The AI bottleneck is moving beyond the chip

Chips still matter. But access to compute alone cannot tell you whether an AI investment is ready to scale. My view: founders need to bring three questions into the planning conversation much earlier.

• Can the infrastructure support the ambition?

Power availability and grid connections can constrain data centre expansion. IEA For businesses using hosted AI, the practical question is dependency. What happens to your critical workflows if your provider cannot deliver the capacity or availability you need?

• What return will we actually measure?

Choose a business outcome before choosing the next model.
Then account for the full cost of delivering it, including integration, human review and rework. A faster first draft means little if correcting it consumes the time you saved.

Define the baseline. Decide what improvement would justify the investment.

• Can the agent complete the work reliably?

A polished demo gives you a reason to test.
Production requires evidence across realistic tasks, including incomplete inputs and failed tool calls. Agent evaluations should examine the system as a whole, with measures such as error rates and cost per task. anthropic.com
Set approval boundaries and define when a person takes over.
Before you plan the next model upgrade, put these questions in your planning document. Give each one an owner, an evidence requirement and a decision threshold.

That gives buyers and investors something concrete to assess: the operating conditions behind your promise.
Which is hardest to demonstrate in your AI roadmap today: infrastructure readiness, measurable return or agent reliability?