Custom AI Silicon: What Anthropic-Samsung and OpenAI-Broadcom Mean for Your AI Product Strategy

In July 2026, two data points landed within weeks of each other. OpenAI unveiled Jalapeño, its first jointly designed inference chip, built with Broadcom. The chip targets GPT inference specifically and delivers better performance per watt than general-purpose accelerators. Around the same time, reporting emerged that Anthropic had begun early talks with Samsung to explore manufacturing a custom chip on Samsung's two-nanometer process. Anthropic hired Clive Chan, who was an early member of OpenAI's chip team, as an additional signal.

Neither move came out of nowhere. Google has run custom TPUs in production for years. Amazon built Trainium and Inferentia for AWS workloads. Meta designed its first inference chip in 2023. What is new is that the companies building the models that most AI products run on, OpenAI and Anthropic, are now on the same path.

The economics are straightforward at scale. Anthropic's annualized revenue run rate surpassed $30 billion in 2026. At that revenue level, even a 20 percent improvement in inference cost per token compounds into hundreds of millions of dollars annually. Custom silicon that is optimized specifically for a provider's model architecture, rather than for a general matrix multiplication workload, can realistically deliver that. OpenAI's Broadcom relationship and Anthropic's existing long-term deal with Google and Broadcom for roughly 3.5 gigawatts of TPU compute starting in 2027 both point in the same direction: providers are moving from renting compute to owning or co-designing it.

Inference costs for frontier models have been falling at roughly 50 to 70 percent per year since 2022. GPT-4-class capabilities that cost $60 per million tokens in early 2023 now cost under $5 in 2026, and Gemini-class capabilities on the Google stack cost even less through TPU economics. Custom silicon accelerates this curve.

When a provider runs inference on general Nvidia H100 or H200 GPUs, a significant portion of each compute cycle goes to operations that are not specific to transformer inference: memory bandwidth that generic tensor cores were not optimized for, power draw from cores that sit idle between attention layers, and memory hierarchy that was not designed around the key-value cache access patterns that dominate LLM serving. A chip purpose-built for a specific model architecture eliminates most of this waste.

The Google TPU story is the clearest precedent. Google began using custom TPUs in 2016. By 2023, TPU v4 delivered roughly 2x better performance per dollar than comparable Nvidia A100 clusters for transformer workloads. Gemini's competitive pricing in 2026 is partly a reflection of that infrastructure lead. OpenAI and Anthropic are now on the same trajectory, with OpenAI roughly two years ahead.

Cost is one side of the silicon story. The other is latency. Custom chips optimized for a specific model's architecture can reduce time to first token and tokens per second significantly. Jalapeño's performance-per-watt advantage translates directly into faster responses at the same cost, or the same response speed at lower cost.

For most AI product teams in 2026, latency is the primary UX constraint. Users accept imperfect outputs more readily than they accept waiting. Sub-one-second time to first token has become the threshold for most conversational applications. Streaming helps, but the underlying token generation speed determines how quickly streams start. As providers move to custom silicon, this threshold will become easier to hit at larger context windows and higher concurrency.

The conventional worry about custom silicon is lock-in: if OpenAI's economics improve because of Jalapeño, and Anthropic's improve because of TPUs and eventually Samsung silicon, does that tighten your dependency on specific providers? The answer is more nuanced than a simple yes.

Custom silicon does not change the API surface that your product calls. You still call an HTTP endpoint with a model name and receive a response. The chip change is invisible to your application layer. What it changes is the economic durability of that provider: a provider that controls its inference cost structure is a more stable counterparty for a long-term product bet than one that is at the mercy of Nvidia GPU availability and pricing.

The more relevant dependency risk is model-specific optimization. When OpenAI runs Jalapeño, the chip is tuned specifically for GPT model architectures. Porting that chip to run Anthropic models efficiently would require significant re-engineering. This means providers' chips will be most beneficial for their own models, and switching from one provider to another will not automatically benefit from the same silicon advantages. The performance and pricing gap between providers may widen as each optimizes its own stack.

Custom silicon benefits frontier API providers most directly, but open-weight models gain an indirect advantage. When OpenAI and Anthropic reduce their per-token costs through custom silicon, they gain room to lower API prices. That pricing pressure forces open-weight managed inference providers to compete. The effect is a floor on how cheap hosted AI inference can get, but it also compresses the gap between self-hosted and hosted.

For AI PMs, this matters when evaluating the open-weight path. The traditional argument for self-hosting open-weight models was cost: at high enough token volumes, self-hosted Llama 4 was cheaper than GPT-4.1 at API pricing. As frontier API prices fall with custom silicon, that crossover point shifts. The cost argument for self-hosting weakens at lower volumes, and the operational complexity of self-hosting stays constant. The open-weight case gets stronger on control, privacy, and compliance grounds, but weaker on pure cost at mid-range volumes.

The Anthropic-Samsung chip will not ship tomorrow. Jalapeño is already in production for OpenAI, but its pricing effects will roll out over months, not days. The strategic moves for AI PMs today are about positioning for the infrastructure environment of 2027 and 2028, not about immediate vendor changes.

The bigger picture: the AI infrastructure layer is maturing fast. Two years ago, the defining constraint for AI products was model capability. Today it is cost and latency. In two years, as custom silicon normalizes across every major provider, the constraint shifts to product differentiation, data, and workflow integration. Teams that plan for that shift now will find themselves in a much stronger position than teams that are still optimizing for a cost environment that will not exist.