Apple spent the summer proving that a 27-billion-parameter language model could run entirely on a phone. Qualcomm is about to answer with a chip built to run one nearly twice that size. The two companies are converging on the same goal, an AI assistant that never has to phone home to a data center, but they are getting there by almost opposite routes, and the gap between those routes is about to become the defining story in mobile hardware for the next two years.
The near-term trigger is Snapdragon Summit 2026, running September 22 to 24 in Maui, where Qualcomm is expected to formally unveil a new Hexagon NPU architecture built around agentic AI and Mixture-of-Experts (MoE) models with up to roughly 30 billion parameters running locally on the device. That announcement lands just weeks after Apple’s own on-device story, the demonstration that a compressed 27-billion-parameter model running on the A20 Pro processes tokens at roughly double the rate of last year’s A19 Pro. Two chip roadmaps, two philosophies, one race: who owns the phone as a personal AI computer.
Apple’s Quiet Compression Play: A20 Pro and Bonsai 27B
Apple’s approach has been architectural rather than promotional. The A20 Pro, built on a 2-nanometer process and described in coverage as Apple’s first 2nm iPhone chip, did not ship with a marketing keynote about “30 billion parameters.” Instead, the story broke through independent testing: developers running Bonsai 27B, a compressed derivative of Alibaba’s Qwen3.6-27B model, found it fit inside the memory available to apps on an iPhone 17 Pro and, separately, ran entirely locally on an iPhone 18 Pro carrying the newer A20 Pro silicon.
The compression numbers are the real headline. PrismML, the AI development firm behind the work, says it shrank Qwen3.6-27B from roughly 54GB down to under 4GB while keeping all 27 billion parameters active simultaneously, rather than routing through a subset like an MoE design does. Bonsai 27B’s own reported footprint is about 3.9GB of memory, and independent developers clocked it around 8 to 9 tokens per second on 8GB iPhones, while PrismML reported 11.0 tokens per second on an iPhone 17 Pro Max using Apple’s MLX framework. RunAnywhere, another party that benchmarked the model, put throughput on the iPhone 17 Pro at roughly 10 tokens per second. None of those numbers will win a leaderboard against a data-center GPU, but they were not designed to. They were designed to prove a 27B-class model can live entirely inside a phone’s power and thermal budget, with the model’s full parameter count intact rather than gated behind expert-routing logic.
That distinction, full dense parameters versus routed experts, is the crux of why Apple and Qualcomm are not actually running the same race. It also explains why Apple’s A20 Pro’s 7-core GPU and expanded memory bandwidth matter as much to this story as the neural engine does. Compressing a dense model down to a few gigabytes only pays off if the chip can move that data fast enough, and Apple’s bandwidth gains this generation were built with exactly that kind of workload in mind.
Qualcomm’s Answer: A Hexagon NPU Built for 30B-Parameter Models
Qualcomm’s pitch, based on details the company has already disclosed ahead of Snapdragon Summit, is bigger and blunter. Its next-generation Hexagon NPU is described as the “third pillar” of the upcoming flagship Snapdragon platform, built explicitly around agentic AI and what the company calls “massive parameter models.” The headline capability is support for Mixture-of-Experts models with a total capacity of roughly 30 billion parameters running locally, where only a fraction of those parameters fire on any given token, trading raw capacity for lower active compute per inference step.
Two design choices back that claim up. Qualcomm says the new NPU adds a transformer-focused Element Accelerator plus a 50 percent larger shared memory pool so frequently accessed model weights sit closer to the compute units, and it is promising up to 50 percent faster prefill for INT4-quantized models, aimed squarely at the long-context workloads where the initial prompt-processing stage usually bottlenecks. The chip is also designed to span a wide precision range, a genuine spread rather than a single fixed format:
Hexagon NPU supported precision range (per Qualcomm disclosures):
INT2 -> INT4 -> INT8 -> FP8 -> FP16
(lower precision = smaller footprint, higher precision reserved for sensitive layers)That spread matters because it lets device makers push most of a model down to 2- or 4-bit weights for size and speed, while keeping a handful of layers at higher precision to protect output quality, a more granular version of the same low-bit philosophy Apple and PrismML are using with Bonsai 27B, just applied to a sparse rather than dense architecture.
Mixture-of-Experts vs Compressed Dense Models: Two Different Bets
It helps to be precise about what each company is actually claiming, because “30B on the phone” and “27B on the phone” sound like the same milestone and are not. Qualcomm’s MoE approach means the model’s total capacity, its stored knowledge, spans roughly 30 billion parameters, but only a subset activate per token, which keeps compute and latency down even as stored capacity grows. Apple and PrismML’s approach keeps all 27 billion parameters active on every token, relying instead on aggressive low-bit quantization and a hybrid-attention architecture to shrink the memory and compute footprint of the full model.
Each has a trade-off. MoE scales capacity cheaply but needs a sophisticated router and enough fast shared memory to swap experts in and out without stalling, which is exactly what Qualcomm’s redesigned memory hierarchy is aimed at. Dense compression keeps every inference pass consistent and simpler to optimize on a single chip, but it caps out earlier since every added parameter costs compute on every single token, which is why PrismML’s engineering effort went into shrinking Qwen3.6-27B rather than into scaling it further. Neither approach is objectively superior yet. They are simply the two bets Apple’s vertically integrated silicon-and-software stack and Qualcomm’s NPU-and-partner-OEM model are placing on how phones should host large models.
The GPU Front: Adreno Neural Fusion vs Arm’s Mali G2-Ultra NX
The NPU story is only half the fight. On September 2, 2026, Qualcomm separately disclosed Adreno Neural Fusion, a mobile GPU architecture that places dedicated AI compute units, which it calls Matrix Cores, inside each of the GPU’s three processing slices, co-located with 18MB of Adreno High Performance Memory. The stated purpose is running AI super-resolution and frame generation directly inside the render pipeline, and Qualcomm says Unity and Unreal Engine already have native integration for it, meaning game studios can turn on AI upscaling without writing custom code.
Arm answered six days later. On September 8, the Mali G2-Ultra NX GPU launch put neural accelerators directly inside the shader cores themselves rather than routing AI work to a separate NPU, a design Arm explicitly positioned as a counterpart to Qualcomm’s Neural Fusion approach. Coverage of the launch described it as among the first mobile GPU architectures to treat neural rendering as a hardware-native capability rather than a software add-on. That framing tells you where the industry thinks the next differentiation battle sits: not just in how big a model a phone can hold in memory, but in how tightly a chip can fuse traditional graphics rendering with AI inference in the same silicon block.
Samsung’s Galaxy S27 Ultra Dilemma: Snapdragon or Exynos
Samsung sits in an unusual position in this race, because it is both a chipmaker and one of Qualcomm’s biggest customers. Coverage of the upcoming Galaxy S27 Ultra reports it is expected to ship with Qualcomm’s next flagship Snapdragon variant, commonly referred to in reporting as the Snapdragon 8 Elite Gen 6, carrying the new Hexagon NPU, at least in the markets where Samsung has historically favored Qualcomm silicon over its own Exynos line. That makes Samsung’s flagship one of the first mass-market vehicles for Qualcomm’s agentic-AI pitch, effectively outsourcing its on-device AI competitiveness to a supplier rather than building it in-house for that model year.
The strategic tension is obvious. If Qualcomm’s MoE approach becomes the default way flagship Android phones run large models, Samsung’s own Exynos silicon has to match that capability or cede the on-device AI story on its own hardware entirely, even while its phones carry the benefit through Qualcomm’s chip. It is a repeat of a familiar pattern in mobile silicon, where Samsung alternates between its own chips and Qualcomm’s depending on region and yield, except now the stakes include whether the phone can run an agentic assistant without a network connection at all.
Where Google’s Tensor Fits Into the Race
Google occupies a third position, closer to Apple’s model than Qualcomm’s in one respect: it controls both the silicon and the assistant software that runs on it. Tensor chips have historically prioritized tight integration with Google’s own models over raw benchmark supremacy, which has meant Google was rarely the company setting the pace on headline parameter counts. The question this cycle is whether that integration advantage still holds when Qualcomm is explicitly designing hardware around 30B-class MoE models and Apple has demonstrated a working 27B dense model on a shipping phone. A chip strategy built around good-enough hardware paired with tightly coupled software gets harder to defend the moment two competitors prove that raw on-device capacity at this scale is achievable rather than theoretical.
PrismML and the Rise of Mobile Model-Compression Middlemen
PrismML’s role in this story points to a layer of the industry that gets less attention than the chipmakers: firms whose entire job is squeezing large foundation models down to sizes a phone can hold. PrismML’s compression of Qwen3.6-27B, from about 54GB to under 4GB while keeping all 27 billion parameters live, is the kind of engineering work that turns a chipmaker’s raw silicon capability into an actual usable product. Without that compression step, the A20 Pro’s memory bandwidth and neural engine gains would have nothing to run.
Qualcomm is not relying on a third-party compression specialist for its headline claim, either. By building INT2-through-FP16 precision support and MoE-aware memory management directly into the Hexagon NPU, it is effectively pulling the compression layer into the chip itself, competing with independent firms like PrismML on their own turf. That is a meaningful strategic difference: Apple’s stack depends on a compression partner doing excellent work on top of good silicon, while Qualcomm is betting it can make the compression problem mostly disappear at the hardware level. Which approach ages better will depend on how quickly foundation-model architectures keep changing underneath both of them.
On-Device AI Chip Comparison, September 2026
Here is how the four major mobile AI platforms compare on what has actually been disclosed as of this week, rather than what is rumored:
| Platform | Model architecture | Reported on-device capacity | Key differentiator | Status |
|---|---|---|---|---|
| Apple A20 Pro (2nm) + Bonsai 27B | Dense, hybrid-attention, low-bit compressed | 27B parameters, ~3.9GB footprint | PrismML compression, MLX runtime, vertical integration | Demonstrated on shipping iPhone 17 Pro / 18 Pro |
| Qualcomm next-gen Hexagon NPU | Sparse Mixture-of-Experts | Up to ~30B parameters (MoE capacity) | INT2-FP16 range, 50% larger shared memory, native Unity/Unreal AI GPU path | Disclosed ahead of Snapdragon Summit, Sept 22-24 |
| Samsung Galaxy S27 Ultra | Inherits Qualcomm MoE stack (reported) | Tied to Snapdragon 8 Elite Gen 6 | Distribution scale, dual-sourcing flexibility | Expected, per device reporting |
| Google Tensor | Proprietary, tightly coupled to Gemini stack | Not publicly disclosed at this scale | Software-model integration over raw parameter count | Unconfirmed for this generation |
How Big Is the On-Device AI Market
The competitive intensity makes more sense once you look at the size of the prize. Market research cited in industry coverage puts the on-device generative AI segment specifically at roughly $6.2 billion in 2026, with a trajectory toward $94.8 billion by 2035, an implied compound annual growth rate near 35 percent. A broader on-device AI market, covering more than just generative workloads, is estimated at roughly $10.76 billion in 2025, climbing toward $75.5 billion by 2033. Smartphones are consistently the largest single category inside those numbers, estimated at somewhere between 42 and 47 percent of on-device AI revenue depending on the specific report’s methodology, and machine-learning-based workloads, the category that includes LLM inference, account for roughly two-thirds of on-device intelligence spending overall.
| Metric | 2025/2026 figure | Forecast | Implied CAGR |
|---|---|---|---|
| On-device generative AI market | ~$6.2B (2026) | ~$94.8B by 2035 | ~35.4% |
| Broader on-device AI market | ~$10.76B (2025) | ~$75.5B by 2033 | ~27.8% |
| Smartphone share of on-device AI revenue | ~42-47% (2026) | – | – |
| ML-based workload share of on-device intelligence | ~67.8% (2026) | – | – |
Those numbers explain why neither Apple nor Qualcomm is treating this as a side project. If smartphones already capture close to half of on-device AI revenue and that revenue base is compounding at roughly 30 to 35 percent a year through the early 2030s, owning the reference architecture for how a phone runs a large model, dense and compressed versus sparse and routed, is not a features race. It is a bid to define the default technical standard the rest of the industry builds against.
From Siri to Sub-Cloud LLMs: A Brief History
On-device AI on phones did not start this year. Apple shipped a dedicated Neural Engine as far back as the A11 Bionic in 2017, initially for Face ID and camera processing rather than language models. Qualcomm’s Hexagon DSP line has carried some form of dedicated AI acceleration since the mid-2010s, long before “on-device LLM” was a marketing phrase. What changed is the target workload. Early neural accelerators were built for narrow tasks: photo classification, voice-trigger detection, camera scene recognition. The shift toward running general-purpose, multi-billion-parameter language models locally, models capable of open-ended reasoning rather than a single classification task, is only becoming practical now because two things arrived together: chips with enough memory bandwidth and low-bit compute support, and compression techniques aggressive enough to fit a useful model into that budget. The A20 Pro and the next Hexagon NPU are best understood as the point where those two curves finally crossed, not as a sudden leap.
Developer Tooling: Unity, Unreal Engine and MLX
Chips only matter if developers can actually target them without heroic engineering effort, and this is where the two camps diverge again. Apple’s path runs through MLX, its own machine-learning framework, which Bonsai 27B uses to run natively across Mac, iPhone, and iPad, alongside a separate CUDA path for the same model family on NVIDIA GPUs. That gives Apple a single, tightly controlled runtime spanning its whole device lineup, consistent with its historical preference for owning the full stack rather than depending on cross-platform standards.
Qualcomm’s play is broader by design. Confirming that Unity and Unreal Engine already have native integration for Adreno Neural Fusion means millions of existing mobile game projects can enable AI-based upscaling and frame generation without a developer writing custom inference code, turning on-device AI from an opt-in SDK integration into a checkbox in an engine every studio already uses. That is a meaningfully different distribution strategy than Apple’s controlled-runtime approach: Qualcomm is betting on reach through existing developer tools, while Apple is betting on depth through a runtime it fully owns.
The Economics of Skipping the Cloud
There is a blunt economic argument underneath all of this engineering. Every prompt that a phone can answer locally is a prompt that never touches a cloud GPU, which means the chipmaker that wins the on-device inference race is quietly reducing the inference-cost burden that would otherwise fall on Apple, Google, or a phone maker’s cloud partner. Qualcomm’s own public framing around its NPU work has leaned into that logic directly. As Qualcomm’s Durga Malladi, Senior Vice President and General Manager for Technology Planning and Edge Solutions, put it in the company’s announcement on bringing generative AI to devices, “With Snapdragon 8 Gen 3 for smartphones and Snapdragon X Elite for PCs, we sparked commercialization of on-device AI at scale,” a claim about prior hardware generations that the company is clearly aiming to extend with its next Hexagon NPU (Qualcomm).
Apple’s version of the same logic is architectural rather than declarative. The company has built Private Cloud Compute as a fallback for workloads too heavy to run locally, a tacit admission that not everything will fit on-device even with aggressive compression, while pushing as much inference as possible onto the phone itself first. Apple’s own presentation on the system put it plainly: “We have created Private Cloud Compute,” describing Apple silicon servers built specifically to extend on-device processing guarantees to the cases where local hardware is not enough (Constellation Research). That two-tier design, phone first, private server second, only makes financial sense if the phone tier keeps absorbing a growing share of requests, which is exactly what compressed 27B and MoE 30B-class models are meant to enable.
What Industry Voices Are Saying
Apple Silicon executive Doug Brooks has described the company’s chip strategy in similarly direct terms, telling one outlet, “We’re seeing tremendous momentum with how people are using Apple products and Apple silicon, particularly in on-device AI workflows” (The Deep View). In separate remarks about how Apple designs hardware around AI workloads, Brooks summarized the company’s philosophy this way: “If you want to build a great device for AI, you need to build a great computer” (tbreak). That comment, made in the context of Apple Silicon’s broader AI push, captures the core of Apple’s bet in this race: rather than chasing a single headline parameter count, the company is treating on-device AI as a systems problem spanning chip, memory, compression software, and runtime together, the same integrated approach that produced the A20 Pro’s bandwidth gains and the Bonsai 27B compression work in the same product cycle.
Five Predictions for On-Device AI Through 2027
- MoE goes mainstream on Android. If Qualcomm’s Hexagon NPU ships as disclosed at Snapdragon Summit, expect most 2027 Android flagships, not just the Galaxy S27 Ultra, to advertise 20B-plus MoE capacity as a standard spec line, the way octa-core CPUs became table stakes a decade ago.
- Apple keeps pushing dense compression rather than switching to MoE. Apple’s vertical control over silicon, OS, and runtime favors a single consistent inference path per token. A full pivot to sparse MoE would complicate the MLX runtime, so a more likely move is deeper collaboration with compression specialists like PrismML on the next model generation.
- GPU-embedded neural cores become the next battleground. With both Qualcomm’s Adreno Neural Fusion and Arm’s Mali G2-Ultra NX now treating neural rendering as hardware-native, expect Apple and Samsung’s in-house GPU teams to disclose comparable embedded-AI GPU designs within the next one to two product cycles.
- Samsung’s Exynos line faces a credibility test. If flagship Galaxy phones keep leaning on Qualcomm for headline on-device AI capability, pressure builds on Samsung to either match that capability in Exynos or narrow Exynos to lower tiers where AI performance matters less.
- Model-compression vendors become acquisition targets. As chipmakers pull precision-format support directly into silicon, independent compression specialists like PrismML face a choice between deeper OEM partnerships or consolidation, since the hardware layer is increasingly capable of doing part of their job natively.
Why This Race Actually Matters to Consumers
It is easy to read chip-architecture disclosures as an industry-insider story, but the practical effect lands on ordinary phone use within a year or two. A phone that can run a 27B or 30B-class model locally can hold a longer, more coherent conversation with an assistant, summarize a document, or draft a reply without sending that content to a server, and it can do it without a network connection at all. That has real privacy implications, since data that never leaves the device cannot be intercepted in transit or retained in a cloud log, and real latency implications, since local inference skips the round-trip that makes cloud-based assistants feel slower on a spotty connection. It also has battery implications that cut the other way: running a large model locally, even a compressed one, draws more power than a simple network request, which is part of why Qualcomm’s NPU work leans so heavily on low-bit precision formats and Apple’s Bonsai 27B numbers are reported in tokens per second rather than headline throughput alone.
The Bottom Line for Buyers Choosing Between Platforms
For anyone shopping in this cycle, the practical takeaway is less about parameter counts and more about ecosystem fit. iPhone buyers get Apple’s tightly controlled, already-demonstrated Bonsai 27B path running on shipping A20 Pro hardware today. Android buyers, particularly on Samsung’s high end, are betting on a Qualcomm roadmap that has been disclosed in detail but has not yet shipped in a consumer device as of this week, with the full picture due at Snapdragon Summit on September 22. Both bets are credible. Neither is finished. The next six months, spanning Qualcomm’s formal unveiling, Samsung’s Galaxy S27 Ultra launch, and any response from Google’s Tensor team, will determine whether dense compression or sparse MoE becomes the default architecture for the next generation of phone-based AI.
Frequently Asked Questions
What is the difference between Apple’s A20 Pro AI approach and Qualcomm’s new Hexagon NPU?
Apple’s approach, paired with PrismML’s Bonsai 27B model, compresses a dense 27-billion-parameter model down to roughly 3.9GB while keeping every parameter active on each token. Qualcomm’s upcoming Hexagon NPU is built for Mixture-of-Experts models with roughly 30 billion parameters of total capacity, where only a subset of those parameters activate per token.
Has Apple officially confirmed the A20 Pro can run a 27B parameter model?
Apple has not made an official marketing claim about this specific benchmark. The capability comes from third-party developer testing and from PrismML, the company behind the Bonsai 27B compression, running the model on iPhone 17 Pro and iPhone 18 Pro hardware.
When will Qualcomm officially unveil its next-generation Hexagon NPU?
Qualcomm is expected to formally detail the full platform at Snapdragon Summit 2026, scheduled for September 22-24 in Maui, though the company has already disclosed several architectural details ahead of the event.
Will the Galaxy S27 Ultra use Qualcomm’s new chip or Samsung’s own Exynos silicon?
Device reporting indicates the Galaxy S27 Ultra is expected to ship with Qualcomm’s next flagship Snapdragon platform, commonly referenced as the Snapdragon 8 Elite Gen 6, carrying the new Hexagon NPU, at least in the markets where Samsung has historically used Qualcomm chips.
What is Mixture-of-Experts (MoE) and why does it matter for phones?
MoE is a model architecture where only a subset of the model’s total parameters activate for any given input, rather than running every parameter on every token. It lets a chip advertise a large total parameter count, such as Qualcomm’s roughly 30 billion, while keeping the actual compute cost per token lower than a fully dense model of the same size.
How big is the on-device AI market in 2026?
Industry market research puts the on-device generative AI segment specifically at roughly $6.2 billion in 2026, projected toward $94.8 billion by 2035, while the broader on-device AI market is estimated near $10.76 billion in 2025, growing toward $75.5 billion by 2033, with smartphones capturing the largest single share of that revenue.
Does running a large language model locally use more battery than using a cloud-based assistant?
Running a compressed model locally, even at low-bit precision, generally draws more power in the moment than sending a simple network request, which is why chipmakers are focused so heavily on low-bit precision formats like INT2 and INT4 to keep on-device inference power-efficient.
Are Unity and Unreal Engine getting AI features from this chip race?
Yes. Qualcomm has confirmed that Unity and Unreal Engine have native integration for its Adreno Neural Fusion GPU architecture, allowing AI-based upscaling and frame generation to be enabled in existing game projects without custom inference code.
