Insight

Nope. Thank you for coming to my Ted Talk. Okay, but seriously, this one will have a lot of source references.

The current AI business model is not sustainable. No major AI provider has publicly demonstrated a profitable business model. Training costs remain enormous, and the cost of inference has never been confirmed as profitable by any provider. If it were, they would say so. Consumer subscriptions are heavily subsidised; the actual compute cost to Anthropic of serving a heavy Claude user significantly exceeds the $20 monthly price. Uber burned through its entire 2026 AI tools budget in four months, with the COO publicly stating the link between AI usage and measurable business outcomes was "not there yet." That is a company with 5,000 engineers, 95% of them using AI tools monthly.

The infrastructure required to sustain the industry is also under serious strain. Nearly half of planned US data center builds for 2026 have been canceled or delayed, held back by power infrastructure shortages and supply chain failures. More than 75 projects worth $130 billion were blocked in Q1 2026 alone, with opposition groups active in 49 US states, driven by electricity costs, water consumption, noise, and scale. Data centers take 18 to 24 months to build, and chips are purchased at the start of construction, meaning the hardware can already be a generation behind before the facility goes live. Oracle is building data centers for OpenAI under a $300 billion agreement contingent on OpenAI reaching revenue targets that HSBC describes as unrealistic, with OpenAI projected to burn $27 billion in 2026 and $63 billion in 2027.

Public sentiment is shifting. Hundreds of Stanford graduates walked out of their 2026 commencement ceremony when Google's CEO took the stage. Although not all was due to AI. Community opposition to data center construction has moved from fringe to mainstream, driven by the real costs these facilities impose on the areas that host them.

For AI to become commercially sustainable, two things likely need to happen: a meaningful hardware breakthrough that makes training and inference significantly less energy intensive, and on the software side, a leap in inference efficiency that produces the same outputs from substantially less compute. Neither is guaranteed or imminent. Apple has a structural advantage most providers do not: the ability to run models locally on billions of devices at near-zero marginal cost. Enterprises are likely to look at deploying models on their own hardware, where costs are controllable and predictable. What becomes of the dominant providers in their current form we don't know. The ones that survive will be those that can close the gap between what it costs to serve a request and what a customer is willing to pay. Right now, that gap is billions of dollars wide.

Still have questions? Give us a call or send us a message to understand more without any pushy sales tactics.