Canva's forecast cut turns model routing into a product line item

Inference is the rare cost line whose quantity is set by people who never see a contract. A user presses generate, a default inside the feature picks the model, and the invoice tracks adoption rather than procurement. Canva's own account of what went wrong, as relayed by SiliconANGLE, was that it had relied too heavily on expensive third-party frontier models ^[3]. The remedy says more than the diagnosis: not a renegotiated rate card, but in-house models, Leonardo.AI and task-level routing ^[4].

The crude arithmetic is worth doing, because nobody else has. A company above $900 million a quarter ^[2] is running above $3.6 billion annualised, so ten points of growth is on the order of $360 million of 2026 revenue withdrawn from the forecast ^[14]. In relative terms, a third of the promised growth rate went to an input cost ^[13].

The savings figures deserve the same treatment. Canva's video and image models were reportedly 17 and 30 times cheaper than frontier alternatives ^[5], which is a 94% and a 97% reduction respectively ^[15]. The blended result was a cut of roughly 90% in the cost of an AI task ^[4]. That gap is not a rounding artefact: it implies a meaningful share of traffic still lands on the costlier option ^[16]. Which is what a working router looks like. The cheap model takes what it can handle, and the residue is escalated at frontier prices. The distance between 90% and 97% is not an infrastructure problem, it is a decision about which tasks are permitted to escalate, and that decision lives in the feature spec.

The other cases in SiliconANGLE's account share an uncomfortable property: the control arrived after the bill. Uber reportedly burned a year of AI budget in a quarter, then reset defaults and routed work to cheaper models ^[6]. Lindy discovered Anthropic had become a larger expense than payroll before it moved traffic elsewhere ^[7]. Alex Karp's complaint about tokenmaxxing, as SiliconANGLE renders it, is that firms optimise the vendor's bill instead of the value they keep ^[10]. Reading the same evidence, the sharper failure is that none of these organisations appear to have known their cost per completed task while the feature was still being designed.

Microsoft sits on both sides. It sells Microsoft, partner and customer-controlled deployments through Foundry ^[9] while stating that it wants to reduce and eventually eliminate what it pays Anthropic ^[8]. And Switzerland's sovereign programme still required a large cheque to Microsoft ^[11]. Sovereignty language, in SiliconANGLE's own examples, does not by itself lower the invoice; routing and telemetry do ^[12].