# Context pack: Qualcomm > You are a structural analyst. The material below is from PlexusGraph — a knowledge-graph research publication. Reason with the user grounded in it: surface the structure, the feedback loops, the chokepoints and flywheels, and the non-obvious connections. When you make a claim from it, you can point to the sources. **In one line:** Qualcomm: The Brilliant Designer Who Rents the Only Factory in Town Source: https://plexusgraph.dev/companies/qualcomm ## Brief *Based on 42 related nodes across 7 research explorations in the semiconductors sector* --- Imagine you design the world's best custom furniture — intricate, beautiful, engineered to perfection. You have no workshop of your own, so you rent space at the one factory in the world capable of building your designs to the tolerances you need. That factory sits on a small island that two superpowers are quietly arguing over. That is Qualcomm's situation in one sentence. Qualcomm is what the industry calls a "fabless" company. It designs chips but does not make them. The making happens at TSMC, a Taiwanese manufacturer that is, for practical purposes, the only place on earth capable of producing the most advanced chips at commercial scale. This arrangement has made Qualcomm enormously successful — designing chips is where the high-margin intellectual work happens, and outsourcing manufacturing means Qualcomm does not have to maintain the most expensive industrial facilities in human history. But it also means Qualcomm's entire product line depends on a factory it does not own, cannot replicate, and cannot protect. Understanding Qualcomm requires holding two ideas at once: the company is genuinely well-positioned in several important ways, and it is also exposed to risks it fundamentally cannot control. --- ## What Qualcomm Actually Does Most people know Qualcomm because its chips power a huge share of the world's Android smartphones. The Snapdragon processors inside billions of phones handle everything from cellular connectivity to the small AI tasks your phone runs locally — face recognition, voice processing, photo enhancement. This is Qualcomm's established home turf. But the company is making a significant bet on a second arena: the data centers where AI systems actually run. When you ask a chatbot a question, that query goes to a massive computer somewhere running specialized chips. Until recently, those chips were almost entirely made by NVIDIA. Qualcomm wants a piece of that market with a new product called the AI200 (and its successor, the AI250). So Qualcomm is simultaneously the dominant player in edge AI — the AI that runs on devices in your hand — and a newcomer trying to break into the AI infrastructure business that currently belongs to NVIDIA. --- ## The Strengths Worth Understanding **Qualcomm found the door NVIDIA left unlocked.** AI chips do two fundamentally different jobs. "Training" is when a model learns — you feed it enormous amounts of data and it figures out patterns. This requires massive, brute-force computation and the software that runs it is deeply tied to NVIDIA's proprietary system called CUDA. Nobody credibly competes with NVIDIA at training. The door is locked, the moat is full, and the drawbridge is up. "Inference" is different. Inference is what happens when you actually use a trained model — when it answers your question, generates your image, or summarizes your document. The requirements here are different: efficiency matters more than raw power, and the software dependency on NVIDIA's CUDA is considerably weaker. This is the door Qualcomm is trying to walk through. The research data gives this structural opening the highest importance rating of any Qualcomm-related finding — both Qualcomm's inference chip nodes connect to this "inference bifurcation" concept at the maximum weight of 9 out of 10. The analysts behind this data consider it the most significant structural opportunity available to Qualcomm. **Qualcomm has memory advantages that look significant on paper.** The AI200 chip is designed to hold an extraordinary amount of data close to the processor — 768 gigabytes, compared to NVIDIA's flagship H100 at 80 gigabytes and the H200 at 141 gigabytes. Why does this matter? Modern AI inference is often constrained not by processing speed but by how quickly you can move data in and out of memory. A chip that holds ten times as much data locally can, in theory, run AI tasks far more efficiently and cheaply. This is a real hardware advantage — if it translates into production deployments, which remains unproven. **Qualcomm is everywhere in a way NVIDIA never will be.** Qualcomm's chips run in billions of devices. Every time a phone recognizes your face or transcribes your speech, there is a reasonable chance a Qualcomm processor is doing that work. This gives Qualcomm a scale of deployment at the "edge" — meaning on actual devices, not in data centers — that cloud computing companies cannot replicate. As AI moves from the cloud toward your devices (partly for speed, partly for privacy, partly for cost), Qualcomm's existing installation base becomes increasingly valuable. No other company in this analysis has meaningful presence at both the device level and the data center level simultaneously. **Qualcomm benefits when others fail.** Samsung, which used to compete with TSMC for Qualcomm's manufacturing business, has been struggling badly with its most advanced chip production — yields stuck around 50% when industry standard is closer to 90%. As a result, Qualcomm and most other major chip designers have migrated entirely to TSMC. This was not a strategic masterstroke by Qualcomm; it was the obvious response to Samsung's problems. But it means Qualcomm is now among TSMC's most established and trusted long-term customers, which matters when TSMC has more demand than capacity. --- ## The Vulnerabilities That Keep Strategists Up at Night **The factory problem.** The research data identifies TSMC concentration as the most pervasive vulnerability in the entire dataset — 10 separate connections between Qualcomm-related concepts and the TSMC geopolitical risk concept. If Taiwan were disrupted by conflict, natural disaster, or political crisis, Qualcomm cannot make chips anywhere else at comparable quality or scale. There is no plan B that works. This is not unique to Qualcomm — NVIDIA, Apple, AMD, and most other leading chip companies face the same exposure — but it means Qualcomm's fate is partially in the hands of geopolitical forces that have nothing to do with how well Qualcomm's engineers design chips. **The lesson from Intel's expensive failure.** Intel built a chip called Gaudi 3 that, on paper, matched NVIDIA's best products in performance. It barely registered in the market. The reason: customers had already built their AI systems around NVIDIA's software tools, and switching required rewriting enormous amounts of code. Nobody wanted to do that, even for equivalent hardware. The research data specifically flags this as the most important warning for Qualcomm's data center ambitions — the must-avoid outcome is explicitly labeled "Intel Gaudi3 Software Ecosystem Collapse." Hardware specs are not enough. Qualcomm needs software that makes it easy for developers to use its AI200 chips, and building that software ecosystem from scratch, against NVIDIA's decade-long head start, is genuinely hard. **Double dependency on one supplier.** Advanced chips require not just manufacturing but also specialized packaging — the process of assembling chip components into a final product. TSMC dominates both. Qualcomm depends on TSMC for wafer fabrication and likely for the advanced packaging that makes its high-memory AI200 design work. This is two layers of dependency on one company in one geopolitical location. The research data notes that TSMC's advanced packaging capacity is running at 100% utilization — meaning even if Qualcomm wants to scale up AI200 production rapidly, the bottleneck may be physical capacity rather than demand. **The tariff trap.** Because Qualcomm manufactures nothing in the United States, any tariff regime that favors domestic manufacturing directly hurts Qualcomm relative to competitors. Current US policy is moving in exactly this direction. Companies with US-based fabs benefit; companies that import chips from Taiwan pay the tariffs. Qualcomm cannot easily fix this — building fabs takes a decade and tens of billions of dollars, and Qualcomm's entire business model is predicated on not doing that. --- ## The Non-Obvious Findings The research surfaces a genuinely surprising dynamic: Qualcomm is simultaneously competing with NVIDIA in inference chips and partnering with NVIDIA in a program called NVLink Fusion, which connects different types of processors within NVIDIA's ecosystem. This is not contradictory — it is a hedge. If Qualcomm's AI200 inference chips struggle to gain traction, the NVLink Fusion partnership keeps Qualcomm relevant inside the AI infrastructure market through a different route. If the AI200 succeeds, the partnership can coexist with competition at the product level. The research does not resolve the tension, but the dual-track strategy limits Qualcomm's downside. The other non-obvious finding concerns the failed 2024 acquisition attempt. Qualcomm made a serious effort to acquire Intel when Intel was in deep financial trouble. The deal collapsed. This matters not just as corporate history but as a signal: Qualcomm was willing to fundamentally transform its business model, potentially becoming a company with its own manufacturing capabilities (through Intel's fabs). That option is gone for now, but Intel remains weak, and the structural pressure that made the acquisition attractive has not disappeared. --- ## What the Data Does Not Tell Us The research is unusually candid about its own gaps. The largest missing piece is Qualcomm's exposure to China. Historically, more than half of Qualcomm's revenue comes from Chinese customers — phone manufacturers, technology companies, and others. If US-China trade restrictions tighten further, or if China accelerates its push toward domestically designed chips, a significant portion of Qualcomm's revenue base is at risk. The current research dataset does not have nodes that capture this exposure, which means the vulnerabilities section of this analysis is probably understated. There is also no data on what Qualcomm is actually investing in software to support the AI200. The hardware story is visible; the software strategy is not. --- ## The Bottom Line Qualcomm is a well-run company with a genuine structural opportunity, meaningful established advantages, and a set of risks that are largely not its fault and largely beyond its control. The opportunity — the inference market opening as AI shifts from training to deployment — is real and the research assigns it the highest confidence rating in the dataset. Qualcomm's hardware credentials for that market are credible. The edge AI installed base is a genuine differentiator that nobody else has at comparable scale. The execution challenge is software, and the warning from Intel's Gaudi 3 failure is the most important single data point in this analysis. Qualcomm has to build a developer ecosystem around its datacenter chips fast enough to matter before NVIDIA's software advantage compounds further. This is doable but not guaranteed. The systemic risk — TSMC concentration, Taiwan geopolitics, US tariff policy — is real and significant, but it is shared across the entire semiconductor industry. Qualcomm is not uniquely exposed; it is exposed in proportion to its success as a fabless company, which is to say, substantially. The company's position is best described as: strong fundamentals, credible upside bet, uncontrollable tail risks, and one critical execution dependency (software) that will determine whether the upside bet pays off or becomes an expensive lesson in the limits of superior hardware. --- *Node weights reflect research-assigned importance on a 0-10 scale. Connection counts indicate analytical proximity across graph explorations. Inferences from structural patterns are noted as such.* ## Deep analysis **Sector:** Semiconductors | **Date:** April 2026 **Data basis:** drawn from 42 related concepts and 224 connections uncovered across seven separate research runs --- ## Structural Position Qualcomm shows up in the research in three distinct guises at once: a fabless chipmaker dependent on other companies' factories, a new entrant challenging the incumbent in datacenter AI inference, and a passive beneficiary of rivals' troubles — Samsung's yield collapse and Intel's foundry crisis among them. The most telling signal is what kind of ideas Qualcomm connects to most. Its two strongest connections — Intel's foundry yield-versus-volume bind and TSMC's role as a geopolitical chokepoint — aren't Qualcomm initiatives at all. They're industry-wide conditions Qualcomm has to navigate as a company that doesn't own its own fabs. That points to something important: much of Qualcomm's strategic risk is systemic, not company-specific. It's exposed to chokepoints across the whole industry that it doesn't control and can't really hedge against. Its third-strongest connection is different in kind. Here Qualcomm is the one acting: its AI200/AI250 datacenter inference push and its performance-per-watt inference approach both directly threaten NVIDIA's dominance of GPU economics, and both link to the split between AI training and inference hardware markets at the highest strength recorded anywhere in Qualcomm's part of the research. Three distinct Qualcomm efforts show up in the data: | Initiative | Strength | Character | |---|---|---| | AI200/AI250 datacenter inference entry | 7.5 | Strategic bet, late entrant | | AI200 performance-per-watt inference wedge | 7.0 | Competitive mechanism | | Hexagon NPU edge AI inference dominance | 6.5 | Established structural position | That descending strength — 7.5, then 7.0, then 6.5 — suggests the research sees Qualcomm's datacenter ambition as plausible but unproven, while its edge position is viewed as solid but lower-stakes next to the size of the datacenter opportunity. --- ## Key Strengths **1. The training/inference split as an entry point (durable, if executed well)** The single most important pattern for Qualcomm: both of its inference initiatives connect to the split between AI training and inference hardware at the highest strength recorded for any Qualcomm-linked finding. The research describes NVIDIA's training position as an "unassailable fortress," while inference is characterized as an "open market" that runs on different competitive rules. Qualcomm isn't trying to storm NVIDIA's training moat — it's going after the inference market, where NVIDIA's software lock-in is weakest and power efficiency matters more. **2. Memory capacity as a differentiator (fragile — unproven at scale)** The AI200's 768GB of memory dramatically outsizes NVIDIA's H100 (80GB) and H200 (141GB), and AMD's MI300X (192GB). The research directly frames this as a memory arms race against AMD's own memory-focused inference strategy: Qualcomm is trying to out-memory the company that pioneered the memory-moat approach. But this advantage exists only on paper until real deployments prove out yield, reliability, and software compatibility. **3. A genuinely large edge AI installed base (durable structural position)** Qualcomm's dominance in edge AI inference through its Hexagon NPU is a real, differentiated asset — a deployment scale, across billions of devices, that no cloud GPU vendor can match. Chips like the Snapdragon X2 Elite (80 TOPS, arriving first half of 2026) and the Snapdragon 8 Elite Gen 5 put Qualcomm in the device-level tier of the three-tier AI inference landscape. This position sits upstream of revenue but downstream of manufacturing risk — Qualcomm collects device royalties no matter what happens in cloud inference. **4. The NVLink Fusion partnership (strategically ambiguous, but potentially durable)** NVIDIA's "embrace, extend, co-opt" interconnect strategy lists Qualcomm as an explicit data-center CPU partner, alongside Fujitsu, Marvell, and Alchip. Notably, AMD and Intel are absent from that partner list. So Qualcomm sits inside NVIDIA's ecosystem even as its own AI200 competes against NVIDIA in inference — a dual-track approach that caps the downside if the AI200 doesn't gain traction, while preserving upside if it does. **5. Beneficiary of Samsung's foundry exodus (durable, but externally driven)** Samsung's 3nm yield crisis — yields stuck around 50% versus TSMC's 90%+ — explicitly drove Qualcomm, alongside Apple, AMD, and NVIDIA, to migrate to TSMC. This wasn't a Qualcomm strategic initiative; it was a reactive move. It eliminated Qualcomm's exposure to Samsung's ongoing failure, but it also deepened Qualcomm's dependence on TSMC. --- ## Structural Vulnerabilities **1. The fabless cliff — systemic and unhedgeable (immediate, outside Qualcomm's control)** Qualcomm is named explicitly, alongside NVIDIA, Apple, and AMD, as a company that designs chips but manufactures nothing. This vulnerability rests on TSMC's role as a geopolitical chokepoint — one of the strongest links found anywhere in this cluster of the research — and that chokepoint is Qualcomm's single most heavily connected concept in the entire dataset, making it the most pervasive vulnerability Qualcomm faces. A Taiwan disruption scenario would trigger a projected $2.7 trillion first-year hit to global GDP and a collapse in AI-relevant power capacity — either of which would immediately halt Qualcomm's entire product line. This risk is systemic, not company-specific, and the research finds no evidence that Qualcomm is pursuing meaningful manufacturing diversification. **2. The software ecosystem problem — the Gaudi3 warning (immediate, partly within Qualcomm's control)** Qualcomm's AI200/AI250 push carries an explicit warning attached to it in the research: don't repeat the collapse of Intel's Gaudi3 software ecosystem — the single strongest constraint attached to Qualcomm's datacenter ambitions. The lesson from Gaudi3 is that raw hardware performance isn't enough to unseat NVIDIA without a competitive software ecosystem behind it. Gaudi3's specs matched NVIDIA's H100, but it captured negligible market share because NVIDIA's CUDA software moat held. Qualcomm is entering datacenter inference with a mobile-chipset background, not a GPU software toolchain — making this the company's most controllable vulnerability, but also the one most dependent on flawless execution. **3. Dependence on advanced packaging (long-term, outside Qualcomm's control)** TSMC's near-monopoly on advanced chip packaging is a second concentration risk layered on top of manufacturing concentration generally. Qualcomm's AI200, given its 768GB memory configuration, almost certainly needs advanced packaging to hit that specification — meaning Qualcomm depends on TSMC twice over, for both fabrication and assembly. Packaging capacity is already running at full utilization, with demand outstripping supply. **4. Intel Foundry is a non-option (long-term structural constraint)** A structural trust barrier keeps fabless companies from ever handing next-generation chip designs to Intel's foundry business, because Intel's own product division competes directly with them — and Qualcomm is explicitly named among those companies. That forecloses the only credible US-based manufacturing alternative to TSMC. Even if Intel's 18A process hits commercial yield targets during its 2026–2027 make-or-break window, Qualcomm is structurally unlikely to manufacture there given the competitive exposure of its designs. Apple's own deal to use Intel's 18A process is flagged as a potential exception that could undermine this trust barrier — but Apple has leverage Qualcomm lacks: differentiated consumer products and no competitive overlap with Intel's chip business. **5. Tariff exposure (immediate, outside Qualcomm's control)** A proposed 100% tariff regime on imported chips, which research flags as self-harming for the US, would hit Qualcomm hard, and a related finding shows the tariff policy actively amplifying the fabless cliff vulnerability described above. As a fabless company importing everything it sells from Taiwan, Qualcomm faces the tariff's full impact. A parallel finding shows the tariff policy creating cost advantages specifically for companies that manufacture domestically in the US — an advantage Qualcomm can't access because it owns no fabs. --- ## Competitive Dynamics **vs. NVIDIA** The research frames this rivalry as deliberately narrow in scope. Qualcomm's AI200/AI250 threatens NVIDIA's monopoly economics, and its performance-per-watt approach erodes it further — but both attacks are confined to inference. The logic is explicit: Qualcomm isn't attacking NVIDIA's training moat, where its software lock-in is unbreakable; it's targeting the inference market, where power and memory economics create a real opening. At the same time, Qualcomm's NVLink Fusion partnership makes it an NVIDIA ecosystem contributor too — a tension the research doesn't resolve. **vs. AMD** AMD is Qualcomm's primary inference competitor: the AI200/AI250 directly competes with AMD's MI300X memory-moat strategy, which pioneered the high-memory approach to inference that Qualcomm is now trying to exceed. AMD brings an established (if imperfect) software ecosystem in ROCm and existing hyperscaler relationships that Qualcomm doesn't have. The research treats this as a direct fight and doesn't give Qualcomm any structural edge over AMD — only a hardware-specification edge. **vs. Intel (Foundry)** Qualcomm's collapsed September 2024 bid to acquire Intel crystallized Intel's vulnerability while cementing Qualcomm's role as an outside actor rather than a partner. The trust barrier described above permanently rules out Intel's foundry as a manufacturing option for Qualcomm, making the two companies structural non-partners despite that acquisition history. **vs. Apple** Apple isn't a direct datacenter inference rival, but it's a useful parallel. Apple's unified-memory architecture is shown amplifying the power-capacity constraints on AI data centers, positioning Apple as a competitor to cloud inference providers at the on-device tier — the same tier where Qualcomm's Hexagon NPU operates. On mobile silicon, the two companies compete directly; at the edge-inference tier, they run parallel architectures serving different device ecosystems. **vs. Hyperscalers (custom silicon)** This is the most tangled competitive relationship in the data. Hyperscalers are simultaneously Qualcomm's target customers for AI200/AI250 and its competitors, via their own internal chip development. NVIDIA's ecosystem strategy is shown partially neutralizing the threat that hyperscaler custom silicon poses to GPU vendors generally — which indirectly helps Qualcomm's competitive position in inference, to the extent that NVLink Fusion anchors hyperscaler deployments to an ecosystem Qualcomm is part of. --- ## Regulatory Exposure The research contains limited direct regulatory detail specific to Qualcomm; what follows is inferred from its structural position rather than pulled from dedicated findings. **Tariff regime (high, immediate exposure).** As a pure fabless company with all its manufacturing offshore, Qualcomm faces maximum exposure to the proposed chip tariff. The research notes this tariff is explicitly designed to favor domestic US manufacturers — a category Qualcomm doesn't belong to. Intel, TSMC's Arizona plant, and Samsung's Austin plant are the beneficiaries; Qualcomm bears the cost. **CHIPS Act (indirect, limited).** Qualcomm owns no fabs, so it can't directly access CHIPS Act manufacturing subsidies. The research identifies political risk to the CHIPS Act as a constraint on Intel's national-champion strategy specifically — a dynamic that doesn't directly help Qualcomm. **Export controls (sector-wide exposure).** Escalating US export controls appear in the broader research but aren't directly linked to Qualcomm in this data. Given that Qualcomm has historically drawn 60%+ of its revenue from Chinese customers, this is a real risk that the current research simply underrepresents. **Antitrust (historically significant, but absent from this research).** Qualcomm's substantial history of antitrust exposure — FTC proceedings, fines from Korea's competition authority, EU investigations into its modem patent licensing — doesn't appear in this research at all, because these research runs focused on supply chain and AI compute dynamics rather than IP licensing. --- ## Strategic Leverage Points **1. The inference split — maximum leverage, maximum execution risk** The single highest-strength findings tied to Qualcomm both point to the split between AI training and inference hardware. This is Qualcomm's single highest-leverage structural position. The research confirms that inference is exactly where NVIDIA's software lock-in matters least, where power efficiency favors leaner architectures, and where memory bandwidth — not raw computing power — determines who wins. The AI200's 768GB memory spec speaks directly to the memory-bottleneck problem that constrains inference economics at scale. If Qualcomm nails its software ecosystem, it simultaneously chips away at NVIDIA's monopoly, rides the growing demand created by cheaper, more efficient inference, and secures its place in the fragmented three-tier inference landscape. But the Gaudi3 warning means this leverage point has one hard prerequisite: real software toolchains, not just faster silicon. **2. Spanning edge and datacenter — a genuine structural differentiator** No other company in the research has credible presence at both the edge-device tier (Hexagon NPU, billions of devices) and the datacenter tier (AI200/AI250). Apple has scale on-device through iPhone and Mac silicon but no datacenter inference ambition. NVIDIA dominates training and GPU-based inference but has no edge presence anywhere near Qualcomm's scale. The research identifies this dual-tier position as a real differentiator but arguably undervalues it — Hexagon NPU's assigned strength (6.5) is lower than its strategic significance would suggest, possibly because the story of tying edge and datacenter together into one continuum hasn't yet been proven out. **3. NVLink Fusion as ecosystem insurance** Qualcomm's participation in NVLink Fusion functions as a hedge against the AI200 falling short commercially. If the datacenter inference bet doesn't gain software traction, Qualcomm's position supplying data-center CPUs within NVLink Fusion keeps it relevant in AI infrastructure regardless. It's a lower-upside, lower-risk path that runs alongside the higher-upside, higher-risk AI200 bet. **4. Deep TSMC ties — a passive advantage** Having migrated away from Samsung's failing 3nm process, Qualcomm sits favorably in TSMC's customer priority stack. The broader exodus from Samsung — Google, Qualcomm, AMD, Apple, and NVIDIA all shifting to TSMC — deepens TSMC's customer concentration, which is itself a systemic risk, but it also means Qualcomm is among the manufacturers with the deepest process familiarity and design integration at TSMC. That's a passive advantage that doesn't require active management, but would be costly for a rival to replicate. --- ## Open Questions **1. China revenue exposure.** Qualcomm's historical dependence on China for 60%+ of revenue is conspicuously missing from this research. Broader findings about supply-chain splitting, tightening export controls, and geopolitical manufacturing lock-in all suggest China-linked revenue is structurally at risk. Whether Qualcomm's business model survives a full US-China decoupling is the most important financial question this research doesn't answer. **2. How much is Qualcomm actually investing in AI200 software?** The research flags the requirement to avoid Gaudi3's software-ecosystem collapse, but it contains nothing quantifying Qualcomm's actual software investment, SDK development, or customer engagement for AI200. This is the critical unknown that will determine whether the inference hardware bet succeeds or follows Gaudi3 into irrelevance. **3. Is Qualcomm diversifying its manufacturing beyond TSMC?** The research identifies TSMC concentration as Qualcomm's single most pervasive vulnerability, but contains nothing about chiplet strategies, Samsung trials, or Intel 18A exploration for non-competing product lines. Whether any manufacturing diversification is underway — even for older chip designs — is unaddressed. **4. ARM licensing dynamics.** Qualcomm's Hexagon NPU dominance rests heavily on Snapdragon's ARM-based architecture. ARM's licensing terms, and Qualcomm's own legal history with ARM (settled in 2022 but subject to renewal risk), represent a foundational dependency the research doesn't capture. **5. How does Qualcomm actually make money on inference?** The research covers Qualcomm's hardware positioning but never explains how Qualcomm plans to monetize AI200/AI250 deployments — outright hardware sales, cloud partnerships, or recurring licensing. A separate finding describes the broader inference-as-a-service market as increasingly commoditized, but doesn't clarify where Qualcomm's own go-to-market fits into that picture. **6. Is the Intel acquisition question really closed?** The collapsed 2024 acquisition bid is documented, but the research doesn't say whether that pressure has actually resolved or is simply dormant. Intel's continuing weakness — operating losses at its foundry business and a make-or-break window in 2026–2027 — keeps the door open to a renewed acquisition attempt, or to a defensive restructuring by Intel that would reshape the competitive picture entirely.