Palantir CEO Alex Karp just fired a direct shot at OpenAI and Anthropic, declaring their token-based pricing model fundamentally broken. Speaking to CNBC, Karp argued that skyrocketing inference costs are forcing enterprises to abandon proprietary AI models in favor of open-weight alternatives - a shift that could reshape the entire AI industry's economic foundation. "Something has gone completely wrong," he said, coining the term "tokenmaxxing" to describe companies burning cash on per-token billing without clear ROI.
Palantir CEO Alex Karp isn't pulling punches. In a sharp rebuke delivered to CNBC, he declared that OpenAI and Anthropic have built AI pricing models that are driving enterprises away - and he's got the receipts to prove it.
"Something has gone completely wrong," Karp told reporters, zeroing in on the token-based billing system that's become standard across frontier AI labs. His argument cuts straight to the tension enterprises face: every API call to GPT-4 or Claude racks up costs that scale unpredictably, making budget forecasting nearly impossible and ROI calculations a nightmare.
Karp coined a new term for what he sees happening across corporate America: "tokenmaxxing." It's his shorthand for companies that chase impressive demos and bleeding-edge capabilities while burning through millions in inference costs without clear business outcomes. According to the Palantir chief, this dynamic is pushing rational enterprises toward open-weight models - AI systems you can download, modify, and run on your own infrastructure without per-token fees.
The timing of this critique matters. OpenAI recently raised its enterprise API pricing across several tiers, while Anthropic continues positioning Claude as a premium product with corresponding price tags. Meanwhile, open-weight alternatives from Meta (Llama 3), Mistral, and others have closed the capability gap enough that cost-conscious CTOs are taking a second look.
Palantir itself has been vocal about efficiency in AI deployment. The company's Foundry platform emphasizes optimization and cost control, positioning it as the anti-tokenmaxxing solution for enterprises that need AI to actually pencil out. Karp's comments reveal what his sales team is probably hearing in every customer conversation: "We love the technology, but we can't afford to scale it."
The economics are hard to ignore. A single large language model inference through OpenAI's API can cost fractions of a cent for simple queries but dollars for complex, multi-turn conversations with lengthy context windows. Multiply that across thousands of employees or millions of customer interactions, and you've got a budget line that makes CFOs nervous. Open-weight models, by contrast, have upfront compute costs but zero marginal fees per token - making them dramatically cheaper at scale.
Karp's criticism lands differently than typical startup griping about incumbents. Palantir works with defense contractors, intelligence agencies, and Fortune 500 clients that have both deep pockets and strict cost accountability. When Karp says enterprises are choosing efficiency over cutting-edge capabilities, he's describing actual procurement decisions happening right now across sectors from finance to manufacturing.
But there's nuance here. OpenAI and Anthropic would argue their pricing reflects massive infrastructure costs and continuous model improvements. Training runs for GPT-4 or Claude cost hundreds of millions, and inference at scale requires data center capacity that rivals cloud giants. Their token-based model, they'd say, aligns incentives: pay for what you use, and benefit from ongoing upgrades without managing infrastructure.
The split Karp's describing - proprietary APIs versus open-weight models - mirrors historical divides in enterprise software. Think Oracle versus MySQL, or Salesforce versus self-hosted CRMs. Companies with sophisticated engineering teams and predictable usage patterns often choose the open route. Those prioritizing speed to market and offloading complexity stick with managed services despite higher long-term costs.
What makes this moment different is velocity. AI adoption is happening faster than previous platform shifts, and pricing models are still being figured out in real-time. Microsoft offers consumption-based Azure OpenAI pricing. Google bundles Vertex AI into cloud contracts. Amazon positions Bedrock as a multi-model marketplace. Nobody's cracked the perfect formula, and Karp's frustration suggests Palantir's enterprise clients are demanding better options.
The "tokenmaxxing" framing is particularly cutting because it implies irrationality - that companies are optimizing for the wrong metric. It's reminiscent of debates around vanity metrics in SaaS or cloud spending waste. Karp's essentially saying: stop chasing the fanciest model and start asking whether AI is actually making you money.
If Karp's reading of the market is right, OpenAI and Anthropic face a strategic dilemma. Lower prices to keep enterprise customers, or maintain margins while accepting that open-weight alternatives will capture cost-sensitive segments? The former hurts unit economics; the latter cedes market share to Meta, Mistral, and whoever else releases competitive open models.
There's also a competitive angle. Palantir benefits directly if enterprises shift spending from API fees to infrastructure and integration work - exactly what Foundry provides. Karp's criticism serves his business interests, but that doesn't make it wrong. The best critiques usually come from people with skin in the game.
Karp's broadside against OpenAI and Anthropic reflects a broader reckoning in enterprise AI: impressive capabilities mean nothing if the math doesn't work. As open-weight models close the performance gap and eliminate per-token fees, proprietary labs face real pressure to justify their pricing or risk becoming niche players serving only the most capability-obsessed customers. The tokenmaxxing era might be shorter than anyone expected, and the shift to efficiency-first AI could redefine which companies dominate the next decade of enterprise technology. Watch how OpenAI and Anthropic respond in their next pricing announcements - silence will speak volumes.
