BLOG // 2026.04.13 // 02:01 SGT

Open-Source AI Was Just Subsidized CAC

As Meta and Anthropic pull up the ladder on their models, founders are learning a hard truth—the open-source AI era wasn't a paradigm shift, just a heavily subsidized customer acquisition cost.

4 MIN READSYS.ADMIN // BRYAN.AI

The Open Source Charade Collapses

I spent years at Amazon and building ShopBack. If there is one thing I learned about platform economics, it is that corporate benevolence is just customer acquisition cost by another name. We all watched the frantic model release frenzy over the last two years. We cheered for the commoditization of intelligence.

We were naive.

The bill is finally coming due. Meta has reversed its open-source AI promise, officially restricting access to Llama. They hooked the developer ecosystem, established the standard, and are now pulling up the ladder. Meanwhile, Anthropic is aggressively capping OpenClaw—behaving more like the Nintendo of AI with its walled-garden approach—while simultaneously fighting battles like Anthropic v. U.S. Government in the lobbying trenches.

Do you really think big tech was going to subsidize your compute and IP forever?

A stark, brutalist server rack room with a heavy steel door shutting, symbolizin

Founders built entire roadmaps assuming foundation models would remain free, open, and endlessly accessible. That assumption is dead. The moat was never the model—it was the infrastructure, the distribution, and the inevitable API lock-in. If your startup’s gross margin relies on Meta or Anthropic acting like a charity, you do not have a business. You have a compounding technical debt bomb.

Prompts are Demos. Workflows are Deployments.

Stop building wrappers. The era of the single-shot prompt is over.

When I look at enterprise deployments today, I apply a simple filter: is this a parlor trick, or does it do actual work? As Tanuj Garg rightly points out, the future is workflow orchestration, not simple prompts. A prompt is stateless. It requires a human in the loop to initiate, evaluate, and act. A workflow is stateful—it compounds leverage by chaining actions across fragmented systems.

But this shift to autonomous agents introduces an entirely new class of operational friction. We are no longer building web interfaces for humans. We have to reconsider how AI agents see your website and how to build for them. If your DOM is a mess of dynamic Javascript and hidden elements, agents will fail.

Abstract visualization of complex data pipelines intersecting with glowing nodes

Worse, when you give agents read/write access to your databases and APIs, the blast radius of a failure increases by orders of magnitude. We are already seeing the fallout. AI agents gone rogue are becoming the next cybersecurity nightmare. An intern making a mistake deletes a table; a rogue agent iterating in a tight loop can exfiltrate your entire customer database and burn through your AWS credits in four minutes. We are optimizing the web for non-human consumers, and most legacy architectures are completely unprepared for the resulting traffic patterns and security vectors.

The Adoption Reality Distortion Field

If you spend all day on tech Twitter, you would think every enterprise is fully automated. Step outside the bubble. Down here in the APAC operator trenches, the reality is much colder—and much slower.

Look at the macro indicators. Despite the hype, Australia is not ready for AI adoption. There is a massive gap between the capability of the models and the capacity of legacy enterprises to ingest them. Integration takes time. Security audits take time. Change management takes time.

A dense Asian metropolis skyline at twilight, emphasizing the concrete economic

And time is the ultimate constraint. You are always balancing across three domains: career, family, and finance. Right now, the financial constraint is biting hard. The Residential Property Price Index jumped 6.1 per cent in Q4 2025.

Why does property inflation matter to an AI builder? Because it dictates the risk appetite of your buyers. Your AI tool doesn't exist in a vacuum—it competes for finite budget against rising operational costs, inflationary pressure, and exhausted engineering teams. When rent goes up 6 percent in a single quarter, CFOs do not sign off on experimental AI pilots. They buy tools that demonstrably cut headcount or accelerate revenue today.

Stop selling the theoretical future. Build orchestration that solves a bleeding neck problem right now, secure it against rogue execution, and charge for the outcome. The free ride is over. Act accordingly.