Graphic & Text | Sister Tang

For someone with minimal programming knowledge, who only uses computers for basic office tasks and gaming, products like Claude Code offer a truly astonishing experience.

Initially, my goal was simple: connect Claude to the yfinance database to query historical stock prices, performance metrics, and real-time price data for Hong Kong and U.S. stocks. After a few straightforward dialogues and some copy-pasting, I achieved this. Later, when I noticed significant anomalies in certain stock indicators, I considered downloading the raw financial report data for stocks in my portfolio and those I closely monitor, and having Claude transcribe them into the database.

I accomplished all this without even starting to use Code. Claude’s conversational suggestions prompted me that I could directly use Code to modify my local server.py file, calculate various indicators based on the original financial reports, and supplement and revise yfinance. This is why the software has garnered such high praise in various AI circles and why its code leak caused such a stir among developers worldwide—it's just so effortless and user-friendly.

Claude provided me with all the Python code, which I neither understand nor need to understand. I downloaded the financial report PDFs for companies like Meituan and Alibaba myself, placed them in a folder, and instructed Code to input them. It read the PDFs, extracted tables, built a database, and occasionally prompted me for permission, which I granted with a click of 'Allow,' and it continued working. At this stage, all my actions involved downloading files and granting permissions.

Next, I didn't even bother opening the browser. I simply said, 'Go to the company's official website and scrape the raw financial data yourself.' Code autonomously opened the browser, navigated to the IR page, downloaded the PDFs, read the tables, created folders on my hard drive to archive them by company and year, and then input the data into the database. That's how it completed the full dataset for the 'Magnificent Seven' U.S. stocks. When websites had anti-scraping measures or required manual permission handling, I took care of it, but for the most part, I just sat in my chair watching the text scroll in the terminal, occasionally answering a confirmation question or clicking 'Allow' with my mouse.

After completing all these operations, I suddenly realized something: I hadn't done anything myself. I didn't need to open the browser, use Excel, drag and drop files, or even create new folders. All my actions were just conversing with it. If I wanted to, I could even forgo the keyboard and just use voice commands.

So, I began to wonder: after AI becomes capable of fully taking over computer operations in the future, is there still a need for graphical operating systems like Windows? The desktop, taskbar, Start menu, window switching—all of these are designed for 'humans operating computers.' Now, it's AI operating the computer, and it doesn't look at the screen, click the mouse, or drag windows. Windows is still running, but the one sitting in front of the screen is no longer human.

01 Who's Looking at the Screen?

Of course, declaring Windows unnecessary is an overly dogmatic judgment. But what Windows represents—all software designed for 'humans operating in front of a screen,' from operating systems to Office to Salesforce to every SaaS product—needs to answer a question after AI becomes widespread:

Is the cost users pay daily for this layer of design still worth the price? To see this clearly, we need to break down software value into three layers.

The bottom layer is infrastructure: computing power, storage, networking, and deployment environments. This layer won't disappear. The middle layer is business logic: how data is processed, how workflows operate, how rules are executed. This is where the core value lies. The top layer is the interface: translating the capabilities of the bottom two layers into graphical presentations that humans can understand and operate. Buttons, menus, tables, charts, drag-and-drop, dropdown boxes—all these are products of the interface layer.

Over the past forty years, the evolution of the software industry has been about thickening this third layer. From DOS command lines to Windows graphical interfaces, to desktop software and mobile touchscreens, every paradigm shift has lowered operational barriers and improved interface experiences. The business logic behind this direction is clear: the better the interface, the more users, the more revenue. According to the subscription models of the entire SaaS industry, every additional person who needs to 'look at the screen and click buttons' means an additional fee, essentially pricing the interface layer.

Anyone who has played games understands this. Games are just code, a bunch of if-else statements and data tables. But after the computer runs this code, what appears on the screen is a gorgeous 3D world, dazzling special effects, and exquisite UI. Most of the GPU's computing power is spent on 'rendering for you to see.' In this sense, Windows is no different from a game—except instead of slaying monsters, you're being productive.

The computing power and energy consumption spent on the desktop's translucent frosted glass effect, the real-time preview when dragging windows, and the bouncing icons on the taskbar—this is the interface tax.

How heavy is this tax? When Windows 11 is idle, it typically consumes 3 to 6 GB of memory, while a Linux server without a graphical interface only needs a few hundred MB when idle. Even after accounting for drivers and system service overhead, the additional memory consumption purely due to the graphical interface is still over 1 to 2 GB—enough for a local AI model to remember thousands more tokens of context.

The situation is even tighter on the GPU side: Windows' Desktop Window Manager (DWM) continuously consumes GPU resources for interface composition and animation rendering. Running an 8-billion-parameter local model on a consumer-grade GPU with 8 GB of VRAM is already pushing it to the limit. A few hundred MB more for the interface can mean the difference between running and not running.

When the primary operator of the computer changes from human to AI, the direction of software evolution reverses for the first time. Instead of continuing to thicken the interface to make it more usable for humans, the goal becomes bypassing the interface to let AI directly access underlying capabilities.

This isn't speculation. Looking at Microsoft's FY2025 numbers, it's clear they're already adapting to and even driving this trend. With total revenue of $281.7 billion, the intelligent cloud (led by Azure) accounted for 38%, productivity and business processes (including M365) accounted for 43%, and personal computing (including Windows) accounted for 19%.

Microsoft moved Windows Commercial from the personal computing segment to the productivity segment (where M365 resides) in FY2025, reducing Windows' presence as a standalone product in its organizational structure. Microsoft isn't a victim threatened by AI; it's a driver of the interface layer's value collapse, actively shifting value from the interface layer to the data and infrastructure layers.

Microsoft has benefited greatly from the AI wave, but the winner is Azure, not Windows. Jensen Huang famously said, 'AI won't recreate tools; it will maximize efficiency with existing ones.' Many interpret this as arguing for the irreplaceability of the current software ecosystem, but the way to maximize efficiency is precisely to bypass the interface and directly access underlying capabilities. AI wants the data itself and the logic to process it, not the Excel spreadsheet you can see.

In other words, computing power and business logic won't disappear, but the premium pricing of the interface layer may not hold. If the world's most successful software company is itself moving away from the interface layer, does this trend really need further explanation?

02 What Remains After Stripping Away the Interface?

The devaluation of the interface tax is already clear from Microsoft's financials. But the more crucial question is: Why?

It's not because AI does a prettier job. Claude Code is hideous—just a black terminal with white text. It replaces Excel not because of a better interface but because all the calculation functions Excel can perform can be encoded into methodologies and taught to AI, allowing it to operate directly at the database level—ten times faster than my manual clicking, and with a much lower error rate for calculations relying on local data.

The reason software interfaces exist is that 'humans need them to operate.' When the methodologies behind operations can be encoded into instructions for AI, this reason disappears. The interface shifts from 'necessary' to 'optional,' and a core part of the product's value becomes a cost that can be bypassed.

This applies to my personal investment research workflow just as much as it does to a fifty-person sales team. Salesforce's core methodologies—lead scoring, pipeline management, quarterly revenue forecasting, etc.—are all codifiable rules and processes. Once encoded, AI can directly operate the underlying CRM database, rendering the fifty individual login interfaces redundant.

Salesforce has clearly realized this. Its Agentforce now charges $2 per conversation, Intercom charges $0.99 per issue resolved, and Kustomer has fully shifted to a hybrid model of AI usage-based pricing on top of subscriptions. In highly standardized business scenarios, the underlying assumption of per-seat pricing is being replaced by the idea that each methodology only needs one agent to execute.

So which software companies should be nervous? The criterion is simple: What remains after stripping away the interface?

If what remains can be encoded into methodologies and taught to AI, then it's just an interface shell. The most typical example is Microsoft's own Office 365—Word's formatting, Excel's data processing, and PowerPoint's presentation creation are all teachable methodologies. That's why Microsoft introduced M365 Copilot with an additional $30 monthly fee, attempting to shift the pricing anchor from the interface to AI.

But if I can use Claude Code to directly generate formatted docx files and process data to output results, Office shifts from a productivity tool to a file format compatibility layer. It can still exist—docx and xlsx are de facto standards, and you'll probably still need these software to present work results to clients and colleagues—but from a productivity standpoint, it's hard to regain its former dominance.

If what remains is data that no one else has, that's a different story. Methodologies can be taught to AI, but data cannot. AI can learn analytical methods but without data, it's like a skilled chef with no ingredients. Nobody cares that Bloomberg Terminal's interface has been ugly for thirty years because you're paying for its data. Palantir's interface isn't its selling point either—governments and militaries pay for its ability to integrate sensitive data sources.

These companies possess unique data barriers, which may actually benefit them in the AI era: the cost of the interface layer is reduced by AI, while the data barriers remain intact, boosting profit margins.

An interesting side note: Chinese enterprise SaaS has long suffered from low valuations, with the mainstream explanation being Chinese companies' low willingness to pay. But from another perspective: China has never developed a habit of paying for desktop software interfaces. Pirated Windows has been used for twenty years, and the penetration rate of genuine Office remains low even now. This isn't stinginess; it's a natural insensitivity to the interface layer.

If the global software industry is experiencing a collapse in interface layer value, Chinese companies have no sunk costs to digest—they've never paid much of a premium for this layer in the first place. Conversely, China has never had its own desktop operating system ecosystem, which used to be a weakness. But if the next generation of systems is designed for AI rather than humans, this represents a massive innovation opportunity.

03 Conclusion

AI doesn't just help you bypass a software interface; it helps you bypass the entire operational chain from downloading financial reports to inputting data to calculating ratios to making peer comparisons. But the time saved should be spent on things AI can't do: go to the front lines, talk to people, even visit malls and supermarkets you haven't been to in a while to observe things that don't exist in any terminal or database.

Because the best thing AI gives you is precisely the least valuable thing in investing: consensus. The interface tax AI saves you should be spent searching for non-consensus.

The same logic applies to investing in software companies. When you buy water, you should focus on what's in it—what ingredients, minerals, and trace elements—not whether the cup is glass or crystal. The interface is the cup; the data is the solution. Don't pay a premium for the craftsmanship of the crystal cup—the composition of the solution is what truly matters.

As for judgment, that's your own taste buds. AI can't give you that, and neither can the interface.