Enterprise AI is entering an accountability phase, and this week made it impossible to ignore.
On August 5, 2026, HubSpot reported Q2 results that beat analyst expectations on every headline metric: revenue of $911.7 million, up 20% year over year; GAAP operating income of $43.3 million; adjusted earnings of $3.26 per share against a consensus of $3.02. The stock fell 20% anyway. The reason was not the quarter. It was the transition to outcome-based AI pricing, and what that transition signals about the broader enterprise AI market.
The same week, the Linux Foundation launched the Tokenomics Foundation, a 30-member industry consortium tasked with creating vendor-neutral standards for measuring what AI actually costs and what it actually returns. Founding members include IBM, SAP, ServiceNow, JPMorganChase, Oracle, Accenture, and Hitachi. Notably absent: OpenAI, Anthropic, Microsoft, and Google. The buyers, not the sellers, are building the accounting framework.
And in a Constellation Research analysis published August 9, Palantir CEO Alex Karp named the problem directly: “tokenmaxxing.” He said enterprises “understand broadly that tokenmaxxing is at their own cost, and certainly understand that tokenmaxxing is leading to them transferring their data, their prompts, the way they run their business, their expertise to a third party.”
Karp is talking his book. He is also right.
What Tokenmaxxing Means for Enterprise AI
Tokenmaxxing describes a feedback loop that most enterprises have entered without realizing it: as AI tools become easier to deploy, teams send more tokens to frontier models, reasoning budgets expand without discipline, and total AI spend grows faster than any measurable return. The term was originally used by model providers to encourage high-volume consumption. It has become, for enterprise buyers, a liability.
Uber’s CFO Balaji Krishnamurthy provided a concrete example of what breaking the loop looks like in practice. In the same August 9 analysis, he was quoted: “We are optimizing token spend by setting better defaults for different use cases, moving certain tasks to lower-cost or open-weight models, and letting employees more clearly understand and manage their spend. As a result, cost per token has declined over the past several months, even as adoption has continued to increase, allowing us to keep overall AI spend broadly stable.”
Uber did not reduce AI use. It enforced routing discipline, model selection by task, and employee spending visibility. Token cost fell. Coverage grew. This is the blueprint enterprise AI teams should be following.
Anthropic has now acknowledged the problem from the vendor side. The company released a guide for IT administrators to control Claude Enterprise costs, framing the shift explicitly: “It’s helpful to measure AI’s cost-per-outcome instead of token consumption as the primary metric of value.”
The HubSpot Signal: What Happens When Vendors Make the Shift
HubSpot did not stumble on pricing. It made a deliberate set of choices in Q2 2026 to accelerate AI transformation by changing the commercial model: introducing outcome-based pricing for its Prospecting Agent (credits charged when a meeting is booked) and Customer Agent (credits charged when a ticket is resolved), cutting entry prices for AI agents, and shifting from seat licenses to credit consumption.
The results were mixed in ways the equity market found deeply uncomfortable. Revenue grew 20%. But net new customer additions came in at 7,000 against a company target of 9,000 to 10,000. Full-year revenue guidance was trimmed by roughly $22 million, or 0.6% at the midpoint. Q3 guidance of $924 to $925 million came in below the Street’s $942.3 million estimate.
CEO Yamini Rangan explained the mechanism on the earnings call: “Predictability has become a defining theme in AI adoption. Businesses have been hit with unpredictable token costs. And they want pricing that is transparent and tied to value.” She described AI deal cycles lengthening, with larger buying committees and more C-suite and board approvals required, especially in the mid-market.
The market punished HubSpot twice in 2026 for the same move. The stock fell roughly 19% in May when the outcome-based pricing announcement was made. It fell another 20% in August after the first full quarter of data. Twice, public investors looked at seats converting to credits and marked down the stock.
The signal for enterprise builders is different from the signal for investors. What the equity market is pricing in is near-term revenue unpredictability. What enterprise buyers are receiving is something they actually want: AI spend that turns off when outcomes do not materialize.
What the Tokenomics Foundation Is Building
The Linux Foundation’s new group is not trying to compete with model providers. It is building the measurement infrastructure that makes vendor-neutral AI accountability possible.
The Tokenomics Foundation’s roadmap, as described in the official press release, includes:
- A shared definition of token value and token density, covering input, output, reasoning, and cached tokens separately.
- A reference model for total AI cost, not just token spend: compute, storage, databases, and engineering labor included.
- A “cost to serve” standard expressed as cost per call rather than cost per token, so the unit maps to work actually performed.
- A value measurement framework that starts with the share of work completed without human involvement, compared against the process cost baseline.
- Token Cost Telemetry improvements to the FOCUS v1.5 billing specification, which already serves as the substrate for cloud cost normalization across providers.
- A “Big-T Framework” for classifying workloads by token complexity before routing to the lowest-cost model that can handle them.
The founding board convened July 30. Technical working groups are forming now. This is standards-body velocity, not product velocity, so enterprises should expect initial specifications in Q4 2026 and adoption in 2027. The value is in the vocabulary it creates: a shared language that procurement departments can use to demand accountability from vendors without needing to understand inference architecture.
The Current State of AI Pricing Models
The enterprise AI pricing landscape is in active flux. Here is where the major models stand:
| Pricing Model | Unit of Charge | Predictability | Enterprise Risk |
|---|---|---|---|
| Per-token (standard) | Input + output tokens | Low | Tokenmaxxing spiral, cost surprise |
| Per-token (reasoning) | Base + reasoning tokens | Very low | 3 to 10x cost for complex tasks |
| Credit-based | Pre-purchased credits consumed | Medium | Unused credit waste, volume renegotiation |
| Outcome-based | Per resolved outcome | High | Vendor absorbs volatility; may increase unit price |
| Seat + token hybrid | Seat fee + consumption overage | Medium | Seat cost plus surprise overages |
| On-premise / open-weight | Infrastructure cost only | High | Capital expenditure, team capability required |
Most enterprise AI contracts today are token-based with credit pre-purchase. The industry is moving toward outcome-based models, but the transition is not clean. HubSpot’s experience shows that even enterprises that want outcome-based pricing need time to build the evaluation infrastructure to trust it.
What Enterprise AI Teams Should Do Now
Measure cost per outcome, not cost per token. Identify the three to five AI use cases consuming the most tokens in your organization. For each, calculate what the task cost before AI and what it costs now per completed unit of work. If cost per completed unit is not falling, the problem is not the model: it is the integration design.
Implement model routing before expanding AI coverage. Tasks that require simple classification, data extraction, or formatting do not need frontier reasoning models. Databricks Unity AI Gateway and Snowflake Cortex AI Gateway both provide routing layers that enforce model selection by task type, cap spend by team, and give finance teams the observability they need to audit AI costs the same way they audit cloud spend.
Set token budgets before deploying agents, not after. Agentic workflows are especially exposed to tokenmaxxing because agents can call tools repeatedly and expand their reasoning traces across multi-step tasks. Budget per task, not per month. Uber’s lesson: defaults matter. The token spend that most enterprises assume is being optimized by the model is actually being driven by default prompt configurations and tool-call chains that no one has reviewed.
Use the Tokenomics Foundation roadmap as a procurement checklist. When evaluating AI vendors in 2026 and 2027, ask: Can they report cost per call, not just cost per token? Can they provide a total cost of AI including compute, storage, and labor? Can they tie spend to business outcomes in your reporting system? Vendors who cannot answer these questions are not ready for enterprise production at scale.
The pricing revolt is not a crisis for enterprise AI adoption. It is the moment the industry moves from experimentation budgets to operational accountability. The companies that build measurement discipline now will be the ones that scale AI spend without the cost surprises that are rewriting Q3 guidance across the software industry.
If your organization is navigating the shift from AI pilots to production-grade deployment with accountable economics, talk to Enera.