallel math involved in teaching a model. But inference — serving that trained model to millions of users — is a different workload with different bottlenecks. Custom chips tuned specifically for inference can dramatically improve throughput and latency.
The Apple Precedent
The best template for what’s happening came from Apple, not the AI industry. When Apple announced its transition from Intel to custom Apple Silicon in 2020, skeptics questioned whether a phone company could build competitive laptop processors. Five years later, Apple Silicon dominates the laptop market with performance-per-watt metrics that Intel and AMD still can’t match.
The lesson was straightforward: when you control both the hardware and the software stack, you can optimize at levels that pure hardware buyers never reach. Apple didn’t just build a chip — it rebuilt its entire software ecosystem around that chip’s strengths.
OpenAI appears to be following a similar playbook. Jalapeño isn’t just a chip — it’s infrastructure designed specifically for how GPT models run inference. The integration with Broadcom gives OpenAI the manufacturing muscle, while the custom design ensures the silicon matches OpenAI’s actual workload patterns.
Who Else Is Building?
OpenAI isn’t making this move in isolation. The custom silicon wave is sweeping through every major tech company with serious AI ambitions.
Google has been running custom Tensor Processing Units (TPUs) since 2016. The latest TPU v6 (Trillium) powers Gemini and Google Cloud’s AI services. Google’s TPUs aren’t just a hedge against Nvidia — they’re a core part of Google’s AI infrastructure strategy, and they’ve been proven at scale for nearly a decade.
Apple continues to expand its custom silicon beyond laptops and phones. The M-series chips now power everything from the Mac Studio to the Vision Pro, and Apple’s investment in on-device AI processing is directly enabled by controlling its chip architecture.
SpaceX is developing custom chips for its aerospace applications — a reminder that the demand for specialized AI hardware extends far beyond consumer tech. When you’re running autonomous docking algorithms on a spacecraft, off-the-shelf GPUs aren’t an option.
Amazon (through AWS) has its Trainium and Inferentia chips, designed specifically for ML training and inference in the cloud. Microsoft is reportedly developing its own AI chips as well, though details remain scarce.
What This Means for Nvidia
Nvidia isn’t going away anytime soon. The company’s CUDA software ecosystem — the programming framework that makes Nvidia GPUs usable for AI — is deeply entrenched. Thousands of AI libraries, frameworks, and tools are built on CUDA, and that ecosystem creates massive switching costs.
But the threat isn’t immediate revenue loss. It’s more subtle and potentially more damaging: the gradual erosion of Nvidia’s position as the default choice.
When every major AI company is building its own chips, Nvidia’s pricing power weakens. When custom alternatives prove viable for inference workloads, Nvidia’s largest customers — the ones driving billions in GPU purchases — start diversifying their hardware supply chains.
For investors, the signal to watch isn’t Nvidia’s quarterly revenue (which will likely stay strong for years). It’s the rate at which companies announce custom silicon programs. Every new entrant is a vote of no confidence in the single-supplier model.
The Philippine Angle
For Filipino developers and businesses, this shift has practical implications. Local data centers and cloud providers are already weighing whether to invest in Nvidia’s latest hardware or wait for custom alternatives that might offer better economics.
If custom silicon delivers on its promise of lower inference costs, the price of running AI workloads in the Philippines could drop significantly. That matters for startups building AI-powered products, for enterprises deploying chatbots and automation, and for developers experimenting with large language models on limited budgets.
The flip side is more complexity. Instead of optimizing for one hardware platform (Nvidia), developers may need to target multiple chip architectures. That’s the trade-off when a monopoly breaks up — more choice, but more work to support those choices.
What This Changes for AI Development
The custom silicon wave reshapes the AI landscape in three concrete ways:
First, inference costs will drop. Competition drives prices down. When OpenAI can run inference on its own chips instead of renting Nvidia’s, the savings get passed along — either as lower API prices or as reinvestment in model capabilities.
Second, the software stack fragments. The era of “write once, run anywhere on Nvidia” is ending. AI frameworks will need to support multiple hardware targets, which means more testing, more optimization, and more deployment complexity.
Third, new players emerge. Custom silicon lowers the barrier for companies that want to build AI-specific hardware. Startups like Groq (with its inference-optimized LPU) and Cerebras (with its wafer-scale chip) are proving that Nvidia isn’t the only path to high-performance AI compute.
The Bigger Picture
The custom silicon wave isn’t about replacing Nvidia overnight. It’s about reducing a single point of failure in the AI supply chain. For an industry building the most powerful technology in human history, having one company control the foundational hardware was always a fragile arrangement.
The real question isn’t whether custom chips will succeed — Apple already proved that model works. It’s how fast the rest of the industry can follow that playbook. And when they do, the economics of AI will shift in ways that benefit everyone building on top of these systems.
For developers in the Philippines and around the world, that’s worth paying attention to. The hardware layer is about to get a lot more competitive, and the effects will trickle down to every API call, every inference request, and every AI-powered product you build.