OpenAI spent a decade renting its compute from Nvidia and Microsoft. On June 24, 2026, it started building some of its own. OpenAI and Broadcom unveiled JalapeÃąo, OpenAI’s first custom-designed chip, in a joint announcement that put a name and a set of numbers behind months of rumors about an in-house processor codenamed “Titan.” Broadcom CEO Hock Tan said early testing showed roughly 50% better cost efficiency than standard AI GPUs, with performance he described as “on par with Nvidia’s Blackwell chips and Google’s tensor processing units,” according to CNBC’s coverage of the announcement.
The claim matters because it lands at the exact moment Nvidia’s data center business is putting up numbers few companies in history have matched, and because Broadcom is now the common thread behind custom silicon programs at Google, Meta, and ByteDance. It also lands eight months after the original OpenAI-Broadcom agreement, the October 2025 pact to co-develop 10 gigawatts of custom AI accelerators, sent Broadcom shares up roughly 9% in a single trading day, according to CNBC, a preview of how closely investors have been tracking every step of this partnership since. JalapeÃąo is not a training chip and it is not shipping at scale yet. But it is the clearest signal so far that OpenAI wants to stop being just a customer of AI hardware and start being a designer of it. Here is what was actually announced, what it means for the AI chip market, and what to watch through 2028.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What JalapeÃąo Actually Is
JalapeÃąo is what OpenAI calls an “Intelligence Processor,” a chip built specifically to run trained AI models rather than train them. That distinction, inference versus training, is the whole point of the design. Training a frontier model requires chips that can move enormous batches of data through repeated calculations over weeks. Inference is different: it means serving one user’s ChatGPT prompt, or one developer’s API call, as fast and as cheaply as possible, over and over, billions of times a day. That is not theoretical anymore: as of June 2026, JalapeÃąo is already running GPT-5.3-Codex-Spark workloads at production-level performance targets, Yahoo Finance reported, giving OpenAI its first real-world read on whether the chip’s economics hold up outside a lab.
Broadcom handled the silicon implementation, networking, and connectivity. Systems integrator Celestica built the boards and racks around it. Broadcom CEO Hock Tan has drawn a sharp line between JalapeÃąo and the GPUs it’s meant to compete with, a distinction he elaborates on further below, describing it as a ground-up inference design rather than a repurposed graphics chip.
What OpenAI and Broadcom have not published is just as notable, even though the hardware itself is no longer hypothetical: Broadcom shipped the first JalapeÃąo samples to OpenAI on June 24, 2026, the same day the partnership went public, according to Bloomberg. There is still no disclosed memory configuration, no confirmed process node, and no independently verified benchmark sheet. Winbuzzer’s reporting on the unveiling notes that final benchmarks remain pending, meaning every performance and cost claim in circulation right now comes from the two companies building the chip, not from a neutral third party.
The Math Behind Broadcom’s 50% Cost Claim
The headline number is a roughly 50% cost-efficiency improvement over “standard AI GPUs” in initial testing, a figure Bloomberg reported directly from the companies’ disclosures on June 24, 2026. OpenAI separately told TechCrunch that same day that JalapeÃąo delivers “significantly” better performance per watt than current state-of-the-art alternatives, which is really the same claim from a different angle: inference cost is mostly a power and utilization problem, not a raw-speed problem.
Every large AI lab pays for compute in two ways: the upfront cost of the chip and the ongoing cost of the electricity and cooling to run it. A chip stripped down to do one job, LLM inference, and nothing else can skip the transistors and features that a general-purpose GPU needs for training, graphics, and scientific computing. That specialization is where the savings are supposed to come from. It is the same argument Google made for TPUs a decade ago and the same argument Amazon makes for Trainium today.
The catch is that “50% cheaper” is a claim from the seller, not an audited result. Broadcom has strong incentive to make custom silicon look good since it is now building chips for four of the highest-profile names in AI. Until JalapeÃąo ships in volume and outside researchers get to run their own workloads on it, the number is best treated as a target OpenAI is designing toward rather than a settled fact.
Why OpenAI Is Building Its Own Silicon Now
OpenAI’s compute bill is arguably the biggest cost overhang in the entire AI industry. The company has stacked up infrastructure commitments that include a $500 billion Stargate buildout plan and a $300 billion cloud-compute agreement with Oracle signed in 2025, on top of a separate deal with AMD covering 6 gigawatts of GPU capacity. Nvidia GPUs sit at the center of nearly all of it.
That concentration creates two problems. The first is price: Nvidia’s data center margins are among the highest in the chip industry, and every dollar OpenAI spends on Nvidia hardware is a dollar it doesn’t control. The second is supply: when every AI lab, cloud provider, and government wants the same Blackwell-class GPUs at the same time, allocation becomes as important as price. Owning a custom inference chip, even a partial one built with a silicon partner, gives OpenAI a second source that Nvidia does not gate.
OpenAI CEO Sam Altman framed the move in broader terms than a simple cost play, tying it to the company’s push to make compute itself less scarce. His full comments on that strategy appear later in this article. The subtext is that OpenAI sees compute scarcity, not model quality, as the constraint standing between it and its next stage of growth.
Nine Months From Blank Slate to Silicon
The most surprising number in the entire announcement might be the timeline. OpenAI President Greg Brockman said the chip went from concept to tape-out in nine months, with OpenAI’s own models used to help accelerate parts of the design process. Chip development normally runs two to three years from architecture decisions to a working part, so a nine-month cycle, even for a first-generation design, is unusually fast by semiconductor industry standards.
Tan explained the design philosophy behind that speed: “JalapeÃąo was designed from the ground up for LLM inference using detailed insights from our close collaboration with OpenAI researchers,” he said, adding that the two teams “optimize the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models,” according to remarks from the companies. Rather than adapting an existing Broadcom design, the teams reportedly built around OpenAI’s actual production workloads from the start.
The deployment schedule is deliberately staged. OpenAI plans initial deployment by the end of 2026, a timeline Reuters confirmed in its June 24, 2026 coverage of the announcement. Broadcom has since put numbers behind that ramp: the company committed 1.3 gigawatts of JalapeÃąo compute capacity for 2027 on its June 2026 earnings call, according to Yahoo Finance, building on OpenAI’s own October 13, 2025 disclosure that the broader 10-gigawatt buildout would represent enough capacity to power the equivalent of more than 8 million U.S. homes, per Reuters, with racks beginning deployment in the second half of 2026 and scaling toward full capacity by 2029. Tan has described small-scale development late this year, a significant ramp-up through 2027, and full-scale operations only arriving in the first half of 2028. That is not a chip replacing Nvidia GPUs next quarter. It is the first rung of a multi-year ladder.
JalapeÃąo vs. Blackwell and the Rest of the Custom Chip Field
JalapeÃąo is not the first custom AI chip built with Broadcom, and OpenAI is not the first AI company to decide that renting Nvidia GPUs forever is too expensive. The table below lines up the major inference and training silicon programs active in mid-2026, based on what each company has actually disclosed.
| Chip / Program | Developer | Silicon Partner | Primary Workload | Status as of June 2026 |
|---|---|---|---|---|
| JalapeÃąo | OpenAI | Broadcom (design), Celestica (systems) | LLM inference | Unveiled June 24, 2026. Deployment starts end-2026, full scale H1 2028 |
| TPU (Ironwood / v7 family) | Broadcom | Training and inference | In production across Google Cloud and internal workloads | |
| MTIA | Meta | Broadcom | Recommendation systems and inference | In production. Broadcom deal extended toward 2029 |
| Trainium3 | Amazon / AWS | Amazon (Annapurna Labs) | Training and inference | In customer deployment, including a 2026 Uber adoption |
| Maia | Microsoft | Microsoft in-house | Training and inference | In production for internal Azure and Copilot workloads |
| Blackwell | Nvidia | Nvidia (TSMC-fabricated) | Training and inference | Shipping at scale. Industry’s default performance benchmark |
The pattern is hard to miss. Every major AI buyer with enough scale to justify the engineering cost is now building, or paying someone to build, a chip tuned to its own workloads. Amazon’s pitch for Trainium3 made a similar cost argument when Uber adopted the chips earlier in 2026, positioning them at roughly half the cost of comparable Nvidia GPU instances. JalapeÃąo is OpenAI joining a race that Google, Amazon, Microsoft, and Meta already started.
The Broadcom Common Denominator
What sets JalapeÃąo apart from Trainium and Maia is who built it. Broadcom does not build its own AI models or run its own cloud. It sells design expertise, networking silicon, and manufacturing coordination to companies that do. That business model puts Broadcom in an unusual position: it is now the common contractor behind custom chips for Google’s TPU program, Meta’s MTIA chips, ByteDance, and now OpenAI.
That concentration is a feature for Broadcom’s revenue and a genuine question for everyone else. Four of the world’s most important AI companies are now relying on the same silicon design house to help them escape dependence on a single GPU vendor. If Broadcom’s engineering pipeline becomes the bottleneck instead of Nvidia’s GPU allocation, the diversification argument gets more complicated. For now, Broadcom has stated no dollar figure specific to the JalapeÃąo deal, though it sits inside a broader OpenAI-Broadcom collaboration that OpenAI’s own blog first sized at 10 gigawatts of custom accelerator capacity running from October 2025 through December 2029, a scope Yahoo Finance reported Broadcom reaffirmed when it disclosed the JalapeÃąo contract terms in June 2026, and one the Motley Fool confirmed again on June 26, 2026 was still tracking toward deployment from the second half of 2026 through 2029. The Motley Fool had already flagged the arrangement back in March 2026, naming OpenAI as Broadcom’s “newest chip customer” in what it described as a multiyear, 10-gigawatt custom accelerator partnership.
Broadcom’s AI Revenue Boom, By the Numbers
The JalapeÃąo reveal landed three weeks after Broadcom reported its fiscal second-quarter results, and the trend line explains why Wall Street was already paying attention. AI semiconductor revenue has roughly doubled year over year for two straight quarters, and the company’s own guidance implies it keeps climbing through 2027.
| Period | AI Semiconductor Revenue | Year-Over-Year Growth | Source |
|---|---|---|---|
| Q1 FY2026 | $8.4 billion | +106% | Broadcom earnings coverage |
| Q2 FY2026 (reported June 3, 2026) | $10.8 billion | +143% | Broadcom Q2 FY26 release |
| Q3 FY2026 (guidance) | $16 billion | N/A | CNBC earnings report |
| Full FY2026 (guidance) | $56 billion | N/A | Company guidance |
| FY2027 (target) | More than $100 billion | N/A | Hock Tan, Q2 FY26 earnings call |
Lay the numbers end to end and the growth curve looks almost like a rounding error turned into a business line. Broadcom’s full Q2 FY2026 results also showed total consolidated revenue of $22.2 billion, up 48% year over year, with adjusted EBITDA of $15.2 billion representing 69% of revenue. Custom AI silicon, the category JalapeÃąo now sits inside, has gone from a side business to the fastest-growing part of Broadcom’s income statement in under two years.
Broadcom AI semiconductor revenue run-rate, FY2026:
Q1 FY26actual -> $8.4B (+106% YoY)
Q2 FY26actual -> $10.8B (+143% YoY)
Q3 FY26guide -> $16.0B
FY26 fullguide -> $56.0B
FY27 target -> >$100.0B
---------------------------------------------
Implied 2-year growth (Q1 FY26 to FY27 target): roughly 12xNvidia Still Sets the Pace
None of this dethrones Nvidia yet. The company most recently reported fiscal Q1 2027 revenue of $81.6 billion, up 85% year over year, with data center revenue alone hitting $75.2 billion, up 92% from a year earlier and ahead of Wall Street’s $73.5 billion estimate. That single quarter of Nvidia data center revenue is close to Broadcom’s entire full-year AI semiconductor guidance.
The quarter before that, Nvidia’s data center segment posted a record $51.2 billion, up 25% sequentially and 66% year over year, according to the company’s official results. Analyst estimates of Nvidia’s share of the data center GPU market range roughly from the mid-70s to more than 90%, depending on how narrowly “GPU” is defined against custom accelerators like JalapeÃąo and TPUs. No single figure is authoritative, but every estimate agrees on the same basic point: Nvidia is not losing the AI chip market this year. It is being asked, for the first time at real scale, to share it.
That is why JalapeÃąo’s framing matters more than its shipping volume. Tan said the chip should “efficiently execute most of our most important workloads close to the hardware’s theoretical limits,” a claim aimed less at replacing Nvidia this year than at giving OpenAI leverage the next time it negotiates a GPU order.
From “Titan” to “JalapeÃąo”: How the Rumor Became a Product
JalapeÃąo did not appear out of nowhere. The Financial Times, via Reuters, reported as early as September 2025 that OpenAI and Broadcom were already targeting 2026 for the first shipments of a jointly designed chip line. Reports as early as March 2026 sharpened that picture, describing an OpenAI project internally referred to as “Titan,” a first in-house processor built with Broadcom and paired with Samsung HBM4 memory. That earlier reporting on OpenAI’s custom chip strategy laid out the same basic structure that JalapeÃąo now formalizes: Broadcom handling the compute logic, Samsung supplying high-bandwidth memory as a separate track, and TSMC fabricating the finished design.
What changed between March and June was specificity. The rumor stage had a codename and a memory supplier. The June announcement added a public name, named executives on record, a nine-month development claim, and a staged rollout through 2028. Memory supply remains the part furthest from resolved. HBM4 capacity is already tight across the entire AI industry, and OpenAI’s chip ambitions add one more large customer competing for the same limited output from Samsung and SK Hynix.
What Executives Are Saying
Public comment on JalapeÃąo so far has come almost entirely from the two companies that built it, which is worth keeping in mind when reading the quotes below. Independent analyst commentary with named benchmarks has not caught up to the announcement yet.
“JalapeÃąo is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems.”
Sam Altman, OpenAI CEO, source
“JalapeÃąo is a blank-slate design for modern LLM inference, not a general-purpose accelerator adapted from earlier AI workloads.”
Hock Tan, Broadcom CEO, source
Tan’s other comments focused on the collaborative design process rather than raw specifications, repeatedly crediting OpenAI’s production workloads as the reason the chip could be tuned as tightly as it was. That framing is consistent with how Broadcom has described its other hyperscaler partnerships: it builds the silicon, but the customer’s own traffic patterns shape the architecture.
Market Impact: What Cheaper Inference Could Mean for AI Pricing
Inference cost is the line item that determines whether AI products make money. Training a model is a one-time (if enormous) expense. Serving that model to hundreds of millions of ChatGPT users every day is a recurring cost that scales with usage, and it is the primary reason AI companies have struggled to turn consumer-facing products profitable even as revenue climbs.
If JalapeÃąo’s cost claims hold up under real production load, the most direct beneficiary is OpenAI’s own margin structure, not necessarily consumer pricing right away. A 50% reduction in inference cost on even a portion of OpenAI’s traffic changes the unit economics of every ChatGPT subscription and every API call running on that hardware. Cheaper inference does not guarantee cheaper API pricing for developers, but it does widen the gap between what OpenAI charges and what OpenAI pays, which is the gap investors in a $500 billion infrastructure buildout are watching most closely.
There is a second-order effect worth tracking too. If custom inference silicon becomes a viable path for one AI lab, competitors that lack Broadcom-scale partnerships face a widening gap against rivals who can design around their own workloads. Smaller AI labs without OpenAI’s balance sheet will likely remain GPU renters for years, even as the largest players increasingly build.
Historical Context: A Decade of Hyperscalers Building Their Own Chips
OpenAI is late to this trend, not early. Google started building TPUs internally in the mid-2010s, well before the current generative AI boom, originally to handle search and translation workloads at a cost Nvidia GPUs couldn’t match at Google’s scale. Amazon followed with its Inferentia and later Trainium chips through Annapurna Labs, the chip design firm it acquired in 2015. Microsoft introduced its Maia accelerator in 2023, and Meta has run MTIA chips inside its recommendation and ranking systems for a similar stretch.
What makes OpenAI’s entry different is that it is not a cloud infrastructure company with decades of data center engineering behind it. It is a model company that grew from a research lab into one of the largest compute consumers on the planet in roughly a decade, and it is now trying to backfill the hardware expertise that Google, Amazon, and Microsoft built up over 15 years. Leaning on Broadcom as a silicon partner, rather than building an in-house chip team from scratch the way Google and Amazon did, is the shortcut that makes a nine-month development cycle plausible at all.
5 Predictions for OpenAI’s Chip Strategy Through 2028
- Nvidia dependence stays high through 2027. With full-scale JalapeÃąo operations not targeted until the first half of 2028, OpenAI’s near-term GPU orders from Nvidia and AMD are unlikely to shrink much before then.
- Independent benchmarks arrive within a year, and they will be contested. Expect third-party labs and rival chip vendors to publish their own JalapeÃąo performance comparisons once early units reach developers, and expect Broadcom’s 50% figure to be argued over rather than simply confirmed.
- Broadcom’s custom-silicon customer list grows again. With Google, Meta, ByteDance, and now OpenAI on the roster, at least one more major AI lab or large enterprise buyer is likely to sign a similar custom-chip arrangement with Broadcom before the end of 2027.
- Memory, not compute logic, becomes the next bottleneck. HBM4 supply from Samsung and SK Hynix is already tight. As JalapeÃąo, TPUs, Trainium, and MTIA all scale production in the same window, memory allocation is likely to matter as much as chip design.
- Nvidia leans harder into its own roadmap rather than price cuts. Expect Nvidia to respond to custom silicon competition by accelerating its Rubin-generation release cadence and performance-per-watt claims rather than by discounting Blackwell-class hardware.
What This Means for Developers and Enterprise AI Buyers
For developers building on the OpenAI API, JalapeÃąo is not something to plan around yet. Nothing about the announcement changes pricing, rate limits, or model availability today, and the staged rollout means most production traffic will keep running on existing Nvidia infrastructure well into 2027. The more useful signal is directional: OpenAI is telling enterprise customers, investors, and competitors that it intends to control its own hardware destiny rather than remain a permanent renter.
Enterprise buyers evaluating AI vendors should treat JalapeÃąo as a medium-term cost signal rather than an immediate one. If the cost claims hold up in production, the earliest visible effect is more likely to show up in OpenAI’s ability to sustain aggressive API pricing under margin pressure than in a specific announced price cut tied to the chip itself.
Frequently Asked Questions
What is OpenAI’s JalapeÃąo chip?
JalapeÃąo is OpenAI’s first custom-designed chip, built with Broadcom and unveiled on June 24, 2026. OpenAI calls it an “Intelligence Processor,” built specifically to run trained AI models rather than train new ones.
What is the difference between an inference chip and a training chip?
Training chips handle the initial process of teaching a model using massive datasets, a process that can run for weeks. Inference chips handle the much more frequent job of running an already-trained model to answer a real user’s prompt, where speed and cost per query matter more than raw throughput.
Did OpenAI design JalapeÃąo itself, or did Broadcom design it?
It was a collaboration. Broadcom handled silicon implementation, networking, and connectivity, while OpenAI supplied the architectural insights from its own production workloads. Systems integrator Celestica built the surrounding boards and racks.
Is JalapeÃąo replacing Nvidia GPUs at OpenAI?
Not immediately. OpenAI’s infrastructure commitments to Nvidia and AMD remain in place, and JalapeÃąo’s own rollout is staged through 2028. It supplements OpenAI’s GPU capacity rather than replacing it in the near term.
How much cheaper is JalapeÃąo than Nvidia GPUs?
Broadcom has said early testing showed roughly 50% better cost efficiency versus standard AI GPUs. That figure comes from the companies building the chip, not from an independent benchmark, so it should be treated as a target rather than a confirmed result until third-party testing catches up.
When will JalapeÃąo actually be deployed?
Initial deployment is planned by the end of 2026, with a significant ramp-up in 2027 and full-scale operations targeted for the first half of 2028, according to Broadcom CEO Hock Tan.
Will this make ChatGPT cheaper or faster for regular users?
Not directly or immediately. Lower inference costs improve OpenAI’s own margins first. Any change to consumer or developer pricing would be a separate business decision, and nothing announced alongside JalapeÃąo commits OpenAI to passing savings on to customers.
Who else uses Broadcom to build custom AI chips?
Broadcom has publicly tied its custom AI silicon business to Google, Meta, and ByteDance, alongside OpenAI’s new JalapeÃąo program. Google’s TPU line and Meta’s MTIA chips are the two most established examples.
