Palantir CEO says AI models steal enterprise value as routing cuts costs 97%

Enterprises are shifting from defaulting to the most powerful AI models to routing tasks by cost, a trend that Palantir says can cut inference spending by as much as 97%.

Palantir Technologies Chief Executive Officer Alex Karp said enterprises are "livid" about frontier AI models capturing their proprietary business value, as the company's new routing system cuts inference costs by as much as 97% in some deployments.

"Enterprises are paying for tokens that create no value while handing over their intellectual property and competitive advantage to third parties," Karp said in a CNBC interview Wednesday. He described the dynamic as a "wealth tax" on companies using AI to generate operational returns.

Palantir's Evolve AI routing system, which automatically assigns tasks to the most cost-effective model rather than defaulting to the most powerful one, has reduced inference costs by up to 97% in certain customer deployments, the company disclosed. The system also optimizes prompts and avoids redundant calls. Separately, routing platform OpenRouter raised $120 million in April, underscoring investor appetite for the category.

The shift from model capability competition to cost optimization represents a structural change in enterprise AI. Companies that enable efficient routing — Palantir, Databricks, OpenRouter — stand to capture a growing share of corporate AI budgets, while high-cost frontier model providers face pressure to justify their pricing.

The Cost Crisis Driving Adoption

Enterprise AI spending has accelerated faster than many companies anticipated. Databricks Chief Executive Officer Ali Ghodsi said his company's Unity AI Gateway, now widely used internally, gained traction because organizations are "burning through AI budgets too fast." The gateway allows companies to route queries across models from OpenAI, Google, and Anthropic based on cost and performance thresholds.

Construction firm McCarthy Building reported its AI token usage fell roughly 60% year over year after implementing model scheduling optimization, without a meaningful drop in output quality. Palo Alto Networks has also adopted model switching strategies to reduce AI-related costs, according to the company.

Japanese AI lab Sakana AI demonstrated a multi-model routing system that exhibits a form of expert分工: math problems are preferentially routed to OpenAI models, while scientific queries are more often directed to Google Gemini. The approach mirrors a broader industry recognition that no single model is optimal for every task.

OpenRouter's $120M Bet on Routing Infrastructure

The routing category attracted its largest single investment to date in April, when OpenRouter closed a $120 million funding round. The platform's "auto-router" lets users set a cost-quality preference on a 0-to-10 scale, and the system dynamically selects models accordingly.

Data from the platform shows roughly one-third of requests are routed to Google's lower-cost models, while only about 10% flow to OpenAI's premium offerings — a cost-tier distribution that would be impossible under a default-to-best approach. OpenRouter integrates routing technology from providers including Not Diamond and supports cross-cloud provider calls to optimize latency and pricing.

AI coding startup Cognition also built its own routing system, achieving near-frontier performance on programming benchmarks at roughly 35% lower cost than using a single top-tier model.

Palantir's Financial Case for the Routing Thesis

Palantir reported Q1 2026 revenue of $1.63 billion, up 84.7% year over year, with adjusted earnings per share of $0.33 beating the $0.28 consensus. U.S. commercial revenue reached $595 million, up 133%, and the company closed 206 deals worth at least $1 million with a total contract value of $2.41 billion.

Karp projected the company would generate $15 billion to $18 billion in free cash flow within two years, a figure he said validates the model-plus-application-layer approach despite market skepticism. Palantir's Rule of 40 score hit 145%, a level Karp said is matched only by fellow AI infrastructure companies Nvidia, Micron and SK Hynix.

Still, PLTR trades at about $127, down 34% year to date, with a forward price-to-earnings ratio near 74 times. The valuation disconnect reflects investor uncertainty about whether Palantir's growth can sustain its premium — a question that hinges on whether enterprise customers continue adopting routing and deployment tools at the pace Karp expects.

Nvidia, Palantir's partner in the sovereign AI push, reported Q1 fiscal 2027 revenue of $81.6 billion, up 85% year over year, with data center revenue of $75.25 billion, up 92%. The company guided second-quarter revenue to $91 billion. The Palantir-Nvidia stack gives enterprises a path to run models on their own infrastructure rather than through third-party APIs, directly addressing the data control concerns Karp raised.

For investors, the key question is whether routing technology becomes a standard enterprise AI layer — like API gateways did for cloud computing — or remains a niche optimization tool. If the former, companies like Palantir and Databricks that embed routing into their platforms could see sustained demand growth. If the latter, the $120 million bet on OpenRouter may prove premature. The answer will show up in commercial contract value, deal counts, and customer adoption through the second half of 2026.

This article is for informational purposes only and does not constitute investment advice.