Alex Karp, Frontier Models and the Real Fight for Enterprise AI

Palantir Technologies Inc. Chief Executive Alex Karp’s recent broadside against the frontier model vendors put a knife to the throat of the central enterprise artificial intelligence debate.

Karp’s argument is that frontier model vendors (he didn’t mention Anthropic and OpenAI by name) intend to suck the knowledge out of enterprises and destroy the “alpha” companies enjoy through their proprietary data, processes and underlying business advantage. In our last Breaking Analysis we called this approach “data communism,” where every firm gains access to the same intelligence.

Our counter to data communism is data capitalism, where proprietary advantage remains exclusive to an organization and its broader ecosystem. The familiar graphic below describes how we see the AI software stack evolving. The most important pieces of this stack in our view are the System of Intelligence or SoI and the System of Engagement, the new user/client surface. We believe competitive advantage accrues to firms that can make these two puzzle pieces interact and learn from the reasoning traces of humans. Critical is how they capture tacit enterprise knowledge to take governed and trusted actions both with and without humans.

The key issue we explore in this Breaking Analysis is: Who will own the operating intelligence of the enterprise?

On the surface, the argument is about closed models versus open models, OpenAI and Anthropic versus Nvidia Corp.’s Nemotron (as an example), and whether enterprises should trust frontier labs with their most sensitive data and workflows. But as Nvidia CEO Jensen Huang said at the GTC 2026 conference, “proprietary versus open is not a thing. It’s proprietary and open.”

The deeper issue is not model choice. It is control.

This question goes to the heart of the debate inside theCUBE Research and across the industry. Here are the two ends of the spectrum with two scenarios from theCUBE Research:

Frontier model leadership – One camp believes the frontier model vendors will dominate the stack because their utility, cost curves, research velocity, volume and compute access will outpace everything else.

Dispersed intelligence – The alternative argument is the frontier players lack the mindset, DNA, process knowledge and trust. This camp agrees that the highest-value layer is not the model itself, but the system of intelligence – the enterprise-specific context layer that captures business rules, policies, processes, state and tacit knowledge as governed assets. And the argument is that other players (for example, Palantir, Databricks Inc., Microsoft Corp., Google LLC, Celonis SE and new startups) will cement a more critical position in the AI stack than the frontier model players.

Both views can be true. And both acknowledge that: 1) the system of intelligence and the client surface must be part of the winning vendor’s stack; and 2) trust and proprietary knowledge must be the exclusive property of the customer. But the Karp conversation unveils the tension between them and we believe will define the future enterprise AI industry power structure. In either case, it’s highly likely that silos of intelligence will emerge over the next decade and the industry will remain heterogenous and fragmented.

To put a finer point on the topic, Karp’s statements imply that OpenAI and Anthropic are stealing data from and overcharging their customers. He positions his company, Palantir, as a critical “application layer” to protect business and governments from those newbies and their bad intentions. He is on a mission to convince businesses they need a more trusted partner than Anthropic or OpenAI. Rather than dealing directly with the two AI firms, he’s arguing Palantir (by proxy) should be that intermediary.

The frontier model dominance case

We’ve not published extensively on this point of view and it’s worth taking a moment to do so here. Much of the following is based on scenario and cost modeling work done by theCUBE Research Analyst Emeritus David Floyer. The methodology and framework is detailed below and draws extensively on Wright’s Law, as applied to software. Wright’s Law says that as cumulative production rises, costs fall in a predictable way. In AI, the equivalent is not just cumulative production but cumulative usage, cumulative tokens, cumulative feedback, cumulative training experience, cumulative inference optimization and cumulative compute deployment.

While some of the AI data forecasts in the methodology are need updating, the principles of the forecasting approach still apply.

https://thecuberesearch.com/using-volume-value-velocity-to-forecast-ai/

Here’s the argument.

Think of the frontier model as a “cognitive surface,” the heart of intelligence and the sole source of tokens. It performs reasoning, planning, synthesis and learning. It is built and runs on the most advanced hardware available and continues to improve rapidly. It is capital-intensive, power-dense and scarce by design. Only a small number of organizations can develop and operate frontier models at scale, and enterprises should not attempt to replicate this function.

Directly coupled to the frontier model is a new layer that does not exist in traditional enterprise architectures: the System of Intelligence. This layer manages all inputs to and outputs from the large language models. It shapes intent, context, constraints and semantic grounding before intelligence is invoked. It expresses intelligence into actions, system interactions and multimodal outputs after tokens are produced. It hosts security, policy enforcement, compliance, auditability, latency control and integration with enterprise systems. It evolves in lockstep with the frontier model while remaining external to it, preserving both control and adaptability.

Though frontier model developers must own their cognitive surface architecturally and evolutionarily, they will almost certainly want to own the SoI and also allow the controlled distribution of both. A fundamental assumption in this scenario must be made explicit: Large enterprises will be permitted by the LLM license terms to operate instances of the cognitive surface locally or within sovereign environments to meet latency, security and regulatory requirements.

However, this distribution will occur under strict contractual and technical control. Enterprises will not be able to modify the LLM itself; they will only be able to configure, utilize and integrate it within defined boundaries. But those configurations, and associated data, process logic and underlying enterprise knowledge remain the sole property of the customer. Semantic grounding, safety constraints, interface definitions and evolutionary alignment will remain under frontier model governance. This arrangement preserves enterprise control over data, latency and policy while preventing intelligence fragmentation or semantic drift.

The LLM is also the place where a small number of early and strategically positioned software-as-a-service vendors will negotiate with frontier model providers to license access to semantic definitions, interfaces and selected process and execution logic. Likely candidates are SAP SE, Oracle Corp., Salesforce Inc., ServiceNow Inc., and other leading SaaS providers.

Vendors of platform services and middleware may also seek certified integration points within the cognitive surface. For example, a platform provider such as Oracle is well-positioned to operate a minimal, authoritative retrieval-augmented generation capability within the cognitive surface, in conjunction with a frontier model provider such as OpenAI. Open-source applications may also be integrated, subject to the same semantic, security and governance constraints.

Frontier leaders as this era’s disruptors

This scenario recognizes frontier players have the world’s top AI researchers. They have access to the largest pools of compute. They have massive user volume. They have consumer products and/or consumer-like volumes that act as high-frequency learning loops. They have brand affinity, developer adoption and enterprise pull. And they have the capital to run far ahead of companies that are trying to compete from narrower positions, all while being rewarded and still losing money.

As such, this point of view assumes model players will have the lowest cost, highest volume and most functional product. The outlook projects that companies will scale with less labor and achieve 10X productivity relative to current best practice metrics. This economic advantage, the argument says, will overwhelm Karp’s concerns about trust because a winner-take-most dynamic and software-like marginal economics will accrue to firms that lead in AI adoption.

If the frontier models keep improving at a faster rate than alternatives, and if inference costs keep falling, enterprises will continue to route more work to them. Smaller models and open models will absolutely have a role. They will be used for cost-sensitive tasks, domain-specific workloads, sovereignty requirements, edge deployments and latency-sensitive use cases. But in many workflows, they may be most effective as part of a broader ensemble led by frontier-class models.

The most aggressive version of this thesis says that token costs will fall so dramatically that today’s budget concerns will look temporary. Enterprises may complain about LLM bills now, but the real comparison is not token cost versus token cost. It is token cost versus headcount.

If a company can grow revenue 10x without scaling labor proportionally, tokens become relatively cheap. If agents allow enterprises to compress support, engineering, finance, operations, compliance, analytics and field work into software-driven workflows, the marginal cost of tokens becomes much lower than the marginal cost of people. In that world, the winning providers are the ones that deliver the highest utility per unit of work – not necessarily the lowest nominal token price.

That is the frontier model bull case.

OpenAI, Anthropic and Google may become the lowest-cost providers of high-utility intelligence because they operate at the largest scale. Their compute purchasing power, model optimization, inference infrastructure and usage volume may give them better marginal economics than alternatives. If that happens, model routing becomes less about avoiding frontier models and more about using them intelligently where their incremental utility justifies the cost.

Karp’s argument is really about enterprise sovereignty

Karp’s comments are in a large part a fear campaign against OpenAI and Anthropic. His language was purposefully fiery, and the reported suggestion that frontier model providers could “take the alpha” of a customer’s business and transfer it into their weights is attention-getting. Notably, there is no public evidence cited that Anthropic or OpenAI trains on customer data in violation of their terms; OpenAI has publicly said it does not train on customer data.

But whether the literal accusation is proven is not the main point.

The enterprise fear is real.

Karp nailed the sentiment. Customers are asking: If our most sensitive workflows, data, decisions, policies and proprietary operating knowledge flow through a frontier model vendor, are we building our future on a supplier that could one day intermediate us, compete with us, or extract too much margin from us?

Palantir’s answer is self-serving but thought-provoking: Don’t let the model vendor own the enterprise brain. Let Palantir sit between the customer and the model, govern the interaction, route workloads to the best model, preserve sovereignty and make the model interchangeable.

That is a classic platform move and the one we’ve been putting forth based on George Gilbert’s work. Palantir wants to own the system of intelligence and treat models as pluggable engines. Its message is the model is important, but the enterpr

[truncated for AI cost control]