OpenAI Partners with Qualcomm and MediaTek to Develop Custom Smartphone Processor

OpenAI is no longer content to be a software company. According to a report from TF International Securities analyst Ming-Chi Kuo, OpenAI is working with Qualcomm and MediaTek to develop a custom smartphone processor, with China-based manufacturer Luxshare tapped as the exclusive assembly partner [1][3]. The target: 300–400 million annual smartphone shipments, a figure that would surpass Apple's 2025 iPhone volumes and represent roughly a quarter of the global smartphone market [1][5].

The announcement arrives alongside OpenAI's separate, already-in-progress effort to build server-side inference chips with Broadcom on TSMC's 3nm process [9][10], and its Jony Ive-designed wearable device expected in the second half of 2026 [13]. Together, these initiatives signal a company that believes software alone won't capture the value it sees in AI — and that controlling hardware is the next competitive frontier.

But between a stagnant smartphone market, entrenched OEM relationships, rapidly commoditizing on-device AI, and OpenAI's own $14 billion in projected 2026 losses [8], the gap between ambition and execution is wide.

The Partnership: Who Does What

The division of labor, as reported by Kuo, gives MediaTek and Qualcomm joint responsibility for processor design and fabrication, while Luxshare handles system co-design and manufacturing [3][4]. Specifications are expected to be finalized by late 2026 or Q1 2027, with mass production slated for 2028 [2][3].

The dual-chipmaker arrangement is unusual. MediaTek and Qualcomm are direct competitors — MediaTek held 36% of global smartphone SoC shipments by volume in Q1 2025 versus Qualcomm's 28% [11]. Working with both simultaneously could give OpenAI access to MediaTek's cost-effective mainstream designs and Qualcomm's premium-tier expertise, but it also raises questions about who owns the resulting intellectual property and how design conflicts are resolved.

Source: Counterpoint Research / EE Times Asia

Data as of Jun 1, 2025CSV

No public details have emerged about the financial structure. Custom chip tape-out costs on advanced nodes typically run $200 million or more for a full SoC development cycle including design, fabrication, software, and qualification [9]. TSMC's 2nm wafers alone cost upward of $30,000 each — nearly double the price of 4nm wafers. OpenAI has the capital: the company closed a $122 billion funding round in March 2026 at an $852 billion valuation [7]. But it also projects cumulative losses of $115 billion through 2029 before reaching profitability [8]. Every new hardware venture adds to that burn.

The Shipment Target: Ambitious or Aspirational?

The 300–400 million annual unit target deserves scrutiny. Global smartphone shipments totaled approximately 1.26 billion units in 2025, a market that has been essentially flat since its 2021 peak [5][6].

Source: IDC / Counterpoint Research

Data as of Mar 1, 2026CSV

Apple, the current market leader by revenue and the top vendor by shipment volume in 2025, shipped roughly 243 million iPhones — a record for the company, representing 20% market share [5]. Samsung shipped slightly fewer. To hit 300 million units, OpenAI would need to capture roughly 24% of the entire global smartphone market, exceeding what Apple has achieved after 18 years of iPhone sales.

Source: Counterpoint Research

Data as of Jan 12, 2026CSV

The comparison is misleading, though, because OpenAI is not planning to sell phones directly in competition with Apple. The more plausible reading is that OpenAI envisions its custom processor appearing across multiple Android OEMs — the way Qualcomm's Snapdragon chips already do. Qualcomm's handset segment generated $27.8 billion in fiscal 2025 revenue by powering devices from Samsung, Xiaomi, OPPO, and dozens of others [11]. If OpenAI's chip were adopted by enough of those same OEMs, 300–400 million units is not physically impossible. But there's a significant difference between designing a chip and convincing manufacturers to use it.

What OpenAI Brings to the Silicon Table

The core technical question is what OpenAI adds to a chip that Qualcomm and MediaTek can't already provide on their own.

Qualcomm's Snapdragon 8 Elite Gen 5, announced in late 2025, ships with a Hexagon NPU that is 37% faster than its predecessor for AI tasks [18]. The chip delivers over 100 tokens per second for on-device LLM decode and time-to-first-token of 0.12 seconds on high-resolution images [18]. Apple's A18 chip features a 16-core Neural Engine rated at 35 TOPS (tera operations per second), with power consumption of 3–4W under sustained load [19].

These are already capable AI processors. The Snapdragon 8 Elite runs multimodal models on-device, handles real-time image segmentation, and supports local inference for models in the 1–7 billion parameter range. What hardware bottleneck, specifically, requires OpenAI's involvement?

Kuo's report suggests the answer is architectural, not computational. OpenAI's vision centers on an "AI agent" phone where users interact primarily through natural language rather than apps — telling the phone what they want done and letting AI agents handle execution [3][12]. This requires what Kuo describes as managing a "full real-time state" of the user: continuous environmental sensing, memory hierarchy management, and a hybrid architecture that processes data locally for low-latency private tasks while offloading intensive workloads to the cloud [3].

If accurate, this isn't a conventional NPU benchmark problem. It's a system architecture problem — optimizing the interplay between on-device inference, sensor fusion, memory management, and cloud connectivity. Whether that requires custom silicon or can be achieved through software running on existing chips is the central technical debate.

The Commoditization Problem

OpenAI's timing faces a headwind: on-device AI is commoditizing fast.

Google's Gemma 4 family, released under the Apache 2.0 license, includes edge-optimized models (2B and 4B parameter variants) that run on existing NPUs from Qualcomm, MediaTek, and Google's Tensor chips [15]. Meta's Llama models offer similar on-device capabilities. Google's LiteRT framework delivers 1.4x faster cross-platform GPU performance and streamlined NPU acceleration for these open models [15].

The practical effect: any Android OEM can ship a phone today that runs competitive on-device AI models for free, using commodity chips. Gemma 4's 26B MoE (mixture-of-experts) model — a parameter count in the range of GPT-3 — runs on a single 16GB GPU [15]. The edge-optimized variants run on phones already in consumers' pockets.

This creates a difficult pricing dynamic for OpenAI-branded silicon. If the custom chip's primary value is running OpenAI's proprietary models faster, OEMs must believe that ChatGPT integration commands enough consumer premium to justify switching from their existing chip suppliers. Given that Samsung already ships with Google Gemini integration, Xiaomi is building its own AI assistant [16], and OPPO has partnered with MediaTek on its own on-device AI model called Omni, the addressable market for OEMs willing to cede AI control to OpenAI may be smaller than the headline numbers suggest.

The OEM Problem

No Android OEM has publicly committed to using OpenAI's custom processor. This is the most significant gap in the current narrative.

Qualcomm has multi-year agreements with key OEMs. In May 2025, Qualcomm and Xiaomi expanded their collaboration with a new multi-year deal [14]. Samsung uses both Qualcomm Snapdragon and its own Exynos chips. OPPO and vivo rely heavily on MediaTek for their mid-range and premium devices.

These relationships involve deep co-engineering. Chip suppliers work with OEMs years in advance on board design, antenna integration, modem compatibility, camera ISP tuning, and carrier certification. Switching chip suppliers is not a matter of swapping components — it requires redesigning the phone.

OpenAI has no track record in mobile hardware. It has never taped out a mobile SoC, never navigated carrier certification in 190+ countries, never dealt with the thermal constraints of a 7mm smartphone chassis. The company is simultaneously developing server-side inference chips with Broadcom (codenamed "Titan," using TSMC's 3nm process) that have already slipped from a Q2 to Q3 2026 deployment target [9][10]. Running parallel chip development programs across mobile and server platforms — with different fabrication partners — is an enormous organizational challenge.

Xiaomi's own trajectory illustrates the difficulty. Despite investing over 200 billion yuan in AI R&D over five years and producing its own XRing O1 chip, Xiaomi still relies on Qualcomm and MediaTek for its flagship processors [16]. Building a competitive mobile SoC is a multi-year, multi-billion-dollar endeavor, and even well-resourced companies have struggled. Google's Tensor chips, despite significant investment, still trail Qualcomm's Snapdragon in raw performance benchmarks.

Regulatory and Privacy Implications

If OpenAI succeeds in placing its silicon inside hundreds of millions of phones while also operating the AI services running on that silicon, regulators will take notice.

The EU, US, and UK competition authorities have already flagged concerns about concentrated control of AI infrastructure. A joint statement from the DOJ, FTC, UK CMA, and European Commission identified three specific risks: concentrated control of key inputs like specialized chips, the ability of incumbents to extend power across AI markets, and arrangements among players that reduce competition [17].

OpenAI itself has been on the other side of this argument. In October 2025, the company warned EU antitrust enforcers about "harmful conduct" by Google, Microsoft, and Apple, arguing that entrenched companies use control over operating systems, app stores, and cloud infrastructure to disadvantage AI competitors [17]. A company that controls both the phone's processor and the AI services running on it would face the same structural criticism it leveled at Apple.

The comparison to Apple's regulatory experience is instructive. Apple faced years of antitrust scrutiny in the EU over App Store policies and vertical integration — scrutiny that culminated in the Digital Markets Act's requirements for sideloading and third-party app store access. An OpenAI that owns the silicon, the OS-level AI agent, and the cloud inference backend would present an even more tightly integrated stack than Apple's current model.

China presents a separate challenge. Luxshare, the designated assembly partner, is based in China. OpenAI's preference for non-China manufacturing — its Jony Ive device is reportedly being assembled by Foxconn in Vietnam or the US [13] — suggests geopolitical sensitivity. But Luxshare's involvement in the smartphone project means the supply chain runs directly through Chinese manufacturing, a potential vulnerability given ongoing US-China technology tensions.

The Timeline Problem

Custom silicon typically requires 2–4 years from architecture to mass production. If specifications are finalized by late 2026 or Q1 2027 as Kuo reports [3], mass production in 2028 is aggressive but plausible — assuming no design revisions, fabrication delays, or yield problems.

By 2028, the competitive landscape will look different. Apple will likely be on its A22 or equivalent chip, built on TSMC's 2nm or successor process. Qualcomm will have shipped two more generations of Snapdragon. MediaTek completed its first 2nm tape-out in September 2025 and has mass production scheduled for late 2026 [9]. Google's Gemini and DeepMind's open models will have advanced by multiple generations. AI phones are forecast to account for 54% of global shipments by 2028 — but the AI capabilities will be built into commodity chips, not waiting for OpenAI's custom silicon.

The server-side chip program offers a cautionary parallel. OpenAI's "Titan" inference chip, developed with Broadcom, has already slipped from its Q2 2026 target to Q3 at the earliest [10]. Mobile SoCs are arguably harder than inference accelerators: they require integrating a CPU, GPU, NPU, modem, ISP, and dozens of other IP blocks into a single die that operates within a 4–5W thermal envelope.

The Steelman Case — and Its Limits

The strongest argument for this partnership is strategic positioning, not near-term revenue. OpenAI may not need 300 million units in 2028. What it needs is a credible hardware path that prevents it from being locked out of the device layer as AI shifts from cloud to edge.

If AI agents do replace apps as the primary interaction model — a bet OpenAI is making explicitly [3][12] — then the company that controls the agent runtime has enormous power. Apple controls the iOS runtime today and extracts 15–30% commissions from every transaction. An "AI agent" phone where OpenAI's model handles purchases, bookings, and communications could generate similar platform economics.

The Jony Ive device, targeting H2 2026 with 40–50 million units through Foxconn [13], may serve as a proof of concept. If a screenless, voice-first wearable demonstrates consumer demand for AI-native hardware, it validates the thesis that justifies the far more expensive smartphone bet.

But the limits are real. OpenAI is a company projecting $14 billion in losses this year [8], running at roughly 40% gross margins constrained by compute costs. It is simultaneously developing server chips, consumer wearables, and now a mobile SoC — while also running one of the world's largest AI research operations. The smartphone chipset business is a low-margin, high-volume enterprise where Qualcomm, with decades of modem and RF expertise, earns its returns through scale. OpenAI has none of that infrastructure.

The 300–400 million unit target, repeated in every headline, traces back to a single analyst report from Ming-Chi Kuo [1][3]. No OEM has confirmed participation. No chip specifications have been disclosed. No timeline has been validated by the chipmakers themselves. Qualcomm's and MediaTek's stock prices moved on the news, but neither company has issued a public statement.

What exists today is a plan to make a plan — specifications to be finalized in a year, production to begin in two. In the smartphone industry, that's not a product. It's a signal. Whether that signal attracts the OEM commitments needed to make the numbers real will determine whether this is the beginning of a new hardware era or an expensive footnote in OpenAI's rapid expansion.