In 2011, Marc Andreessen told us software was eating the world. For fifteen years, the advice to every company was the same: hire engineers, build your own software, become a tech company. That advice is about to reverse. Over the next two years, the companies with the best odds won’t be relying on software as their competitive advantage. Instead, they will rely on physical, real-world moats.
I don’t have inside information on any AI lab’s roadmap. However, as an executive coach, I have a privileged view into top-of-mind challenges for leaders in product and technology at the frontier AI labs, AI-native startups in Silicon Valley, and tech-first companies across the US, UK, and Europe. I work with engineering leaders watching their team’s work get automated, product executives getting questions about why OpenAI can’t build their software faster and cheaper, or COOs weighing an AI vendor contract versus building in-house. And here’s what I see when I put their experiences together.
Software is increasingly built by the few, likely working at an AI lab
We are at a point where AI can build software autonomously, earlier than many executives had anticipated. At Anthropic and OpenAI, people inside describe almost 90% of new code as written by the latest models. A PM agent generates the tickets, coding agents make the changes, and specialized review agents check each other’s work before it ships. Humans review by hand only when there’s serious security or privacy risk, or when a change touches hardware and data center infrastructure.
In fact, the AI is building the models themselves. Anthropic’s IPO filing says Claude is now involved in many aspects of researching, training, and testing later model generations, from running experiments and generating training data to automating evaluations and writing and reviewing code (S-1).
As a result, a company no longer needs a 200-person engineering org to ship a competitive product. It needs a small team that directs the models well. The typical team becomes a small group of senior people who frame problems, judge output, and own what ships. Entry-level engineering roles, as we’ve known them, will largely disappear.
If agents can write, review, and ship most code, the main cost of building software stops being engineering salaries and becomes the cost of model access, which the AI labs control and make their competitive advantage over any pure software company.
AI labs are competing for eyeballs and enterprise access
With their lower cost and faster ability to build software, the labs are building products well beyond the model itself. Anthropic recently entered enterprise productivity, with docs and spreadsheets, competing directly with Microsoft Office and the Google Workspace Suite. They are also well-positioned to build and sell security products. Within a month of launching Project Glasswing, Anthropic and its partners had found more than ten thousand high- or critical-severity vulnerabilities in essential software (Techmeme roundup).
At the same time, consumers increasingly start with Claude or ChatGPT instead of Google, whether they’re planning a trip or organizing an event. The younger users, many of whom don’t have pre-existing wiring to open Google search, default to asking AI rather than searching. And as Google knows well, whoever owns that first query owns distribution. They increasingly own the consumer front door, and they’re building applications in the same categories as their enterprise customers.
The scarce make-or-break resource is compute
Now, the cost of building software for the labs comes mostly down to compute costs. The price of each task is falling, but demand is growing so fast that the total bill keeps climbing. The labs are now racing for chips, power, and land. The labs are now racing for chips, power, and land. In October 2025, Sam Altman went to Seoul, where OpenAI signed letters of intent with Samsung and SK Hynix to supply DRAM for its Stargate project and to build data centers in South Korea (TechCrunch). Anthropic committed $50 billion with Fluidstack to build data centers in Texas and New York (TechCrunch), its first major build-out of its own. Press analysis of its S-1 puts total infrastructure obligations above $600 billion, much of it payable whether or not the capacity gets used (VKTR).
As a result of the increasing scarcity of compute, the AI labs are rationing it, and keeping the best for themselves. Anthropic’s S-1 says it divides its compute among model development, serving external customers, and internal use of Claude (S-1). As a customer of Anthropic, your workload competes with the lab’s own products for the same supply. Controlling access to the models that their customers rely on to build software is their competitive advantage. Usage limits for external customers tightened during peak hours in March 2026 and loosened in May after a new compute deal (VKTR). Their own products are built by and run on models their competitors can’t access. The AI labs will increasingly keep the best models for themselves.
What’s left is real-world complexity and accountability
So if the AI labs are building any software more cheaply, what becomes of the current software companies?
The ones that survive will likely have a strong component of physical goods or human interaction. The labs show little appetite for the complexity involved in leaving the software world. They don’t want to get into the complexity of running a warehouse, manning a health clinic, or delivering food. That leaves an opening for companies willing to carry the staff, the liability, and regulatory complexity the labs want to avoid. Logistics, healthcare, financial services, and insurance are likely to stick around.
But even with these businesses, similar to Google search, the AI labs will likely own the power to recommend and charge for customer access.
Winner takes all?
Software history has seen something like this before. Windows dominated operating systems, and Google dominated search. Each held most of its market and had huge pricing power over everyone building on top. The AI Labs are going this direction quickly. OpenAI, Google, and Anthropic together control 84% of the AI agent market (Tech Insider).
What’s different now is scope. Microsoft, Apple, and Google owned the access and platform with Windows, iOS, and Search / Android. Other businesses built the apps. When software is nearly free to build, an AI lab can own the access, the platform, and the apps. The cost advantage of traditional SaaS players disappears because anyone can build quickly, and their distribution weakens as discovery moves into assistants the labs control.
On the other hand, companies hate lock-in. Dependency on one AI lab is a board-level risk, treated the way manufacturers treat single-supplier risk. Most enterprises already run two or three models and route work between them (RaftLabs), and switching at the model layer can be as simple as changing an API call.
Pricing power can also be limited, because every price increase pushes more work to cheaper or open-weight models. Regulators are watching. The European Commission has said it will examine whether certain AI services should be designated as core platform services under the Digital Markets Act (Communications Today).
But none of these arguments was absent in the Windows era. Companies disliked lock-in then too, and regulatory battles were frequent. Microsoft won anyway, because defaults and distribution beat stated preferences. It’s too early to tell where this will land, but the battle isn’t pretty.
A word for tech leaders
As a product or engineering leader, what is most important now? Here are the three discussions I recommend you invest in today:
What is your competitive advantage outside of software? Real-world operations, Regulatory sophistication, or leading humans?
What are the economics of model tiering and model switching for your business? How can you pitch these business cases to your investors?
How will your customers find you when an AI assistant is standing between you and them? How much are you willing to pay for access?
That’s all folks! See you next week at 3:14 pm.
Yue