Here’s a Morphilus perspective on where the money flows in AI, and what it means for small operators
There’s a war being fought at the top of the AI industry right now. It’s a war over compute, capital, and control. Most microbusiness owners aren’t watching it. They probably should be, because the outcome will directly shape what tools they can afford, what services disappear overnight, and whether the current window of “cheap and capable AI” stays open.
Here’s how the pieces fit together.
The Open Source Floor and Why AI Hasn’t Priced You Out……Yet.
The single biggest reason a five-person trade business in Geelong can access world-class AI tools for less than a phone bill is open source. Not charity.
When Meta released Llama, they didn’t do it out of generosity. They did it because their business model doesn’t depend on selling AI, it depends on selling attention. Commoditising the model layer suits them. The GitHub and Hugging Face communities did the rest, fine-tuning, packaging and thus turning research releases into tools anyone can run on modest hardware.
Then came the Chinese models. DeepSeek R1 was the moment the Western AI establishment had a quiet panic. A model competitive with GPT-4 class performance, trained at a fraction of the reported cost, released open source. Qwen, Yi, Baidu’s ERNIE — the Chinese open source ecosystem is now a genuine forcing function on Western pricing. When a capable model is free to download and self-host, the premium models have to justify their price tag.
For small operators, this creates a floor. There is now a baseline of “good enough” AI capability that cannot be priced away — because it’s free. That’s not nothing. That’s actually everything.
But there’s a catch: Free models still need compute to run. And compute is about to get complicated. Not to mention that free madels require a higher level of self reliance.
The Capital Trap When Your AI Provider Needs Billions to Stay Relevant
OpenAI. Anthropic. Google DeepMind. Microsoft and XAI, aren’t software companies in the traditional sense anymore. They’re infrastructure companies with software interfaces.
Training a frontier model, the kind that moves the needle on capability, now costs hundreds of millions of dollars. The next generation may cost potentially billions. That money has to come from somewhere.
For Anthropic, it’s coming from Enterprise customer revenue (currently running at at a rate of US$30b/yr but growing rapidly towards US$80b/yr ) It is also getting investment from Amazon (a reported $4 billion commitment) and Google. For OpenAI, it’s a complex web of Microsoft investment, consumer subscriptions, and an IPO that’s been talked about for years. For Google, it’s mainly in-house, being cross-subsidised by a search advertising machine that generates $200 billion a year and despite the initial code red jeopardy of its advertising model when ChatGPT launnced, is also showing positive growth.
The main course over the next 12 months will be the planned IPOs of SpacexAI, Anthropic and OpenAI. There is no “All you can eat” buffet. The big VIP diners will get their fill. Potentially leaving only scraps of liquidity for other parts of the US economy let alone for players in other countries.
What this means structurally: the frontier model companies need enterprise contracts to justify their capital raises. Not your $20 a month subscription. Not mine. They need the Westpacs, the Telstras, the Fortune 500 procurement departments signing multi-year agreements with committed spend.
That’s the race happening right now, Anthropic and OpenAI pulling enterprise clients away from traditional SaaS providers like Salesforce, ServiceNow, and SAP. Those legacy vendors built their moats on workflow lock-in. AI is dissolving that lock-in faster than anyone expected. Why buy a $150,000 enterprise CRM when an AI agent network under your direct control can replicate the output for a fraction of the cost?
The enterprise cash is flowing toward AI. The question is what happens to everyone else as a result, and just when might the cost incentive to go Open source with the Chinese models offset concerns about sovereign risk.
The Compute Crunch and The Physics Problem No One Can Spend Their Way Out of Fast Enough
Here’s where it gets genuinely uncomfortable.
AI demand is not growing linearly. Every new capability unlock creates new use cases, which creates new demand, which requires more compute. Inference, ie actually running models to answer questions, is now consuming more GPU capacity than training. Multiply that by the agentic AI wave coming, where models aren’t answering one question but running multi-step workflows autonomously, and the compute demand curve starts to look near-vertical.
NVIDIA’s H100 and H200 clusters have lead times that stretch into the future measured in years. Microsoft, Google, Amazon, and Meta are collectively committing over $300 billion in data centre investment in 2024-2025 alone. The plan is close to US$800b investment in 2026/27.
While in China, data centres are tapping remote and very cheap excess clean energy sources, to run mega data centres utilising relatively large numbers of prior generation GPUs.
Conclusion:
The AI Economy is growing fast on a monthly basis, and is accelerating because agentic capabilities are spreading to more users and all users are learning how to spawn multiple agents.
Although chip speed and manufacturing capability improvements will be accelerated by AI, the reality is that the complexity of chip making is such that next generation equipment and manufacturing technologies take years to implement.
Having liquidity to finance growth is crucial for big tech firms. Large enterprises will be the key driver for AI economics, leaving small enterprises to ponder accessibility, justifiable use cases and their own security issues.
Where willl you be dining? Will you be booking a VIP table or sharing dim sum with the crowd? Will you even be able to book a table? We’d love to get your comments.