Ramp data shows AI spending stalling just as Anthropic nears its IPO

Corporate spending on artificial intelligence stalled over late summer despite surging query volumes, according to corporate spend management platform Ramp. While token usage among businesses expanded rapidly, aggregate spend rose by just 1.5% across August and September compared to a 42% month-over-month jump recorded in July, as detailed in TechCrunch's report. The findings indicate that aggressive provider price cuts, enterprise discounting, and internal engineering optimizations have combined to curtail top-line growth at major model providers including OpenAI and Anthropic.

This report examines the underlying transaction data tracked across venture-backed enterprises, detailing the divergence between processing volumes and enterprise bills. It provides an operational overview for financial analysts, software engineering leaders, and procurement directors tracking generative model economics, billing volatility, and pre-IPO valuations. The metric changes highlight shifting pricing power as enterprise customers deploy lighter inference infrastructure and cost-control systems to execute production workloads at a fraction of prior expense.

The deal behind Ramp data shows AI spending stalling just

The transactional shift emerged directly from enterprise card and invoice activity across venture-backed high-growth technology companies, which constitute almost the entire addressable corporate spend tracked in Ramp's corporate ledger. Ramp lead economist Ara Kharazian noted that the most aggressive tier of corporate users, representing the top 1% of adopters, experienced a stutter in their monthly expenditure across August and September. These organizations had peaked during early summer at nearly $10,000 per employee each month, but their spending dropped just below that boundary as summer closed, sufficient to depress overall spending growth across the wider provider market.

Median corporate customers tracked by Ramp spend between $12 and $15 per employee each month, while firms in the top 10% bracket allocate between $600 and $1,000 per employee per month. Although these middle-tier and upper-decile companies recorded modest budget increases, their additional dollar allocations were not large enough to counterbalance the plateauing numbers among the top 1% cohort. While token spending did record a brief bounce of 8.9% between September 23 and October 23, Kharazian noted that structural headwinds on sustained spending growth remain dominant as consumption efficiency overtakes unconstrained budget growth.

Why Ramp data shows AI spending stalling just raised

The slowdown in spending does not stem from declining enterprise interest or collapsing token workloads, because raw processing volumes climbed steadily throughout the entire measurement period. Instead, the market has undergone structural model price cuts and aggressive discounting, transforming enterprise purchasing into a decisive buyer's market. In September, OpenAI cut API access costs in half for both GPT-6 Sol and GPT-6 Luna, putting direct pricing pressure across commercial software pipelines. Anthropic responded by reducing rates on its Fable 5.1 and Opus 5.5 models, disclosing in a regulatory filing that its pricing reductions slashed top-line revenue growth by approximately $600 million.

Enterprise contract structures have accelerated this revenue suppression, as both leading providers negotiated multi-year enterprise volume commitments designed to secure customer workloads that inadvertently compressed actual billings. Unlike ride-hailing services such as Uber and Lyft, which subsidized early rides before consolidating market power and raising consumer rates, foundational model developers face persistent downward price pressure due to alternative infrastructure providers. Kharazian observed that these vendor subsidies cannot easily convert into unilateral price hikes because enterprise customers retain access to competitive commodity alternatives across multiple ecosystems.

What the Ramp data shows AI spending stalling just money

Engineering teams have introduced aggressive internal frugality measures that sharply decrease the financial return on rising operational queries. Companies increasingly configure automated model routers that prevent expensive foundational models from processing long-running, routine business tasks. Instead, systems redirect baseline queries to ultra-lightweight architectures or specialized routing utilities such as Jev, an emerging model-selection tool designed to evaluate simple binary outcomes. Shifting simple operational flows to targeted tools allows organizations to cut inference expenses on tasks that previously ran across multi-tier frontier networks.

Internal corporate policies have also placed hard budget caps on individual employee seat licenses, preventing experimental workloads from inflating monthly financial statements. Notion co-founder Akshay Kothari corroborated the dynamic with internal operational data, highlighting that rapid increases in system tokens were accompanied by substantial drops in total operational costs. As engineering teams master prompting efficiency, token consumption climbs while the net dollar yield returned to model providers contracts, creating a widening structural disconnect between underlying compute delivery and realized software revenue.

Competitive context for Ramp data shows AI spending stalling just

Competitive alternatives from Meta and Nvidia have further reinforced budget conservatism by supplying open-source and low-cost enterprise model distributions to commercial developers. Model developers based in the United States and China have introduced a range of lightweight open alternatives that allow internal teams to bypass commercial API paywalls altogether. By combining open releases with specialized routing layers, companies satisfy production requirements without dedicating significant budget allocations to proprietary inference providers, preventing commercial vendors from capturing outsized margins on basic task automation.

This billing contraction has introduced turbulence across private secondary markets and impending public offerings for model developers. Bloomberg reported that Anthropic's expected valuation dropped from an estimated $2 trillion in early October down to between $1.2 trillion and $1.3 trillion, while credit rating agencies prepared to issue speculative junk grades on its debt. Simultaneously, OpenAI trimmed its proposed valuation target from $1.4 trillion down to $1.25 trillion and scaled back its fundraising ambitions by roughly $10 billion, reflecting investor reassessments of rapid expansion theses.

Open questions on Ramp data shows AI spending stalling just

A critical unanswered question for model providers is whether new application paradigms can expand non-software enterprise spending up to the lucrative levels maintained by high-growth engineering firms. While autonomous coding software established commercial viability across financial and software sectors, specialized professional environments such as medical clinics and automotive dealerships face fundamental operational limits that preclude $10,000 monthly seat expenditures. Without broad vertical expansion into non-technical corporate environments, model creators remain dependent on a saturated base of venture-backed software developers who are actively automating their own cost-reduction strategies.

Investor expectations now hinge on whether autonomous personal agents can catalyze a second sustained spending boom comparable to the commercial emergence of coding assistants. Providers such as OpenAI, SpaceXAI, and Meta have rolled out personal agent frameworks alongside specialized startups like Instinct, seeking to create consumer and administrative enterprise demand. However, economic analysts express skepticism regarding whether personal agents can match the high revenue productivity delivered by coding tools, leaving significant uncertainty over whether novel product categories can restore hypergrowth to frontier artificial intelligence developers.