The Custom Silicon Breakaway: OpenAI’s Jalapeño, Anthropic’s Samsung Gambit, and the Talent Heist

July 15, 2026

On June 24, 2026, OpenAI’s Greg Brockman and Sam Altman stood alongside Broadcom CEO Hock Tan, holding the first engineering samples of a chip that hadn’t existed nine months earlier. The chip is called Jalapeño. A week later, reports surfaced that Anthropic had entered early-stage talks with Samsung to manufacture its own custom AI processor on a 2nm process. And somewhere in between, one of the earliest engineers on OpenAI’s hardware team quietly left to join Anthropic.

If the AI wars had a hardware front, it just opened wide.

The Nine-Month ASIC That Shocked the Industry

Jalapeño is an inference-optimized ASIC, purpose-built for running large language models—the kind of chip that strips away the general-purpose baggage of GPUs and focuses entirely on transformer workloads. It’s fabbed on TSMC’s 3nm node, designed in partnership with Broadcom, and integrated into systems by Celestica. Engineering samples are already running production-class models like GPT-5.3-Codex-Spark at target power and performance levels.

But the real headline isn’t the silicon. It’s the timeline. OpenAI went from concept to tape-out in nine months. That’s roughly one-third the time it took Google’s first TPU, and an order of magnitude faster than traditional chip houses. The company calls it “the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors.” Broadcom CEO Hock Tan added fuel, claiming Jalapeño matches Nvidia’s Blackwell and Google’s TPUs on inference, delivering roughly 50% cost improvements per token or per kilowatt.

On Hacker News, where the story piled up 428 points and 279 comments, the reaction was a mix of awe and raised eyebrows. “Nine months from design to tape-out is either revolutionary or misleading,” one commenter wrote. The 50% figure, currently backed by zero independent benchmarks, didn’t help. “Show me the numbers,” another added. Skepticism, for now, is the default.

The development speed was enabled by two factors. First, OpenAI used its own models to assist in chip design and optimization—a recursive loop that some engineers called poetic and others shrugged off as marketing. Second, the hardware team is led by Richard Ho, a former core engineer on Google’s TPU program, who joined OpenAI in late 2023 and has built a team of about 40 people.

Anthropic’s Counter: Samsung, 2nm, and a Talent Heist

Anthropic’s chip effort is at a much earlier stage, but the signals are impossible to ignore. The company is in talks with Samsung Electronics to manufacture a custom AI chip, eyeing Samsung’s 2nm process and advanced packaging capabilities. Discussions are preliminary—no detailed design work has begun, and Anthropic could still walk away. But the mere existence of the talks, combined with a key talent grab, suggests the company is serious.

That talent grab came in the form of Clive Chan. He was the second hire on OpenAI’s custom chip program and helped build Jalapeño from the ground up. In June 2026, just before the chip’s public debut, Chan announced his departure on X. “I want to climb a new mountain from the bottom again,” he wrote. Within a day, his bio read Anthropic.

The move stung, not just because of the timing, but because it was part of a pattern. Over a dozen former OpenAI staffers now hold key roles at Anthropic, including co-founder Andrej Karpathy, who joined the pretraining team earlier this year. If OpenAI had aggressive non-compete clauses, one Hacker News commenter noted, “Anthropic might not even exist.” The talent front is as heated as the silicon front.

Anthropic already runs Claude on three chip families—Nvidia GPUs, Google TPUs, and Amazon Trainium—and has committed to roughly 4.5 gigawatts of TPU capacity starting in 2027. Its annualized revenue run rate crossed $30 billion earlier this year, more than tripling from late 2025. At that scale, owning even a slice of the silicon stack becomes a spreadsheet no-brainer.

Why the GPU Divorce Is Getting Expensive

Inference at scale is brutal. GPUs offer massive parallelism and shine at training, but during inference they suffer from chronic underutilization—wasting power, die area, and capital. As OpenAI’s models grew larger, that cost-to-performance gap became untenable. Jalapeño is tuned specifically for LLM kernels, memory movement, and serving patterns. It’s less flexible than a GPU, but potentially far more efficient on the narrow workloads that dominate OpenAI’s API calls and ChatGPT traffic.

The supply chain angle is equally urgent. Nvidia controls an estimated 74% of the AI chip market. For frontier labs, that means supply constraints, cost volatility, and a roadmap dictated by someone else. Custom silicon offers a path to direct control over the infrastructure that underpins the entire product.

Nvidia isn’t sitting idle. It reportedly struck a high-dollar licensing deal with inference startup Groq, effectively blocking OpenAI from a potential alternative partnership. It’s also providing credit backing for the massive data center loans that keep its own GPUs in play. The message: you can try to leave the ecosystem, but the exit is expensive and the gatekeeper is well-armed.

The 10GW Bet and the Ghost of the Cold War

OpenAI and Broadcom’s ambition extends far beyond a single chip. The partnership targets 10 gigawatts of OpenAI-designed accelerators by the end of 2029, with the first phase consuming 1.3 GW and costing an estimated $18 billion in chip production alone. The full program could run north of $180 billion.

The chosen site for this compute campus is a former uranium enrichment plant in Piketon, Ohio—a place that once fueled Cold War nuclear weapons. Now it will house AI inference hardware, powered primarily by new natural gas plants. It’s a messy, capital-intensive symbol of how strategic this compute buildup has become. SB Energy, a SoftBank subsidiary, is developing the power infrastructure, with the first 800 MW slice targeted for 2028.

For Anthropic, the path is less defined. Samsung’s 2nm yields hover around 55%, below the 60% threshold typically needed for mass production. But with TSMC’s 2nm capacity booked until 2029, Samsung may be the only fab willing to play ball on the timeline Anthropic wants. It’s a calculated risk—one that could either accelerate Anthropic’s independence or leave it with a very expensive silicon souvenir.

The Flexibility Trap and What Comes Next

Every inference ASIC is a bet that model architectures won’t change too radically, too fast. If the industry pivots away from Transformers, Jalapeño’s tightly optimized logic becomes a liability. Community sentiment on this was split. Some pointed out that inference workloads are relatively stable; others warned that “ASIC” and “future-proof” rarely appear in the same sentence without irony. Anthropic’s chip, still in the concept phase, could bake in more programmability—or could be designed for a specific Claude architecture that hasn’t been locked down yet.

For now, no independent benchmarks exist for Jalapeño. Broadcom’s claims remain claims. Anthropic’s Samsung talks could evaporate. The 10 GW campus is a decade-long undertaking. But the direction of travel is set. Frontier AI labs are no longer content to be software-only tenants on Nvidia’s land. They’re pouring foundations, hiring chip architects, and poaching each other’s talent to own the ground beneath their models.

A Hacker News commenter captured the shift better than any press release: “The direct change with Jalapeño is that OpenAI goes from solely buying computing power to participating in defining that computing power.” What that power ultimately costs—and who ends up holding it—is the question neither chip has answered yet.