The AI infrastructure bubble question — and why your software should stay portable

There's a loud debate running through AI right now: is the money pouring into infrastructure meeting real demand, or inflating a bubble? You don't have to pick a side to protect yourself from the answer.

The numbers driving the debate

The buildout is staggering. On July 1, Together AI raised $800M at an $8.3B valuation — more than double its value 16 months earlier — led by Aramco Ventures, with annual bookings crossing $1.15B. Days earlier, Blackstone pledged $30B toward AI data centers in Japan over three to five years, eyeing sites above a gigawatt. Blackstone's COO put the bull case bluntly: the risk of too little compute outweighs bubble fears.

The bear case is quieter but sharper. Models keep getting cheaper to run as efficiency improves — and if inference costs fall faster than demand rises, a chunk of today's spending math gets rewritten. So does the pricing you're quoted.

Why it matters for your business

Here's the translation for a company under 500 people: you feel this through your vendor's invoice. If inference costs collapse, prices should fall — but only if you're free to move. If your operation is welded to one provider's proprietary API, you inherit their pricing, their outages, and their roadmap, whether or not it still serves you.

The defense isn't guessing the market. It's architecture. When we build automations and internal tools, we keep the expensive, swappable parts — the AI model, the compute — behind a clean boundary. Your business logic lives in your infrastructure, not rented inside someone else's platform. If a cheaper or better model shows up next quarter, you switch a config, not your whole stack. That's the difference between owning the shape of your software and paying rent on it forever.

Moves that hold up regardless of which side of the debate wins:

• Negotiate 12-month terms with renewal caps — don't lock in multi-year at today's prices.
• Keep workloads portable so you can switch providers if inference gets cheaper.
• Treat current vendor pricing as temporary, in both directions.

Sources: TechCrunch, Nikkei Asia.