The Apple–Alibaba Qwen Gambit: A Forensic Breakdown of the Centralization Trade That Just Reshaped Crypto AI

On August 8, China's Cyberspace Administration posted a register of approved generative AI services. It listed Huawei's Xiaoyi assistant. It listed OPPO's AndesGPT. And buried in the same block—a fact that should have hit crypto's AI narrative like a waterfall liquidation—it listed Apple Intelligence, powered by Alibaba's Qwen models. Not Baidu's ERNIE, which months of rumors had flagged as the front-runner. Not ByteDance's Doubao, the consumer favorite. Alibaba's Qwen, integrated at the operating-system level.

The anomaly is the non-reaction. Alibaba's equity barely twitched. Render, Bittensor, Akash, all flat. Crypto's AI token sector didn't even blink. The market has been conditioned to treat China AI headlines as noise, and that reflex is exactly what a battle trader learns to fade. When an announcement changes the structural ownership of inference supply in the world's second-largest economy, and token prices fail to reprice, you're looking at a mismatch between narrative and reality. This isn't about a smartphone feature set. It's about who controls the fastest-growing compute sector on the planet, and what the winning architecture—centralized, compliant, walled—says about a thousand whitepapers that claimed the future would be decentralized, open, and permissionless.

Chaos is not a bug; it is the raw material. And this deal just minted a lot of it.

Context: The Fortress Just Opened a Back Door

Apple Intelligence was announced at WWDC 2024 as Apple's answer to the generative AI wave. The architecture was, by design, conservative: on-device inference first, using Apple's custom neural engines on the A17 Pro and M-series silicon; a new layer called Private Cloud Compute for requests that exceed on-device capability; and an operating philosophy of data minimization baked into the platform. The design principle was unambiguous—Apple owns the full stack, from silicon to orchestration, and no third party touches your prompts.

Then China happened.

China is Apple's second-largest market and a strategic battleground where Huawei's resurgence has been eating into iPhone share since the Mate 60 surprise in 2023. But Chinese law under the Interim Measures for the Management of Generative AI Services requires that public-facing generative AI services be registered with the CAC before operation. Apple cannot deploy its proprietary foundation models in China without clearing that gate. And clearing it with self-owned models is politically delicate and technically suboptimal for Chinese-language performance—a language domain where domestic open-weight models still hold meaningful advantages. So Apple made a decision that the old, vertically integrated Apple would have fought: contract the model layer of its flagship AI product to a Chinese partner.

The partner is Alibaba. The model family is Qwen—Tongyi Qianwen in full branding—an open-weights family that has topped Chinese-language benchmark leaderboards in head-to-head matchups against Baidu's ERNIE, ByteDance's Doubao, and even DeepSeek in several evaluation suites. The integration is registered and operational. Chinese users get Qwen-powered features inside Apple's system-level AI interface, without switching apps, without installing third-party clients.

The timing is the first tell: the CAC register entry and Apple's announcement landed the same day. This was not a random press drop; it was a coordinated regulatory clearance event. Both parties wanted approval on the public record before the marketing machine started moving. That coordination carries a second, deeper signal—a Chinese regulator and an American tech giant synchronized their communication cadence. In crypto terms, it's as if a major exchange license and a token listing announcement were pre-cleared by the same authority before going live. Unusual. And it tells you just how much pre-negotiation happened behind closed doors to make this work.

Core: Reading the Architecture the Way I Read Bytecode

The Boundary Is the Adversary

Let's get into the technical stack, because that's where the real information lives. The most important sentence in the entire announcement is not "Apple collaborates with Alibaba." It's the claim that Chinese users get Qwen capabilities without switching apps. That means system-level integration. That means Apple's AI orchestration layer—the one that decides which requests run on-device, which go to Apple's Private Cloud Compute, and which are handed to third parties—now has a routing rule for the China region: send complex reasoning to Qwen.

Here's the architectural tension nobody was paid to unpack. Private Cloud Compute was built to be Apple's isolated compute enclave. Its security guarantees rest on hardware attestation: the PCC clusters are Apple-controlled, Apple-audited, and cryptographically verifiable through published transparency logs. The entire premise of PCC is that no external jurisdiction, no vendor, and no cloud operator can intercept the payload. When you bolt a Chinese third-party cloud onto the back end of that pipeline, you have broken the attestation boundary. The Qwen portion of China's Apple Intelligence cannot be running inside Apple's PCC. It has to be running inside Alibaba Cloud, in a CAC-supervised environment, subject to Chinese data sovereignty law. The "contextual awareness" that Apple advertises across the boundary becomes a mixed-trust system: basic Siri-level requests stay with Apple; complex generative requests land on Alibaba's side. Apple isn't going to publish a transparency log for Alibaba's inference nodes. It can't. The two security models are philosophically irreconcilable.

This isn't a flaw in execution; it's a boundary that was designed on day one of this collaboration, because there is no way to design around it. The user prompt passes through Apple's front-end layers, then gets wrapped, transported to Alibaba's cloud, run through Qwen, and the response flows back through Apple's interface. At the boundary, there is a data transfer, a request translation, a drop into a third-party trust domain. This should be deeply uncomfortable for anyone who bought "Apple guarantees your data stays private" as a blanket statement.

I flash back to the 2020 Uniswap arbitrage grind whenever I see a boundary like this. My team ran over five thousand arbitrage trades in three months on Ethereum mainnet, and the edge was always at the boundary—the mempool boundary, the gas-pricing boundary, the reorg boundary. Every bridge in DeFi is a boundary, and every boundary is a vulnerability. Here you have Apple, which built a verified-security model, bridging to a boundary it cannot verify, with zero cryptographic integrity between the two domains. In DeFi we'd call that an upgradeable bridge contract with no timelock and an admin key we can't see. The fact that consumers won't feel the risk doesn't mean it doesn't exist; the LUNA collapse taught me that the market never feels a structural flaw until the exact moment it does.

The engineering problem is deeper than trust. Apple has tuned years of user-experience polish around deterministic, low-latency interactions. Siri's failure modes are a running joke, but at least the infrastructure is first-party; Apple can tune the entire stack. With Qwen, Apple's team is integrating against a third-party model API whose version, quantization, and deployment schedule Apple does not control. Notice that the official announcement refuses to specify which Qwen variant is deployed. That matters enormously. Qwen-2.5's 72B configuration is a fundamentally different product from the mobile-scale 1.5B distilled variant. Capability differences are non-linear: reasoning depth, tool-calling accuracy, and multilingual nuance all jump between sizes. If Apple deployed the smaller model to control latency, Chinese users get a degraded Apple Intelligence experience relative to US users. If Apple deployed the large variant, the request routing becomes aggressively cloud-dependent, which reintroduces latency and cost.

Let me do the latency math. China's average mobile network round-trip to an in-country inference endpoint is not the problem—data must remain in-country, so the endpoint sits on Alibaba's mainland infrastructure. But a complex prompt hammering a 72B-parameter model at FP8 precision yields time-to-first-token on the order of 400 to 800 milliseconds on strong infrastructure. Add the network hop, add Apple's orchestration layer, add contention under load. Under realistic usage, users can absolutely see 1.5-second perceived cold responses. Compare that to an on-device distilled model where first token lands in 30 to 80 milliseconds. Users experience this difference viscerally, and Apple's marketing machine has just tied its China brand to a backend it cannot tune, cannot inspect, and cannot even announce changes to.

In my 2021 NFT floor-sweeping days, I learned a simple rule: when you buy an asset you cannot inspect beyond the metadata, you are not an investor—you are a bag-holder with extra steps. Apple is now integral to a model pipeline it doesn't own. The CFTC-style, regulated, vertically integrated Apple narrative is over. In its place is a bridging structure that looks suspiciously like a DeFi composability stack: front-end contract, external oracle, one admin key. And we all know how those stories end when the admin key rotates without warning.

The Compute Demand Is a Fiscal Event

Now the second-order effect, which is where my P&L instincts kick in. Alibaba Cloud is China's largest cloud provider and one of the few private operators with substantial GPU clusters. But Apple China is a different order of magnitude from any existing Alibaba inference customer. Let's build a conservative forecast.

Apple's installed base in China is estimated at roughly 200 to 250 million active iPhones, plus iPads and Macs. Assume the full Apple Intelligence experience is limited to iPhone 15 Pro and newer hardware, since the on-device accelerator requirements are strict. That's maybe 45 to 60 million eligible devices at launch. Now assume 20% of eligible users actively engage with Qwen-powered features, and each active user generates five inference-demanding requests per day. That yields 6 million users multiplied by 5 requests—about 30 million complex inference calls per day. If the average request involves a 7B-parameter model, each call costs roughly 40 to 100 generated tokens plus system overhead. But if Apple routes a meaningful fraction of requests to 72B-class models, the per-call compute cost multiplies by a factor of ten.

Let's be concrete. A 72B model serving under concurrent load at 4-bit quantization requires roughly 40 gigabytes of GPU memory just to host weights, spans two to four GPUs depending on architecture, and generates responses at a per-node throughput of maybe 30 to 80 tokens per second under batch. Thirty million daily long-form calls at a modest average of 200 output tokens each comes to 6 billion output tokens per day. At 200 tokens per second sustained per GPU with efficient batching, a single GPU serves roughly 17 million tokens per day. Divide. You need a minimum of 350 GPUs for the absolute low-end base case, and realistically 2,000 to 5,000 GPUs dedicated solely to Qwen inference for Apple China—before redundancy, before headroom for peaks, before model updates and retraining pipelines. That is a multi-thousand-GPU commitment just to satisfy the conservative scenario.

Here's the market signal buried in that arithmetic. Alibaba must either repurpose its existing mainland GPU fleets—heavily invested in A100/H100-class hardware before export controls tightened, plus an escalating portfolio of domestic alternatives like Huawei Ascend—or it must allocate fresh capital to expand. There are two hard constraints. First, US export controls mean H100-class access for mainland data centers is capped; the effective ceiling on cutting-edge Nvidia capacity in China is fixed by third-country sourcing and grey-market channels. Second, domestic alternatives such as Ascend 910B have improved but still lag in tensor-core efficiency for large-model inference workloads. This is a real supply squeeze embedded inside the deal.

The pattern is eerily familiar to anyone who watched Layer-2 fee markets after the Dencun upgrade. The ecosystem celebrated EIP-4844 for bringing near-zero blob gas fees. But blob data capacity is finite; as more rollups ship and each expands its block production, blob space saturates and fees climb back toward meaningful levels. Same sequence: cheap, abundant capacity; optimistic onboarding; capacity exhaustion; the return of scarcity pricing. The Apple-Alibaba arrangement is the same playbook applied to inference in China. Initially, usage will grow as the features get discovered. But infrastructure sized for 5% adoption will be swamped by 20% adoption, and the upcharge dynamic—latency spikes, retry storms, capacity rationing—will resemble blob gas pressure almost exactly.

Now connect the dots to crypto. The decentralized AI infrastructure thesis—Render, Bittensor, Akash, IO.net, Hyperbolic, and a dozen others—has sold one story: idle GPUs everywhere will coalesce into a viable alternative to hyperscale clouds, and AI inference demand will route toward the most efficient, permissionless capacity. The Apple-Alibaba arrangement is the strongest possible evidence against that thesis in its current form. When the world's largest consumer technology company needed guaranteed, regulated, low-latency, low-risk inference in the world's second-largest market, it didn't go to the open market. It went to a sovereign-compliant hyperscaler with an existing regulatory relationship. Institutional AI demand does not want cheaper compute. It wants legally sanitized compute. The procurement gatekeeper for massive enterprise inference demand is compliance, not price-to-performance. Permissionless GPU networks structurally cannot win an institutional bidding process that starts with regulatory vetting.

Here is the core insight: decentralized GPU networks are not competing with Amazon, Microsoft, or Alibaba on price—they are disqualified on compliance before the technical evaluation even starts.

I have run this playbook before. When my team audited Terra's contracts in 2022, we didn't need to simulate the death spiral to predict the outcome; the stablecoin's worst-case collateral logic was visible in the contract itself, and the market-cap-to-reserve ratio was the tell. The "reserve" in this case is regulatory relationship. Apple's compute procurement is now inseparable from Alibaba's CAC-approved status. The moat isn't throughput. It's license. And crypto AI's token markets are still trading as if throughput were the product.

The Regulatory Template Is the Real Product

Now the dimension most crypto-native analysts ignore, which is precisely why it moves markets last: the regulator.

The same CAC register that cleared Apple Intelligence also cleared Huawei's Xiaoyi and OPPO's AndesGPT. Three mobile AI assistants, same batch. That is not an accident; it is a deliberate regulatory event. China is formally creating a new licensing category: the OS-embedded generative AI assistant, distinct from standalone chatbot applications. The legal framework that previously targeted chat apps has been extended to the operating-system layer. The message to every device manufacturer in China is simple: you are now a regulated AI distribution platform, and your model partners must be registered, auditable, and compliant.

The implications run in two directions. First, the global tech industry: Apple's acceptance of a CAC-registered third-party model is a precedent that Samsung, Google where legally permissible, and every hardware vendor with China ambitions will study. The playbook for doing AI business in China is now written: pick a registered domestic model provider, contract under CAC rules, announce only after regulatory clearance. Second, the crypto industry: think about what this means for the entire "AI on blockchain" value proposition. Decentralized models are designed to be uncensorable, and this register just demonstrated that China treats the AI stack as a licensing domain—like banking, not like an open protocol. Whatever your views on censorship resistance, the marginal cost of compliance for building AI products in China just went up dramatically, and the marginal benefit of "permissionless" went down for any business that touches mainland users.

I built and launched an AI-agent trading protocol in 2025 with institutional clients, and part of the workflow involved Chinese-language models for sentiment extraction. The operational reality of running those models in a regulated environment is the difference between shipping a product in three quarters and waiting eight. Nobody on a token Discord is modeling for that. They should be, because the licensing pattern is spreading globally—the EU AI Act, India's advisory frameworks, Brazil's first drafts. The Apple-Alibaba deal is the first verifiable proof that "system-level AI integration plus regulator plus market leader" is a replicable template. Every time it replicates, the surface area for crypto's permissionless AI shrinks by another coat of paint.

There's a darker edge to this template. A registered, OS-level AI assistant is also a censorship vector that makes life harder for any decentralized alternative. When the assistant is embedded on the phone, the phone is the compliance point. Enforcement stops being about blocking websites; it becomes about attestation at the silicon level. If this template spreads over the next 24 months, the retail devices that crypto networks rely on become endpoints of regulated AI infrastructure, and the regulatory lens flips from "who is mining" to "who is reasoning."

Privacy: Apple Didn't Even Call a ZK-Proof

Let's focus on the privacy dimension, because that's where new information is hiding in plain sight. Apple has spent a decade building a privacy brand on the claim "we can't see your data even if we wanted to." The Secure Enclave, differential-privacy layers, on-device processing, App Tracking Transparency—this is a fortress of privacy marketing, and much of it was genuine. Private Cloud Compute was supposed to extend that fortress into the cloud.

The Qwen integration introduces a third-party model provider that, under Chinese law, has specific access and retention obligations. The user's Chinese-language prompts—often intimate: health queries, relationship arguments, work documents—will transit through Alibaba's servers. Does Apple have an anonymization contract? Almost certainly. Does Apple's privacy white paper for China explain exactly what happens to the data? Not yet. Does the architecture provide any cryptographic proof that user prompts are not being logged for training? Absolutely not.

And here is the part that should stab every "AI plus crypto" bull right in the thesis: we already have the cryptographic tools to prove inference integrity—zkSNARKs, TEEs, federated learning, verifiable inference protocols. None of them appear to be part of this arrangement. The technology that could reconcile Apple's privacy brand with a third-party inferencing partner exists. It runs on centralized hardware and, in principle, on decentralized networks. Apple, with an essentially unbounded R&D budget, chose a plain network API call.

The honest caveat is that verifiable inference has a long way to go before it can serve tens of millions of users at consumer latency. But Apple's choice is not just a technical limitation; it's a value statement about the priority of provable privacy relative to shipping speed. After the Terra audit, I became allergic to "we have a white paper describing a security model" as a source of truth. I want to see the attestation log. I want to perform the challenge-response. I now run per-inference attestation hooks in the agent protocols I deploy, not because I'm a decentralization romantic, but because my P&L proved that unverifiable oracle feeds get exploited. When Apple's China users cannot verify which model processed their prompts, they are in exactly the same position as a DeFi user trusting an oracle without checking the deviation thresholds. The latency problem that keeps me awake in DeFi—the gap between the source and the book—is structurally identical here: Apple is the interface, Alibaba is the book, and the user just hopes the feeds are honest.

Why Qwen Won, and What the Selection Tells Us

Now let's address why Alibaba landed the deal, because the selection process is itself a market signal.

The rumor mill originally had Apple talking to Baidu. Later reports suggested Apple was in talks with Baidu, Alibaba, and ByteDance simultaneously. The final choice was Alibaba. The tell is in the platform capabilities. Baidu's ERNIE models are strong in Chinese-language benchmarks, but Baidu Cloud's infrastructure is not in the same league as Alibaba's. ByteDance has Doubao, the consumer app leader, but ByteDance's enterprise cloud business is less mature. Alibaba had the only combination of a top-tier open-weights model, hyperscale cloud with GPU depth, early CAC compliance, and a developer ecosystem that is the closest thing China has to Hugging Face plus AWS combined. No one else matched that pile of attributes.

The consequence is not just that Alibaba won a contract; it means "model plus cloud plus compliance" is now the winning product package for enterprise AI supply. That's a bearish signal for pure-play model companies with no cloud, and a bullish signal for vertically integrated stack owners.

The Baidu angle is genuinely painful to watch. If Baidu did lose this deal after reports that it was at the table, the strategic damage goes beyond missed revenue. It's a second lost era for a company that held some of China's earliest AI patents but has a track record of fumbling consumer distribution. For the broader AI landscape, it's a reminder that open ecosystems win developer mindshare, and developer mindshare influences procurement decisions at consumer scale. Qwen's open-weights strategy—dismissable to some as commoditization—turned out to be a commercial weapon. It created availability and integration case studies that closed ERNIE or Doubao couldn't match. In 2017 I chose to audit unknown ICO contracts precisely because I could read the bytecode before trusting the pitch; Qwen won this deal in part because developers could actually run and test the models before the procurement committee voted. Open weights are the competitive moat of the Chinese AI stack.

There's a second competitive question that matters more for crypto: is this deal exclusive? The announcement language is ambiguous. In Apple's App Store, multiple search engines coexist in the browser; Apple has historically avoided single-vendor lock-in where it weakens negotiating leverage. But for the Chinese AI experience, a single model is cleaner, and for Alibaba's stock, exclusivity would be worth billions of dollars of narrative premium. For crypto AI, this distinction is existential. The multi-model integration pattern—an "App Store for models"—is exactly the future that decentralized routing networks claim to be building. If Apple ships a centralized model picker inside iOS before any decentralized network achieves meaningful external adoption, the "decentralized model routing" narrative loses its beachhead.

Where the Trade Actually Is

Let's translate all of this into tradable signals.

For Alibaba, this is an acceleration of the cloud-plus-AI narrative. Alibaba Cloud has needed an anchor case for its infrastructure story, and Apple is the anchor. For the planned Alibaba Cloud IPO—a thesis that has been circulating for two to three years—this deal provides a "global Tier-1 customer" storyline that no other Chinese tech company can claim. But the nuance is that short-run revenue contribution will be tiny relative to Alibaba's overall scale. Inference-as-a-service at consumer scale is real revenue, but priced at heavily negotiated rates; Apple treats this as cost of goods, not as a partnership of equals. Apple controls the front end; Alibaba becomes a premium commodity supplier. The strategic value is to Alibaba's future narrative, not its current income statement.

For Apple, this is defensive. The alternative is losing China's high-end phone market to Huawei and other AI-accelerated domestic flagships. Apple's AI story in the West is "Apple Intelligence will transform the phone." In China, without a local model, that story couldn't exist. The market reaction—flat Apple, flat Alibaba—reflects the fact that both companies' valuations had already discounted the regulatory breakthrough through months of rumors. The actual trading alpha is in derivative reads: Chinese AI supply-chain names, GPU procurement announcements, and the cloud infrastructure supply chain around Alibaba.

This is where I bring back my NFT floor-sweeping rule: price moves happen when a belief is validated by data, not when a rumor is confirmed by an announcement. The rumors have been circulating for months; the announcement is a non-event for anyone who was paying attention. The trade lives in the follow-ons. I'm tracking three signals over the next quarter. First, Alibaba Cloud's GPU procurement and capacity announcements—if they issue fresh purchase tenders or expand a data center footprint in a thousand-rack-plus block, that's a "compute demand confirmed" signal. Second, Apple China's technical support documentation—model version disclosure, data-handling details, latency SLAs; any sign of a large Qwen variant deployment is a "complex reasoning demand" signal. Third, Alibaba's next earnings call—any upward revision to Cloud/AI revenue guidance. Each of these signals is more actionable than the original announcement itself.

Contrarian: The Bearish Read Is Too Lazy

Now let me invert the thesis, because Chaos is not a bug; it is the raw material.

The obvious bearish read is: Apple plus Alibaba equals centralization wins, decentralized AI dies. I think that is too lazy, and lazy conclusions are how you lose money on the next rotation.

Here is the contrarian take: this deal is the most powerful catalyst the decentralized AI privacy narrative has ever received, because it collapses Apple's privacy fortress brand in the world's largest consumer market. For a decade, Apple could claim privacy as a product differentiator. When Chinese users' prompts now transit Alibaba's cloud—and when the actual data-handling mechanics get disclosed or leaked—the contrast with what is cryptographically possible becomes inevitable. A huge cohort of users will suddenly understand, for the first time, that "your data stays on your device" was actually "your data stays on your device unless you use this one really useful feature." And that one feature is now the most compelling AI capability on their phone.

That sudden awareness is demand for a new trust architecture. Not one where a corporation promises, but one where a verifiable system proves.

The second inversion: the compute-squeeze math I laid out earlier is actually the strongest bullish argument for decentralized GPU infrastructure in the mid-term. Alibaba's sanctioned, compliant capacity is finite. Apple-scale demand does not fit in that box. If Qwen-powered Apple Intelligence experiences latency problems, retry storms, or capacity degradation during peak hours—and any 30-million-calls-per-day launch will hit some wall—the "open market for compute complementing hyperscaler capacity" story gets its first institutional attention. I've seen this movie in institutional trading: when the regulated venue hits capacity, smart money discovers the alternative venues. Decentralized GPU supply doesn't need to beat Alibaba for the Apple contract. It needs to be standing at the exit when Apple's demand overflows.

The third inversion: the regulatory template cuts both ways. If CAC-style licensing becomes the global model for AI regulation, that is a massive headwind for permissionless AI. But it is also an accelerant for sovereign AI stack competition. The European and Indian responses to "US big tech plus Chinese big tech" alliances will be to sponsor their own model-plus-infrastructure stacks. Sovereignty-driven demand is more diverse than profit-driven demand, and it carves out political niches for open, self-hosted, and eventually decentralized infrastructure. The EU AI Act is not a death sentence for open AI; it's a forcing function for compliant open infrastructure. Crypto AI projects that actually lean into "auditable, sovereign, verifiable"—rather than "rent-a-GPU"—will find the institutional aisle opening.

The fourth and simplest inversion: the market hasn't priced any of this precisely because the token market reacted with a shrug. When an event produces zero priced-in response but changes the structural competitive position of an entire sector, you have a setup for a violent catch-up move in one direction or the other the moment real data shows up—the first latency complaints, the first model-version revelation, the first GPU procurement rumor. My claim is not that crypto AI tokens go to zero or to the moon. My claim is that the current price is a latent-volatility setup, and the side you take depends on whether you're trading the first-order effect (centralization, bearish) or the second-order effects (awareness, overflow demand, sovereignty, mixed-to-bullish for specific subsectors).

Takeaway

We don't trade headlines; we trade the gap between what's announced and what's deployed.

The Apple-Alibaba Qwen registration is a verified fact. What is not verified: which model version, which data boundary, which revenue split, whether the deal is exclusive, whether Alibaba's infrastructure can handle the load. I'd put the probability at 60 to 70 percent that early user experience reports show meaningful degradation under heavy load, and I'd put the probability at 80 percent that Apple's China privacy documentation—when it is finally published—will describe the data handoff in ways that uncomfortably complicate the global privacy narrative.

For investors, the checklist is straightforward: track Alibaba's GPU procurement, Apple China's technical documentation, and the next CAC register expansion. For crypto, understand that the Apple-Alibaba deal is not a death blow to decentralized AI; it is the most visible demonstration to date that the centralized AI supply chain is about to hit capacity and trust limits simultaneously. The next bull run in AI tokens will be led by projects that can show actual institutional-grade verifiable inference, not GPU-resale dreams.

Speed is the only currency that doesn't depreciate while you're still reading the press release. And in this market, the fastest edge is reading the architecture, not the banner.