Capital moves towards computing power
Amazon, Microsoft, Alphabet and Meta were expected to spend roughly USD 600 billion on AI infrastructure in 2026, according to a summary cited by The Jerusalem Post. Data centres, servers, chips, networks and energy are the physical foundation behind seconds of work in a browser.
The biggest platforms build capacity while smaller companies rent and pay by consumption. An Israeli start-up may have a world-class algorithm, yet its economics depend heavily on model, cloud and transfer prices.
Oracle illustrates pressure on the balance sheet
The report said Oracle had cut about 21,000 jobs, roughly 13 per cent of its workforce, amid financing for AI expansion and large capacity commitments. This does not tie each job to a server but shows infrastructure's budget priority.
Debt, power contracts and future compute revenue create different risks from licence software. Slower AI demand or falling compute prices could change the return on expensive centres, making the same boom that opens an application market a source of fixed obligations.
An Israeli product must watch unit economics
Software firms need to know the cost of every model response and whether customers pay for it. User growth can enlarge losses without sound pricing, so teams must track context length, model choice, caching, query frequency and limits.
The capital wave also creates opportunities in cyber security, optimisation, data infrastructure and model controls,Israeli strengths. Success will mean measurable value at a price covering an increasingly visible compute bill.
Every request has a variable cost
Traditional software has low marginal user cost, while generative models compute every request according to input, output, model and repeated steps. An agent reading documents and calling services may cost many times more than text completion, so economics must be measured by feature.
Averages hide heavy users. Routing, limits and quality measurement should choose a smaller model when sufficient and expensive computation only where it adds real value.
Architecture and provider contracts shape margin
A company can buy an API, host an open model or combine providers, trading capital, speed, data control and price exposure. Long contracts may lower unit cost and create lock-in, so application logic should permit ongoing price and quality comparisons.
Regions, quotas, availability, data handling and responsibility for model changes matter alongside token price. Evaluation suites and prompt versioning reduce regression and complaint costs and are financial controls as well as technical ones.
Outcome pricing needs a measurable benefit
SaaS vendors may charge per resolved request, document or action, but must define success. A wrong answer should not be billed as a correct one; quality measures, disputes and transparent audits are necessary.
Some outcomes depend on human decisions and external events. A pilot should measure time, errors and eliminated steps before pricing so the customer keeps part of the gain and the vendor covers compute risk.
Costs should be visible during feature design
Developers need per-scenario budgets, customer dashboards and anomaly alerts before a monthly bill. Context limits, caching, batching and model selection can cut cost without hurting quality; late optimisation can turn a popular feature into a business problem.
Czech buyers should require price caps, usage export and an off switch for expensive automation. Rising infrastructure cost is not an argument against AI but a reason to manage it like cloud, logistics or energy.
Financial discipline must reach the product
The best optimisation minimises cost per correctly completed task, combining technical telemetry with errors, complaints and user time. A cheap model that creates repeats and review can cost more overall; vendors should expose limits and quality choices.
Concentration among a few clouds and models gives providers power. Data portability, evaluation suites and the ability to switch models are commercial insurance. Czech firms should contract for price changes, availability and exit before agents enter critical workflows.
