TL;DR: In May 2021 we named consumption pricing risk as structural: the model transfers cost risk to the customer, and unexpected usage spikes produce backlash, not loyalty. Roughly 4.8 to 5.1 years later, primary sources corroborated the mechanism, from a FinOps industry survey to Uber’s own COO. What matters now is the diagnostic behind the call.
In May 2021, while the market was declaring subscription pricing dead, we published a piece arguing the opposite. The core point named the consumption pricing risk everyone was ignoring: the model transfers cost risk to the customer, and that transfer produces friction, not loyalty.
The exact language from that 2021 article:
“Consumption pricing essentially transfers all cost risk to the customer, which can negatively affect the software company. As the customer assumes more risk, they take longer to evaluate their purchase and figure out how to budget for it, slowing the sales cycle.”
And:
“Unexpected spikes in utilization can cause significant customer backlash and churn. As higher use equates to higher costs, some customers will balk when utilization spikes, particularly when those spikes are out of their control. Think of the reaction during the pandemic if Zoom prices were based on the number of video calls.”
Consumption pricing risk is the structural exposure a vendor creates when its pricing shifts cost uncertainty onto the buyer. The customer cannot estimate the bill in advance, usage spikes produce backlash instead of loyalty, and the resulting churn arrives late and is hard to trace back to the pricing itself.
Being right this early wasn’t luck. It came from a single question we ask of any pricing decision, and that same question points somewhere specific for AI pricing now.
What we said in 2021, and why it cut against the grain
The 2021 market narrative was directionally clear and almost uniformly positive about usage-based pricing. A January 2021 TechCrunch piece declared subscriptions dead. Insight Partners published a piece celebrating usage-based-pricing IPOs as evidence the model was winning. “Usage-based” had acquired the feel of an industry consensus rather than a strategic choice.
SPP’s counter-position wasn’t that consumption pricing fails. It was narrower and more specific: most companies rushing to consumption models hadn’t done the underlying metric work, and the billing mechanism was being mistaken for the strategy.
Consumption versus subscription is a false binary. Both are payment structures wrapped around a more fundamental decision: the value metric, the unit a price attaches to. The value metric is the upstream-most decision in the pricing architecture. Every packaging and pricing-model choice downstream inherits whatever the metric got right or wrong.
A company can build a consumption model on a well-chosen metric and do fine. A company can build a subscription model on a badly chosen metric and bleed. The 2021 piece argued that the rush to consumption pricing was a rush to change the billing mechanism without interrogating the metric underneath it.
It was a direct structural objection to what the market was celebrating.
Five years of confirmation, in the market’s own words
The evidence that followed didn’t arrive as a single event. It accumulated at two distinct layers: the practitioner organizations managing spend, and the C-suite executives accounting for what that spend produced.
The practitioner layer: FinOps goes from afterthought to mandate
In February 2026, the FinOps Foundation’s annual industry survey of 1,192 respondents reported that 98% of those companies now actively manage AI spend, up from 31% two years earlier. That figure is self-reported industry data, not a peer-reviewed finding.
That is not incremental growth. A governance category that covered less than a third of the market two years ago now covers nearly all of it. The FinOps Foundation also updated its own mission in February 2026. It broadened its charter from “Advancing the People who manage the value of Cloud” to “Advancing the People who manage the Value of Technology.” That charter change is telling. A governance discipline redefined its own scope to keep pace with the problem.
The lead time from the 2021 warning to this datapoint is approximately 4.8 years.
The C-suite layer: a brand-name buyer says the spend isn’t legible
In June 2026, Uber COO Andrew Macdonald made two on-the-record statements about AI spend that bear out the 2021 warning about cost risk and buyer resistance.
On a TechCrunch-covered podcast on June 2, 2026, Macdonald said it is “very hard to draw a line” between Uber’s AI spending and new consumer features. On June 17, 2026, wheresyoured.at reported him saying the link between AI costs and useful consumer features “is not there yet.”
Both statements came as Uber capped employee AI spending after exceeding budget within four months. The COO of one of the world’s most data-sophisticated companies was describing his own organization’s AI spend as hard to trace to delivered value.
The lead time from the 2021 warning to this datapoint is approximately 5.1 years.
What this evidence proves, and what it does not
These two data points confirm the mechanism the 2021 piece named. When cost risk sits with the buyer and the value isn’t legible, buyers organize to manage and resist that risk. The 31%-to-98% FinOps jump is that organizing, at scale. Macdonald’s public statements are the C-suite version of the same dynamic.
What this does not prove is that SPP forecast any specific 2026 event. The Uber spend cap, the FinOps mission update, any named vendor’s repricing decision: none of those were called in advance. The mechanism was identified. The specific expressions of it were not.
One note on sourcing. The closest literal match to the 2021 language, the widely discussed AI developer-tooling bill-shock stories, isn’t cited here: the reporting tracing it to primary sources doesn’t meet this piece’s citability bar. That the strongest-sounding example is also the least sourced says something about how early the accountability cycle still is.
Why the call held up: a diagnosis, not a guess
The 2021 analysis didn’t require predicting AI. It required one diagnostic habit: when a pricing decision moves cost risk from vendor to buyer, ask who bears the consequence when usage is unpredictable, and how long before that shows up as friction.
That question doesn’t expire. The answer changes as markets change, which is why a 2021 answer is still legible in 2026.
The architectural framing underneath it: the licensing model is the pairing of a value metric with entitlement rules. Consumption pricing is a licensing-model decision, not a billing preference. When that decision is made without deliberate metric selection, the billing mechanism carries a weight it was never designed for.
Buyers can’t estimate their spend. Vendors can’t predict their revenue. The mismatch produces exactly the governance overhead and C-suite doubt now visible in the FinOps and Uber data.
The direction is not ours alone. Peer-reviewed modeling of software pricing under competition reaches a compatible conclusion. In competitive markets, fixed pricing can be more profitable than usage-based pricing once the cost of monitoring usage outweighs the benefit of acquiring low-usage customers. The 2021 objection matched where economic theory already pointed.
This is the discipline we built into Continuous Monetization and deliver through LevelSetter: pricing architecture revisited on the cadence the product ships, using real transaction behavior instead of a one-time bet. The 2021 diagnosis held in 2026 because that underlying question, is the metric right and who bears the risk, was asked when the market was skipping it.
What the same method says about AI pricing’s next move
The same diagnostic applies forward. At each stage of AI pricing’s evolution, ask who bears the cost risk when usage is unpredictable, and whether the buyer can estimate the bill in advance. That question has already driven one visible transition, and it is driving another.
The value metric has already shifted once, and the shift isn’t finished
AI pricing started with raw consumption: tokens, API calls, inference requests. That structure transferred cost risk directly to the buyer and made spend estimation nearly impossible. Token counts are opaque to most buyers, variable by query complexity, and subject to model updates the buyer doesn’t control. We have documented one downstream effect separately: when the bill is unpredictable, buyers ration usage and suppress the exploration a new product depends on.
The industry’s first response was credit-based pricing: assign credits to actions, sell credit packs, abstract the token count away. Credits tried to relocate the risk into a more legible unit. In practice, they mostly relocated it into opacity, and the buyer still can’t estimate the bill in advance or trace credit consumption back to anything they understand.
The six structural problems with credit-based pricing are a separate analysis. The point here: a credit is a surrogate unit the vendor defines and controls, not a value metric the buyer can estimate against before committing.
The current movement toward outcome-based and resolution-based metrics is the market’s second attempt to answer the same question. Intercom’s per-resolution pricing for Fin, HubSpot’s Breeze structure, and Sierra’s agent-outcome framing are all attempts to put the value metric where it belongs, attached to something the buyer can observe and verify. The risk still exists, but it is at least attached to a unit with a legible value claim.
The diagnostic for anyone evaluating these: can the buyer estimate spend before committing, and does the metric track value the buyer controls? If the answer to either is no, the 2021 mechanism applies again.
Cost-control architecture is becoming a procurement criterion
In 2021, buyers “balked” informally at usage spikes. By 2026, the balking has hardened into something more structured.
Our cost-control architecture analysis documents what this looks like at the procurement layer. Buyers now ask vendors about enforcement delay: how long between a spend-cap breach and actual enforcement. They ask about liability during that delay, and about audit-trail depth. A spend cap that takes 24 hours to enforce is not a spend cap in any meaningful sense if a model inference loop can generate a month’s worth of charges in an afternoon.
This is the direct, current-day descendant of the 2021 warning. Buyer-side resistance has moved from informal frustration to contractual specification. By mid-2026, the AI-cost-drift risk is being marketed to boards by generalist analyst firms as a governance gap.
SPP holds a different frame than those firms. Their answer is workflow monitoring and audit trails. The durable answer stays at the architecture layer: a deliberately chosen value metric plus contract guardrails. Monitoring a badly chosen metric more closely does not fix the metric.
Billing consolidation as a symptom of the same fight
The consolidation of usage-billing infrastructure documented in the billing consolidation analysis reads through the same lens. As metering and billing platforms consolidate, the strategic question was never which platform runs the meter. The question is who controls what the meter measures.
A vendor that owns the meter definition owns the risk allocation. A buyer who accepts the vendor’s meter definition without scrutiny has accepted a cost-risk transfer without negotiating it.
The forward-looking read: a dated, checkable claim
As of mid-2026, SPP’s read is this: cost-control architecture will become a standard line item on enterprise AI vendor scorecards before the end of 2027. By that we mean buyers will routinely require documented enforcement delay, liability caps, and audit-trail access as procurement conditions, not as negotiated exceptions.
The second claim: outcome-based and resolution-based value metrics will move from early-adopter positioning to table-stakes expectation among customer-facing AI-agent vendors within the same window. Vendors still pricing on raw consumption or credits in those categories by late 2027 will face the same buyer-side resistance and governance overhead visible in the FinOps data today, at larger scale.
These are SPP’s read as of 2026, not established market consensus. The same discipline that generated the 2021 observation will revisit these in 2027 and say plainly whether they held.
The bridge to Continuous Monetization
A single correct diagnosis across five years is a useful credibility signal. An operating discipline that keeps re-asking the same diagnostic question, on the cadence the product changes, is what protects a company from the next version of the same mistake.
The value metric question doesn’t resolve once. It resurfaces when the product changes, when buyer usage patterns shift, and when the cost structure underneath the model moves. For most AI-adjacent products in 2026, all three of those conditions are moving at once and on an accelerating schedule.
Continuous Monetization is the discipline we built for exactly that condition, delivered through LevelSetter. Pricing architecture is treated like product development, with every decision a hypothesis vetted against real transaction behavior rather than a one-time bet. The cadence isn’t quarterly or annual. It matches the cadence at which the product itself ships.
The 2021 piece asked a diagnostic question at a moment when the market wasn’t asking it. The discipline is to never stop asking it.
If your value metric hasn’t been stress-tested since your last major model update, start here. When you want a second read on the metric before your next release, book a working session.
