AI Gateways and the AI OS Value Chain: A Deep Dive

Core Conclusions

• Over the next two to three years, no single player will "fully capture" the AI gateway; instead it will form a two-layer control structure. The bottom layer (operating systems, browsers, search, and super-apps) owns distribution, permissions, and default placement; the top layer (personal AI assistants and agent platforms) competes for control over "intent interpretation, context integration, and task execution." On current share, browsers remain highly concentrated in Chrome, Safari, and Edge, and search is still dominated by Google, so for an AI-native gateway to truly displace the traditional gateway it must win the default slot, the permission slot, or a high-frequency workflow slot. StatCounter shows that in April 2026 global browser share was Chrome 68.02%, Safari 17.04%, and Edge 5.53%; global search share was Google 90.02% and Bing 5.14%. The EU's DMA already treats Android and iOS/iPadOS as core platform services under priority scrutiny, which also means OS-level gateways remain the strongest infrastructure.

• The most likely first controllers of the "AI OS layer" are not pure model companies, but three classes of incumbent platforms that already hold permissions and context. First, Google/Apple/Microsoft, the owners of OS, browser, and productivity platforms. Second, Meta/Amazon/Tencent, platforms with super-distribution scenarios and user graphs. Third, OpenAI/Anthropic/Perplexity, the AI-native companies; but unless they win system permissions, default distribution, or high-quality connectors, they can only serve as "upper-layer coordinators" first and will struggle to become the underlying OS directly. Apple's Apple Intelligence explicitly emphasizes on-device processing and Private Cloud Compute; Google is connecting Gemini to Gmail, Photos, YouTube, Search, and Chrome; Microsoft is embedding Copilot deeply into Windows, Microsoft 365, the Agents SDK, and its enterprise permission system.

• The key to upgrading a personal AI assistant from a "Q&A tool" to a "cross-app execution layer" is not better chatting, but four things: obtaining personal context, obtaining callable tools/apps, obtaining system-level permissions, and establishing approval and security governance. OpenAI has put the Apps SDK, MCP access, and computer use into its product stack; Anthropic provides standardized tool calling through MCP and computer use; Apple exposes app actions to Siri and Apple Intelligence through App Intents; Google pulls workflows such as Gmail, Calendar, Flights, and Shopping into Gemini through Connected Apps, Chrome Gemini, and Personal Intelligence.

• The biggest difference between an AI OS and a traditional OS is not that it replaces the kernel, but that it adds an "intent–context–permission–execution" orchestration layer. The core objects of a traditional OS are applications, files, windows, processes, and permissions; the core objects of an AI OS become user intent, long-term memory, cross-app actions, model routing, approval flows, and result write-back. The moat of an AI OS is therefore not just model capability, but local permissions, the personal data graph, enterprise connectors, developer protocols, and security governance. Apple's Private Cloud Compute, Samsung's Personal Data Engine, Google's Personal Intelligence, and Microsoft's Copilot/Agent mode are all early forms of this layer.

• What lands first and carries the highest revenue certainty is not the "fully automated personal agent," but three scenarios:

Enterprise AI gateways (Microsoft 365 Copilot, Salesforce Agentforce, ServiceNow, Glean, Atlassian Rovo), advertising and subscriptions for AI search/answer engines (Google Search AI Overviews, ChatGPT Search, Perplexity), consumer-subscription AI assistants (ChatGPT, Gemini, Claude, Perplexity Pro, Alexa+). Among these, the monetization certainty of enterprise gateways and search advertising is clearly higher than that of pure consumer agents.

• The companies already showing clear "AI gateway monetization signals" are concentrated mainly in platform giants and a handful of enterprise software platforms. Google holds 350 million paid subscriptions on one side and is expanding Search/Shopping ads within AI Overviews on the other; Microsoft's annual report disclosed that the Copilot family exceeds 100 million MAU across commercial and consumer endpoints combined; Meta AI has surpassed 700 million MAU, and Meta has announced it will use users' interactions with AI for content and ad personalization; Salesforce disclosed Agentforce ARR of 800 million USD with 29,000 deals signed; Glean disclosed ARR above 200 million USD; ServiceNow disclosed that the number of customers with Now Assist annual ACV over 1 million USD grew more than 130% year over year.

• Apple, Meta, Amazon, and Samsung currently look more like they are using AI to protect existing gateways than to create a separately and cleanly carvable new profit pool. Apple Intelligence still centers on hardware experience, privacy, and ecosystem retention; the strongest "personal context + screen awareness + App Intents" capabilities on its official pages remain on the roadmap of future software updates. Meta AI's direct-payment evidence is weaker than its user scale, and at this stage the more important increment lies in content distribution and the ad-training loop. Alexa+ bundles via Prime, clearly prioritizing protection of the e-commerce/membership and home gateway. Samsung Galaxy AI still mostly serves hardware differentiation and defensive functions on the Android side.

• What AI-native challengers truly threaten most is not the phone OS itself, but the traditional "open app, search manually, switch pages manually" traffic-distribution logic. OpenAI, Perplexity, Anthropic, The Browser Company, Brave, and DuckDuckGo are merging chat, search, browsing, files, and web operations into a single interface. Perplexity Comet has explicitly defined the browser as a "personal assistant" that can manage email, shopping, calendar, and travel; Brave Leo and Duck.ai treat "privacy-first in-browser AI" as their differentiator; OpenAI and Anthropic are pushing computer use toward a "general software execution layer."

• The track with the greatest revenue elasticity is not the most science-fiction one, but the one closest to existing profit pools: First, search advertising migrating to AI results pages; second, enterprise software migrating to agent licensing and outcome/usage-based billing; third, cloud inference and model calls; fourth, consumer AI subscriptions; fifth, transaction commissions from agents calling e-commerce/payment/local services. The best margins still usually come from high-margin software, subscriptions, and advertising, not from the hardware itself. Apple's Services gross margin in FY2025 was 75.4%, well above the 36.8% product gross margin; Meta's 2025 advertising revenue was 196.2 billion USD; Google Services explicitly places search advertising, subscriptions/platforms/devices, and Google Play within the same revenue framework.

• The most stable platform-core positions are still Alphabet, Microsoft, Apple, Meta, and Amazon; among AI-natives the strongest are OpenAI, Anthropic, and Perplexity, but their "entry rights" remain weaker than the OS/browser default slot. To capture the profit pool over the long run, AI-native companies must upgrade from a "chat box" to "connectors + permissions + app calls + developer platform + transaction loop." OpenAI is walking this path along the Apps SDK, MCP, computer use, and Enterprise apps; Anthropic builds the protocol and tool layer through MCP and computer use; Perplexity builds the browser/search/transaction loop through Comet, Instant Buy, the Publisher Program, and its API.

• What the market most easily overrates at this stage is companies with a "gateway narrative but no evidence of retention, payment, developer ecosystem, or permission completeness." Especially: those that launched only a chat gateway with no system/enterprise permissions; those with trials but no paid upgrades; those with plugin ambitions but no standardized tool protocol; those with DAU growth but no enterprise repurchase or ARR; those with only a browser shell or search shell but no default slot, security policy, or distribution channel. For such companies one should strictly distinguish five stages: "product launch," "user trial," "retention/payment validation," "revenue landing," and "scaled adoption." The monetization evidence of Anthropic, Perplexity, and Glean is now clearly stronger than that of most pure-concept agent companies, yet still weaker than Big Tech's system-level moats.

• AI assistants will erode part of the value of traditional search, browsers, the App Store, SaaS front ends, and phone apps, but not in a synchronized, linear, or wholesale way. The first to be compressed are the information-retrieval layer, low-value-added front ends, weak-workflow apps, and long-tail SEO traffic; the last to be replaced are products requiring strong interaction, strong trust, heavy regulation, strong payment liability, and a strong professional UI, such as professional design, professional trading, and complex enterprise admin consoles. OpenAI's and Anthropic's computer use, Google Chrome's automated browsing, and Samsung's Cross App Action all prove "cross-page execution" is becoming feasible, but large-scale unsupervised automatic execution is still constrained by permissions, payment, legal liability, and security issues.

• The six biggest catalysts over the next twelve to twenty-four months: Siri truly shipping "personal context + cross-app actions"; Gemini's default penetration across Android/Chrome/Search and the expansion of AI advertising; scaled call data for ChatGPT apps / enterprise apps / agents; proof of high-frequency Copilot usage in Windows and Microsoft 365; whether AI browsers like Perplexity/Comet can win a substantive default slot or transaction commissions; and regulatory landings in the US, Europe, and Japan on the App Store, browsers, default search, and data portability.

• The biggest risk is not falling behind on models, but gateway monetization coming in below expectations: users unwilling to pay continuously, OS-level free bundling, still-high inference costs, mounting privacy and antitrust pressure, rising content/IP litigation, and developers ultimately choosing to "build around existing OS/office/browser platforms" rather than around an independent AI gateway. The Chrome-divestiture proposal in the Google search case, the Apple smartphone monopoly case, and the continued push of the DMA and Japan's MSCA on App Store and browser rules all show that gateway-type profit pools will be revalued by both AI innovation and regulation going forward.

Value Chain Panorama and Gateway Control

Gateway control is shifting from single-point traffic distribution to multi-layer orchestration

From an investment perspective, the battle for the AI gateway is essentially not about "whose model is smarter" but about who holds stronger default distribution, system permissions, user context, developer interfaces, and a commercial loop. Today's control landscape still tilts heavily toward incumbent platforms: the browser gateway is held by Chrome/Safari/Edge, the search gateway is firmly dominated by Google, the mobile OS is controlled by Android and iOS, and the enterprise knowledge-work gateway is concentrated in existing software platforms such as Microsoft 365, Google Workspace, Salesforce, and ServiceNow. AI-native companies lead at the interaction layer but still lag on permissions and default placement.

My judgment is: The short-to-mid-term ranking of control is roughly operating systems/browsers > search and office platforms > super-apps/social and e-commerce platforms > AI-native assistants/agent platforms. But at the "upper intent-gateway" layer, ChatGPT, Gemini, Copilot, Claude, and Perplexity can gradually rewrite user behavior from "find a webpage/find an app" to "state intent/demand a result/execute in batch." The real winners are usually not those who replace the old gateway, but those who re-tax the old gateway's profit pool.

AI gateway value chain panorama

Key product positioning matrix

Business Models, Cost Structure, and Scenario Analysis

How AI gateways make money

The core business model of AI gateways has gradually expanded from "a single subscription" to a comprehensive platform model of subscription + advertising + enterprise licensing + API + app distribution + transaction commissions + cloud consumption + hardware pull-through.

The two surest categories of money are still the AI-ification of existing profit pools:

The first is search advertising. Google does not treat AI Overviews as a replacement for search ads, but upgrades it into the new interface for search advertising, and has already extended Search/Shopping ads to AI Overviews on desktop. In other words, AI search does not necessarily destroy advertising; it may instead raise monetization efficiency for high-commercial-intent queries.

The second is enterprise software licensing and platform fees. Microsoft, Salesforce, ServiceNow, Glean, and Atlassian are all turning AI into a higher-ARPU work gateway; the only difference is the billing method: Microsoft leans toward "per user/per workspace + Azure," while Salesforce and ServiceNow are moving toward a hybrid of agents, outcome units, tokens, and ACV. Salesforce has publicly disclosed Agentforce ARR of 800 million USD; ServiceNow disclosed that the number of customers with Now Assist annual ACV over 1 million USD grew more than 130% year over year.

The third category is consumer subscriptions. ChatGPT, Gemini, Claude, Perplexity, and Alexa+ all prove that heavy users will pay continuously for efficiency, research, code, and voice experiences. OpenAI, Google, and Anthropic have all put AI features into tiered pricing; Amazon treats Alexa+ as a Prime value-add, prioritizing membership stabilization.

The fourth category is app calls and transaction commissions. This is the profit pool most worth tracking over the next two years, but the evidence is still insufficient. Perplexity has launched in-app Instant Buy with PayPal; Meta AI is clearly leaning toward shopping recommendations and content commercialization; Amazon has the strongest potential for an integrated "AI assistant–product catalog–payment fulfillment" loop.

Comparison of business models

On balance, the long-term-best investment is not a single business model, but a platform model driven by multiple engines across advertising/subscription/enterprise/cloud/transactions. For this reason, the certainty of Alphabet, Microsoft, Amazon, and Meta is generally higher than that of single-product companies.

AI gateway value and cost structure

The cost structure below is a research estimate, used to understand value distribution; it does not represent official company figures. The estimates draw mainly on the API pricing of OpenAI, Anthropic, and Perplexity, and the public statements of Amazon/Microsoft/Apple and others on infrastructure, privacy, and agent execution.

From the user-value angle, the operating costs an AI gateway most easily lowers are search cost, information-integration cost, app-switching cost, form/web operation cost, and repetitive knowledge-work cost. The revenue it most easily adds is subscriptions, ad clicks, enterprise licensing, cloud calls, and e-commerce conversion. The links most easily automated are travel planning, shopping comparison, document organization, customer-service queries, knowledge retrieval, and code and office pipelines; the hardest to automate are payment, contract signing, healthcare, law, child privacy, and processes requiring clear accountability.

Three scenarios

The scenarios below are research judgments, not company guidance. Their logic is based on current OS/browser/search share, AI-assistant MAU and the pace of enterprise adoption, the rhythm of AI-search advertising pilots, and the paid landing of enterprise agents.

My base-case judgment is: enterprise AI gateways and AI search advertising realize first; the personal AI OS and general cross-app agents realize later. Therefore, what is most worth researching now is not "who will build the strongest AI," but "who can deposit AI behavior into revenue and permissions."

Deep Dive by Track

Below, the 30 tracks the user raised are compressed into a single research matrix. Scores are research judgments out of 10, weighing revenue certainty, platform barriers, cost, and regulation together.

The core conclusion behind this matrix is: the best track is not the same as the newest track. What truly deserves heavy research weighting is: AI search, enterprise AI gateways, agent platforms, developer protocols and tool calling, browser agents, and transaction-type AI gateways. The "pure AI OS narrative," "pure personal memory narrative," and "pure AI app-store narrative," if lacking evidence of permissions, connectors, and retention, easily stay at the conceptual layer.

Company Tiering, Scoring, and Valuation

Company tiering and investment priority

Scoring model

Using the weights the user suggested, with moderate adjustments:

• Direct exposure to AI gateway revenue: 20%

• User gateway, data, and ecosystem barriers: 25%

• Developer platform and app-calling capability: 15%

• Monetization maturity and retention: 15%

• Financial quality and margins: 10%

• Market space and growth elasticity: 10%

• Valuation reasonableness: 5%

Reverse-risk scoring:

• Insufficient user adoption and retention: 20%

• Free-bundling and platform built-in risk: 20%

• Privacy and antitrust regulatory risk: 20%

• Inference cost and gross-margin pressure: 15%

• Risk of being squeezed by the OS or model platform: 15%

• Overvaluation: 10%

Total-score ranking of key companies

Valuation and market expectations

The valuation snapshot below uses market data around 2026-05-19, mainly real-time finance data and the latest public financial disclosures. Because forward PE and EV/EBITDA are not reported on a uniform basis globally, this table focuses on market cap, current PE, and the latest revenue/profit trend, with other metrics marked "needs further verification."

Companies that already fairly fully reflect AI-gateway expectations: Apple, ServiceNow, some AI-native unicorns, Perplexity (private market), and some AI-browser startups. Companies where an expectation gap may still exist: Alphabet, Amazon, Meta, Alibaba, Baidu, and Tencent. The reason is simple: these companies currently hold ready-made profit pools, yet the market tends to focus more on "model/chat-product buzz" while underestimating the revaluation power of the "gateway tax."

Key Public and Private Company Lists

Public companies most worth continued deep research

The table below focuses on the 15 public companies with the most sufficient evidence and the clearest "AI gateway benefit path." It is not a buy/sell recommendation, but a priority list for follow-up research.

Deep snapshots of key public companies

The following are "snapshot-version deep analyses" of 15 public companies, focused on answering the "AI gateway benefit path" and the "AI gateway disruption path."

Alphabet

Alphabet is the most complete AI gateway platform in this cycle, because it owns Search, Chrome, Android, Workspace, YouTube, Google Play, and Google Cloud simultaneously. The Gemini app has passed 750M MAU, total paid subscriptions reached 350M, and Gemini Enterprise sold 8 million seats within four months of launch; AI Overviews ads have expanded to desktop. In other words, Alphabet is not building a new product line, but AI-ifying five profit pools of the old gateway at once. Its biggest moat is query-intent data + browser + phone OS + ad network + Workspace enterprise data. The main risk is antitrust: the US search case once proposed a Chrome divestiture, and although the court ultimately did not require divesting Chrome/Android, it has imposed constraints on distribution and data sharing. On balance, Alphabet is the company among the "platform-type winners" that should least be understood simply as "defensive."

Microsoft

Microsoft's AI gateway strategy centers not on consumer chat but on making Copilot the enterprise work OS. The 2025 annual report disclosed that the Copilot family exceeds 100M MAU across commercial and consumer endpoints combined; Microsoft 365 Copilot puts chat, search, create, notebooks, and agents into one experience; Copilot Chat is free for eligible Microsoft 365 users, but agents require an Azure subscription, which makes it easy for Microsoft to convert front-stage AI use into back-stage cloud consumption. Its biggest moat is Windows + M365 + Entra + Security + developer tools. The risk is that consumer brand mindshare still trails ChatGPT, and whether usage frequency is enough to upgrade Copilot from an "office plugin" to a "daily gateway" still needs validation.

Amazon

Amazon's AI gateway is not on the PC or phone, but in the home, membership, and transactions. From 2026, Alexa+ is priced at 19.99 USD/month for non-Prime users in the US and free for Prime members, accessible via Alexa.com, the app, and compatible devices. Amazon's advantage is that it has not only voice and home hardware but also Prime, the product catalog, payment, fulfillment, and AWS. If users really start handing "shopping, home management, local services, and simple task execution" to an AI assistant, Amazon will be one of the companies most likely to turn "assistant traffic" directly into transaction commissions and membership retention. The risk is that Alexa has long failed to become a strong general gateway, and its phone/PC ecosystem is clearly weaker than Google/Apple/Microsoft.

Meta

Meta AI's strength is not at the permission layer but in distribution and advertising. Meta AI surpassed 700M MAU in 2025 and launched a standalone app emphasizing preference memory and context; Meta has also announced it will use users' interactions with AI for content and ad personalization. For investment, this means Meta AI does not need to prove subscription ARPU first the way ChatGPT does, and can benefit by turning AI into a "higher-conversion content/ad interface." Meta's weakness is weak system permissions and limited cross-app task execution, making it hard to become a general AI OS; but within the "social content–recommendation–advertising" loop, it may be the AI gateway with the most profit elasticity.

Apple

Apple is the AI gateway player with the highest potential moat but the slowest realization. Apple Intelligence's design fits the ideal form of an AI OS very well: on-device processing, Private Cloud Compute, understanding of personal context, and calling app actions through App Intents. But Apple's official pages still clearly state that Siri's key capabilities relying on personal context "are in development and will be delivered in a future software update." Therefore, Apple currently looks more like it is using AI to consolidate iPhone/Mac/Services than to create separately carvable new revenue in the short term. With Services FY2025 revenue of 109.158 billion USD and a 75.4% gross margin, Apple's most likely path remains hardware upgrades + Services strengthening + App Store/advertising/cloud-services defense. Its biggest risk is not falling behind on models, but user and developer patience.

Salesforce

Salesforce is one of the few companies among enterprise agent platforms that has completed the leap "from story to ARR." FY2026 Q4 disclosed Agentforce ARR of 800 million USD, up 169% year over year, with 29,000 cumulative deals signed, and it publicly discloses tokens and agentic work units as operating metrics. For investment, the key for Salesforce is not that "it has AI," but that it has CRM data, processes, permissions, channels, and outcome-measurement units. This makes it easier than most enterprise software companies to upgrade AI from a "feature enhancement" to a "work execution layer." What truly bears watching is whether Agentforce can expand from customer service/sales to broader workflows, and whether outcome billing can support higher ARPU.

ServiceNow

ServiceNow's moat comes from enterprise-process standardization + permission and audit capability. In Q1 2026 the company disclosed that the number of customers with Now Assist annual ACV over 1 million USD grew more than 130% year over year; meanwhile, large-customer ACV remained strong. ServiceNow's biggest opportunity is not to build a general chat assistant, but to build enterprise "constrained-execution agents." Such scenarios are less flashy than consumer AI, but carry higher willingness to pay, retention, and security demand. Its main risk is valuation and realization speed: if AI only raises the value of existing seats rather than opening a new ACV dimension, the stock's AI premium will come under pressure.

Oracle

Oracle looks more like an AI gateway picks-and-shovels player than a gateway itself. It can benefit through OCI, databases, and the enterprise back-end data layer, but it is hard for it to control the user's main gateway. Oracle's strategic value lies in this: if in the future many agents no longer "open SaaS front ends" directly but call back-end databases and APIs, the importance of the back-end platform will rise. The catch is that this is closer to an infrastructure revaluation than a user-gateway revaluation.

Adobe

Adobe's AI is more of a product upgrade and defense. Features like Firefly can raise creative-workflow value, shorten production time, and improve subscription stability, but Adobe does not have the first-tier gateway capability of search/browser/OS/super-app. The place most worth researching is not the general AI gateway, but whether, as agents take over more of the "asset generation–editing–distribution" flow, Adobe can upgrade from a "front-end application" to a "creative-workflow API and execution layer."

Alibaba

Alibaba is one of the few companies in China with a consumer gateway + cloud + commercial loop AI platform. Quark is officially defined by Alibaba as a strategic innovation business and was upgraded in 2025 to a flagship AI super-assistant; Alibaba Cloud continued its high growth in the latest quarter, driven by AI demand. Unlike overseas giants, Alibaba's biggest value lies not only in models or chat, but in whether it can string Quark/Qwen → Taobao/Alipay/Amap/Fliggy → Alibaba Cloud into a complete AI gateway loop. If this chain holds, Alibaba will be one of China's companies closest to a "Google + Amazon hybrid" AI gateway platform. The risk lies in whether high-frequency consumer-AI mindshare is sufficiently concentrated on Quark, and whether cloud/consumer synergy can form clear disclosures.

Baidu

Baidu is the most direct public-market sample of China's AI search transformation. Reuters reporting shows Baidu's 2026 Q1 AI-powered business revenue was 13.6 billion RMB, up 49% year over year, exceeding half of core revenue for the first time; the annual report explicitly breaks Baidu's core new AI business into AI infrastructure, AI applications, and AI-native marketing services, and emphasizes that many AI applications use a high-stickiness subscription model. For investment, the key for Baidu is not "whether it can do search," but whether, as traditional advertising declines, AI subscriptions, cloud, and AI marketing can rebuild the profit model. This is the closest real case among Chinese internet companies of "transitioning from traditional search to AI search and answer engines."

Tencent

Tencent's biggest opportunity is not the Yuanbao app itself, but WeChat, payment, mini-programs, and enterprise collaboration. In Tencent's latest disclosure, Tencent Cloud productivity AI agent solutions grew rapidly, and WorkBuddy was called China's largest-DAU productivity agent; meanwhile, Weixin/WeChat combined MAU has reached 1.418 billion. In other words, Tencent's core chips are still super-distribution and transaction scenarios, not a standalone large-model brand. If the WeChat agent, mini-program agents, and payment and merchant connectors open at scale, Tencent's gateway value may be repriced by the market. The weakness is that consumer AI product mindshare started relatively late, and in-house models and a standalone assistant brand still need to catch up.

Samsung Electronics

Samsung is one of the most active phone makers in building an "on-device AI OS surface layer." The Galaxy AI page emphasizes security and privacy, and Samsung Newsroom further introduces the Personal Data Engine, emphasizing on-device, cross-app personal data protection; the Galaxy S25's Cross App Action/Seamless Action begins stitching complex tasks across apps. Samsung's problem is not the product, but platform power: it still sits within the Android ecosystem, with weaker system-layer control than Google and a weaker software ecosystem than Apple. Its benefit path therefore tilts more toward premium-model markups and retention than toward directly eating the big advertising/search/subscription pools the way Google does.

SAP

SAP is likely an important beneficiary of Europe's enterprise AI gateway, because it naturally owns the ERP/process/enterprise-data backbone. But this round of research did not systematically verify the revenue breakdown, seats, and customer penetration of its Joule and AI agents, so this report only places SAP in a high-priority secondary research pool and offers no stronger quantitative conclusion. Needs further verification.

NAVER

NAVER simultaneously holds search, local-life, and content-platform foundations, giving it a unique gateway advantage in South Korea. If the fusion of AI search and local services accelerates, NAVER will be a very important regional sample. But this round did not systematically verify its AI product disclosures and financial breakdown. Needs further verification.

Important private companies and primary-market watch pool

Risks, Catalysts, and Final Conclusions

Risk analysis

The biggest risk in this theme is not that "AI is unimportant," but that the importance of the AI gateway is correctly seen by the market while its commercial realization is overestimated.

First is user willingness-to-pay risk. Consumer AI assistants are growing DAU and weekly actives fast, but those who can pay continuously are usually heavy users. Once OS/browser vendors bundle AI for free, the paid space for independent AI apps gets compressed. Apple, Google, and Microsoft have all put AI directly into the system or core suites.

Second is inference cost and gross-margin pressure. The API documentation of OpenAI, Anthropic, and Perplexity shows that the unit cost of multi-step search, deep research, and computer use is clearly higher than pure chat. If the eventual business model cannot pass costs on to advertisers, enterprises, or premium subscribers, many AI-native gateways will first hit a gross-margin bind.

Third is privacy, security, and accountability. Once an agent gains access to email, calendar, files, web, and payment, prompt injection, malicious websites, data leakage, mis-execution, child safety, and financial/medical liability all become primary problems, not secondary ones. Anthropic's computer use, OpenAI's computer use, Windows Recall's privacy controls, and Samsung's Personal Data Engine all show vendors are already putting "security governance" up front.

Fourth is antitrust and platform rules. The US DOJ search case once proposed a Chrome divestiture; the US DOJ has sued Apple for smartphone monopoly; the EU's DMA continues to push app distribution, interoperability, and data portability; Japan's MSCA directly covers smartphone software such as app stores, browsers, and search engines, and forces Apple to open alternative app marketplaces and alternative payments in Japan. Future AI-gateway profit pools are likely to be redistributed under regulation.

Fifth is content and publisher backlash. The more successful AI search and answer engines are, the more they erode traditional web clicks and the long-tail content business model. Perplexity has launched a publisher-compensation model, and Google is also moving ads into AI Overviews, but whether this is enough to stabilize content supply remains unsettled.

Metrics most worth tracking over the next twelve months

What is most worth tracking is not the "large-model leaderboard," but the hard metrics below that prove the gateway and profit pools are being restructured:

• The DAU/WAU/MAU and paid users of ChatGPT, Gemini, Meta AI, Copilot, and Claude.

• The advertising, affiliate, and transaction conversion of Google AI Overviews, ChatGPT Search, and Perplexity.

• The ARR/ACV, large-customer count, and connector usage of Microsoft 365 Copilot, Agentforce, Now Assist, Glean, and Rovo.

• Whether the local memory, cross-app actions, and privacy controls of Apple/Samsung/Windows truly enter the main flow.

• The MCP/tool-calling/computer-use call volume of OpenAI, Anthropic, and open-source agent frameworks.

• Changes in default-slot, distribution, payment, and browser rules from the EU DMA, the US DOJ, the UK CMA, and Japan's MSCA.

Final conclusions

The AI operating system and the personal AI assistant have shifted from "better chatbots" to "a higher-level software orchestration layer." This layer will not immediately replace iOS, Android, Windows, or the browser kernel, but it will gradually become the new user-intent gateway, app-execution layer, personal-context layer, and agent-platform layer.

This report's five most important conclusions:

First, the five segments most worth watching are: AI search/answer engines, enterprise AI gateways, agent platforms and tool calling, browser agents, and transaction-type AI gateways. These five are either closest to existing profit pools or most likely to control the future permission and app-calling layer.

Second, the 10 public companies most worth deep research are: Alphabet, Microsoft, Amazon, Meta, Apple, Salesforce, ServiceNow, Alibaba, Baidu, Tencent. Among them, Alphabet and Microsoft are the most complete platform-type names; Amazon and Meta are the profit-pool re-raters most likely to be undervalued; Apple has a deep moat but slow realization; Salesforce/ServiceNow are the pure-software samples realizing enterprise agents fastest; Alibaba/Baidu/Tencent are the AI-gateway migration paths most worth watching in China.

Third, the 10 private companies most worth tracking are: OpenAI, Anthropic, Perplexity, Glean, The Browser Company, Brave, DuckDuckGo, Notion, Rewind/Limitless, and execution-layer infrastructure companies like Browserbase/Zapier/n8n. Among these, OpenAI/Anthropic/Perplexity determine the upper-layer AI-gateway migration; Glean/Notion determine the knowledge-work gateway; The Browser Company/Brave/DuckDuckGo determine browser restructuring; Rewind/Limitless determine whether personal memory can become an independent moat.

Fourth, the five points most easily misunderstood by the market are:

• An AI gateway is not the same as a model gateway;

• An AI OS is not a new kernel, but a new orchestration layer;

• Large DAU does not equal high profit;

• The value of enterprise agents is often higher than consumer-side showmanship;

• The real moat is not the "chat experience," but permissions, context, connectors, audit, and developer protocols.

Fifth, the narrower research directions most worth prioritizing over the next six to twelve months: I suggest focusing on one or two of the following five to keep drilling down: AI search, AI browsers, personal AI assistants, MCP tool calling, and enterprise AI gateways. If I could only pick one, I would prioritize AI search and browser agents; if I could only pick the one most certain to deliver on results, I would prioritize enterprise AI gateways and agent platforms.

Open questions and limitations

This report uses public materials available through 2026-05-19 as far as possible, but three classes of limitations remain: First, many private companies and some Asian public companies do not disclose AI-gateway revenue, MAU, developer counts, and call volumes on a uniform basis; Second, valuation metrics such as forward PE, EV/EBITDA, and EV/AI revenue are not reported on a uniform basis globally, so this report only provides valuation snapshots for companies with sufficient evidence; Third, some segments are still at the "product launch/user trial" stage rather than the "scaled adoption" stage, so judgments need continuous validation via retention, payment, and call volume. For these companies, this report uniformly marks "not disclosed" or "needs further verification" and adds no fabricated supplements.

This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.