OpenAI's Donut Is a Side Story. Its Inference Chip Is the Real One.

OpenAI and Broadcom unveiled a custom LLM inference chip called Jalapeño in June, developed from initial design to tape-out in nine months — the fastest ASIC development cycle ever achieved in high-performance semiconductors. Early lab testing shows performance per watt substantially better than current state-of-the-art. The chip is designed to deploy at gigawatt scale with data center partners starting in 2026.

Then, two days ago, Bloomberg reported that OpenAI's first consumer device — a doughnut-shaped smart speaker without a display, battery-powered, camera-equipped, priced at $300-$400 — is close to shipping. The design comes from Jony Ive's studio, LoveFrom, which OpenAI acquired for $6.4 billion in 2025.

The market is going to fixate on the donut. It shouldn't. The chip is what actually matters for the investable thesis.

The inference transition, and who builds the road

The AI compute market is splitting along a structural fault line that isn't getting enough attention from most retail investors. Training still runs on general-purpose GPUs — Nvidia's CUDA ecosystem remains the default for building new models. But inference is where the economics are rewriting the hardware landscape, and inference favors custom application-specific integrated circuits, or ASICs, over general-purpose accelerators.

An ASIC is a chip designed from scratch for one specific workload — unlike a GPU, which is a flexible processor that can run many types of compute. The trade-off is that ASICs can't adapt to new architectures mid-cycle, but for the specific task of running large language models on inference workloads, they deliver 30–50% lower total cost of ownership and superior performance-per-watt. At the scale that hyperscalers and model companies operate, those margins compound into billions of dollars in savings.

This is why every major AI company is now building custom inference silicon. Microsoft announced its Maia 200 chip running GPT-class models internally. Google's TPUs, designed through Broadcom, are becoming a viable alternative to Nvidia for inference at scale. Anthropic trains across a mix of Amazon Trainium, Google TPUs, and Nvidia GPUs, and just signed a multi-gigawatt deal for next-generation TPU capacity starting in 2027.

Jalapeño puts OpenAI directly inside this transition. It is what OpenAI calls an "Intelligence Processor" — a blank-slate design optimized for modern LLM inference, not adapted from existing GPU architecture. The chip was accelerated by OpenAI's own AI models during the design process, creating a feedback loop where the company's software improves the hardware that will run its next generation of software.

Broadcom is the enabler you can buy

Here's the part that connects to a tradable thesis: Broadcom designs virtually every hyperscaler custom AI chip on the market.

Broadcom's AI semiconductor revenue hit $8.4 billion in Q1 FY2026 — a 106% year-over-year jump, exceeding the company's entire AI revenue for all of fiscal 2024 ($12.2 billion). The company guided to $10.7 billion for Q2 FY2026, implying 140% year-over-year growth. CEO Hock Tan declared a line of sight to more than $100 billion in AI chip revenue by fiscal 2027, backed by a $73 billion committed customer backlog and TSMC production capacity secured through 2028.

Broadcom now serves six major custom AI chip customers — up from three two years ago. Google is the anchor, with a long-term design partnership running through 2031 for TPUs. Meta is deploying multiple gigawatts of Broadcom-designed XPUs starting in 2027. Anthropic is accessing up to 3.5 gigawatts of TPU capacity through Google-Broadcom infrastructure. And OpenAI is Broadcom's newest major customer, developing its first custom inference ASIC through a reported $10 billion partnership targeting over 1 gigawatt of compute capacity starting in 2027.

Broadcom holds an estimated 70%+ market share in custom AI accelerator design services, compared to roughly 25% for secondary player Marvell. The company also captures adjacent networking revenue — Tomahawk switches and Jericho routers — estimated at $0.40 to $0.60 for every dollar customers spend on accelerators.

What this means: Broadcom is sitting at the intersection of the most important structural shift in AI hardware. It's not just selling chips. It's the design factory that enables every major AI company to escape Nvidia's pricing power for inference. As the custom ASIC market grows at a projected 44.6% compound annual growth rate through 2033, Broadcom is the company that profits regardless of which customer's chip wins — because they all use Broadcom.

The donut is an inference endpoint, not a product play

The consumer device matters only as evidence of OpenAI's full-stack strategy, not as a standalone hardware bet. OpenAI is pursuing a two-layer architecture for its own inference ecosystem: Jalapeño at the data center for heavy model runs, and a custom on-device chip for the consumer endpoint running compact local models.

Reuters reported last December that OpenAI is exploring custom processors for on-device inference — chips optimized for tight power constraints and single-user workloads, unlike server chips designed for parallel throughput across millions of users. The goal is to handle the majority of follow-up tasks locally, reducing the need to stream a user's entire life to the cloud.

The timeline here is important. Lighter, cloud-based devices — like the donut — are coming first. The privacy-sensitive, always-on devices relying on local inference chips will arrive later, potentially taking a few years for on-device technology to mature. OpenAI is building the architecture in stages.

This is the hardware-to-software value migration in reverse: OpenAI started as a software company and is now building the physical infrastructure underneath its models. The donut is just the visible tip. Jalapeño is the foundation.

The Apple lawsuit is a real risk signal

Here's the counter-signal. Apple filed a trade secrets lawsuit against OpenAI on July 10, accusing the company of "systematically stealing confidential data" including hardware product information, technical specifications, and supply chain details. The complaint names OpenAI, its foundation, io Products (Jony Ive's design firm, though Ive himself is not named), and several former Apple employees now at OpenAI.

Apple is seeking a preliminary injunction specifically aimed at stopping OpenAI from developing AI devices or products based on Apple's technology. The filing was escalated on August 4 with expanded allegations naming 11 additional former Apple employees. OpenAI responded on its blog on August 3, calling the lawsuit "careless, aggressive and oddly personal" and disputing Apple's timeline — including claims that Apple's lawyers initially emailed the wrong person and fabricated a discussion with OpenAI's general counsel.

The device launch, already pushed from the hoped-for H2 2026 to 2027, could face further delay if an injunction is granted. But critically, the lawsuit targets device development, not chip design. Jalapeño's data-center deployment trajectory is on a separate track, built with Broadcom and Celestica, and already has engineering samples running workloads at production target frequency and power in lab environments.

The financial picture: Broadcom vs. Nvidia

Broadcom trades at a $2.02 trillion market cap with a forward P/E of 104x, compared to Nvidia at $5.38 trillion and a forward P/E of 34x. Broadcom's revenue growth of 32% year-over-year and operating margin of 43% are excellent, and its gross margin on AI chip sales expanded to approximately 65%.

But P/E is the wrong lens for this comparison. Nvidia's valuation reflects its scale and entrenched position in training — the CUDA moat is real, and Nvidia still commands roughly 70% of the overall AI chip market. Broadcom's premium multiple reflects the option value of the custom ASIC transition: if the shift from general-purpose GPUs to custom inference accelerators continues at current pace, Broadcom's $100 billion AI revenue target in fiscal 2027 — which would represent more than 58% of the company's projected total revenue — would fundamentally reprice the stock.

The risk is execution. Scaling simultaneous ASIC designs for six hyperscaler customers while managing TSMC advanced-node capacity is a real operational challenge. And a rapid shift in AI architectures away from transformers could render some custom silicon suboptimal given the 18–24 month design cycles involved.

Where should capital go?

I believe OpenAI's full-stack hardware push — from Jalapeño at the data-center level to on-device inference at the consumer edge — is the clearest signal yet that the inference era is not a phase but the dominant architecture of the next three to five years. The market is still pricing Nvidia as though training workloads will carry the cycle indefinitely. They won't.

The investable thesis is Broadcom. It's the infrastructure layer that profits from every customer's move toward custom inference silicon. The $73 billion backlog, six major hyperscaler relationships, long-term agreements running through 2031, and 70%+ share in the custom ASIC design market create a compounding platform.

However, Broadcom trades rich. A forward P/E of 104x assumes that $100 billion AI revenue target materializes faster than most analysts' base cases suggest — and even management's own $100 billion line of sight may be back-half weighted into late 2027 or early 2028. I think a smaller allocation — not a 10% position but something closer to 3–5% — is the right sizing given the execution risk and the premium the market has already baked in.

The debate is not whether Broadcom stays important in the AI cycle. It is whether the return profile at this valuation is still more compelling than Nvidia at 34x earnings with a $5.38 trillion revenue base, or whether the custom ASIC transition justifies paying a significant premium for the asymmetry. In my opinion, the transition is real, but much of the upside is priced into Broadcom already. The allocation should reflect that — present and growing, but not dominant.

The break condition for this thesis is straightforward: if hyperscalers stop shifting inference workloads to custom ASICs and instead double down on Nvidia GPUs — or if Broadcom's execution on simultaneous multi-customer designs slips materially — the premium evaporates. If the shift accelerates as projected, the $100 billion run rate becomes the new baseline and the multiple compresses as growth becomes expected rather than exceptional. Either way, the custom silicon transition is the axis the entire AI hardware market is rotating around right now. The donut is just decoration.