In 1996, a corporate chief executive could sign a purchase order for $40-a-seat Netscape Navigator licenses, install the software across thousands of office desktops, and proudly inform the board of directors that the company was now an internet-native enterprise. It was a golden era for shrink-wrapped software. Enterprise customers routinely wrote six-figure cheques so their staff could browse the World Wide Web, while retail shoppers walked into high-street electronics stores to buy cardboard boxes containing a single floppy disk of Navigator for $39. The employees opened their browser windows, marveled briefly at static text pages indexed on AltaVista, and went straight back to running the business on physical memos, spreadsheets, and fax machines.
Today, enterprise leadership is executing the exact same ritual with Claude and ChatGPT. Chief information officers buy $20-a-month subscriptions, watch staff use isolated browser tabs to rewrite polite decline emails, and declare that the firm has adopted artificial intelligence. It is the identical category error. Corporate leaders are confusing access to a new communications medium with the difficult, unglamorous work of rebuilding how the organisation actually operates.
The Velocity Illusion #
To understand the fragility of this moment, it helps to remember just how invincible Netscape looked at the dawn of the web.
Netscape was the fastest-scaling software company in corporate history. Founded in 1994, it grew from $1.4 million in its first year to an inflation-adjusted $1 billion annual revenue run rate in under forty months. Its August 1995 initial public offering kicked off the dot-com boom, valuing a sixteen-month-old startup at nearly $3 billion on its first day of trading. By late 1995, its market capitalisation peaked past $6.5 billion. It seemed to have captured the future.
Anthropic makes Netscape look like a horse and cart. Anthropic exploded from $1 billion in annualized run rate in late 2024 to $65 billion by July 2026, driven by enterprise API consumption and developer tools like Claude Code.
Staggering top-line velocity creates a dangerous corporate illusion. Generating massive revenue at the dawn of a platform shift does not mean you own the platform. Often, it merely means you were the first legible window onto a protocol that nobody yet understood how to use.
Netscape discovered this lesson with brutal speed. In the final quarter of 1996, Netscape booked nearly $60 million in browser license fees alone. Five quarters later, after Microsoft bundled Internet Explorer into Windows for free, Netscape’s browser license revenue collapsed to literally $0.
“Windows Is Dead” versus “SaaS Is Dead” #
In 1995, Marc Andreessen famously declared that Netscape Navigator would reduce Windows to a mundane set of poorly debugged device drivers. The logic seemed compelling: if every application lived on the open web, the underlying operating system would become irrelevant.
Andreessen was catastrophically wrong. Windows was not just software; it was an entrenched commercial ecosystem of hardware drivers, enterprise network permissions, IT deployment contracts, and deeply ingrained human habits.
Today, venture capitalists and foundation model founders are reciting Andreessen’s exact speech almost word for word. Enterprise software is dead, they announce. The model is the new computer, and systems of record like Salesforce, Workday, and ServiceNow will be reduced to commoditised SQL databases with buttons.
This thesis repeats the exact same blunder. Enterprise applications are the complex digital constitutions of modern corporations. They hold the audit trails, the compliance rules, the messy integrations, and the institutional compromises negotiated over thirty years. The operating system did not die thirty years ago; it simply absorbed the browser. Today, enterprise software is quietly absorbing the model.
The Two-Front Pincer #
Netscape died because it occupied an intermediate layer that could not defend itself against attacks from above and below.
Bill Gates recognised the threat in his famous 1995 memo, The Internet Tidal Wave, and deployed Microsoft’s ultimate weapon: giving away an identical product for free. Microsoft bundled Internet Explorer directly into the operating system, stripping Netscape of its retail pricing power overnight. Simultaneously, open-source developers released the Apache web server, destroying Netscape’s high-margin enterprise server business from below.
Anthropic sits in the exact same squeeze.
From above, platform incumbents are commoditising distribution. Microsoft, Google, and Apple are embedding baseline reasoning directly into Windows, Office 365, Google Workspace, and mobile operating systems for zero incremental dollars. When a corporate employee already has capable intelligence baked into Word, Excel, Chrome, and their phone, convincing a chief financial officer to renew a standalone $20-a-month subscription becomes an impossible enterprise sales pitch.
From below, the value of proprietary model weights is evaporating. Open-weight distillation replicates 95% of closed frontier capabilities within months for pennies on the dollar. When engineering teams can run open weights on private cloud infrastructure with zero licensing markups, the commercial pricing power of proprietary API endpoints collapses toward the floor.
The Captive Application Handicap #
Anthropic’s attempt to escape this trap by moving up the stack into developer tools reveals a structural contradiction.
If a company genuinely wants to own the enterprise application layer, its software must be ruthlessly optimized for the user. It should route every single subtask to whatever engine delivers the highest intelligence, the fastest throughput, and the best outcome. Anthropic cannot do this because its applications exist to consume its own compute.
Developers buy tools like Claude Code on monthly subscriptions, expecting the ultimate coding agent. Yet Claude Code is hardwired exclusively to Anthropic’s own models, like Claude Sonnet 5, locking developers into a single vendor’s limitations.
In practice, alternative models are often vastly superior for specific programming tasks. On the Artificial Analysis Intelligence Index, GLM-5.3 Flash scores 42 against Sonnet’s 38, delivers dramatically higher token throughput, and consumes a fraction of the compute, costing $0.25 per task compared to $5.09 for Sonnet.
In interactive software development, latency and throughput are everything. A software engineer waiting fifteen seconds for an agent to generate a multi-file diff will quickly abandon a tool if an alternative model can complete the same work in two seconds. Forcing a developer to wait for a slower model because it carries the corporate logo actively degrades the product.
An independent developer tool can integrate faster, higher-scoring models tomorrow. Anthropic’s applications remain held hostage by corporate vanity, unable to route to superior rival models without admitting that their own weights are not the universal answer to computing.
The Window versus the Machine #
The final hurdle is architectural. A standalone chat interface is inherently incapable of running a business.
It possesses no shared operational memory across teams, leaving each employee’s interaction trapped in an isolated browser silo. It has no persistent relational data model, no multi-user permission hierarchy, and no capacity to execute cross-system workflows without a human remembering the magic prompt.
Netscape proved the web was real, made billions for its early investors, and watched its core business get wiped out in eighteen months before being acquired by AOL as an afterthought.
Anthropic has proved that machine reasoning is miraculous. Yet by selling floating-point weights through an API window, it risks repeating the oldest tragedy in the technology industry: delivering a platform shift to the world, while the platform incumbents quietly take the profits.