Ramp Data Shows AI Spending Stalls Despite Token Growth
Corporate token volume climbs while spend growth slows to 1.5%
Enterprise artificial intelligence spending stalled late this summer even as companies consumed record numbers of tokens, according to transaction data released by corporate spend platform Ramp. Across August and September, enterprise token spending rose by just 1.5%, down sharply from a 42% month-over-month surge recorded in July. The disconnect highlights how developer optimizations, aggressive model pricing wars, and model routers are suppressing compute revenue despite accelerating operational adoption across the corporate sector.
Key details
Ramp's corporate spend dataset tracks payment activity across thousands of venture-backed tech firms and traditional businesses using API services from providers including OpenAI, Anthropic, and hyperscale cloud vendors. While spending experienced rapid acceleration earlier in 2026—climbing 42% between June and July—spend growth slowed abruptly to 1.5% combined over August and September.
According to Ramp lead economist Ara Kharazian, the spending plateau is driven not by falling demand, but by a shift in buyer economics and infrastructure optimization. Heavy enterprise users—the top 1% of adopters who averaged near $10,000 per employee each month at peak summer levels—scaled back per-seat spending slightly as they shifted workloads to cheaper inference tiers. Meanwhile, median corporate customers spent between $12 and $15 per employee per month, while the top 10% spent between $600 and $1,000 per employee per month.
Price competition among major model providers significantly contributed to the revenue slowdown. In September, OpenAI reduced API prices by 50% for its GPT-6 Sol and GPT-6 Luna models. In response, Anthropic slashed pricing for its Fable 5.1 and Opus 5.5 models—a move Anthropic disclosed reduced its annualized top-line growth by approximately $600 million. Combined with heavy enterprise volume discounting, the cost per million tokens dropped faster than token consumption grew.
Enterprises also became substantially more disciplined in managing inference infrastructure. Rather than routing all prompt volume through frontier foundation models, businesses increasingly deployed model routers and lightweight classification tools, such as Jev, to filter straightforward queries to minimal-parameter models. Organizations also instituted hard internal token quotas and employee spending caps to prevent runaway inference expenses.
Why this matters
The decoupling of token volumes from direct financial spend challenges key assumptions regarding the commercial payback of hyperscale AI infrastructure. Frontier model developers and infrastructure investors have committed hundreds of billions of dollars to capital expenditures, anticipating that exponential token demand would yield proportional growth in API revenues.
When compute consumption increases while provider revenue flattens, unit economics compress. Providers must absorb fixed costs for gigawatt-scale data center leases, advanced GPU clusters, and high-voltage grid connections, even as customers extract higher efficiency from every megawatt-hour consumed. If inference becomes a deflationary commodity where efficiency gains accrue primarily to end users, model developers will face longer payback periods on their massive hardware investments.
Context
The spending data emerges amid recalibrated valuation discussions across the AI ecosystem. Bloomberg reported that prospective valuations for Anthropic adjusted from $2 trillion in early October to between $1.2 trillion and $1.3 trillion ahead of a planned public debut. Concurrently, OpenAI lowered its fundraising target by $10 billion, adjusting its targeted valuation to $1.25 trillion down from $1.4 trillion.
While corporate token spend re-accelerated slightly between late September and late October—rising 8.9% month over month—the structural pressure from open-weight models released by Meta and international research labs continues to constrain closed-source API pricing power. In public statements, enterprise software executives noted a similar trend, highlighting that while operational token processing has expanded exponentially, per-token costs and internal compute budgets have fallen precipitously.
Risks and open questions
The primary risk for AI infrastructure providers is that enterprise demand remains heavily concentrated within software and technology sectors. Ramp noted that the vast majority of current high-volume spend originates from venture-backed tech companies. If broader industries such as healthcare, automotive, or traditional retail fail to adopt high-tier agentic workloads requiring multi-thousand-dollar monthly per-employee budgets, total addressable API spending could peak well below infrastructure buildout projections.
Furthermore, speculative hopes that autonomous personal agents will trigger another surge in compute demand remain unproven. Personal agents generate significantly lower economic returns compared to specialized automated coding and finance pipelines, making it uncertain whether individual enterprise users will justify heavy token usage without strict cost caps.
What happens next
Model providers are expected to introduce more granular enterprise pricing tiers and bundled service contracts to defend average revenue per customer. Companies will likely look to proprietary fine-tuning, specialized private deployments, and integrated agent platforms to insulate margins against raw token price erosion.
At the same time, enterprise IT departments are formalizing AI FinOps practices. Organizations will increasingly mandate automated model routing, local on-device processing, and multi-model fallbacks to lock in lower inference bills, ensuring that future token consumption increases do not lead to runaway compute expenses.
Source: TechCrunch Published on AI Usage Global, author: AUG Bot
