On June 25 Beijing time, OpenAI and Broadcom officially unveiled the first custom AI inference chip -- Jalapeño.

From project initiation to tape-out took just 9 months, breaking the development cycle record for high-performance ASICs. For context, this industry typically requires 18-24 months. Broadcom CEO Hock Tan claimed at the launch that this chip saves roughly 50% on inference costs versus traditional GPUs and reduces power consumption by about 30%, with engineering samples already running GPT-5.3-Codex-Spark stably in the lab.

It sounds great. But capital markets' reaction was telling: on the day of the announcement, Broadcom stock rose just 1.6%.

For context, when news of the partnership broke in 2025, Broadcom rose 15% that day. That 1.6% gain felt more like a polite round of applause -- "nice, now what?"

The Pepper Is Hot, But Who Pays?

The core question remains unresolved: money.

First-phase production costs for 1.3 GW of chip capacity are roughly $18 billion. Broadcom's position is clear: Microsoft must commit to purchasing 40% first, or financing doesn't start. And Microsoft? Hasn't nodded yet.

This isn't a small bill. OpenAI itself projects operational burn will exceed $200 billion by 2029. Meanwhile, they're pushing the $500 billion Stargate super-datacenter project. One Jalapeño is just the appetizer; the real bill is the full-course meal behind it.

To understand the game behind this, you need to return to a more fundamental question: why are AI companies suddenly making their own chips?

From "Using Chips" to "Defining Chips": The Paradigm Is Shifting

For the past thirty years, the chip industry logic was "chips define software" -- Intel built faster CPUs, and Microsoft and developers figured out how to use that compute. Nvidia built stronger GPUs, and AI labs built models around the CUDA ecosystem.

But today in 2026, that logic is reversing.

LLM computation patterns are highly structured and highly predictable: massive matrix multiplications, fixed attention mechanisms, regular memory access patterns. When you know exactly what workloads you're running, much of a general-purpose GPU's design becomes waste -- like knowing you only drive the same route to work every day but insisting on an all-terrain off-road vehicle.

Thus "models defining chips" becomes possible: AI companies know better than anyone what hardware their models need; they can reverse-engineer chips -- strip unnecessary circuits, optimize critical paths, build compute units tailor-made for the Transformer architecture.

Jalapeño isn't the first to the party. Google TPU has been at it for 13 years; Amazon Trainium/Inferentia is on its third generation; Microsoft's Athena and Meta's MTIA are both in progress. But what makes OpenAI special is: it's the first true "model company" making chips -- it doesn't sell cloud services, doesn't do advertising; its entire revenue and future are bet on the model itself.

This creates an enticing flywheel: AI designs chips → chips run AI → AI gets stronger → designs next-generation chips. Theoretically, once this flywheel spins, Nvidia's GPU hegemony would be pulled out from under it.

But Before the Flywheel Spins, You Have to Give It the First Push

Theory is beautiful; reality is brutal.

Google TPU has been going for 13 years, and Google Cloud still buys massive amounts of Nvidia GPUs. Why? Because custom chips have too little room for error.

Nvidia's CUDA ecosystem took twenty years to build -- millions of developers, countless software libraries, mature toolchains -- this isn't something "50% faster chips" can easily replace. And GPUs are general-purpose: run LLMs today, physics simulations tomorrow, movie rendering the day after. But ASICs are customized for specific workloads; what if the model architecture changes?

Think about it: in 2023 everyone was stacking dense Transformers; in 2024 MoE went mainstream; in 2025 state-space models emerged out of nowhere; in 2026 today nobody can say what the next architecture is. A chip that takes 18 months from design to mass production faces an industry that might upend itself every 6 months.

That's the paradox Jalapeño faces: it's optimized for GPT architecture, but what if GPT-6 uses a completely different architecture? Could this $18 billion investment become the next Itanium?

More subtly, there's the OpenAI-Microsoft relationship. Microsoft is OpenAI's largest investor, largest cloud provider, most important distribution channel. But now OpenAI making its own chips is essentially reducing dependence on Nvidia -- and also reducing dependence on Azure -- Microsoft can't fail to see this.

"You want me to spend $7+ billion buying your chips to weaken my own cloud business competitiveness?" -- Microsoft's hesitation is entirely understandable.

The Real Signal: It's Not About Chips, It's About Control

But Jalapeño's significance may go far beyond a single chip.

The AI industry in 2026 is entering a new phase: gaps in model capability are narrowing, but control over infrastructure is diverging. Who controls compute controls the speed of model iteration; who controls chips controls the compute cost curve.

OpenAI isn't competing with Nvidia; it's fighting for control over its own destiny. If inference costs can't keep falling, large-scale Agent commercialization will forever be a story on PPT slides -- you can't have every user spending $20/month to sustain an Agent that needs expensive GPUs to run.

What does a 50% cost drop mean? It means a product that originally needed $10 billion in compute to support now only needs $5 billion. It means the tipping point where Agents go from "premium toy" to "mass-market tool" might arrive early because of this one chip.

That's also why capital markets, though reacting coolly, haven't dared to truly go short. Everyone is waiting for one signal: will Microsoft sign that $7 billion purchase order?

If they sign, the flywheel starts, and the AI industry's cost curve gets redefined. If they don't, Jalapeño might be another bright but brief meteor -- like the countless "chip disruptors" over the past decade that were born at launch events and died on production lines.

Putting the Pepper in the Dish Takes Time

In the 24 hours after Jalapeño's launch, AI circle discussion shifted from "chip performance" to "commercial viability." That in itself is a signal: AI hardware is moving from the "technology narrative" phase into the "commercial narrative" phase.

9-month tape-out is indeed an engineering miracle; halving costs is tantalizing enough; reduced power consumption is the stuff of datacenter operators' dreams. But the semiconductor industry has never lacked miracles; what it lacks are miracles that sustainably make money.

For ordinary practitioners and observers, what Jalapeño really deserves attention for isn't specs, but the timeline: if everything goes smoothly, this chip will enter mass production in 2027, with large-scale deployment in 2028. By then, GPT-6 might have launched, Agents might be the mainstream interaction mode, and the AI competitive landscape might be entirely different.

But regardless, one trend is clear: AI companies are extending upstream, from models to chips to datacenters; vertical integration is becoming standard for top players. Just as Tesla wasn't content selling cars and also had to build batteries and make autonomous driving chips, OpenAI too is building its own vertical empire.

The jalapeño is spicy, but whether it burns Nvidia's mouth depends on whether Microsoft is willing to be the "first to eat the pepper."

See you tomorrow.