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

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.